text
stringlengths 4
2.78M
| meta
dict |
|---|---|
-30pt plus 1pt minus 1pt
\
[**Structure function of the nucleus in the perturbative QCD with $N_c\rightarrow\infty$\
(BFKL pomeron fan diagrams)**]{}\
M.Braun\
Department of High Energy physics, University of S.Petersburg,\
198904 S.Petersburg, Russia
Equation for the sum of BFKL pomeron fan diagrams is rederived by direct summation and solved numerically for rapidities $y\leq 50$. At high rapidities $y>20$ the resulting cross-sections for the scattering of a longitudinally polarized $q\bar q$ pair on the nucleus cease to depend on its transverse dimension and tend to a constant limit 0.1768 $R_A^2$, which corresponds to scattering of a colour dipole on a black disk. Thus the unitarity is restored and the singularity in the $j$ plane is reduced to a simple pole at $j=1$. The nuclear structure function at small $x$ behaves as $Q^2\ln(1/x)$. The found gluon density has a soliton-like form in the $\log k$ space: its form is close to Gaussian, independent of rapidity, the center moving towards higher $\log k$ with a nearly constant velocity as rapidity increases.
Introduction
============
In the framework of the colour dipole model of A.H.Mueller \[1,2\] it follows that in the high-colour limit $N_c\rightarrow\infty$ the scattering on a heavy nucleus is exactly described by the sum of fan diagrams constructed of BFKL pomerons, each of them splitting into two \[3\]. This sum seems to be unitary by itself. It is important that no splitting into three or more pomerons, as introduced in \[4\], occurs, although formally they give contributions of the same order. Because of this fact, once the splitting vertex is known, construction of the amplitude for the interaction with the nucleus becomes straightforward, reducing to summing BFKL pomeron fan diagrams. This procedure has been well-known since the times of the old Regge-Gribov theory \[5\].
In the perurbative QCD a pioneering step was taken also many years ago by L.Gribov, E.Levin and M.Ryskin, who summed fan diagrams in the double log approximation and wrote their well-known non-linear GLR equation \[6\]. In the framework of the BFKL dynamics the necessary tool for constructing fan diagrams is the corresponding triple pomeron vertex, which was found in the colour diplole approach for $N_c\rightarrow\infty$ by A.H.Mueller and B.Patel \[2\] and in the $s$-channel unitarity approach for any number of colours by J.Bartels and M.Wuesthoff \[7\]. The equivalence of both results was shown in \[8\].
The equation for the sum of BFKL fan diagrams with this splitting vertex was written by I.Balitsky \[9\] in his original operator expansion formalism and by Yu.Kovchegov \[10\] in the colour dipole framework (with a somewhat unconventional form of the coupling to the target) . Its perturbative solution in the region of small non-linearity (outside the saturation region) was studied in \[11\]. Asymptotic estimates of the solution were presented in \[12\].
In the present paper we first rederive the BFKL fan diagram equation by direct summation using the standard form of the pomeron-target coupling. It has a form of a simple (and elegant) evolution equation in rapidity $y$ for a wave function $\phi(y,q)$ in the momentum space: =-H(y,q)-\^2(y,q), where $H$ is the BFKL Hamiltonian for the so-called semi-amputated function (a similar form was also obtained in \[11\]).
In spite of its tantalizing simplicity, Eq. (1) does not seem to allow for an analytical treatment except by perturbative methods, not valid in the most interesting region of strong non-linear effects, or by qualitative asymptotic estimates. The bulk of this paper is correspondingly devoted to its numerical analysis. We numerically study evolution of the wave functuion in rapidity, starting from an appropriately chosen initial function. The results are then used to find the structure function of the nucleus at small $x=e^{-y}$ and various virtualities $Q^2$. Our results show that for a heavy nucleus the longitudinal part of structure function saturates at $x\rightarrow 0$ to a universal function $F_{2L}^{(as)}(Q^2)$, independent of the nucleus atomic number but strongly dependent on $Q^2$ in the whole range of $Q^2\leq 10^5$ (GeV/c)$^2$ explored. In fact $F_{2L}^{(as)}(Q^2)$ is just proportional to $Q^2$. This reflects the behaviour of the scattering cross -section of a longitudinally polarized $q\bar q$ pair on the nucleus: at very small $x$ it becomes independent of both the pair size and $x$. The latter property indicates that the unitarity is restored and the leading singularity in the complex momentum $j$ is reduced to a simple pole at $j=1$. Note that the limiting cross-section is in accordance with a picture in which the longitudinally polarized photon with a certain probability splits into colour dipoles, which then scatter on the nucleus as on a black disk. For the leading transverse part of the structure function this probability results infinite. Due to this fact both the cross-section and the structure function continue to grow at $x\rightarrow 0$ approximately as $\ln (1/x)$. These results are in agreement with predictions made in \[11,12\].
As to the gluon density, we have found it to have a Gaussian shape as a function of $\xi=\ln k$ with a center at $\xi_0\propto y$. So with the growth of rapidity it propagates towards higher momenta, practically preserving its form, very much like a soliton wave. Obvious limitations on computing time and memory allow to follow its movement up to momenta not higher than of the order $10^{10}$ GeV/c. However we expect that this behaviour persists in the model until arbitrary high values of momenta. As a result, at any fixed value of $k$ the density eventually goes to zero as $y\rightarrow\infty$. In this sense we have the maximal saturation of the density possible.
The paper is organized as follows. In section 2 we present derivation of the BFKL fan diagram equation by direct summation. Section 3 is devoted to the numerical solution of Eq. (1). In Section 4 the nuclear structure function and gluon density are calculated. Section 5 contains our conclusions. In the Appendix we compare our BFKL fan diagram equation with the Kovchegov’s one \[10\].
Fan diagram equation for the BFKL pomerons
==========================================
We start with a single scattering contribution to the forward ampitude for the interaction of the projectile particle with a nucleus. In the BFKL framework, at fixed impact parameter $b$, it has a well-known form \_1(y,b)=isg\^4AT(b)d\^2rd\^2r’(r)G(y,r,r’) \_N(r’).Here $\rho$ and $\rho_N$ are the colour densities of the projectile and the target nucleon respectively. Function $G$ is the forwrad BFKL Green function \[13\]: G(y,r,r’)= \_[n=-]{}\^[+]{}e\^[in(-’)]{} \_[-]{}\^(r/r’)\^[-2i]{}, where $\phi$ and $\phi'$ are the azimuthal angles and ()=2(\_sN/2)((1)-[Re]{}(1/2+i)) are the BFKL levels. Due to the azimuthal symmetry of the projectile colour density one may retain only the term with zero orbital momenta $n=0$ in (3). Separating the projectile part, the single scattering term may be written in the form \_1(y,b)=2isd\^2r(r)\_1(y,b,r), where \_1(y,b,r)=g\^4AT(b)d\^2r’G(y,r,r’)\_N(r’).
The double scattering contribution has been calculated in \[14\]. In the limit $N_c>>1$ one finds $${\cal A}_2(y,b)=-is\frac{g^2N_c}{4\pi^3}\int d^2r \rho(r)
\int_0^y dy_2 \int\prod_{i=1}^3 d^2r_i
\delta^2(r_1+r_2+r_3)$$ G(y-y\_2,r,r\_1)\_1(y\_2,b,r\_2)\_1(y\_2,b,-r\_3). Presenting ${\cal A}_2$ in terms of $\Phi_2$ similarly to (5) we find $$\Phi_2(y,b,r)=-\frac{g^2N_c}{8\pi^3}
\int_0^y dy_2 \int\prod_{i=1}^3 d^2r_i
\delta^2(r_1+r_2+r_3)$$ G(y-y\_2,r,r\_1)\_1(y\_2,b,r\_2)\_1(y\_2,b,-r\_3).
The whole set of fan diagrams will be evidently summed by the equation which is graphically illustrated in Fig. 1: $$\Phi(y,b,r)=\Phi_1(y,b,r)
-\frac{g^2N_c}{8\pi^3}
\int_0^y dy_2 \int\prod_{i=1}^3 d^2r_i
\delta^2(r_1+r_2+r_3)$$ G(y-y\_2,r,r\_1)(y\_2,b,r\_2)(y\_2,b,-r\_3). The impact parameter $b$ appears here only as a parameter and the dependence on it will be implicit in the following. In terms of $\Phi$ the total forward scattering amplitude on the nucleus will be given at fixed $b$ by (y,b)=2isd\^2r(r)(y,b,r).
One can rewrite Eq. (9) as an evolution equation in $y$. To this end we present $\Phi$ as an integral over $\nu$’s similar to (3): $$\Phi(y,r)=\int d\nu r^{1-2i\nu}\Phi(y,\nu).$$ Then from (9) we find $$\Phi(y,\nu)=\frac{g^4}{64\pi^2(\nu^2+1/4)^2}e^{\omega(\nu)y}
\int d^2 r'r'^{1+2i\nu}\rho(r')AT(b)-$$$$\frac{g^2N_c}{32\pi^5(\nu^2+1/4)^2}e^{\omega(\nu)y}
\int_0^y dy_2 e^{-\omega(\nu)y_2}\int\prod_{i=1}^3 d^2r_i
\delta^2(r_1+r_2+r_3)$$ r\_1\^[1-2i]{}(y\_2,r\_2)(y\_2,-r\_3). Multiplying both parts by $e^{-\omega(\nu)y}$ and taking a derivative in $y$ we obtain $$\left(\frac{\partial}{\partial y}-\omega(\nu)\right)\Phi(y,\nu)=
-\frac{g^2N_c}{32\pi^5(\nu^2+1/4)^2}
\int\prod_{i=1}^3 d^2r_i
\delta^2(r_1+r_2+r_3)$$ r\_1\^[1-2i]{}(y,r\_2)(y,-r\_3). Returning to the $r$ space we take into account that ()r\^[1-2i]{}=-r\^[1-2i]{},where $\hat{H}$ is the BFKL Hamiltonian \[13\]. Then (12) transforms into (+)(y,r)= - \_[i=1]{}\^3d\^2r\_i \^2(r\_1+r\_2+r\_3) G(0,r,r\_1)(y,r\_2)(y,-r\_3). Using G(0,r,r\_1)= \_[n=-]{}\^[+]{}e\^[in(-\_1)]{} \_[-]{}\^d (r/r\_1)\^[-2i]{}=\^2(r-r\_1), we simplify (14) to a form (+)(y,r)= - \_[i=2]{}\^3d\^2r\_i \^2(r+r\_2+r\_3) (y,r\_2)(y,-r\_3).
The initial condition is determined from (9) to be (y,r)\_[y=0]{}=\_0(r) with \_0(r)=g\^4AT(b)d\^2r’G(0,r,r’)(r’). Except for the initial condition and different variables, Eq. (16) coincides with the one constructed by Kovchegov in the colour dipole approach in \[10\] (see Appendix).
Now we go to function (y,r)=(y,r) and pass to the momentum space. For $\phi(y,q)$ Eq. (16) reads (+r\^2)(y,q) =- (y,q)(y,q), where we have used that $\phi(y,q)$ in fact depends only on $|q|$, which follows from the initial conditon \_0(y,q)\_[y=0]{}=e\^[-iqr]{}\_0(r) with $\Phi_0$ given by (18) and depending only on $|r|$ due to the azimuthal summetry.
The Hamiltonial $r^{-2}\hat{H}r^2$ which appears on the left-hand side is the standard forward Hamiltonian for the semi-amputated functions r\^2=H= \[q\^2+r\^2 -2(2 +(1))\]. Indeed the eigenfunctions of $\hat{H}$ $\Phi_{n,\nu}(r)=r^{1-2i\nu}e^{in\phi}$ go over into the eigenfunctions of $H$ after division by $r^2$.
This brings us to nearly the final form of our equation (+H)(y,q) =-\^2(y,q). It remains only to appropriately rescale $H$, $\phi$ and $y$ to obtain the BFKL fan diagram equation in the final form (1). We introduce H=, =y, [and]{} (,q)=2(,q), where =r\^2+q\^2-2(2+(1)). The equation for $\tilde{\phi}$ takes the form (1) (+\_0) (,q)=-\^2(,q) with the initial condition (,q)\_[=0]{}= d\^2rd\^2r\_1e\^[iqr]{} G(0,r,r\_1)\_N(r\_1).
Numerical solution of the BFKL fan diagram equation
===================================================
To solve numerically Eq. (26) we first transform the BFKL Hamiltonian $H$ to more convenient variables. In the momentum space one can write action of $H$ on a function $\phi(q)$ as H(q)=-2\_0\^kdk. We transform it to variables $u,v$ which take values in \[0,1\]: q=, k=, where $q,k$ are in GeV/c and $M_{1(2)}$ is a lower(upper) integration limit in $\ln k$. Then (28) goes into H(u)=-2(M\_1+M\_2)\_0\^1dv, where f(u)=.
Now we discretize the interval \[0,1\] in $u$ and $v$ into $n$ equidistant points $u_0,u_1,...u_n$ and $v_0,v_1,...v_n$ to convert the evolution equation (26) into a set of $n$ ordinary 1st order non-linear differential equations in $y$. This set of equations can be solved by standard methods. We used the simplest 2nd order Runge-Kutta algorithm. The standard values of $n$ and of the number of iterations to evolve in two units of $\tilde{y}$ were 800. We have studied evolution of $\phi(y,q)$ from the initial value at $y=0$ up to $\tilde{y}=10$. The (fixed) value of the strong coupling $\alpha_s=g^2/4\pi$ has been chosen to be 0.2. With this choice our maximal rapidity is around 50.
An evident difficulty which one meets are the values of $H\phi$ at endpoints $u_0$ and $u_n$, at which the introduced cutoffs make the results not reliable. To overcome this difficulty we calculated $H\phi$ at these points by extrapolation from the neighbouring points. The stability of this procedure was checked by comparing the results for double and quadruple values of $n$.
The initial function is determined by the colour density of the nucleon $\rho_N(r)$ according to (27). To simplify our calculations we have taken the Yukawa form for $\rho_N(r)$: \_N(r)=, where $\mu$=1/0.7 fm has a meaning of the inverse nucleon radius. This choice may look a bit arbitrary, but our results show that with the growing rapidity the system quickly forgets not only the form but even the absolute magnitude of $\rho_N$, so that the solution becomes independent of the initial function and governed only by the internal dynamics of Eq. (26) itself.
Doing the integrations over $r$ and $r_1$ in (27) and over $\nu$ inside the BFKL Green function (3) we find for $q>\mu$ \_0(q)\_[q>]{}=B {1- \_[n=0]{}(-1)\^n()\^n } and for $q<\mu$ $$\tilde{\phi}_0(q)_{q<\mu}=B
\Big\{\frac{1}{4}[(2\ln\frac{q}{\mu}+\psi(3/2)-\psi(1)-1)^2+
\psi'(3/2)+\psi'(1)+1]-$$ \_[n=0]{}(-1)\^n()\^n }, where the dimensionless coefficient B= carries all the information about the nucleus. Its maximal value with $\alpha_s=0.2$ is about 0.12 for the central scattering on lead ($b=0$). The asymptotic behaviour of $\phi_0(q)$ at $q\rightarrow\infty$ and $q\rightarrow 0$ is governed by the first terms in (33) and (34), which come from the poles of the BFKL Green functions at $\nu=\pm i/2$: $$\phi_0(q)\sim 1/q^2,\ \ q\rightarrow\infty,$$\_0(q)\~\^2q, q0. We have checked that our results parctically do not change if the complicated function which multiplies $B$ in (33) and (38) is substituted just by $\ln^2(q/(\mu+q)$ with the same asymptotic behaviour (36).
The initial function depends on the form of the nucleus profile function $T(b)$. For our numerical calculation we have chosen T(b)=2/V\_A corresponding to a finite nucleus of radius $R_A$ and volume $V_A$ with a constant density. We have taken $R_A=A^{1/3}R_0$ with $R_0=1.2$ fm.
Our results for $\phi(q)$ for two values of $B=0.12$ and 0.02, corresponding to a central and very peripheral collisions off lead, respectively, are presented in Figs 2,3. One observes that the solution $\phi$ at rapidities $\tilde{y}>2$ evolves to a very simple and universal form, practically independent of the initial function. Crudely speaking it linearly falls with $\xi=\ln k$ until it meets the $x$-axis wherefrom it stays equal to zero. The slope of the falling part is exactly equal to 1, so that very crudely ()=\_1(y,b)-, [for]{} \_1(y,b). The value of $\xi_1(y,b)$ and hence the interception point with the $x$ axis grow linearly with $y$, so that with the growth of $y$ the picture simply shifts to the right. In reality the curve for $\phi$ of course has no break: the two straight lines of which it is formed join smoothly in the vicinity of $\xi_1$. As will be clear later, the physically important region is precisely this vicinity, where $\phi$ is not trivial.
The results for $\phi(k)$ do not depend on the chosen cutoffs, provided they are taken to cover the region around the interception point $\xi_1$. Otherwise the evolution stops as soon as $\xi_1$ touches the upper cutoff $M_2$. So if one wants to study evolution up to high values of $y$ the upper cutoff should be taken correspondingly high. In our calculations we chose $M_1=10$ and $M_2=20$, having verified that further raising of either $M_1$ or $M_2$ does not change the results.
In conclusion we find that $\phi$ does not possess any finite limit as $y\rightarrow\infty$ so that Eq. (26) does not lead to any saturation of the wave function $\phi$ at high rapidities, contrary to naive expectations. However in the next chapter we shall see, that such saturation indeed occurs for physical quantities. The point is that function $\phi$ by itself has no physical meaning. It is its derivatives which matter.
Nuclear structure function and the gluon density
================================================
The nuclear structure function is obtained in the standard manner as F\_2(x,Q\^2)=(\_T+\_L), where $\sigma_{T,L}$ are the total cross-section for the scattering on the nucleus of a virtual photon with transversal (T) or longidunal (L) polarization. Both cross-sections can be found from the imaginary part of the forward scattering amplitude (10). In terms of $\tilde{\phi}$ we find \_[T,L]{}=4d\^2b d\^2r\_[T,L]{}(r)r\^2 (y,r,b). Here we explicitly indicated the dependence of $\phi$ on the impact parameter; $\rho_{T,L}(r)$ are the well-known colour densities of the virtual photon split into a $q\bar q$ pair (see e.g. \[15\]). With massless quarks \_T(r)= \_[0]{}\^[1]{}d(\^[2]{}+(1-)\^[2]{})\^[2]{}[K]{}\_[1]{}\^[2]{} (r)) and \_[L]{}(r)= Q\^2 \_[0]{}\^[1]{}d\^2(1-)\^2[K]{}\_[0]{}\^[2]{}(r), where $\epsilon^2=Q^{2}\alpha (1-\alpha)$ and $Z^2$ is a sum of squares of quark electric charges in units $e$.
Passing to momentum space we find \_[T,L]{}=4d\^2b (q,y,b)w\_[T,L]{}(q), where w\_[T,L]{}(q)=d\^2r r\^2\_[T,L]{}(r)e\^[iqr]{}. Straightforward calculation leads to the following expressions for $w_{T,L}(q)$. For the transverse density one finds[^1] w\_T(q)= \_[0]{}\^[1]{}d(\^[2]{}+(1-)\^[2]{})\_q\^2\[(q\^2/2 +\^2)J(q,)\], where J(q,)= . The longitudinal density $w_{L}$ is given by the same expression with substitutions $$\alpha^{2}+(1-\alpha)^{2}\rightarrow \alpha(1-\alpha),\ \
q^2/2+\epsilon^2\rightarrow -4\epsilon^2$$
With the found numerical values for the function $\phi$ in the range $0\leq\tilde{y}\leq10$ we evaluated the nuclear structure function of lead ($A=207$) for various $Q^2$ between 3 and 10$^5$ (GeV/c)$^2$. The results are most instructive for the the cross-section $\sigma_{L}$ for the scattering of a longitudinally polarized virtual photon on the nucleus. In Figs. 4,5 we present it (with $e^2\rightarrow 1$) as a function of $y$ at fixed $Q^2$ and as a function of $Q^2$ at fixed $y$ respectively (in GeV$^{-2}$). From Fig. 4 one clearly sees saturation in rapidity: for any value of $Q^2$ the cross-section tends to the same limit of $228.8$ (GeV/c)$^{-2}=
0.1768 R_A^2$ as $\tilde{y}$ goes beyond 5. The resulting constant cross-section is evidently consistent with the unitarity restrictions. In terms of the complex angular momentum $j$ our results indicate that the original cut at $j>1$ is reduced to a simple pole at $j=1$. Fig. 5 illustrates an unusual behaviour in $Q^2$ which sets in at high $y$: instead of going down as $1/Q$ in the standard BFKL approach, the cross-section becomes independent of $Q^2$ to a very high precision.
These results for the longitudinal cross-section can be conveniently interpreted in terms of scattering of colour dipoles off the nucleus. The density $\rho_L(r)$, appropriately normalized, can be interpreted as a probability distribution for the longitudinal photon to split into colour dipoles of transverse dimension $r$. The normalization factor $$D=\int d^2r \rho_L(r)=\frac{e^2N_cZ^2}{12\pi^2}=0.028145\,e^2$$ can be considered as a total probability for the photon to split into dipoles. Then the dipole-nucleus cross-section is found by dividing $\sigma_L$ by factor $D$, which gives 8131.1 (GeV/c)$^{-2}$ independent of $Q^2$, that is, of the dipole dimension. This value exactly equals $2\pi R_A^2$, corresponding to scattering off a black disk.
The transverse part of the structure function does not admit this interpretation, due to the fact that $\rho_T(r)$ is not normalizable. As a result both the cross-section and the structure function do not saturate at large $y$ but continue to grow nearly linearly in $y$. This is illustrated in Figs. 6,7 where the total structure function (including the much smaller longitudinal part) is shown as a function of $x=e^{-y}$ and $Q^2$ respectively.
We finally come to the gluon density. Although, strictly speaking it is not a physical quantity, its properies have been much discussed recently in connection with its saturation for the large nucleus \[16\]. It can be related to our function $\phi$ via the standard expression for the structure function in its terms (see e.g \[14\]) F\_2(x,Q\^2)= d\^2b d\^2r\[\_T(r)+\_L(r)\] F(x,r,b), where F(x,r,b)=k\^2 (1-e\^[-ikr]{})(1-e\^[ikr]{}) and $\int d^2b (\partial xG(x,k^2,b)/\partial k^2)$ is up to a factor the gluon density in the momentum space: = = d\^2b. In the following we shall study the double density in momentum and impact parameter, which is just $\partial xG(x,k^2,b)/\partial k^2$. Comparing (47), (48) with the corresponding expression in terms of $\phi$, which follows from (39) and (40) we find a relation F(x,r,b)=r\^2(,r,b). Taking a Fourier transform of (48) and neglecting the term proportional to $\delta^2(k)$ we obtain = k\^2\_k\^2(,k,b). This is the desired relation between our function $\phi$ and the gluon density in the combined momentum and impact parameter space.
Applying $k^2\nabla_k^2$ to the found function $\phi$ we thus find the gluon density up to a trivial numerical factor evident from (51)( $\sim 0.76$ with $\alpha_s=0.2$). Function $h(k)=k^2\nabla_k^2\tilde{\phi}(y,k,b)$ for different values of $y$ and $B=0.12$ and 0.02 is shown in Figs. 8,9. Its form at different $y$ results quite remarkable. As one observes, at any given rapidity the found density has the same, roughly Gaussian shape in variable $\xi=\ln
k$, centered at the point $\xi=\xi_0(y)$ very near to $\xi_1$ at which the straight line (38) crosses the $x$-axis (see Section 2). With the growth of $y$ the distribution moves to the right with a nearly constant velocity practically preserving its form. Approximately the distribution can be described by h(k)=h\_0 e\^[-a(-\_0(y))\^2]{}, where $h_0$ and $a$ are practically independent of $y$ and $\xi_0(y)$ linearly grows with it: h\_00.3, a0.3 \_0(y)=\_[00]{}+2.23 . The only quantity which clearly depends on the initial distribution is the starting position $\xi_{00}$, so that for different initial functions the picture in Figs. 8 and 9 shifts along the $\xi$ axis as a whole.
Evidently at a given value of $k$ the density stays always limited, irrespective of the form of the initial distribution (and on the atomic number $A$, in particular). In this sense we have saturation as discussed in \[16\]. However with the growth of $y$ the strongly peaked density moves away toward higher values of $k$ so that the density at a fixed point tends to zero at high values of rapidity. We thus have “supersaturation”: with $y\rightarrow\infty$ the gluon density at an arbitrary finite momentum tends to zero.
Comparing Figs. 8 and 9 one can see how the memory about the initial distribution (except for $\xi_{00}$) is gradually erased in the course of the evolution in $y$. For a very peripheric collison off lead ($B=0.02$) at the initial stages of the evolution the density is correspondingly much smaller than for a central collision ($B=0.12$). However already at $\tilde{y}=2$ the form of the distribution is practically indistinguishable from the central collision, the only remaining difference being the shift along the $\xi$ axis.
Conclusions
===========
The BFKL fan diagram equation has been solved numerically in the large range of rapidities up to $y=50$. The main results are the following.
The idea that the fan diagrams themselves satisfy the unitarity condition has been supported by the fact that the found cross-sections for the scattering of a $q\bar q$ pair off the nucleus tend to a constant value at high rapidities. Since the found cross-sections do not rise with energy, the leading $j$-plane singularity turns out to be a simple pole at $j=1$. The limiting cross-sections prove to be universal: they do not depend on $Q^2$, that is, on the transverse dimension of the $q\bar q$ pair, nor on the initial colour distribution inside the nucleon. Their dependence on the target nucleus thus reduces to a scale factor $R_A^2$. Physically they correspond to scattering of a colour dipole off a black disk.
The nuclear structure function does not saturate at high rapidities, due to the singularity of the transverse distribution $\rho_T(r)$ at $r=0$, which makes it non-normalizable. At large $y$ it continues to grow nearly linearly in $y$.
These results fully agree with predictions made in \[11\] on the basis of the found perturbative solution of the BFKL fan diagram equation and asymptotic estimates made in \[12\].
A completely novel result concerns the gluon density of the nucleus. At sufficiently high rapidities, greater than 10, the gluon density aquires a form of the soliton wave in $y-\ln k$ space, which, with the growth of $y$, moves along the $\ln k$ towards greater $k$ preserving its nearly Gaussian shape. Thus at any finite $k$ the gluon density eventually goes to zero at high enough $k$
Finally we have to stress that all these properties begin to be clearly visible only at very high rapidities and momenta: $y>10$ and $k> 100$ GeV/c with $\alpha_s=0.2$. With smaller $\alpha$’s these values grow correspondingly.
Acknowledgements
================
The author is grateful to Dr.G.P.Vacca who draw his attention to a possibility to simplify the BFKL fan diagram equation. He is also grateful to Prof. Bo Andersson for his interest in this work and stimulating discussions. He thanks the Department for Theoretical Physics of the Lund University for hospitality during his stay in Lund, where this work was completed. The author is finally most thankful to Dr. Yu.Kovchegov, fruitful discussions with whom helped to find and correct an error in the expression for $w_T$ (Eq. (45)) in the first version of this paper.
Appendix. Colour dipole approach
================================
To compare our BFKL fan diagram equation to that of Kovchegov \[10\] we present here a short derivation of Eq. (1) from the colour dipole approach, which will make clear the difference between the two equations.
In the colour dipole approach the single scattering term (2) is presented in the form (y,b)=-2isAT(b)d\^2r(r)d\^2r\_1n\_1(r,r\_1,y)(r\_1), where (r\_1)= -g\^4d\^2r’G(0,r\_1,r’)\_N(r’) and $n_1(r,r_1,y)$ is a single dipole density at rapidity $y$ introduced by A.H.Mueller; $r_1$ and $r$ are the dipole lengths at rapidity $y$ and at $y=0$ respectively, so that n\_1(r,r\_1,y)\_[y=0]{}=\^2(r-r\_1). (Note that as in \[10\] our $n_1$ is A.Mueller’s one divided by $2\pi
r_1^2$.)
To introduce the multidipole densities A.H.Muller constructed a generating functional $Z(r_1,r_0,y|u)$, where $u=u(\rho_i,\rho_f)$ is a function of two dipole endpoints in the transverse space (which, for brevity, Iwe denote by a single symbol $\rho$ for). The functional $Z$ satisfies the following nonlinear equation $$Z(r_1,r_2,y|u)=u(r_1,r_0)e^{2y\omega(r_{10})}+$$\_0\^y dy’e\^[2(r\_[10]{})(y-y’)]{}d\^2r\_2 Z(r\_1,r\_2,y’|u)Z(r\_2,r\_0,y’|u), where $r_{10}=r_1-r_0$ etc and $\omega(r)$ is the gluon trajectory $\omega(q)$ in which the momentum $q$ is substituted by $r$: (r)=- with $\epsilon$ the cutoff at small $r$ (in the ultraviolet). The functional $Z$ is normalized according to Z(r\_1,r\_0,y|u)\_[u=1]{}=1. The $k$-fold inclusive dipole density is given by a $k$-fold derivative of $D$ with respect to $u$ at $u=1$: n\_k(r\_1,r\_0,y;\_1,...\_k)=\_[u=1]{}. At this point we make our first comment as to the comparison with \[10\]. Our form of the functional equation for $Z$ is the same as in the original pater of A.H.Mueller \[1\] and in \[10\], except for a slightly different choice of dipole coordinates and for the order of arguments in the 2nd $Z$ in the non-linear term: their form would correspond to $Z(r_0,r_2,y'|u)$. Comparing with the BFKL equation for the single dipole density, one can verify that our choice is better. However this point is irrelevant for the following, since in the interaction with the nucleus only even functions of $r_{10}$ appear.
The dipoles are to interact with the nucleus target with a zero transferred momentum. If the dimension of the dipole is smaller than the internucleon distance in the nucleus then it will interact with a single nucleon as a whole. This picture lies at the basis of the standard fan diagram approach, where each pomeron finally interacts with a single nucleon. We used precisely this picture in our derivation of Eq. (26) in Section 2. In this picture the only trace of the nucleus will be an additional factor $AT(b)$ multiplying the interaction with the nucleon $\tau(r)$ (Eq. (54)) at a given impact parameter $b$. In \[10\] a different idea is exploited: it is assumed that each of the dipoles can interact with many nucleons. The latter interaction is assumed to have an eikonal form. So the interaction with the nucleus they consider is a two-stage one: first the projectile generates many colour dipoles (BFKL fan diagrams) and then each dipole multiply interacts with the nucleus [*a-la*]{} Glauber. Although technically it is not difficult to take into account the final eikonalization of the interaction, appropriately changing the single scattering term $\tau(r)$ in (55) and the following formulas, we do not think it is reasonable. On the one hand, as we shall see, in the interaction with the nucleus the densities are considerably damped at large distances as compared to the usual BFKL behaviour. The confinement should further restrict their spatial dimensions. So for a nucleus with a large internucleon distance it does not seem reasonable to assume simultaneous interaction of a dipole with two or more nucleons. On the other hand, should such interactions be really important, one cannot expect to correctly describe the interaction with the nucleus of a dipole of a given (and fixed) dimension by the Glauber formula. Note that the expression used in \[9\] for it does not correspond to the BFKL picture for the scattering on a single nucleon. So we take $\tau$ as given by (55) and this is the main difference between our derivation and that of \[10\].
With a chosen $\tau$, the interaction with the nucleus will be described by densities \_k(r\_1,r\_0,y)=\_[j=1]{}\^k(d\^4\_j (\_j)AT(b)) n\_k(r\_1,r\_0,y;\_1,...\_k). Here $\tau(\rho)=\tau(\rho_f-\rho_i)$ depends only on the dipole length and is given by (55).
Differentiating (57) and using (61) one easily obtains $$\nu_1(r_1,r_0,y)=AT(b)\tau(r_{10})
e^{2y\omega(r_{10})}+$$\_0\^y dy’e\^[2(r\_[10]{})(y-y’)]{}d\^2r\_2 \[\_1(r\_1,r\_2,y’)+\_1(r\_2,r\_0,y’)\] and for $k>1$ $$\nu_k(r_1,r_0,y)=
\frac{g^2N_c}{8\pi^3}\int_0^y dy'e^{2\omega(r_{10})(y-y')}\int d^2r_2
\frac{r_{10}^2}{r_{12}^2r_{20}^2}
[\nu_k(r_1,r_2,y')+\nu_k(r_2,r_0,y')$$+\_[j=1]{}\^[k-1]{} \_j(r\_1,r\_2,y’)\_[k-j]{}(r\_2,r\_0,y’)\]. We suppress the evident dependence on the impact parameter $b$. From the structure of the equations and the form of the inhomogeneous term it follows that the densities $\nu(r_1,r_0,y)$ depend only on the initial dipole length $r_{10}$. As a result, the two terms separated from the sum over $j$ on the right-hand side give the same contribution. One then finally finds equations \_1(r\_[10]{},y)=AT(b)(r\_[10]{}) e\^[2y(r\_[10]{})]{}+ \_0\^y dy’e\^[2(r\_[10]{})(y-y’)]{}d\^2r\_2 \_1(r\_[20]{},y’) and for $k>1$ $$\nu_k(r_{10},y)=
\frac{g^2N_c}{4\pi^3}\int_0^y dy'e^{2\omega(r_{10})(y-y')}\int d^2r_2
\frac{r_{10}^2}{r_{12}^2r_{20}^2}
\nu_k(r_{20},y')+$$\_0\^y dy’e\^[2(r\_[10]{})(y-y’)]{}d\^2r\_2 \_[j=1]{}\^[k-1]{}\_j(r\_[12]{},y’)\_[k-j]{}(r\_[20]{},y’).
The total forward scattering amplitude on the nucleus will be given by the expression (54) in which the single dipole interaction $\int d^2r_1n_1(r,r_1,y)\tau(r_1)$ is substituted by the sum of all multidipole interactions $\sum_k\nu_k(r)$. Presenting the amplitude in the form (10) we have (r,y)=-\_[k=1]{}\_k(r,y). Summing Eqs. (64) and (65) over $k$ we obtain an equation for $\Phi$: $$\Phi(r_{10},y)=
-AT(b)\tau(r_{10})e^{2y\omega(r_{10})}+
\frac{g^2N_c}{4\pi^3}\int_0^y dy'e^{2\omega(r_{10})(y-y')}\int d^2r_2
\frac{r_{10}^2}{r_{12}^2r_{20}^2}
\Phi(r_{20},y')-$$\_0\^y dy’e\^[2(r\_[10]{})(y-y’)]{}d\^2r\_2 (r\_[12]{},y’)(r\_[20]{},y’).
Comparing this equation with the one derived in \[10\], apart from a different inhomogeneous term, which was discussed earlier, we find difference in the spatial arguments of the function $\Phi$. Our $\Phi$ depends only on one such argument: the dipole dimension $r_{12}$. The equivalent Kovhegov’s function $N$ depends on two spatial arguments: it depends not only on the dipole dimension but also on its center-of-mass coordinate $b_0$, not to be confused with the impact parameter $b$ which does not enter his equation at all. The dependence on $b_0$ should be governed by the inhomogeneous term, which seems to be independent of $b_0$ (Eq. (6a) of \[10\]). Then the dependence of $N$ on the 2nd argument seems to disappear and Kovchegov’s equation coincides with (67). However this contradicts his initial formula (Eq. (4) of \[10\]) in which one integrates over all $b_0$. Modulo all these (small) inconsistencies and a different inhomogeneous term, our Eq. (67) coincides with Kovchegov’s.
References
==========
1\. A.Mueller, Nucl. Phys.,[**B415**]{} (1994) 373.\
2. A.Mueller and B.Patel, Nucl. Phys.,[**B425**]{} (1994) 471.\
3. M.A.Braun and G.P.Vacca, Eur. Phys. J [**C6**]{} (1999) 147.\
4. R.Peschanski, Phys. Lett. [**B409**]{} (1997) 491.\
5. A.Schwimmer, Nucl. Phys. B94 (1975)445.\
6. L.V.Gribov, E.M.Levin an M.G.Ryskin, Nucl. Phys. [**188**]{} (1981) 555; Phys. Rep. [**100**]{} (1983) 1.\
7. J.Bartels and M.Wuesthoff, Z.Phys., [**C66**]{} (1995) 157.\
8. M.A.Braun, Eur. Phys. J [**C6**]{} (1999) 321.\
9. I.Balitsky, hep-ph/9706411; Nucl. Phys. [**B463**]{} (1996) 99.\
10. Yu. Kovchegov, Phys. Rev [**D60**]{} (1999) 034008.\
11 Yu. Kovchegov, preprint CERN-TH/99-166 (hep-ph/9905214).\
12.E.Levin and K.Tuchin, preprint DESY 99-108, TAUP 2592-99 (hep-ph/9908317).\
13. L.N.Lipatov in: “Perturbative QCD”, Ed. A.H.Mueller, World Sci., Singapore (1989) 411.\
14. M.A.Braun, Eur. Phys. J [**C6**]{} (1999) 343.\
15. N.N.Nikolaev and B.V.Zakharov, Z.Phys. [**C64**]{} (1994) 631.\
16. A.Mueller hep-ph/9904404; 9906322; 9902302.\
Figure captions
===============
Fig. 1. The equation summing fan diagrams. Lines represent pomerons.\
Fig. 2. $\phi$ as a function of momentum at different $y$ (in units $\pi/\alpha_sN$) for central collisions on lead ($B=0.12$, Eq.(359)).\
Fig. 3. $\phi$ as a function of momentum at different $y$ (in units $\pi/\alpha_sN$) for peripheral collisions on lead ($B=0.02$, Eq.(35)).\
Fig. 4. Cross-section $\sigma_L$ (with $e^2\rightarrow 1$) as a function of $y$ (in units $\pi/\alpha_sN$) for different $Q^2$.\
Fig. 5. Cross-section $\sigma_L$ (with $e^2\rightarrow 1$) as a function of $Q^2$ for different $y$ (in units $\pi/\alpha_sN$).\
Fig. 6. The structure function of lead $F_{2A}(x,Q^2)$ as a function of $x$ for different $Q^2$.\
Fig. 7. The structure function of lead $F_{2A}(x,Q^2)$ as a function of $Q^2$ for different $x$.\
Fig. 8. The gluon density (in units $\pi^2N/2\alpha_s$) as a function of momentum at different rapidities $\tilde{y}$ for central collisions on lead ($B=0.12$, Eq. (35)).\
Fig. 9. The gluon density (in units $\pi^2N/2\alpha_s$) as a function of momentum at different rapidities $\tilde{y}$ for peripheral collisions on lead ($B=0.02$, Eq. (35)).
[^1]: An error in this formula lead to some erroneous conclusuions about the $y$-behaviour of the structure function in the original version
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Metrics on the space of sets of trajectories are important for scientists in the field of computer vision, machine learning, robotics and general artificial intelligence. Yet existing notions of closeness are either mathematically inconsistent or of limited practical use. In this paper we outline the limitations in the existing mathematically-consistent metrics, which are based on [@schuhmacher2008consistent], and the inconsistencies in the heuristic notions of closeness used in practice, whose main ideas are common to the CLEAR MOT measures [@keni2008evaluating] widely used in computer vision. In two steps we then propose a new intuitive metric between sets of trajectories and address these problems. First we explain a natural solution that leads to a metric that is hard to compute. Then we modify this formulation to obtain a metric that is easy to compute and keeps all the good properties of the previous metric. In particular, our notion of closeness is the first that has the following three properties: it can be quickly computed, it incorporates confusion of trajectories’ dentity in an optimal way and it is a metric in the mathematical sense.'
author:
-
bibliography:
- 'tracking\_metric.bib'
title: |
A metric for sets of trajectories that is\
practical and mathematically consistent
---
Introduction {#sec:intro}
============
Consider two sets $A$ and $B$ each with multiple trajectories. We denote the set of all trajectories as $\mathcal{T}$ and the set of all finite sets of trajectories as $\mathcal{S}$. For concreteness, we assume that $A = \{A_1,...,A_k\}\in \mathcal{S}$ and $B = \{B_1,...,B_l\}\in \mathcal{S}$ where each trajectory, say $A_1$, is a set of state-time pairs $\{(t_1,x_1),...,(t_n,x_n)\}\in \mathcal{T}$ with $t_i \in \mathbb{N}$ and $x_i \in \mathbb{R}^p$, which we can also represent as $x_i = A_1(t_i)$. In this paper we propose a new distance measure $\mathcal{D}(A,B)$ for $A,B \in \mathcal{S}$.
Measures of distance between sets of trajectories are common, for example, in the field of computer vision tracking. In this case $A$ contains the trajectories of several objects and $B$ is the set of approximate trajectories that a tracking algorithm reproduces from video-data. A state-time pair $(t,x)$ can be, for example, the position and velocity of an object at time $t$. $\mathcal{D}(A,B)$ is important to evaluate the performance of multi-object tracking algorithms and to distinguish a good tracker, $\mathcal{D}(A,B) =$ “small”, from a bad, tracker $\mathcal{D}(A,B) =$ “large”.
Machine learning is another example where a measure of closeness between sets of trajectories is important. Imagine that each datapoint in a data set is a set of the trajectories followed by all players in a team during a football play. Now consider the following three tasks: the unsupervised task of clustering football plays, the supervised task of building a classifier for the different plays and the information retrieval task of finding plays similar to a reference play. We can solve these tasks using algorithms that only require a measure of distance on the sets of trajectories, e.g. [@ganti1999clustering; @kleinberg2002approximation; @yianilos1993data] respectively.
Our goal is that $\mathcal{D}(A,B)$ is both mathematically consistent and useful in practice. A common mathematical inconsistency of similarity measures used in practical applications involving trajectories is that they are not a mathematical metric. In Section \[sec:limitations\_MOT\] we show this is the case with the CLEAR MOT measures [@keni2008evaluating], widely used in computer vision, and hence also with many other distance measures that use heuristics similar to the CLEAR MOT. On the contrary, the measure we introduce satisfies the properties of a metric: for every $A,B,C \in \mathcal{S}$ we have (i) [**(non-negativity)**]{} $\mathcal{D}(A,B) \geq 0$, (ii) [**(coincidence)**]{} $\mathcal{D}(A,B) = 0$ iff $A = B$, (iii) [**(symmetry)**]{} $\mathcal{D}(A,B) = \mathcal{D}(B,A)$ and (iv) [**(sub-additivity)**]{} $\mathcal{D}(A,C) \leq \mathcal{D}(A,B) + \mathcal{D}(B,C)$.
In the context of tracking, it is easy to see why not dealing with a mathematical metric can lead to inconsistencies. Imagine that under a certain distance $\mathcal{D}$ both tracker1’s output, $O_1$, and tracker2’s output, $O_2$, are close to the ground truth, $GT$. In this setting, one intuitively expects that $O_1$ and $O_2$ are also close to each other. This expectation is related to the sub-additivity property $\mathcal{D}(O_1,O_2) \leq \mathcal{D}(O_1,GT) + \mathcal{D}(O_2,GT)$ and we consider its violation an inconsistency. The need for symmetry, non-negativity and the coincidence property is intuitive. Note that the machine learning algorithms mentioned above, namely[@ganti1999clustering; @kleinberg2002approximation; @yianilos1993data], also require $\mathcal{D}$ to be a metric.
In addition to seeking mathematical consistency, our goal is also to define a useful measure. In the examples that follow we again focus on vision tracking because in this context it is clear why the existing mathematically consistent measures, which are based on [@schuhmacher2008consistent], are not useful in practice.
We obviously want $\mathcal{D}(A,B)$ to have a computation time that scales well with the number and length of the trajectories in $A$ and $B$. We show our metric is fast to compute in Section \[sec:num\_res\_run\_time\]. In addition, to be useful, $\mathcal{D}(A,B)$ must behave as expected in scenarios with intuitive answers as, for example, in the scenarios we now explain. These help us understand how a consistent and practical $\mathcal{D}$ should look like, help us introduce relevant previous work, some notation, and also help understanding why our task is not easy.
![Scenarios where defining $\mathcal{D}(A,B)$ is intuitive. The red cross markers [ $\times$]{} are the trajectories from $B$, the blue circle markers [ $\circ$]{} are the trajectories from $A$ and the green square markers [ $\Box$]{} from $C$.](./figures/simple_example_1.eps "fig:"){height="2.9cm"} (-53,65)[**(a)**]{} (-53,-0)[[Time]{}]{} (-100,35) ![Scenarios where defining $\mathcal{D}(A,B)$ is intuitive. The red cross markers [ $\times$]{} are the trajectories from $B$, the blue circle markers [ $\circ$]{} are the trajectories from $A$ and the green square markers [ $\Box$]{} from $C$.](./figures/simple_example_2.eps "fig:"){height="2.9cm"} (-53,65)[**(b)**]{} (-53,-0)[[Time]{}]{} (-100,35)\
![Scenarios where defining $\mathcal{D}(A,B)$ is intuitive. The red cross markers [ $\times$]{} are the trajectories from $B$, the blue circle markers [ $\circ$]{} are the trajectories from $A$ and the green square markers [ $\Box$]{} from $C$.](./figures/simple_example_3.eps "fig:"){height="2.9cm"} (-53,65)[**(c)**]{} (-53,18)[[ $A_1$]{}]{} (-73,38)[[ $A_2$]{}]{} (-53,28)[[ $B_2$]{}]{} (-53,38)[[ $B_1$]{}]{} (-53,-0)[[Time]{}]{} (-100,35) ![Scenarios where defining $\mathcal{D}(A,B)$ is intuitive. The red cross markers [ $\times$]{} are the trajectories from $B$, the blue circle markers [ $\circ$]{} are the trajectories from $A$ and the green square markers [ $\Box$]{} from $C$.](./figures/simple_example_4.eps "fig:"){height="2.9cm"} (-53,65)[**(d)**]{} (-53,-0)[[Time]{}]{} (-80,60)[[ $A_1$]{}]{} (-80,27)[[ $A_2$]{}]{} (-53,52)[[ $B_1$]{}]{} (-53,30)[[ $B_2$]{}]{} (-23,46)[[ $C_1$]{}]{} (-23,36)[[ $C_2$]{}]{} (-100,35)
\[fig:simple\_examples\_about\_metric\]
[**Scenario 1:**]{} In Fig. \[fig:simple\_examples\_about\_metric\]-(a) the set $A$ has only one trajectory $\{(1,-1.0)$, $(2,-0.6)$, $(3,-0.2)$, $(4,0.2)$, $(5,0.6)$, $(6,1.0)\}$ and the set $B$ has only one trajectory $\{(1,-0.9)$, $(2,-0.54)$, $(3,-0.18)$, $(4,0.18)$, $(5,0.54)$, $(6,0.9)\}$. One intuitive way to define $\mathcal{D}(A,B)$ in this case is to form the vectors $[-1.0$, $-0.6$, $-0.2$, $0.2$, $0.6$, $1.0]$ and $[-0.9$, $-0.54$, $-0.18$, $0.18$,$0.54$, $0.9]$ and compute the sum of the distances between each of their components. This gives $\mathcal{D}(A,B) = 0.36$. More generally, we can use any vector norm between these two vectors and, as long as both $A$ and $B$ only have one trajectory of equal length, this procedure defines a metric, $$\mathcal{D}(A,B) = \|A-B\|,$$ where $A$ and $B$ on the right hand side are the vector representation of the trajectories $A$ and $B$.
[**Scenario 2:**]{} In Fig. \[fig:simple\_examples\_about\_metric\]-(b) the sets $A$ and $B$ still only have one trajectory but now the trajectories have “holes”. To define $\mathcal{D}$ for this scenario we can reduce it to Scenario $1$ by computing the total distance between each of the components of the vectors $[-1.0$, \*, $-0.2$, $0.2$, $0.6$, $*]$ and $[*$, $-0.54$, $-0.18$, $0.18$,$0.54$, $*]$ and using the rule that $|x - *| = M > 0$ if $x \in \mathbb{R}$ and that $| * - *| = 0$. For example, if $M = 0.1$ we get $\mathcal{D}(A,B) = 0.3$. We can interpret the symbol \* as total uncertainty in a tracker $A$ regarding the position of an object $B$ at a particular point in time, in which case we penalize its performance by a fixed amount. If tracker $A$ outputs \* because the object $B$ does not exist in that time instant (e.g. it is outside the tracker’s range) we do not penalize its performance, hence $|* - * | = 0$.
More generally, let $A^+ = [...,A^+(t),...]$ and $B^+ = [...,B^+(t),...]$ be the vector representation of the single trajectories in $A$ and $B$, now with each point living in $\mathbb{R}^p$, and padded with the symbol $*$ and let $d^+:\mathbb{R}^p\cup\{*\} \times \mathbb{R}^p\cup\{*\} \mapsto \mathbb{R}$ be an extension of some point-wise metric $d:\mathbb{R}^p \times \mathbb{R}^p \mapsto \mathbb{R}$ such that for every $x,y \in \mathbb{R}$ we have $d^+(x,y) = \min\{2M,d(x,y)\}$, $d^+(x,*)= d^+(*,x)= M > 0$ and $d^+(*,*) = 0$. One can show that $d^+$ is a metric (c.f. Lemma \[th:extdismetric\] in Appendix \[app:proof\_of\_th\_d\_nat\_is\_metric\]) and hence so is $$\mathcal{D}(A,B) = \sum_t d^+(A^+(t),B^+(t)).$$
[**Scenario 3:**]{} For Fig. \[fig:simple\_examples\_about\_metric\]-(c) we cannot define $\mathcal{D}(A,B)$ as easily as in Scenario 1 and 2 because $A$ and $B$ have multiple trajectories. We can try the same approach and reduce $\mathcal{D}$ to a norm in a vector space by (a) writing $A$ as the concatenation of $A_1 = [-0.90,-0.78,-0.66,-0.54,-0.42,-0.30]$ and $A_2 = [-0.7, -0.42, -0.14, 0.14, 0.42, 0.70]$ and (b) writing $B$ as the concatenation of $B_1 = [-1.00, -0.64, -0.28, 0.08, 0.44, 0.80]$ and $B_2 = [-0.60,-0.56, -0.52, -0.48, -0.44,-0.40]$ and (c) computing the total distance between the components of the vectors $A$ and $B$ thus formed. However, this procedure is not well defined because if we concatenate $[A_1,A_2]$ and $[B_1,B_2]$ we get $\mathcal{D}(A,B) = 6.40$ and if we concatenate $[A_2,A_1]$ and $[B_1,B_2]$ we get $\mathcal{D}(A,B) = 1.68$. Effectively we have defined $\mathcal{D}(A,B)$ by comparing each trajectory in $A$ against one trajectory in $B$ but since $A$ and $B$ are unordered sets we do not know if we should compare $A_1$ with $B_1$ or with $B_2$.
One way to resolve this ambiguity is to choose the smallest of the values computed, $1.68$ in our example. That is, we choose the most favorable comparison between trajectories. The authors in [@ristic2011metric] show that this procedure results in a metric. In this paper we call this metric OSPA because it is basically an extension of the OSPA metric of [@schuhmacher2008consistent] from $A$ and $B$ being sets of points to $A$ and $B$ being sets of trajectories.
To write a general expression for OSPA we first extend $A$ and $B$ such that each have $m$ trajectories of length $T$. First we make all trajectories have the same length $T$ by padding them with state-time pairs $(*,t)$. Padded trajectories are represented by $A^+_i$ and $B^+_i$. Afterwards, if $A$ has $k$ trajectories and $B$ has $l$ trajectories we extend $A$ to $A^+ = \{A^+_1,...,A^+_k,A^+_{k+1},...,A^+_m \}$ and we extend $B$ to $B^+ = \{B^+_1,...,B^+_l,B^+_{l+1},...,B^+_m\}$ where $m = k + l$ and where the extra trajectories $A^+_i$, $i > k$, and $B^+_i$, $i > l$, are all of the kind $\{(1,*),...,(T,*)\}$. We write the set of all extended trajectories as $\mathcal{T}^+$. Let $\Pi$ be the set of all permutations of $[m] \equiv \{1,...,m\}$. We represent any $\sigma \in \Pi$ as an ordered set $(\sigma_1,...,\sigma_{m})$ where $\sigma_i \in [m]$.
The OSPA metric is defined as $$\mathcal{D}(A,B) = \min_{\sigma \in \Pi} \sum_{i,t} d^+(A^+_i(t),B^+_{\sigma_i}(t)).$$
This metric requires a procedure to match trajectories in $A$ with trajectories in $B$ for comparison and matching (parts of) trajectories is also a central theme in our paper. Indeed, early on researchers identified finding such a match as a central task in defining a metric for sets of trajectories [@fridling1991performance; @drummond1992ambiguities].
In the context of computer vision tracking it is important that $A$ and $B$ are extended to have have $m = l+k$ trajectories each. Let $A$ be the ground-truth trajectories and $B$ be the tracker’s reconstructed trajectories. We can match any of the $k$ original trajectories in $A$ to a trajectory $B^+_i = \{(1,*),...,(T,*)\}$, $i > l$, to represent that the tracker missed a trajectory completely and we can match any of the $l$ original trajectories in $B$ to a trajectory $A^+_i = \{(1,*),...,(T,*)\}$, $i > k$, to represent the tracker producing a spurious trajectory.
The next scenario shows why OSPA-based metrics can be uninformative in practice.
[**Scenario 4:**]{} For the scenario in Fig. \[fig:simple\_examples\_about\_metric\]-(d), we cannot use an OSPA-based metric to obtain a $\mathcal{D}(A,B)$ that is practical in applications such as performance evaluation in computer vision tracking. Indeed, if we use a procedure that matches full-trajectories to full-trajectories as described in Scenario 3, and after we compute the sum of the absolute values of the difference between components of the matched trajectories, we obtain a value of $\mathcal{D}(A,B) = 7.2$. We obtain this value regardless of whether we compare $A_1= \{(1,1.00)$, $(2, 0.60)$, $(3, 0.20)$, $(4, -0.20)$, $(5, -0.60)$, $(6, -1.00)\}$ with $B_1= \{(1,1.00)$, $(2, 0.60)$, $(3, 0.20)$, $(4, 0.20)$, $(5, 0.60)$ and $A_2 = \{(1,-1.00)$, $(2, -0.60)$, $(3, -0.20)$, $(4, 0.20)$, $(5, 0.60)$, $(6, 1.00)\}$ with $B_2 = \{(1,-1.00)$, $(2, -0.60)$, $(3, -0.20)$, $(4, -0.20)$, $(5, -0.60)$, $(6, -1.00)\}$, or we compare $A_1$ with $B_2$ and $A_2$ with $B_1$. Note however that if we compute $\mathcal{D}(A,C)$ or $\mathcal{D}(B,C)$, where $C$ is an object that is stopped at position $0$, we also get $7.2$. If we interpret $A$ as the trajectories of two objects and $B$ and $C$ as the output of two different trackers, the OSPA-based metric is saying that the trackers are equally good, yet, most people would probably resist equating both because the tracker that outputs $B$ is actually doing a good job at estimating positions but simply makes one confusion of identity of the objects $A_1$ and $A_2$ as they pass by each while the other tracker miss estimates the objects’ positions all the time.
OSPA-based metrics tend to over penalize the quality of a tracker in all situations like in Fig. \[fig:simple\_examples\_about\_metric\]-(d) where one trajectory in $A$ follows closely one trajectory in $B$ but later follows another trajectory in$B$. When this happens we say we have an ***identity switch*** and the fact that OSPA-based metrics cannot deal with identity switches makes then useless in many important applications. For example, imagine that we are tracking hundreds of bats, just like the authors [@betke2007tracking; @betke2008thermal; @wu2009tracking] did, and we use a tracker that has the impressive ability to simultaneously track all the bats flying around a cave very accurately for many minutes. Only once in a while the tracker confuses the identity of two bats that pass by very close to each other. According to OSPA, the performance of this tracker is very poor. However, probably most people agree that its performance is above that of most existing trackers.
It is because of examples as the above that the computer vision tracking community mostly uses measures of closeness like the CLEAR MOT to assess tracking performance. The CLEAR MOT are in fact two measure of closeness, MOTP and MOTA. Although they are not a metric, they do allow to distinguish good trackers from the bad trackers in realistic scenarios. In particular, they allow the user to control how much identity switches should penalize the value of $\mathcal{D}(A,B)$.
The main idea behind the CLEAR MOT is to use a simple heuristic to match *parts* of trajectories in $A$ *to parts* of trajectories in $B$ instead of matching full trajectories in $A$ to full trajectories in $B$ like in the OSPA-base metrics. Let $\Pi^T$ be the set of all sequences of length $T$ of permutations of $[m]$ objects. In other words, if $\Sigma \in \Pi^T$ then $\Sigma = (\sigma(1),\sigma(2),...,\sigma(T))$ where $\sigma(t) \in \Pi, \forall t$. The CLEAR MOT heuristic sequentially builds a sequence $\Sigma_{\text{MOT}} = (\sigma_{{\text{MOT}}}(1),\sigma_{{\text{MOT}}}(2),...,\sigma_{{\text{MOT}}}(T)) \in \Pi^T$ in such a way that $\mathcal{D}(A,B)$, as defined in Definition \[def:MOT\_metric\], is small and at the same time the number of times $\sigma(t)$ is different from $\sigma(t+1)$ is also small.
\[def:MOT\_metric\] The CLEAR MOT’s distance measure for evaluating tracking precision is called MOTP and is defined as $$\mathcal{D}(A,B) = \sum_{i,t} d^+(A^+_i(t),B^+_{\sigma_{\text{MOT}_i}(t)}(t)).$$
The heuristic focuses on simplicity to guarantee that we can compute $\mathcal{D}(A,B)$ fast. Unfortunately, at the expense of being computable, it inherits serious limitations as we explain in Section \[sec:limitations\_MOT\]. It also depends on a user-defined parameter $thr$ that controls how easily $\sigma_{\text{MOT}}(t)$ changes from one time instant to the other. Intuitively, $thr$ allows the user to say how close two objects must be for MOTP to have confidence that they should be compared.
\[def:MOT\_association\] The CLEAR MOT matching heuristic defines $\Sigma_{\text{MOT}}$ sequentially as follows.
1. Initialize $\sigma_{\text{MOT}}(1)$ such that $\sum_{i} d^+(A^+_i(1),B^+_{\sigma_{\text{MOT}_i}(1)}(1))$ is minimal;
2. For each $t>1$ do: for all $i,j\in[m]$ such that $\sigma_{\text{MOT}_i}(t-1)=j$ and $d^+(A^+_i(t),B^+_{j}(t)) < thr$ fix $\sigma_{\text{MOT}_i}(t)=j$. We call such matches as ***anchored***. Define the non-fixed components of $\sigma_{\text{MOT}}(t)$ such that $\sum_{i} d^+(A^+_i(t),B^+_{\sigma_{\text{MOT}_i}(t)}(t))$ is minimal.
What is the value of MOTP for an example like in Fig. \[fig:simple\_examples\_about\_metric\]-(d) if $thr = 0.19$? First we need to compute $\Sigma_{\text{MOT}}$. Let us focus on the pair $A$ and $B$. For $t=1$ we minimize distance by matching $A_1$ to $B_1$ and $A_2$ to $B_2$. This is mathematically equivalent to setting $\sigma_{\text{MOT}}(1) = (1,2)$. For $t=2$ the previous association is anchored because $d^+(A^+_1(1),B^+_{\sigma_{\text{MOT}_1}(1)}(1)) = d^+(A^+_2(1),B^+_{\sigma_{\text{MOT}_2}(1)}(1)) = 0 < thr$ and so $\sigma_{\text{MOT}}(2) = (1,2)$. The same happens for $t=3$. For $t=4$ the previous association cannot be anchored because that would lead to distances larger than $thr$ so we need to re-minimize the distance. This leads to $\sigma_{\text{MOT}}(4) = (2,1)$. For $t=5$ and $t=6$ this association is anchored. In short, we get $\Sigma_{\text{MOT}} = ((1,2),(1,2),(1,2),(2,1),(2,1),(2,1))$. Using this value we compute $\mathcal{D}(A,B) = 0$. We also get $\mathcal{D}(A,C) = \mathcal{D}(B,C) = 7.2$. In other words, MOTP says that, apart from an identity switch, tracker $B$ is better than tracker $C$, which is the intuitive answer we expect. Recall that, as we said before, although MOTP defines a measure of closeness between sets of trajectories that is useful, it is mathematically inconsistent. In other words, there is still no measure that is both mathematically consistent and useful in measuring the distance between sets of multiple trajectories. Our main contribution is to show how to capture all the good properties of all the definitions above, and none of their limitations, by using a definition of the form $$\begin{aligned}
\label{eq:slopydefDnat}
\mathcal{D}(A,B) \hspace{-0.1cm}=\hspace{-0.1cm}
\min_{\Sigma \in \Pi^T} \Big\{ \mathcal{K}(\Sigma)
\hspace{-0.1cm}+\hspace{-0.1cm}\sum^T_{t=1} \sum^{m}_{i=1}d^+(A^+_i(t),B^+_{\sigma_i(t)}(t)) \Big\}\end{aligned}$$ where $\mathcal{K}: \Pi^T \mapsto \mathbb{R}$ is a function that penalizes $\sigma(t)$ changing with time. With this we fill an important gap in the literature,
More comprehensively, our contribution is the following:
1. We show that the CLEAR MOT association heuristic can be unintuitive;
2. We show that MOTP does not define a mathematical metric;
3. We use to define a similarity measure that we prove is a metric and that deals with identity switches optimally;
4. We modify to define the first similarity measure that is fast to compute, is a metric, and is useful in practice;
We now present each of our contributions in order. We start however with an extensive literature review to clearly delineate the gap we will fill.
Note that in this paper we repeatedly make reference to the CLEAR MOT measures and OSPA-based metrics. Not because there are no other metrics beside these two but because they are excellent representatives of the two categories in which most the work done in this area falls: similarity measures that are practical (although even these have problems) but are not a metric and similarity measures that are a metric but are not practical.
Related work {#sec:relwork}
============
In this section we review existing work on metrics for sets of trajectories. Among all the work we review we find no similarity measure between two sets of trajectories $A$ and $B$ that is mathematically consistent (a metric) and, at the same time, is useful and can deal with identity switches (it looks for similarities between parts of trajectories of $A$ and parts of trajectories of $B$). As we argue in the introduction, these are two desired characteristic that we are the first to incorporate and we focus our discussion around them. Several other ideas in the work we review can be easily incorporated into our work and define new metrics that compete with past work beyond these two characteristics. The discussion of these variant metrics is beyond the scope of this paper.
Our work is related to the problem of defining a distance between two sets $A$ and $B$, a topic too vast to review in this paper. In [@deza2009encyclopedia] the reader can find many of these distances. In this paper, the sets $A$ and $B$ are sets of trajectories, and this limits the scope of our discussion. In the simplest case, when trajectories have only one point/vector, typical definitions involve computing an average of the distance between all pairs of elements from $A$ and $B$, e.g. [@fujita2013metrics], or the sum of distance between a few pairs of elements from $A$ and $B$ obtained by some procedure that matches elements of $A$ with elements of $B$, e.g. [@schuhmacher2008consistent; @gardner2014measuring]. In general however, each trajectory is composed by a set of vectors indexed by time, which limits our discussion even more. To the best of our knowledge the most rigorous works on metrics for sets of trajectories are based on the ideas proposed in [@schuhmacher2008consistent]. In [@schuhmacher2008consistent] the authors propose a distance between sets of vectors, the optimal sub-pattern assignment metric (OSPA), and explain its advantages for evaluating the performance of multi-object filters compared to other distances between sets. In particular, the OSPA metric has better sensitivity to cardinality differences between sets than the Hausdorff metric and does not lead to complicated interpretations as the optimal mass transfer metric (OMAT) that [@hoffman2004multitarget] propose to address the limitations of the Hausdorff metric. The OSPA metric between two sets of vectors is well defined for any metric between vectors.
All the spin-offs of [@schuhmacher2008consistent] focus on defining a metric between two sets of trajectories $A$ and $B$ for the purpose of evaluating the performance of tracking algorithms. We recall however that many applications in machine learning and AI apart from tracking benefit if we work with a metric, for example like the ones we define in this paper, rather than a similarity measure that is not a metric. It is crucial to note that all the spin-offs of [@schuhmacher2008consistent] only compare full trajectories in $A$ to full trajectories in $B$ and hence suffer from the same limitations that we describe in Example 4 in the Introduction.
The authors in [@ristic2011metric] define the OSPA-T metric in two steps. In the first step they solve an optimization problem that optimally matches full tracks in $A$ to full tracks in $B$ while taking into account that tracks have different lengths and might be incomplete. In the second step, they assign labels to each track based on this match, they compute the OSPA metric for each time instant using a new metric between pair of vectors that considers both the vectors’ components as well as their labels and they sum all the OSPA values across all time instants. Although the optimization problem of the first step defines a metric, [@vu2014new] point out that the full two-step procedure that defines OSPA-T can violate the triangle inequality.
The authors in [@vu2014new] define the OSPAMT to be a metric and to be more reliable than OSPA-T when we evaluate the performance of multi-target tracking algorithms. The OSPAMT metric also computes an optimal match between full trajectories in $A$ and full trajectories in $B$ but unlike OSPA-T allows to match one full trajectory in $A$ to multiple full trajectories in $B$ (and vice-versa). The authors make this design choice not to penalize a tracker when it outputs only one track for two objects that move closely together.
Some extensions to OSPA incorporate the uncertainty in the measurements. The Q-OSPA metric defined in [@QOSPA2013] incorporates uncertainty by weighting the distance between pairs of points by the product of their certainty and by adding a new term that is proportional to the product of the uncertainties. The H-OSPA metric defined in [@nagappa2011incorporating] incorporates uncertainty by using OSAP with distributions as elements of $A$ and $B$ instead of vectors and using the Hellinger distance between distributions instead of the Euclidean distance between vectors. The authors of both works focus only on the simpler case where the sets $A$ and $B$ contain vectors/points and not trajectories. However, combining their work with that of [@ristic2011metric] or [@vu2014new] to obtain a metric between sets of trajectories is immediate.
The papers above are fairly recent and the search for similarity measures between sets of trajectories that are a metric is not older. However, researchers in the field of computer vision have been interested in defining similarly measures for sets of trajectories to evaluate the performance of tracking algorithms much prior to these works. It is impossible to review all work done in this area. Specially because the evaluation of the performance of trackers has many challenges other than the problem of defining a similarity measure. See [@ellis2002performance; @milan2013challenges] for some examples of these other challenges. Nonetheless, we mention a few works and point out ideas in them that relate to our problem. We emphasize that none of the following works defines a metric mathematically.
Most of our paper is about solving the problem of forming a good match between elements in $A$ and $B$. In [@fridling1991performance], which is expanded in a subsequent paper [@drummond1992ambiguities], the authors are one of the first to identify this as the central problem in defining a similarity measure, although some of their ideas draw from the much earlier Ph.D. thesis of one of the authors [@Drummond75]. They propose a one-to-one association between the different points of $A$ and $B$ but this association is optimally computed independently every instant and there is no discussion about the number of changes in matching that this might create.
One of the reasons why the CLEAR MOT metrics are widely used is because they create a simple association between $A$ and $B$, i.e., the association between $A$ and $B$ does not change often in time. It appears that [@colegrove1996performance] is one of the first works that describes how to control the number of association changes when computing a similarity measure. The authors do not associate $A$ and $B$ independently every time instant but rather use a sequential matching procedure that tries to keep the association from the previous time instant if possible. This is similar to the procedure used in the much more recent CLEAR MOT measures that we discuss in the Introduction.
The association that [@colegrove1996performance] use at every point in time is not one-to-one optimal like in [@fridling1991performance] or in the CLEAR MOT, rather the authors use a simple thresholding rule to associate neighboring elements of $A$ and $B$. The idea of using a simple threshold rule to compare $A$ and $B$ seems to have survived until relatively recent. For example, in [@yin2007performance] the authors match a full trajectory in $A$ to a full trajectory in $B$ if they are close in space for a sufficiently long time interval. The authors in [@bashir2006performance] use a similar thresholding method to match $A$ and $B$.
Shortly after the papers [@fridling1991performance; @drummond1992ambiguities] some of the same authors discuss again the problem of associating $A$ and $B$. In [@drummond1999methodologies] they propose four different methods to solve this problem. In the first method, $A$ and $B$ are optimally matched independently at each time instant, possibly generating different associations every time instant. In the second method, the association between $A$ and $B$ cannot change with time. This is similar to what happens in the OSPA-based metrics. It creates simple associations but can also lead to problems as we discuss in the Introduction. In the third method, the authors allow the association between $A$ and $B$ to change in time but only in special circumstances. Unfortunately, as they point out, this leads to a NP-hard problem. Finally, in the fourth method, the authors discuss some ad-hoc ideas to try to minimize the number of changes in association. Some of these ideas resemble the idea used in the CLEAR MOT of keeping the association from the previous time period if possible.
Not all challenges in defining a measure of performance for trackers are related to defining a similarity measure between trajectories. In addition, not all challenges in defining a similarity measure between trajectories are related to defining a way to compare and associate the trajectories in $A$ with the trajectories in $B$. Even if we assume $A$ and $B$ are already matched, there is the question of computing some quantity from this match. A typical quantity researchers compute is the sum of the distances between the different points of matched trajectories, averaged over all pairs of matched trajectories. As explained later, our metric is a combination of this quantity with the number of confusion of identities in the match between $A$ and $B$.
Apart from other distance-related quantities, researchers also compute many quantities that evaluate the quality of the match itself. For example, the number of trajectories in $A$ that do not match to any trajectory in $B$ or the number of times in which the match between $A$ and $B$ changes. [@rothrock2000performance; @gorji2011performance] list many quantities that can be computed to measure the similarity between $A$ and $B$ and some of these quantities might be useful to define similarity measures for applications outside tracking.
It is worth mentioning a few works that differ from the main stream in this aspect. One of them is [@pingali1996performance], where the authors define a similarity measure based on comparing the occurrence of special discrete events in $A$ and $B$, and another is [@edward2009information], where the authors propose an information theoretic measure of similarity between sets of trajectories. Finally, in [@porikli2004trajectory] the authors propose a similarity measured based on hidden Markov models that does not assume that the temporal sampling rates of the trajectories are equal.
The limitations of the CLEAR MOT {#sec:limitations_MOT}
================================
The CLEAR MOT similarity measures are one of the most commonly used measures that are able to associate parts of trajectories of $A$ with parts of trajectories in $B$ such that the sum of the distance between associated points is small and at the same time the association is simple, i.e. does not change often in time. The CLEAR MOT achieve this using the heuristic procedure in Definition \[def:MOT\_association\]. In Theorem \[th:MOT\_inconsistent\] we show that this heuristic can be unintuitive and in Theorem \[th:MOPT\_no\_metric\] we show that MOTP is mathematically inconsistent.
\[th:MOT\_inconsistent\] For any $\Sigma \in \Pi^T$ and any $A,B \in \mathcal{S}$ without holes and with $m$ trajectories each let $swi(\Sigma) = \frac{1}{T-1}\sum^{T-1}_{t=1} \mathbb{I}(\sigma(t) \neq \sigma(t+1))$ and $dist(\Sigma,A,B) = \frac{1}{T}\sum^T_{t=1} \sum^m_{i=1} \|A_i(t) - B_{\sigma_i(t)}(t)\|$.
For any CLEAR MOT threshold $thr > 0$, there exists $A$, $B$ and $\Sigma$ such that $dist(\Sigma,A,B) < \mathcal{O}(1/T)$, $swi(\Sigma) =0$ and $dist(\Sigma_{\text{MOTP}},A,B) > m\,thr - \mathcal{O}(1/T))$.
In the context of computer vision tracking the theorem above shows that there are situations in which the average tracking performance gets arbitrarily close to optimal with time while the CLEAR MOT association $\Sigma_{\text{MOT}}$ says that it does not. In other words, the CLEAR MOT can be unintuitive.
The following theorem show that, for any $thr > 0$, the MOTP is mathematically inconsistent.
\[th:MOPT\_no\_metric\] MOTP is not a metric.
The proofs of these two theorems are in Appendix \[app:proof\_of\_th\_mota\_bad\_association\].
A natural metric for sets of trajectories {#sec:naturalslowmetric}
=========================================
Despite the large volumes of research about similarity measures for sets of trajectories, largely motivated, at least initially, by the need to assess the performance of tracking systems, our work is the first to propose a measure that is both mathematically consistent and practical.
To build towards our final metric, in this section we define a metric that *captures the best match between the parts of trajectories in $A$ and parts of trajectories in $B$ while tacking into account that simple matches are preferred to complex ones*. As the four scenarios in the introduction suggest, this is a desirable characteristic for a measure of distance and is only partially present in the OSPA and the CLEAR MOT which are representatives of the two currently segregated camps of (a) the consistent metrics and of (b) the practical metrics. Unfortunately the metric we now define is hard to compute and we need an extra section to modify it and reduce computational complexity.
Before we introduce our metric we introduce some new notation.
We denote the identity permutation by $I = (1,2,...,m)$. We also define $\Pi^T$ as all [sequence of permutations]{} of length $T$. In particular if $\Sigma \in \Pi^T$ then $\Sigma = (\sigma(1),\sigma(2),...,\sigma(T))$ where $\sigma(t) \in \Pi, \forall t$. Given $\Sigma \in \Pi^T$, we define $\Sigma^{-1} \in \Pi^T$ as $ \Sigma^{-1} \equiv (\sigma(1)^{-1},\sigma(2)^{-1},...,\sigma(T)^{-1})$ where $\sigma(t)^{-1}$ is the inverse permutation of $\sigma(t)$. In addition, if $\Sigma' = (\sigma'(1),\sigma'(2),...,\sigma'(T))\in \Pi^T$ is another sequence of permutations we define $\Sigma \circ \Sigma' \in \Pi^T$ as $\Sigma \circ \Sigma' \equiv (\sigma(1) \circ \sigma'(1),...,\sigma(T) \circ \sigma'(T))$ where the symbol $\circ$ denotes composition of permutations.
We now introduce a another new definition.
\[th:property\_of\_Sigma\] $\mathcal{K}: \Pi^T \mapsto \mathbb{R}^+_0$ is a map that satisfies the following three properties (i) $\mathcal{K}(\Sigma) = 0$ if and only if $\Sigma$ is constant $\Sigma = (\sigma,\sigma,...,\sigma)$ for some $\sigma \in \Pi$, (ii) $\mathcal{K}(\Sigma^{-1}) = \mathcal{K}(\Sigma)$ and (iii) $\mathcal{K}(\Sigma \circ \Sigma') \leq \mathcal{K}(\Sigma) + \mathcal{K}(\Sigma')$.
We now define a new distance measure.
The [**natural distance**]{} between two sets of trajectories is a map $\mathcal{D}_{nat}: \mathcal{S}\times\mathcal{S}\mapsto\mathbb{R}^+_0$ such that for any $A,B \in \mathcal{S}$ [$$\begin{aligned}
\label{eq:defDnat}
\mathcal{D}_{nat}(A,B) \hspace{-0.1cm}=\hspace{-0.1cm} \min_{\Sigma \in \Pi^T} \Big\{ \mathcal{K}(\Sigma)
+\sum^T_{t=1} \sum^{m}_{i=1}d^+(A^+_i(t),B^+_{\sigma_i(t)}(t)) \Big\}. \end{aligned}$$ ]{}
\[rm:scaling\_factor\] This definition is compatible with scaling the term $\mathcal{K}(\Sigma)$ by some constant $\alpha > 0$. We are not ignoring this possibility but simply put any scaling factor ‘inside’ $\mathcal{K}$. Indeed, if $\mathcal{K}(\Sigma)$ satisfies properties (i), (ii) and (iii) then so does $\alpha \mathcal{K}(\Sigma)$. It is the user’s decision to determine, for the application she has in mind, the relative importance of the terms $\mathcal{K}$ and $d^+$ in .
This definition is related to the CLEAR MOT measure in the sense that it builds an association between parts of trajectories in $A$ and parts of trajectories in $B$ such that the distance between associated trajectories is small and the association does not change often in time. It is the term $\mathcal{K}$ in the definition of $\mathcal{D}_{nat}$ that allows us to control the complexity of the match between the trajectories in $A$ and in $B$. This guarantees that $\mathcal{D}_{nat}$ is as useful in practice as the CLEAR MOT.
This definition is related to the OSPA metric regarding mathematical consistency but is much more general. Indeed, we recover the OSPA metric if we confine ourselves to $\mathcal{K} = \mathcal{K}_{OSPA}$ defined as $\mathcal{K}_{OSPA}(\Sigma) = 0$ if $\Sigma = (\sigma,\sigma,...,\sigma)$ and $\mathcal{K}_{OSPA}(\Sigma) = \infty$ otherwise. Using $\mathcal{K} = \mathcal{K}_{OSPA}$ forces us to compare full trajectories in $A$ to full trajectories in $B$, something that as we explained in the Introduction, leads to unintuitive results.
Strictly speaking our definition does not define a single measure but a family of measures,one for each choice of $d(.)$ and $\mathcal{K}(.)$. The following theorem shows that all the measures in this family are a metric. Its proof is in Appendix \[app:proof\_of\_th\_d\_nat\_is\_metric\].
\[th:Dnaturalismetric\] The map $\mathcal{D}_{nat}$ is a metric on $\mathcal{S}$.
A natural choice for $d$ is the Euclidian metric, that is, $d(x,x') = \|x-x'\|_2$. We now show that several desirable definitions for $\mathcal{K}$ satisfy the properties in Definition \[th:property\_of\_Sigma\]. In the context of computer vision tracking, all these $\mathcal{K}$ allow to penalize trackers that generate many identity switches.
Perhaps the most obvious definition is for $\mathcal{K}$ to count the number of times that the association between $A$ and $B$ changes.
\[th:K\_count\] Let $
\mathcal{K}_{count}(\Sigma) = \alpha \sum^{T-1}_{t=1} \mathbb{I}(\sigma(t+1) \circ \sigma(t)^{-1} \neq I)$, $\alpha > 0$. $\mathcal{K}_{count}$ satisfies Definition \[th:property\_of\_Sigma\].
Another two desirable choices for $\mathcal{K}$ are (a) the function that sums the minimum number of transpositions to go from one permutation to the next and (b) the function that sums the number of adjacent transpositions to go from one permutation to the next. In what follows $k_{Cayley}(\sigma)$ gives the minimum number of transpositions to obtain the identity permutation from $\sigma \in \Pi$ and $k_{Kendall}(\sigma)$ gives the number of adjacent transposition that the bubble-sort algorithm performs when sorting $\sigma$ to obtain the identity permutation. The Cayley distance can be traced back to [@cayley1849lxxvii] and the Kendall distance to[@kendall1938new].
\[th:K\_trans\] Let $\mathcal{K}_{trans}(\Sigma) = \alpha \sum^{T-1}_{t=1} k_{Cayley}(\sigma(t+1)\circ\sigma(t)^{-1})$, $\alpha > 0$. $\mathcal{K}_{trans}$ satisfies Definition \[th:property\_of\_Sigma\].
\[th:K\_adjtrans\] Let $\mathcal{K}_{adjtrans}(\Sigma) = \alpha \sum^{T-1}_{t=1} k_{Kendall}(\sigma(t+1)\circ\sigma(t)^{-1})$, $\alpha > 0$. $\mathcal{K}_{adjtrans}$ satisfies Definition \[th:property\_of\_Sigma\].
The proof of Theorems \[th:K\_count\], \[th:K\_trans\] and \[th:K\_adjtrans\] is in Appendix \[app:proof\_for\_K\_propreties\].
Although many intuitive choices for $\mathcal{K}$ satisfy properties (i), (ii) and (iii) some natural ones to not. For example, given a $\beta \geq 1$ we might want to define $\mathcal{K}_{maxcount}$ as $\mathcal{K}_{maxcount}(\Sigma) = \mathcal{K}_{count}(\Sigma)$ if $\mathcal{K}_{count}(\Sigma) \leq \beta$ and $\mathcal{K}_{maxcount}(\Sigma) = \infty$ if $\mathcal{K}_{count}(\Sigma) > \beta$ (we can replace $\infty$ by some very large number if we want to be technical about the range of $\mathcal{K}$ being $\mathbb{R}^+_0$). This $\mathcal{K}$ basically forces us not to create an association between $A$ and $B$ more complex than a certain amount $\beta$, something natural to desire. The following is proved in Appendix \[app:necessary\_conditions\_for\_D\_nat\_metric\].
\[th:K\_maxcount\_not\_proper\] $\mathcal{K}_{maxcount}$ does not satisfy (iii).
Even if a function $\mathcal{K}$ violates some of the properties it might still be the case that using that function in defines a metric. However, we often find that from a set of associations $\Sigma$, $\Sigma'$ that violate some of the three properties in Definition \[th:property\_of\_Sigma\] we can build three set of trajectories $A$, $B$ and $C$ that violate some of the properties required of a metric. This is the case, for example, for $\mathcal{K}_{maxcount}$ and $d$ the Euclidean metric.
\[th:K\_maxcount\_leads\_to\_not\_metric\] If $\mathcal{K} = \mathcal{K}_{maxcount}$ and $d$ is the Euclidean distance then equation does not define a metric.
The proof of this theorem is in Appendix \[app:necessary\_conditions\_for\_D\_nat\_metric\]. In this sense, the conditions in Definition \[th:property\_of\_Sigma\] are both necessary and sufficient for equation to be a metric.
Some choices for $d$ and $\mathcal{K}$ lead to a metric $\mathcal{D}_{nat}$ that is easy to compute $\mathcal{D}_{nat}$. For example, if we choose $\mathcal{K} = \mathcal{K}_{OSPA}$ and $d$ to be the Euclidean distance then $\mathcal{D}_{nat}$ reduces to OSPA metric which can be solved in polynomial time using, for example, the Hungarian algorithm [@kuhn1955hungarian] (cf. [@schuhmacher2008consistent]). Unfortunately, many desirable choices for $\mathcal{K}$ and $d$ lead to a hard problem. For example, if $d$ is the Euclidean distance and $\mathcal{K} = \mathcal{K}_{count}$ or $\mathcal{K} = \mathcal{K}_{Cayley}$ or $\mathcal{K} = \mathcal{K}_{Kendall}$ then computing $\mathcal{D}_{nat}$ is related to solving a multi-dimensional assignment problem that is HP-hard [@garey2002computers].
We solve this issue in the next section where we modify $\mathcal{D}_{nat}$ to obtain a new metric that is as easy to compute as solving a linear program (recall that all LPs can be solved in polynomial time [@khachiian1979polynomial]).
A natural and computable metric
===============================
First we introduce some notation. Given $A,B \in \mathcal{S}$ we define $A^+$ and $B^+$ just like in Section \[sec:naturalslowmetric\]. We define $T$ and $m$ accordingly. Given $d$ we also define $d^+$ like in Section \[sec:naturalslowmetric\]. Let $\mathcal{P}$ be the set of all doubly stochastic matrices, that is, $\mathcal{P} = \{W \in \mathbb{R}^{m \times m}:
W^{\dagger} {\bf 1} = {\bf 1}, W {\bf 1} = {\bf 1}, 0 \leq W \leq 1\}$. Let $\mathcal{P}^T$ be the set of all sequences of length $T$ of doubly stochastic matrices. Given $A, B \in \mathcal{S}$ and $t \leq T$ we define the matrix $D^{AB}(t) \in \mathbb{R}^{m \times m}$ as $(D^{AB}(t))_{ij} = d^+(A^+_i(t),B^+_j(t))$. Let $\|\|$ be a matrix norm that satisfies the following property for any four matrices $X_1,X_2,Y_1,Y_2 \in \mathcal{P}$ $$\label{eq:propofnormforDcomp}
\|Y_2X_2 - X_2X_1\| \leq \|Y_2 - Y_1\| + \|X_2 - X_1\|.$$ Note that this is analougous to property (iii) that we use for $\mathcal{K}$ in Section \[sec:naturalslowmetric\].
The [**natural computable distance**]{} between two sets of trajectories is a map $\mathcal{D}_{comp}: \mathcal{S}\times\mathcal{S}\mapsto\mathbb{R}^+_0$ such that for any $A,B \in \mathcal{S}$ [ $$\begin{aligned}
\label{eq:defDcomp}
&\mathcal{D}_{comp}(A,B) = \hspace{-0.4cm}\min_{\{W(t)\} \in \mathcal{P}^T} \Big\{ \sum^{T-1}_{t = 1} \|W(t+1) - W(t)\|\nonumber\\
&+ \sum^T_{t=1} \text{\bf tr}(W(t)^{\dagger} D^{AB}(t)) \Big\}. \end{aligned}$$ ]{}
This definition is compatible with scaling the first term by some constant $\alpha> 0$. We are not ignoring this possibility but simply put any scaling factor “inside” the matrix norm. In Section \[sec:num\_res\] we solve for different values of $\alpha$ to obtain a trade-off plot between its two terms.
The motivation for this definition is to obtain a measure that is as similar as possible to $\mathcal{D}_{nat}$ but that we can compute fast. Just like MOTP and $\mathcal{D}_{nat}$, the measure $\mathcal{D}_{comp}$ also builds an association between parts of trajectories in $A$ and parts of trajectories in $B$. This association is represented by the matrices $W(t)$. However, unlike $\Sigma$ in $\mathcal{D}_{nat}$, the $W$ matrices also give us a weight/strength for the different matches between $A$ and $B$. Just like $\mathcal{D}_{nat}$, and as the next theorem shows, $\mathcal{D}_{comp}$ is also mathematically consistent.
\[th:D\_comp\_is\_metric\] The map $\mathcal{D}_{comp}$ is a metric on $\mathcal{S}$.
The proof of this theorem is in Appendix \[app:proof\_that\_D\_comp\_is\_metric\]. Most importantly, unlike $\mathcal{D}_{nat}$, computing $\mathcal{D}_{comp}$ amounts to solving a convex optimization problem. In short, $\mathcal{D}_{comp}$ combines all the advantages of OSPA, MOTP and $\mathcal{D}_{comp}$.
The following Lemma gives sufficient conditions for a metric $\|.\|$ to satisfy property . See Appendix \[app:proof\_that\_norms\_for\_Dcomp\_are\_many\] for a proof.
\[th:many\_norms\_satisfy\_D\_comp\] If $\|.\|$ is a sub-multiplicative norm and $\|W\| \leq 1$ for all $W \in \mathcal{P}$ then $\|.\|$ satisfies property .
This lemma implies, for example, that the 1-norm and $\infty$-norm spectral norm for matrices are valid choices for $\|.\|$. In particular, the use of $\|.\|_1$ in $\mathcal{D}_{comp}$ is extremely useful because it induces the changes of association to be sparse in time and, as the next theorem shows, it reduces $\mathcal{D}_{comp}$ to solving a linear program. The theorem’s proof is in Appendix \[app:proof\_that\_D\_comp\_equals\_LP\].
\[th:dcompisLP\] For any $A, B \in \mathcal{S}$, $\mathcal{D}_{comp}(A,B)$ with $\|.\|$ equal to the matrix 1-norm can be computed (in polynomial time) by solving a linear program.
The metric $\mathcal{D}_{comp}$ allows to consider a scaling factor $\alpha$ in front of the term $\sum_t\|W(t+1)-W(t)\|$, which we call the switch term, $swi$. We call the other term the distance term $dist$. Like the $thr$ value in the CLEAR MOT, $\alpha$ allows us to penalize heavily the error in distances in some applications ($\alpha = small$); and in other applications allows us to penalize heavily the number of associations switches ($\alpha = large$). One can however report how close a set $A \in \mathcal{S}$ is from $B \in \mathcal{S}$ without choosing a specific $\alpha$, simply plot the tradeoff curve between $swi$ and $dist$ as a function of $\alpha$. The closer this curve is to the left/bottom edges of the positive quadrant of the cartesian plane, the closer $A$ to $B$.
The next theorem shows that the tradeoff curve produced by the two terms of $\mathcal{D}_{\text{comp}}$, which we denote by $(swi_{\mathcal{D}_{\text{comp}}}(\alpha),dist_{\mathcal{D}_{\text{comp}}}(\alpha))$, is, in a sense, the best trade-off curve between $dist$ and $swi$ that any metric can generate. To state the theorem we need to define the region $\mathcal{R}$ of pairs of distance-value and switch-value that we can improve upon with some stochastic matrices $\{W(t)\}$ produced by some association procedure of some similarity measure. More concretely, $\mathcal{R} = \{(dist,swi)\in\mathbb{R}^2: swi \geq \sum_t \|W(t+1)-W(t)\| \text{ and } dist \geq \sum_t \text{\bf tr}(W(t)^\dagger D^{AB}(t)) \text{ for all } \{W(t)\} \in \mathcal{P}^T\}$. The proof of following result is in Appendix \[app:proof\_of\_optimal\_trade\_off\_curve\_for\_D\_comp\].
\[th:optimal\_tradeoff\] If $(dis, swi) \in \text{int}(\mathcal{R})$ then there exists $\alpha$, such that $dis + \alpha swi > dis_{\mathcal{D}_{\text{comp}}}(\alpha) + \alpha swi_{\mathcal{D}_{\text{comp}}}(\alpha)$. If $(dis, swi) \in \partial \mathcal{R}$ then $(dis, swi)$ is equal to some point in the curve $(dis_{\mathcal{D}_{\text{comp}}}(\alpha),swi_{\mathcal{D}_{\text{comp}}}(\alpha))$ for some $\alpha$ or is equal to a linear combination of points in the curve.
In the proof of the above theorem we show that $\mathcal{R}$ is convex. Hence, because the boundary of $\mathcal{R}$ is by definition Pareto efficient, this boundary must be a convex non-increasing curve. Since the curve $(dis_{\mathcal{D}_{\text{comp}}}(\alpha),swi_{\mathcal{D}_{\text{comp}}}(\alpha))$ effectively covers this boundary, we can use this property to efficiently estimate the trade-off curve. Basically, we can compute $(dis_{\mathcal{D}_{\text{comp}}}(\alpha),swi_{\mathcal{D}_{\text{comp}}}(\alpha))$ for some discrete set values of $\alpha$ and then we can take the convex hull of these points.
Computing $\mathcal{D}_{comp}$ in practice {#sec:num_res_run_time}
------------------------------------------
There are multiple ways to compute $\mathcal{D}_{comp}$ in practice, all of which amount to solving a convex optimization problem. For this paper we coded a solver in C based on the Alternating Direction Method of Multipliers (ADMM) because the ADMM is known to scale well for large problems and its modular nature makes it easy, in future work, for us to research variants of the formulation $\mathcal{D}_{comp}$ without having to re-write much of our code. For a good introduction to the ADMM, related methods and applications see[@boyd2011distributed]. Our implementation is approximately $\times 50$ faster than using Matlab CVX [@cvx].
In Figure \[fig:profilingcodeforDcomp\] we show the run-time of our code as a function of the total duration $T$ of the input data for a different number of free association variables $W_{ij}(t)$ per time instant $t$. We use the terminology “free association variables” to emphasize that some $W_{ij}(t)$ might be a priory set to zero if the distance between point $i$ and point $j$ at time $t$ is very large, which can save computation time. In the figure we choose the euclidean distance for $d(.,.)$ and the one-norm for matrix norm $\|.\|$ in the switch term. Similar run-times hold for other metrics. The run-time is computed for randomly generated sets of trajectories on a single core of a $1.4$GHz Intel Core i5 MacBook Air. In our tests, the ADMM always converged to $1$% accuracy in less than $150$ iterations.
(-105,23) (-85,-05) (-80,74) (-75,68) (-75,61) (-75,53) (-75,45) (-75,38) (-75,31) (-75,74) (-68,68) (-68,61) (-68,54) (-68,47) (-68,40) (-70,-05) (-12,-05)
To interpret the plots, imagine you want to evaluate the quality of a tracker when tracking $22$ players in a soccer field. Imagine that the tracker operates at $30$ frames per second and also that your tracker is noisy so that it produces a few false tracks that create approximately $10$ extra points per instant. To compute $\mathcal{D}_{comp}$, and after you extend the ground-truth and hypothesis sets from $A$ and $B$ to $A^+$ and $B^+$, you are dealing with distance matrices with about $((22+10)\times 2)^2 = 4096$ variables per instant $t$. Using Fig. \[fig:profilingcodeforDcomp\] you can conclude that it will take you about $40$ seconds to evaluate the accuracy of $800/30 = 26.6$ seconds of data if you use a C implementation of the ADMM on a $1.4$GHz MacBook Air. If you reduce the number of free variables per frame to half by setting $W_{ij}(t)$ to zero if point $i$ and $j$ are larger than a given threshold then you can reduce the time to process $26.6$ seconds of data to about $20$ seconds.
More numerical results {#sec:num_res}
======================
To the best of our knowledge, $\mathcal{D}_{comp}$ is the first similarity measure that is a mathematical metric and deals with identity switches in an optimal way while being computable in polynomial time. In particular, we can use it to generate optimal tradeoff curves between $dist$ and $swi$ (c.f. Theorem \[th:optimal\_tradeoff\]).
To illustrate this point we now use $\mathcal{D}_{comp}$ to assess the closeness between sets $A$ and $B$ both synthetically generated and coming from a tracking application and compare it with MOTP. We make the comparison through $dist$-$swi$ tradeoff plots. For $\mathcal{D}_{comp}$ we use the extended Euclidean metric for $d^+$, the component-wise $1$-norm for $\|.\|$ and the tradeoff curve is parametrized by a multiplying factor $\alpha > 0$ that we put in front of the switch term. For the MOTP, we compute the value for and $swi$ in the same way as for $\mathcal{D}_{comp}$ but using the MOT heuristic association between $A$ and $B$ described in Definition \[def:MOT\_association\]. The MOTP tradeoff curve is parameterized by $thr$.
A direct interpretation of our results is that $\mathcal{D}_{comp}$ is better than MOTP. However, a more important interpretation of our results is that $\mathcal{D}_{comp}$ builds better associations between $A$ and $B$ than the greedy associations that are widely used in the literature and the non-greedy association that do not allow switches, like in OSPA. Because of this, most measures of accuracy, that are dependent on first establishing a correspondence between $A$ and $B$, might be improved if we use the association that $\mathcal{D}_{comp}$ builds. Therefore, although we restrict our numerical comparison to MOTP, indeed it is one of the most widely used metrics in computer vision, we can use our work to also improve, for example, Multiple Object Tracking Accuracy (MOTA), False Alarms per Frame, Ratio of Mostly Tracked trajectories, Ratio of Mostly Lost trajectories, number of False Positives, number of False Negatives, number of ID Switches, number of tracks Fragmentation and many of the measures in [@rothrock2000performance; @gorji2011performance].
Real trackers and real data {#sec:towncenternumerics}
---------------------------
In Figure \[fig:real\_data\_results\]-(a) we show the performance of the trackers in [@benfold2009guiding] and [@benfold2011stable] on the AVG-TownCentre data set. We call these trackers $Tracker09$ and $Tracker11$ respectively. The data set is part of the Multiple Object Tracking Benchmark [@motchallenge] that is widely used in computer vision. It comes from a pedestrian street filmed from an elevated point at a resolution of $1920\times1080$ for $3$ minutes and $45$ seconds and can be downloaded from[@datafortracker09and11]. In Figure \[fig:real\_data\_results\]-(b) we show the performance of the trackers in [@yang2012multi] and [@poiesi2015tracking] on the PETS2009 data set, also part of the Multiple Object Tracking Benchmark. We call these trackers $Tracker12$ and $Tracker15$ respectively. Its resolution is $768\times576$, its duration is $1$ minute and $54$ seconds and it can be downloaded from [@pets2009data; @set]. We produced all the plots using the exact same output that each tracker produced in its respective paper thanks to the authors who provided us with their trackers’ output. For computing we set $M = 20$ for AVG-Towncenter and $M = 50$ for PETS2009.
(-60,95) (-120,35) (-72,01) (-100,34) (-100,28) (-55,22) (-55,16) (-80,64) (-80,58) (-35,64) (-35,58) (-60,95) (-115,35) (-72,01) (-90,24) (-90,18) (-35,46) (-35,40) (-90,44) (-90,38) (-35,69) (-35,63)
\[fig:real\_data\_results\]
As expected from Theorem \[th:optimal\_tradeoff\], the tradeoff curves from MOTP understate the performance of the trackers. In particular, all trackers are achieving substantial smaller number of switches without incurring in larger distance costs that what MOTP reports. Interestingly, for these trackers and datasets $\mathcal{D}_{comp}$ keeps the same relative ordering of performance as MOTP, although it is conceivable that there are situations in which a tracker $1$ is better than a tracker $2$ according to MOTP but not according to $\mathcal{D}_{comp}$. It would be interesting to find such an example. Finally notice that although the difference between the curves is large in absolute value, in relative values the curves are not very different. The absolute values are large because both and $swi$ terms are sums of distances and switches over all tracks and over all time instants and the relative values are small because all trackers are state-of-the-art and perform almost as well as possible.
Random ensemble of trajectories
-------------------------------
In the examples above $A$ and $B$ are relatively close to each other because the trackers are all good trackers. In this section, to better appreciate the behavior of $\mathcal{D}_{comp}$, we randomly generate a set $A$ with $25$ trajectories (slightly more than e.g. the number of players in a football game) and then a set $B$ of trajectories that are a distorted version of the trajectories in $A$. We generate the trajectories in $A$ by randomly starting and ending a trajectory in time and having the object in that trajectory randomly change its velocity’s direction along the way. The trajectories in $B$ are generated by randomly fragmenting the trajectories in $A$, randomly removing some of the resulting trajectories, randomly adding noise to all trajectories and randomly flipping or not the ID of two trajectories if they pass by each other close enough. In the end, $B$ might have more or less than $25$ trajectories. In total we have four knobs to increase/reduce the distance between $A$ and $B$: the amplitude of noise, $AMPnoise$, the probability of fragmenting a track at each point in time, $FRAGprob$, the probability of deleting a points in the track, $DELprob$, and the threshold distance after which we allow to tracks ID to be switched or not randomly, $SWIdist$.
These random trajectories are far more diverse than the trajectories in most publicly available data sets because real objects, like people, have fairly simple trajectories In addition, we do not just test two datasets, like in Section \[sec:towncenternumerics\], but we datasets form about $20$ different levels of distortion for each of the $4$ knobs described above and for each of these levels of distortion we test $30$ random sets $A$ and $B$. Hence, if you will, and in the context of computer vision tracking, it is as if we test $2400$ different data sets of ground-truth and hypothesis trajectories.
We study the similarity between the random sets $A$ and $B$ using $dist$-$swi$ trade-off plots. The smaller the area under the curve of a the trade-off plot, the more similar $A$ and $B$ are. In Figure \[fig:AUC\_evol\_synthetic\_results\] we show the average area under the trade-off curve (AUC) for the random sets $A$ and $B$ under the different knob setting. The AUC is normalized by the largest AUC possible. The largest AUC is the product of the largest distance-error possible with the largest switch-error possible. Each point in the plots is an average over $30$ random pairs $A$ and $B$ with the same knobs setting. In each plot we keep all but one knob constant.
(-108, 45) (-63 , 70)[ MOTP]{} (-50 , 50)[ $\mathcal{D}_{comp}$]{} (-55 , 90)[(a)]{} (-70 , 0)[$AMPnoise$]{} (-108, 45) (-63 , 70)[ MOTP]{} (-50 , 45)[ $\mathcal{D}_{comp}$]{} (-55 , 90)[(b)]{} (-70 , 0)[$DELprob$]{}\
(-108, 45) (-63 , 70)[ MOTP]{} (-50 , 45)[ $\mathcal{D}_{comp}$]{} (-55 , 90)[(c)]{} (-70 , 0)[$FRAGprob$]{} (-105, 45) (-43 , 70)[ MOTP]{} (-50 , 42)[ $\mathcal{D}_{comp}$]{} (-55 , 90)[(d)]{} (-70 , 0)[$SWIdist$]{}
\[fig:AUC\_evol\_synthetic\_results\]
As expected from Theorem \[th:optimal\_tradeoff\], the AUC of $\mathcal{D}_{comp}$ is smaller than the AUC of MOTP. Note that it is incorrect to interpret these results as saying that MOTP is a like a scaled version of $\mathcal{D}_{comp}$. Recall that we use the same $d^+$ to measure the distance between points in space in both MOTP and $\mathcal{D}_{comp}$. We also compute the number of switches that the association of MOTP produces in the same way that we compute the number of switches that the association in $\mathcal{D}_{comp}$ produces. The difference in the curves of Figure \[fig:AUC\_evol\_synthetic\_results\] as a much deeper significance than a simple rescaling: for the $600$ random pairs that we test, which consider different kinds of distortions between $A$ and $B$, and if we interpret each $A$ and $B$ as the ground-truth and output of a tracker, MOTP always says that the tracker is worse than it actually is. The $\mathcal{D}_{comp}$ can see similarities between $A$ and $B$ that the MOTP cannot.
Conclusion
==========
The problem of defining a similarity measure between sets of trajectories is crucial for computer vision, machine learning and general AI. An essential aspect in defining a similarity measure is finding a good and simple association between the elements of the sets. Existing similarity measures that define useful associations fail to be a metric mathematically speaking and the ones that are a metric only consider very primitive associations: full trajectories to full trajectories. The metric $\mathcal{D}_{comp}$ is the first that simultaneously (1) is a metric, (2) compares parts of trajectories to parts of trajectories, (3) allows to control the complexity of the association between trajectories in an globally optimal way and (4) can be computed in polynomial time.
The general idea of defining mathematical metrics for sets of trajectories using convex programs is our greatest overarching contribution. The next step is to explore variants of our metric that allow to incorporate uncertainty, as well as richer comparisons between $A$ and $B$ without losing its good properties.
The limitations of the CLEAR MOT {#app:proof_of_th_mota_bad_association}
================================
We will construct an example for which the theorem is true. Our example proves the theorem for any $1 < thr < 2$ and for $A = \{A_1,A_2\}$ and $B=\{B_1,B_2\}$, i.e. $m=2$, where $A_i$ and $B_i$ are trajectories living in 1D. We explain how to generalize this example to any $thr$ and any $m$ at the end. Consider two sets of one-dimensional trajectories $A = \{A_1,A_2 \}$ and $B = \{B_1, B_2\}$ defined as in Figure \[fig:clearmot\_bad\_metric\] but where the trajectories extend beyond instant $15$ until time $T$.
{height="4.0cm"} (-80,-5)[Time]{} (-154,50) (-35,98) (-35,92) (-35,85) (-35,79) (-35,72) (-35,65)
\[fig:clearmot\_bad\_metric\]
If $1 < thr < 2$ the CLEAR MOT builds a sequence $\Sigma_{\text{MOT}}$ that is equal to $\{(1,2),...,(1,2),(2,1),(2,1),...\}$ because at some instant between time $4$ and time $12$ the initial association $\{A_1 \leftrightarrow B_1, A_2 \leftrightarrow B_2\}$ exceeds $thr < 2$ and is replaced by $\{A_2 \leftrightarrow B_1, A_1 \leftrightarrow B_2\}$ and after instant $12$ this last association is anchored given that $1 < thr$.
The number of times that $\sigma_{\text{MOT}}(t) \neq \sigma_{\text{MOT}}(t+1)$ is $1$ so $swi(\Sigma_{\text{MOT}}) = \frac{1}{T-1} = \mathcal{O}(1/T)$. After instant $12$ we have $\sum_{i} \|A_i(t) - B_{\sigma_{{\text{MOT}}_i}(t)}(t)\| = 2$ so $dist(\Sigma_{\text{MOT}},A,B) > \frac{2(T-12)}{T} = 2 - \mathcal{O}(1/T)$. However, if we choose $\Sigma = \{(1,2),...,(1,2)\}$ we have $swi(\Sigma) = 0$ and $dist(\Sigma,A,B) < \frac{12 \times 7.5}{T} = \mathcal{O}(1/T)$.
We can make the proof hold for an interval around any $thr$ by rescaling space. This changes the bounds on $dist$ by a factor of $thr$ and leaves the bounds on $swi$ unchanged. In addition, we can make the proof hold for any $m$ (even) by extending $A$ and $B$ as follows. If the 1D trajectory $A_i$ above is equal to $\{(1,A_i(1)),...,(T,A_i(T))\}$ define the 2D trajectory $A^{(k)}_i$ to have time-state pairs $\{(1,[A_i(1)); C k],...,(T,[A_i(T); Ck])\}$ where $C$ is a constant large enough to guarantee that trajectories for different $k$s are not close to each other. Define $B^{(k)}_i$ in a similar way. To make the proof hold for any $m$ (even) extend $A$ to $A = \{A^{(0)}_1,A^{(0)}_2,A^{(1)}_1,A^{(1)}_2,...,A^{(m-1)}_1,A^{(m-1)}_2\}$ and extend $B$ in a similar way. To extend the proof for an odd $m$ simply append to $A$ and $B$ two equal trajectories far away from all other trajectories such that they are matched and hence contribute nothing to both $swi$ or $dist$. With these sets the bounds previously computed on $swi()$ change by a factor of $m$ and the bounds on $dist(.)$ change by a factor of $m/2$. In the statement of our theorem statement we are using $\mathcal{O}(m/T) = \mathcal{O}(1/T)$ and $\mathcal{O}(m \, thr/T) = \mathcal{O}(1/T)$.
\[app:proof\_of\_th\_motp\_not\_metric\]
We construct $A,B,C \in \mathcal{S}$ for which the triangle inequality is violated, specifically, $\mathcal{D}(A,B) > \mathcal{D}(A,C) + \mathcal{D}(C,B)$. Our example proves that the theorem holds for $1 < thr < 2$ but by rescaling space the proof follows for any $thr > 0$.
Consider the sets $A = \{A_1, A_2\}$, $B = \{B_1, B_2\}$ and $C = \{C_1, C_2\}$ as in Figure \[fig:clearmot\_bad\_metric\] but where the trajectories extend beyond time $15$ to some large $T$. To make calculations simpler, we work with $\mathcal{D}$ divided by $T$ in Definition \[def:MOT\_metric\].
Let us compute $\mathcal{D}(A,B)$ first. The association $\Sigma_{\text{MOT}}$ for this distance is $\{(1,2),..,(1,2),(2,1),...\}$ because (1) we start with the association $\{A_1 \leftrightarrow B_1,A_2 \leftrightarrow B_2\}$, (2) at some point between instant $4$ and instant $12$ we need to change the association to $\{A_1 \leftrightarrow B_2,A_2 \leftrightarrow B_1\}$ because the initial association exceeds $thr < 2$ and (3) after time $12$ the association $\{A_1 \leftrightarrow B_2,A_2 \leftrightarrow B_1\}$ is anchored because $thr > 1$. This association leads to $\mathcal{D}(A,B) > \frac{2(T-12)}{T}$.
Now we compute $\mathcal{D}(A,C)$. The computation for $\mathcal{D}(C,B)$ is similar so we omit it. The association for $\mathcal{D}(A,C)$ is $\Sigma_{\text{MOT}} = \{(1,2),...,(1,2)\}$ because (1) we start with $\{A_1 \leftrightarrow C_1,A_2 \leftrightarrow C_2\}$, (2) the association $A_1 \leftrightarrow C_1$ is always anchored because the distance between $A_1$ and $C_1$ is always zero and thus always smaller than $thr > 1$ and (3) after some point between instant $4$ and instant $12$, when the distance between $A_2$ and $C_2$ exceeds $thr < 2$, MOTP still keeps the association $A_2 \leftrightarrow C_2$ because $A_1$ and $C_1$ are already anchored. This association leads to $\mathcal{D}(A, C) < \frac{8.5}{T}$.
Therefore, for $T$ large enough we have $\mathcal{D}(A, B) > \frac{2(T-12)}{T} >
\frac{8.5}{T} + \frac{8.5}{T} > \mathcal{D}(A, C) + \mathcal{D}(C, B)$.
A natural metric for sets of trajectories {#app:proof_of_th_d_nat_is_metric}
=========================================
To prove Theorem \[th:Dnaturalismetric\] we need the following Lemma.
\[th:extdismetric\] The map $d^+$ is a metric on $\mathbb{R}^p \cup \{*\}$.
Let $x'',x',x \in \mathbb{R}^p \cup \{*\}$. To prove the coincidence property observe that $d^+(x,x') = 0$ either implies that $x=x'=*$ or, since $M>0$, implies that $d^+(x,x') = d(x,x') = 0$, which, because $d$ is a metric, implies that $x=x'=0$. In other words, $d^+(x,x') = 0 \Leftrightarrow x=x'$. The symmetry property, $d^+(x,x')=d^+(x',x)$, follows directly from the definition. To prove the subadditivity property we need to consider eight different cases. It is trivial to check that the following six cases satisfy the triangle inequality: $x,x',x'' \in \mathbb{R}^p$; $x=x'=*,x'' \in \mathbb{R}^p$; $x''=x'=*,x \in \mathbb{R}^p$; $x=*,x',x'' \in \mathbb{R}^p$; $x''=*,x,x' \in \mathbb{R}^p$; $x=x''=*,x' \in \mathbb{R}^p$. We check the two other cases separately. If $x,x',x'' \in \mathbb{R}^p$ then $d^+(x,x'') = \min\{2M,d(x,x'') \} \leq \min \{2M,d(x,x') + d(x',x'')\} \leq \min\{2M,d(x,x')\} + \min\{2M,d(x',x'')\} = d^+(x,x') + d^+(x',x'')$. If $x'=*, x,x'' \in \mathbb{R}^p$ then $d^(x,x'') = \min\{2M,d(x,x'') \} \leq M + M = d^+(x,x') + d^+(x',x'')$.
Let $A,B,C$ be any three elements in $\mathcal{S}$.\
*Coincidence property*: We show that ${\mathcal{D}_{nat}}(A,B) = 0$ if and only if $A = B$. Remember that $A$ and $B$ are unordered sets of trajectories so $A = B$ means that there is an isomorphism between $A$ and $B$. In other words, they are equal apart tom a relabeling of their elements. If $A = B$ and we set $\Sigma = (\sigma,\sigma,...,\sigma)$ where $\sigma$ is an isomorphism between $A$ and $B$, then the objective in is equal to zero. Since the minimum of must always be non-negative, we conclude that $A = B \Rightarrow {\mathcal{D}_{nat}}(A,B) = 0$. Now assume that ${\mathcal{D}_{nat}}(A,B) = 0$ and let $\Sigma^*=(\sigma^*(1),...,\sigma^*(T))$ be a minimizer of . ${\mathcal{D}_{nat}}(A,B) = 0$ implies that $\mathcal{K}(\Sigma^*) = 0$ and therefore $\sigma^*_i(t) = \sigma^*_i(1)$ for all $t$ and $i$. Since the labeling of the trajectories does not matter in computing ${\mathcal{D}_{nat}}$, we assume without loss of generality that their labeling is such that we can write $\sigma^*_i(t) = i$. ${\mathcal{D}_{nat}}(A,B) = 0$ also implies that, for all $t$ and $i$, we have $d^+(A^+_i(t),B^+_{\sigma^*_i(t)}(t)) = d^+(A^+_i(t),B^+_{i}(t)) = 0$. Since $d^+$ is a metric this in turn implies that $A^+_i(t) = B^+_i(t)$ for all $i$ and $t$, which is the same as saying that $A^+ = B^+$. If $A \neq B$ then $A^+ \neq B^+$ therefore $A^+ = B^+$ implies that $A = B$. To be more specific, $A$ is equal to $B$ apart from a relabeling of its trajectories, which we can do because $A$ and $B$ are unordered sets of trajectories.\
*Symmetry property*: ${\mathcal{D}_{nat}}$ only depends on $A$ and $B$ through $d^+$ and, like all metrics, this is a symmetric function. In addition, if we swap $i$ and $\sigma_i(t)$ in the minimum of remains unchanged. It follows that ${\mathcal{D}_{nat}}(A,B) = {\mathcal{D}_{nat}}(B,A)$.\
*Subadditivity property*: We prove that ${\mathcal{D}_{nat}}(A,C) \leq {\mathcal{D}_{nat}}(A,B) + {\mathcal{D}_{nat}}(B,C)$. Since $d^+$ is a metric, we can write that $$\begin{aligned}
\label{eq:addingBtoproofDnat}
&d^+(A^+_i(t),C^+_{\sigma_i(t)}(t)) \leq d^+(A^+_i(t),B^+_{\sigma'_{\sigma_i(t)}(t)}(t))\nonumber\\
& + d^+(B^+_{\sigma'_{\sigma_i(t)}(t)}(t),C^+_{\sigma_i(t)}(t))\end{aligned}$$ for any $\Sigma' = (\sigma'(1),...,\sigma'(T)) \in \Pi^T$ and for all $i$ and $t$. Now notice that $$\begin{aligned}
\sum^m_{i=1}d^+(B^+_{\sigma'_{\sigma_i(t)}(t)}(t),C^+_{\sigma_i(t)}(t))
= \sum^m_{i=1}d^+(B^+_{\sigma'_{i}(t)}(t),C^+_{i}(t)).\end{aligned}$$ Using this together with we can write $$\begin{aligned}
&\sum^T_{t=1} \sum^m_{i=1} d^+(A^+_i(t),C^+_{\sigma_i(t)}(t))
\leq \sum^T_{t=1} \sum^m_{i=1}d^+(A^+_i(t),B^+_{\sigma'_{\sigma_i(t)}(t)}(t))\\
&+ \sum^T_{t=1} \sum^m_{i=1}d^+(B^+_{\sigma'_{i}(t)}(t),C^+_{i}(t)).\end{aligned}$$ Let us define $\Sigma'' = (\sigma''(1),...,\sigma''(T)) \in \Pi^T$ where $\sigma''_i(t) = \sigma'_{\sigma_i(t)}(t)$. This means that $\Sigma'' = \Sigma' \circ \Sigma$. We can use $\Sigma'$ to rewrite the above expression as $$\begin{aligned}
\label{eq:distanceineqproofDnat}
&\sum^T_{t=1} \sum^m_{i=1} d^+(A^+_i(t),C^+_{\sigma_i(t)}(t))
\leq \sum^T_{t=1} \sum^m_{i=1}d^+(A^+_i(t),B^+_{\sigma''_{i}(t)}(t))\nonumber\\
&+ \sum^T_{t=1} \sum^m_{i=1}d^+(B^+_{\sigma'_{i}(t)}(t),C^+_{i}(t)).\end{aligned}$$ If we use the second and third property that $\mathcal{K}$ satisfies we can also write $$\begin{aligned}
\label{eq:matchineqproofDnat}
&\mathcal{K}(\Sigma) = \mathcal{K}(\Sigma'^{-1} \circ \Sigma' \circ \Sigma) \leq \mathcal{K}(\Sigma'^{-1}) + \mathcal{K}(\Sigma' \circ \Sigma) \nonumber\\
&= \mathcal{K}(\Sigma') + \mathcal{K}(\Sigma'').\end{aligned}$$ Now we add both sides of and and obtain $
\mathcal{K}(\Sigma) + \sum^T_{t=1} \sum^m_{i=1} d^+(A^+_i(t),C^+_{\sigma_i(t)}(t))
\leq \mathcal{K}(\Sigma'') + \sum^T_{t=1} \sum^m_{i=1}d^+(A^+_i(t),B^+_{\sigma''_{i}(t)}(t))
+ \mathcal{K}(\Sigma')+\sum^T_{t=1} \sum^m_{i=1}d^+(B^+_{\sigma'_{i}(t)}(t),C^+_{i}(t)).
$ Finally, we find the minimum of both sides of the inequality over all pairs of $\Sigma$ and $\Sigma'$. Recall that we can choose $\Sigma'$ independently of $\Sigma$. Since for every pair $(\Sigma,\Sigma')$ there is exactly one pair $(\Sigma',\Sigma'')$, finding the minimum over $\Sigma$ and $\Sigma'$ is the same as finding the minimum over $\Sigma'$ and $\Sigma''$. Therefore we get, $$\begin{aligned}
&\min_{\Sigma \in \Pi^T} \mathcal{K}(\Sigma) + \sum^T_{t=1} \sum^m_{i=1} d^+(A^+_i(t),C^+_{\sigma_i(t)}(t))\\
& \leq \min_{\Sigma''\in \Pi^T} \mathcal{K}(\Sigma'') + \sum^T_{t=1} \sum^m_{i=1}d^+(A^+_i(t),B^+_{\sigma''_{i}(t)}(t))\\
& + \min_{\Sigma'\in \Pi^T} \mathcal{K}(\Sigma')+\sum^T_{t=1} \sum^m_{i=1}d^+(B^+_{\sigma'_{i}(t)}(t),C^+_{i}(t)),\end{aligned}$$ which is the same as ${\mathcal{D}_{nat}}(A,C) \leq {\mathcal{D}_{nat}}(A,B) + {\mathcal{D}_{nat}}(B,C)$.
\[app:proof\_for\_K\_propreties\]
The first two properties are immediate to check. To prove the third property it sufficient to prove that $\mathbb{I}((\sigma'(t+1) \circ \sigma(t+1)) \circ (\sigma'(t) \circ \sigma(t))^{-1} \neq I) \leq \mathbb{I}(\sigma(t+1) \circ \sigma(t)^{-1} \neq I) +\mathbb{I}(\sigma'(t+1) \circ \sigma'(t)^{-1} \neq I).$ To prove this we consider different cases. If $\sigma'(t+1) = \sigma'(t)$ and $\sigma(t+1) = \sigma(t)$ then the inequality holds as $0 \leq 0 + 0$. If $\sigma'(t+1) \neq \sigma'(t)$ and $\sigma(t+1) \neq \sigma(t)$ then the inequality holds since the left hand side is at most $1$ and the right hand side in this case is equal to $1 + 1$. If $\sigma(t+1) = \sigma(t)$ and $\sigma'(t+1) \neq \sigma'(t)$ then the inequality holds as $1 \leq 0 + 1$. The final case is when $\sigma(t+1) \neq \sigma(t)$ and $\sigma'(t+1) = \sigma'(t)$. Notice that $\mathbb{I}((\sigma'(t+1) \circ \sigma(t+1)) \circ (\sigma'(t) \circ \sigma(t))^{-1} \neq I) = \mathbb{I}( \sigma(t+1) \circ \sigma(t)^{-1} \neq \sigma'(t+1)^{-1}\circ\sigma'(t))$ and conclude that the inequality holds with $1 \leq 1 + 0$.
We prove the theorem for $\mathcal{K}_{trans}$. The proof for $\mathcal{K}_{adjtrans}$ is very similar. The first two properties are immediate to check from the definition of $\mathcal{K}_{trans}$. To prove the third property it suffices to show that $k_{Cayley}((\sigma'(t+1) \circ \sigma(t+1)) \circ (\sigma'(t) \circ \sigma(t))^{-1}) \leq k_{Cayley}(\sigma(t+1)) \circ \sigma(t)^{-1}) + k_{Cayley}(\sigma'(t+1) \circ \sigma'(t))^{-1}).$ To prove this we use of the following two facts. Fact 1: if $\sigma,\sigma' \in \Pi$ then $k_{Cayley}(\sigma\circ\sigma') \leq k_{Cayley}(\sigma) + k_{Cayley}(\sigma')$. Fact 2: if $\sigma,\sigma' \in \Pi$ then $k_{Cayley}(\sigma\circ \sigma' \circ\sigma^{-1}) = k_{Cayley}(\sigma' )$. Both facts are obvious if stated in plain english. Fact 1: The shortest set of transpositions that takes $\sigma\circ\sigma'$ to $I$ is shorter than any set of transpositions that first takes $\sigma\circ\sigma'$ to $\sigma'$ and then takes $\sigma'$ to $I$. Fact 2: The shortest set of transpositions that takes a permutation of objects to the identity permutation is as short as the set of transpositions that take a relabeling of those objects to the identity permutation.
Let $\sigma_A=\sigma(t)\circ \sigma(t+1)^{-1}$, $\sigma_B = \sigma'(t)\circ \sigma'(t+1)^{-1}$ and let $\sigma_C=\sigma'(t)\circ \sigma_A\circ \sigma'(t)^{-1}$. Observe that the permutation $\sigma_C \circ \sigma_B$ satisfies $$\sigma_C \circ \sigma_B \circ ((\sigma'(t+1) \circ \sigma(t+1)) \circ (\sigma'(t) \circ \sigma(t))^{-1}) = I.$$ Therefore we can apply Fact 2 followed by Fact 1 and obtain $
k_{Cayley}((\sigma'(t+1) \circ \sigma(t+1)) \circ (\sigma'(t) \circ \sigma(t))^{-1}) = k_{Cayley}(\sigma^{-1}_B \circ \sigma^{-1}_C)
\leq k_{Cayley}(\sigma^{-1}_B) + k_{Cayley}( \sigma^{-1}_C) = k_{Cayley}(\sigma^{-1}_B) + k_{Cayley}( \sigma^{-1}_A)
= k_{Cayley}(\sigma(t+1)\circ\sigma(t)^{-1}) + k_{Cayley}(\sigma'(t+1)\circ\sigma'(t)^{-1}).
$
\[app:necessary\_conditions\_for\_D\_nat\_metric\]
We give a counter example that violates property (iii) for $\beta = 1$. It is easy to come up with similar counter examples that violate property (iii) for any value of $\beta \geq 1$. Let $I = (1,2)$ be the identity permutation and let $\sigma_0 = (2,1)$ be the permutation that swaps $1$ and $2$. Let $\Sigma = (I,\sigma_0,\sigma_0)$ and let $\Sigma' = (I,I,\sigma_0)$. We have $\mathcal{K}_{maxcount}(\Sigma) = 1$ and $\mathcal{K}_{maxcount}(\Sigma') = 1$ but $\mathcal{K}_{maxcount}(\Sigma' \circ \Sigma) = \infty > \mathcal{K}_{maxcount}(\Sigma') + \mathcal{K}_{maxcount}(\Sigma)$.
We produce three sets of trajectories $A$, $B$ and $C$ that make violate the triangle inequality. In this proof we assume $\beta = 1$ but it is easy to change $A$, $B$ and $C$ such that the proof holds for any $\beta$,[^1]. Let $I = (1,2)$ be the identity permutation and let $\sigma_0 = (2,1)$ be the permutation that swaps $1$ and $2$. Let $A = \{A_1,A_2\}$ where $A_1 = \{(1, 2), (2, -2),(3, -2)\}$ and $A_2 = \{(1, -2), (2, 2),(3, 2)\}$, let $B = \{B_1,B_2\}$ where $B_1 = \{(1,2), (2,2),(3,2)\}$ and $B_2 = \{(1,-2), (2,-2),(3,-2)\}$ and let $C = \{C_1,C_2\}$ where $C_1 = \{(1, 2), (2, 2),(3, -2)\}$ and $C_2 = \{(1, -2),$ $(2, -2),$ $(3, 2)\}$. Now consider with $\mathcal{K} = \mathcal{K}_{maxcount}$ and $\beta = 1$. The optimization problem $\mathcal{D}_{nat}(A,B)$ has a minimum of $1$ at $\Sigma = (I,\sigma_0,\sigma_0)$. The optimization problem $\mathcal{D}_{nat}(B,C)$ has a minimum of $1$ at $\Sigma = (I,I,\sigma_0)$. When we solve the optimization problem $\mathcal{D}_{nat}(A,C)$ we are only allowed to perform one change in the association between $A$ and $C$, otherwise the term $\mathcal{K}_{maxcount}$ makes us pay a very large cost. With only one change in the association, we incur in a distance of $4$ for $t = 1$ or $t = 2$ or $t = 3$. Hence, $\mathcal{D}_{nat}(A,C) \geq 4 > 1 + 1 = \mathcal{D}_{nat}(A,B) + \mathcal{D}_{nat}(B,C)$.
A natural and computable metric {#app:proof_that_D_comp_is_metric}
===============================
Let $A, B$ and $C$ be any three elements of $\mathcal{S}$.\
*Coincidence property*: If $\mathcal{D}_{comp}(A,B) = 0$ then $W(t) = W(1)$ for all $t$. Hence, if $\mathcal{D}_{comp}(A,B) = 0$ then $D^{AB}(t)_{ij} = 0$ for all $i,j$ such that $W(1)_{ij} > 0$. Now recall that if $W_{ij}$ is a doubly stochastic matrix then there exists a permutation $\sigma \in \Pi$ such that $W_{i\sigma_i} > 0$ for all $i$. Therefore, there exist $\sigma \in \Pi$ such that $D^{AB}(t)_{i\sigma_i} = 0$ for all $t$ and $i$. In other words, $A$ and $B$ are the same apart from a relabeling of their elements.\
*Symmetry property*: Using the properties of [**tr**]{} we have that $\text{\bf tr}(W^{\dagger}(t)D^{AB}(t)) =$ $\text{\bf tr}((W^{\dagger}(t)$ $D^{AB}(t))^{\dagger})$ $= \text{\bf tr}(D^{AB}(t)^{\dagger}W(t))= \text{\bf tr}(D^{BA}(t)W(t))= \text{\bf tr}(W(t) D^{BA}(t))$. Minimizing with respect to $\{W(t)\}$ is the same as minimizing with respect to $\{W(t)^{\dagger}\}$ so $\mathcal{D}_{comp}(A,B) = \mathcal{D}_{comp}(B,A)$.\
*Subadditivity property*: We prove that $\mathcal{D}_{comp}(A,C) \leq \mathcal{D}_{comp}(A,B) + \mathcal{D}_{comp}(B,C)$. First note that, since $d^+$ is a metric we have that $D^{AC}(t)_{ij} \leq D^{AB}(t)_{ik} + D^{BC}(t)_{kj}$ for any $k$. Let $\{W_1(t)\}, \{W_2(t)\} \in \mathcal{P}^T$. We multiply both sides of the previous inequality by $W_1(t)_{ik}W_2(t)_{kj}$ and sum over $ijk$ and obtain $
\sum^{m}_{i,j,k=1} W_1(t)_{ik}W_2(t)_{kj} D^{AC}(t)_{ij}\leq \sum^{m}_{i,j,k=1} W_1(t)_{ik}W_2(t)_{kj} D^{AB}(t)_{ik}\\
+ \sum^{m}_{i,j,k=1} W_1(t)_{ik}W_2(t)_{kj}D^{BC}(t)_{kj}.
$ Note that $W_1(t)$ and $W_2(t)$ are doubly stochastic matrices therefore we have $\sum^m_{j=1} W_2(t)_{kj} = 1$ and $\sum^m_{i=1} W_1(t)_{ik} = 1$ and we can re-write the previous inequality in matrix for as $
\text{\bf tr}((W_1(t)W_2(t))^{\dagger} D^{AC}(t)) \leq
\text{\bf tr}(W_1(t)^{\dagger} D^{AB}(t)) +
\text{\bf tr}(W_2(t)^{\dagger} D^{BC}(t)).
$ From the definition of $\|\|$ we have that $\|W_1(t+1)W_2(t+1) - W_1(t)W_2(t)\| \leq \|W_1(t+1) - W_1(t)\| + \|W_2(t+1) - W_2(t)\|.$ Summing both sides of the last two inequalities and, summing over $t$, and minimizing over $\{W_1(t)\}, \{W_2(t)\} \in \mathcal{P}^T$ we have that [$$\begin{aligned}
&\min_{\{W_1(t)\},\{W_2(t)\} \in \mathcal{P}^T} \sum^{T-1}_{t=1} \|W_1(t+1)W_2(t+1) \\
&- W_1(t)W_2(t) \| +
\sum^T_{t=1} \text{\bf tr}((W_1(t)W_2(t))^{\dagger} D^{AC}(t))\\
&\leq \min_{\{W_1(t)\} \in \mathcal{P}^T}
\sum^{T-1}_{t=1} \|W_1(t+1)- W_1(t) \|+
\sum^T_{t=1} \text{\bf tr}(W_1(t)^{\dagger} D^{AB}(t)) \\
&+\min_{\{W_2(t)\} \in \mathcal{P}^T}
\sum^{T-1}_{t=1} \|W_2(t+1)- W_2(t) \|
+
\sum^T_{t=1} \text{\bf tr}(W_2(t)^{\dagger} D^{BC}(t)).\end{aligned}$$ ]{} Note that for any $W_1(t), W_2(t) \in \mathcal{P}$ we have that $W_1(t) W_2(t) \in \mathcal{P}$ so in the minimization performed on the left hand side we can replace $W_1(t) W_2(t)$ by a single $W(t)$. The subadditivity property follows.
\[app:proof\_that\_norms\_for\_Dcomp\_are\_many\]
Let $X_1,X_2,Y_1, Y_2 \in \mathcal{P}$. By an application of the triangle inequality followed by two applications of the sub-multiplicative property followed by two applications of the norm-bound property we have that $
\|Y_2X_2 - Y_1 X_1\| \leq \|Y_2(X_2 -X_1 + X_1)- Y_1 X_1\|
\leq \|Y_2(X_2 -X_1)\| + \| Y_2 X_1 - Y_1 X_1\|
=
\|Y_2(X_2 -X_1)\| + \| (Y_2 -Y_1)X_1\|
\leq
\|Y_2\| \|(X_2 -X_1)\| + \| (Y_2 -Y_1)\| \|X_1\|
\leq \|(X_2 -X_1)\| + \| (Y_2 -Y_1)\|$.
\[app:proof\_that\_D\_comp\_equals\_LP\]
Let us look at the optimization problem more closely. First notice that the constraints that define $\mathcal{P}$ are a set of linear constraints. Second notice that the first term in the objective is a linear function of $\{W(t)\}$. Third notice that we can replace the term $\sum^{T-1}_{t=1} \|W(t+1) - W(t)\|$ in the objective by $\sum^{T-1}_{t=1} e_t$ if we add the additional constraints that $\|W(t+1) - W(t)\| \leq e_t$ for all $t$. Now we show that these additional constraints can be represented by linear constraints. Since $\|W(t+1) - W(t)\| = \max_{j} \sum_i |W(t+1)_{ij} - W(t)_{ij}|$ each of these constraints can be replaced by $\sum_i h(t)_{ij} \leq e_t$ for all $t$ and $j$ if we add the additional constrain that $|W(t+1)_{ij} - W(t)_{ij}| \leq h(t)_{ij}$ for all $i$, $j$ and $t$. Each of these constrains can be written as $W(t+1)_{ij} - W(t)_{ij} \leq h(t)_{ij}$ and $-W(t+1)_{ij} + W(t)_{ij} \leq h(t)_{ij}$ and both of these constraints are linear so we are done.
\[app:proof\_of\_optimal\_trade\_off\_curve\_for\_D\_comp\]
We first prove that $\mathcal{R}$ is convex. Let $(a,b) \in \mathcal{R}$ and $(c,d) \in \mathcal{R}$. We have that $a \geq \sum_t \|W_1(t+1)-W_1(t)\|$ and $b \geq \sum_t \text{\bf tr}(W_1(t)^\dagger D^{AB}(t))$ and $c \geq \sum_t \|W_2(t+1)-W_2(t)\|$ and $d \geq \sum_t \text{\bf tr}(W_2(t)^\dagger D^{AB}(t))$ for some $\{W_1(t)\},\{W_2(t)\} \in \mathcal{P}^T$. Let $W_3(t) = \beta W_1(t) + (1-\beta) W_2(t)$ for some $\beta \in [0,1]$ and all $t$. Since $\mathcal{P}^T$ is a convex set we have that $W_3(t) \in \mathcal{P}^T$. In addition, since the functions $\sum_t \|W(t+1)-W(t)\|$ and $\sum_t \text{\bf tr}(W(t)^\dagger D^{AB}(t))$ are convex functions of $\{W(t)\}$ we have that $\beta a + (1-\beta) c \geq \sum_t \|W_3(t+1)-W_3(t)\|$ and $\beta b + (1-\beta) d \geq \sum_t \text{\bf tr}(W_3(t)^\dagger D^{AB}(t))$. Therefore, we conclude that $\beta (a,b) + (1 - \beta) (c,d) \in \mathcal{R}$.
We now prove the rest of the theorem. Since $\mathcal{R}$ is convex, we can write it as the following intersection of half-planes $\mathcal{R} = \{(x,y)\in \mathbb{R}^2:x + \alpha y \geq \min_{\{W(t)\} \in \mathcal{P}^T} \alpha u + v \text{ s.t. } u \geq \sum_t \|W(t+1)-W(t)\| \text{ and } v \geq \sum_t \text{\bf tr}(W(t)^\dagger D^{AB}(t)) \text{ for all } \alpha \in [0,\infty] \}$.\
From this representation it directly follows that, for any $\alpha \in [0,\infty]$, any point $(dis,swi) \in \text{int}(\mathcal{R})$ satisfies $
\alpha swi + dis > \text{minimum } \big \{ \alpha u + v \text{ subject to }
W(t)\in \mathcal{P}^T; u \geq \sum_t \|W(t+1)-W(t)\|;
v \geq \sum_t \text{\bf tr}(W(t)^\dagger D^{AB}(t))\big\}
= \alpha \sum_t \|W^*(t+1)-W^*(t)\|
+ \sum_t \text{\bf tr}(W^*(t)^\dagger D^{AB}(t))= \alpha swi_{\mathcal{D}_{\text{comp}}}(\alpha)
+ dis_{\mathcal{D}_{\text{comp}}}(\alpha),
$ where $\{W^*(t)\}\in \mathcal{P}^T$ is the optimum point of the optimization problem of $\mathcal{D}_{comp}$.
From this representation we also conclude that any point at the boundary of $\mathcal{R}$ is either of the form $(\sum_t \|W^*(t+1)-W^*(t)\|, \sum_t \text{\bf tr}(W^*(t)^\dagger D^{AB}(t))) = (dis_{\mathcal{D}_{\text{comp}}}(\alpha),swi_{\mathcal{D}_{\text{comp}}}(\alpha))$ for some optimal $\{W^*(t)\} \in \mathcal{P}^T$ and some $\alpha$ or it is a convex combination of points of this form.
[^1]: It is also not hard to see that we do not really need $d$ to be the Euclidean distance for the proof to hold.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In an era of increasingly advanced experimental analysis techniques it is crucial to understand which phase space regions contribute a signal extraction from backgrounds. Based on the Neyman-Pearson lemma we compute the maximum significance for a signal extraction as an integral over phase space regions. We then study to what degree boosted Higgs strategies benefit $ZH$ and $t\bar{t}H$ searches and which transverse momenta of the Higgs are most promising. We find that Higgs and top taggers are the appropriate tools, but would profit from a targeted optimization towards smaller transverse momenta. MadMax is available as an add-on to Madgraph5.'
author:
- Tilman Plehn
- Peter Schichtel
- Daniel Wiegand
title: 'MadMax, or Where Boosted Significances Come From'
---
Maximum significance {#sec:intro}
====================
The recent Higgs discovery has shown that modern analysis techniques have become standard in high energy physics. Such techniques go beyond simple event counting in phase space regions which have been identified as signal–rich ahead of time. Multi-variate strategies with ten or more kinematic observables seem to make it impossible for the experimental collaborations to provide sufficient information on the behavior of each individual observable and their correlation in the signal and background phase spaces. This means that our research field would benefit from a compact tool to study the leading effects in the computation of quoted significances from multi–variate analyses [@orig].
More specifically, when analyses like Higgs coupling measurements start to be limited by theory uncertainties we need to clearly identify the phase space regions which carry the analysis result [@sfitter]. Moreover, recent progress in boosted Higgs and top studies [@bdrs; @tth; @buckets] has shown that identifying the appropriate phase space patterns can trigger the development of entirely new, specialized analysis objects like fat jets [@top_tagger] or deconstructed parton showers [@deconstruction]. Again, this points to the need of a fast Monte–Carlo tool which can reliably identify those phase space regimes which are critical to separating a given signal from backgrounds [@orig].
In <span style="font-variant:small-caps;">MadMax</span> we will not attempt include all detector effects, because matrix element techniques including appropriate transfer functions are hugely computer intensive. For the same reason, we will limit ourselves to the parton level, assuming that the key phase space patterns of signal and background processes are defined by the hard processes. This allows us to use the <span style="font-variant:small-caps;">Madgraph</span> [@madgraph] framework for most of our event generation. On the other hand, these two approximations should have a clearly defined mathematical effect on the bottom line, in our case stating that our result gives an [ *upper limit*]{} on the significance any full analysis can reach.
The setup of such a tool has been developed for the example process of Higgs production in weak boson fusion with a subsequent decay to muons many years ago [@orig]. It allows us to compute the maximal significance with which we can separate a signal–plus–background hypothesis from the background only based on Monte–Carlo event generation. It has the key feature that it allows for cuts on the contributing phase space and computes this maximum significance as a strictly increasing function when we add more phase space regions. We will rely on this feature to answer the key physics question of this paper: [*how much boost should we target in boosted Higgs and top searches?*]{}
Mathematically, our computation is inspired by the Neyman–Pearson lemma, stating that the likelihood ratio is the most powerful variable to distinguish between a background hypothesis and the signal–plus–background hypothesis [@proof]. This is formally defined as the minimum probability for a false negative outcome given a fixed probability for the false positive signal outcome. If we assume that the signal–plus–background hypothesis is true this implies the lowest probability of mistaking the signal for a background fluctuation.
In experiment we have to measure any multi-dimensional probability density function. In our approach we use the parton–level transition amplitude for signal and background processes to compute the probability density over the full phase space at a given order in perturbation theory [@early_mem]. We use a similar notation to the so-called matrix element method [@mem], but emphasize that within <span style="font-variant:small-caps;">Madgraph</span> the experimental matrix element approach is already supported by <span style="font-variant:small-caps;">MadWeight</span> [@madweight].
For now limiting ourselves to irreducible backgrounds, signal and background processes with identical degrees of freedom in the final state, we can simultaneously probe the signal and background phase space using a vector of random numbers $r$, with or without acceptance cuts, $$\sigma_\text{tot}
= \int_\text{cuts} dr \; M(r) \; d\sigma(r) \; .
\label{eq:int_first}$$ The phase space boundaries are included in the integral, and the differential cross section $d\sigma(r)$ includes all phase space factors and the Jacobian for transforming the integration to the random number basis. The integration over the parton distribution momentum fractions $x_{1,2}$ is included in the phase space integral. A measurement function $M$ can parameterize additional cuts or detector efficiencies. Because $r$ is a basis vector, cuts on observable quantities consistently remove these phase space regions from all processes. All potentially available information is included in the array of event weights $M(r) \, d\sigma(r)$.
A cut analysis defines a signal–rich region and then counts events in that region. The variable that discriminates between signal and background is the number of events ($s,b$) in this region. For counting analyses the likelihood of observing $n$ events assuming the background-only hypothesis is given by the Poisson distribution $\text{Pois}(n|b)=e^{-b} \, b^n/n!$. We can generalize this number counting by introducing a discriminating observables vector $x$. We assume that the background–only hypothesis $H_b$ is described by the normalized distribution $f_b(x)$, while the signal–plus–background hypothesis $H_{s+b}$, assuming no interference, is described by $f_{s+b}(x) =[sf_s(x) + bf_b(x)]/(s+b)$. Following the Neyman–Pearson lemma, the most powerful test statistic is the likelihood ratio. The total likelihood for the full vector $x=\{x_j\}$ can be factorized into the Poisson likelihood to observe $n$ events, and the product of the individual event’s likelihood $f(x_j)$, $$\begin{aligned}
{5}
q(x) &= \log \; \frac{L(x|H_s)}{L(x|H_b)}
= \log \; \frac{\text{Pois}(n|s+b) \; \prod_{j=1}^n f_{s+b}(x_j)}
{\text{Pois}(n|b) \; \prod_{j=1}^n f_b(x_j)} \notag \\
&= \log \left[ e^{-s} \; \left( \frac{s+b}{b} \right)^n \;
\frac{\prod_{j=1}^n f_{s+b}(x_j)}
{\prod_{j=1}^n f_b(x_j)}
\right] \notag \\
&= {-s} + \sum_{j=1}^{n} \log \left( 1+\frac{s f_s(x_j)}
{b f_b(x_j)}
\right) \; .
\label{eq:llr}\end{aligned}$$ The key step of our description is to generalize the observables vector $x$ to all individual phase space points $r$. Following Eq. they are probed by the Monte Carlo generation, which means that we can compute the normalized probability distributions $f(x)$ from the parton–level matrix elements and construct a log–likelihood ratio map of all final state configurations using the normalized probability distributions $d\sigma(r)/\sigma_\text{tot}$ for the signal and background, $$q(r) = -\sigma_{\text{tot},s} \; \mathcal{L} \; + \,
\log \left( 1 + \frac{d\sigma_s(r)}
{d\sigma_b(r)}
\right) \; .
\label{eq:likelimap}$$ ${\cal L}$ denotes the integrated luminosity. To construct the single–event probability distribution $\rho_{1,b}(q)$ we combine the background event weight with the log–likelihood ratio map $q(r)$, $$\rho_{1,b}(q_0) = \int dr \; \frac{d\sigma_b(r)}{\sigma_{\text{tot},b}}
\; \delta \left( q(r) - q_0 \right) \; .
\label{eq:rho_1}$$ For multiple events, the distribution of the log–likelihood ratio $\rho_{n,b}$ can be computed by repeated convolutions of the single event distribution. This convolution we can evaluate using a Fourier transform [@clfft]. The expected log–likelihood ratio distribution for a background including Poisson fluctuations in the number of events $n$ is $$\rho_b(q) = \sum_n \text{Pois}(n|b) \times \rho_{n,b}(q) \; .
\label{eq:rho_b}$$ To compute it from the single–event likelihood $\rho_{1,b}(q)$ we first Fourier transform all functions $\rho(q)$ into complex–valued functions of the Fourier–transformed likelihood ratio $\overline{\rho_{1,b}}(\overline{q})$. The convolution in $q$ space becomes a multiplication in Fourier space, namely $\overline{\rho_{n,b}} = (\overline{\rho_{1,b}})^n$. The sum over $n$ in Eq. has the closed form $\overline{\rho_b} = \exp[b
\; (\overline{\rho_{1,b}} - 1)]$. For the signal–plus–background hypothesis we expect $s$ events from the $\rho_{1,s}$ distribution and $b$ events from the $\rho_{1,b}$ distribution. Similar to the above formula we have $\overline{\rho_{s+b}} = \exp[ b
(\overline{\rho_{1,b}} - 1) + s (\overline{\rho_{1,s}} - 1)]$. A transformation back into $q$ space gives us log–likelihood ratio distributions $\rho_b(q)$ and $\rho_{s+b}(q)$ [@lepstats]. Finally, given a value $q$ we can calculate the background–only confidence level $$\text{CL}_b(q) =\int_{q}^\infty dq' \; \rho_b(q') \; .
\label{eq:clb}$$ To estimate the discovery potential of a future experiment we assume the signal–plus–background hypothesis to be true and compute $\text{
CL}_b$ for the median of the signal–plus–background distribution $q^*_{s+b}$. This expected background confidence level can be converted into an equivalent number of $Z$ Gaussian standard deviations by implicitly solving $\text{CL}_b(q^*_{s+b}) = \left(
1-\text{erf}(Z/\sqrt{2}) \right)/2$.
In general it is clear how to include detector effects in our simulation. However, to determine the maximal significance in a strict sense we should not include detector effects, because they always decrease the significance. In our case lepton and jet directions are usually well measured. The jet energy scale can be an issue for the detailed analysis, but we do not expect it to have a great effect on our results, either. Combinatorics will eventually be an issue, but again it will not be critical for the analyses we present in this paper. In contrast, the experimental resolution of $m_{bb}$ is nowhere close to the physical Higgs width in the Standard Model. We therefore introduce a Gaussian smearing for this one observable. The convolution of the physical Higgs width with this Gaussian we can safely approximate as the Gaussian detector resolution alone [@orig].
We will discuss two specific analyses in this paper, boosted Higgs searches in $ZH$ production [@bdrs] and in $t\bar{t}H$ production [@tth]. They are crucial for a model–independent determination of the heavy quark Yukawa couplings at the LHC, for the measurement of the most sensitive probes for new physics in the Higgs sector [@2hdm; @lecture]. In both cases we limit ourselves to the irreducible backgrounds, the processes where the $H \to
b\bar{b}$ decay is replaced by $Z \to b\bar{b}$ and QCD $b \bar{b}$ continuum production. The additional final state particles we assume to be fully reconstructed. For the case of the $Z$ boson discussed in Section \[sec:zh\] this is clearly realistic, as long as we rely on leptonic $Z$ decays. Since our analysis focuses on the kinematics on the $b\bar{b}$ system our findings can be generalized to other $W$ and $Z$ analysis channels. For the $t\bar{t}H$ analysis presented in Section \[sec:tth\] this approach requires an brief motivation: we know that all backgrounds except for the $t\bar{t}b\bar{b}$ continuum can be targeted with global kinematic cuts [@tth; @buckets]. For the irreducible continuum background we can ideally reconstruct the top momenta using a top tagger [@heptop; @top_tagger]. Because we are mostly interested in different phase space regions for the $b\bar{b}$ pair the assumption of measured top momenta is appropriate, as long as we do not consider the strictly maximum significance a realistic estimate.
Boosted $\mathbf{ZH}$ production {#sec:zh}
================================
The first channel for which we would like to quantify the benefits of specific, boosted phase space regions is $ZH$ production at 14 TeV with a Higgs decay to $b$-quarks [@bdrs]. For the Higgs mass of 126 GeV the corresponding branching ratio ranges around 58% [@hdecay]. Both Higgs decay jets are $b$-tagged. Since $p_{T,bb}$ hardly exceeds 250 GeV for the relevant events, approximately shared between the two tagged $b$-jets, we assume a constant single $b$-tagging efficiency of 60%.
For our statistical analysis we assume the $Z$ decays to $\ell =
e,\mu$ to be reconstructed perfectly. Because all the leading backgrounds also include this $Z$ decay, possible small detector effects will hardly impact our results. As detector efficiencies we include a rough factor 60% for the lepton pair, approximately correcting for the fact that we would probably only use leptons close to the $Z$ pole and that not all leptons end up in the central detector with $p_{T,\ell}>10$ GeV and $|\eta_\ell|<2.5$. These global efficiencies mainly ensure that our integrated maximum significance is not completely unrealistic; they hardly impact our study of the phase space distribution of this significance. For the signal event generation we replace the Breit–Wigner shape of $m_{bb}$ by a Gaussian with the experimental resolution of $\pm 12$ GeV. The strictly speaking appropriate convolution of the Breit–Wigner shape with the physical Higgs width and the Gaussian shape based on the experimental resolution is very well approximated by the Gaussian alone [@orig]. Higher–order corrections to the $ZH$ production rate [@zh_nnlo] are included as a variable global scaling factor of the signal rate.
The main background is continuum $Zb\bar{b}$ production at the (leading) order $\alpha \alpha_s$. A second background is the same final state at (leading) order $\alpha^2$, which includes $ZZ$ production with one decay $Z \to b\bar{b}$. Fake-$b$ backgrounds are negligible in comparison and have no unique phase space features which would force us to consider them beyond a correction to the $b$-tagged continuum QCD backgrounds. A major issue of the QCD continuum background, partly related to the invariant mass of the $b$-jets, is the poor convergence of the total cross section as a series in $\alpha_s$. First, to avoid issues with gluon splitting into two $b$-quarks versus $t$-channel production of two widely separated $b$-jets we require a mass window of $m_{bb} = 114 - 138$ GeV all through our analysis.[^1] In terms of the maximum likelihood this mass window might appear overly conservative, but on the other hand we expect experimental analyses to apply such a window to define clear side bands. Technically, it is trivial to extend this mass window to two or three standard deviation within <span style="font-variant:small-caps;">MadMax</span>. For the signal our window captures 68% of the total cross section. We show the corresponding $m_{bb}$ distributions in the left panel of Fig. \[fig:zh\_production\], illustrating a rather depressing signal–to–background ratio.
![Left: $m_{bb}$ of signal and backgrounds in the range we are considering. Center: transverse momentum of the $b\bar{b}$ system for $m_{bb} = 114 - 138$ GeV. For the QCD background the solid line approximates the merged results by using the $p_{T,bb}$-dependent correction factor of Eq.. Right: angular separation $R_{bb}$ for the signal and background, including the $p_{T,bb}$ correction for the QCD background. The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:zh_production"}](zh-m_bb "fig:"){height="31.00000%"} ![Left: $m_{bb}$ of signal and backgrounds in the range we are considering. Center: transverse momentum of the $b\bar{b}$ system for $m_{bb} = 114 - 138$ GeV. For the QCD background the solid line approximates the merged results by using the $p_{T,bb}$-dependent correction factor of Eq.. Right: angular separation $R_{bb}$ for the signal and background, including the $p_{T,bb}$ correction for the QCD background. The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:zh_production"}](zh-pt_bb "fig:"){height="31.00000%"} ![Left: $m_{bb}$ of signal and backgrounds in the range we are considering. Center: transverse momentum of the $b\bar{b}$ system for $m_{bb} = 114 - 138$ GeV. For the QCD background the solid line approximates the merged results by using the $p_{T,bb}$-dependent correction factor of Eq.. Right: angular separation $R_{bb}$ for the signal and background, including the $p_{T,bb}$ correction for the QCD background. The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:zh_production"}](zh-r_bb "fig:"){height="31.00000%"}\
To simplify the <span style="font-variant:small-caps;">MadMax</span> analysis we generate events for the QCD continuum background using the irreducible $Zb\bar{b}$ process. To get a handle on the perturbative accuracy of this simulation we also compute a $Zb\bar{b}$ event sample with up to two hard additional jets, consistently combined between the hard matrix element and the parton shower using <span style="font-variant:small-caps;">Madgraph</span> [@madgraph] with <span style="font-variant:small-caps;">Mlm</span> multi–jet merging [@mlm]. While such an event sample is not formally improved in fixed–order perturbation theory, it should capture the leading effects from large logarithms as well as from initial states opening only in combination with additional jets in the final state [@zbb_nlo]. The difference between the total leading–order $Zb\bar{b}$ rate to the merged prediction implies correction factors around 2.1. For the continuum QCD background the distribution of the merged sample including up to two hard jets is indeed harder than for the fixed–order $Zb\bar{b}$ process. While we use the simpler, fixed–order event sample in our <span style="font-variant:small-caps;">MadMax</span> analysis, we reweight it to the merged $p_{T,bb}$ distribution using the $p_{T,bb}$-dependent correction factor $$\begin{aligned}
{5}
\log \frac{d\sigma_\text{ME+PS}}{d\sigma_\text{LO}} =
0.65 + 1.1\times 10^{-3}\,p_{T,bb}
+ 4.0\times 10^{-6}\,p_{T,bb}^2 \; .
\label{eq:zh_kfac}\end{aligned}$$ The increasing form as a function of $p_{T,bb}$ we limit by fixing the cross section ratio for all values above $p_{T,bb}>350$ GeV to the maximum value, even though the number of events in this phase space is too small to observe any effects from such a cut-off.
We illustrate the $p_{T,bb}$ distributions for the signal and the backgrounds in the center panel Fig. \[fig:zh\_production\]. The $p_{T,H}$ distribution for the signal is slightly less steep than the corresponding background distributions, owing to the gluon parton densities and a general QCD preference for small invariant masses between jets. Nevertheless, the Poisson factor in our statistical analysis will essentially remove the few events with $p_{T,H} >
250$ GeV for an integrated luminosity in the $50 - 100~\ifb$ range. As a test, we also show the geometric separation of the two $b$-jets in the right panel of Fig. \[fig:zh\_production\]. The $p_{T,bb}$-dependent correction perfectly reproduces the multi-jet merged distribution for this observable.
![Maximum significance for the $ZH$ signal for slices in the reconstructed $p_{T,bb}$ (left) and the geometric separation of the $b$-jets, $R_{bb}$ (right). We only consider events inside the mass window $m_{bb} = 114 - 138$ GeV. The significance is computed for an integrated luminosity of $50~\ifb$.[]{data-label="fig:zh_slices"}](zh-significance-p_bb "fig:"){width="31.00000%"} ![Maximum significance for the $ZH$ signal for slices in the reconstructed $p_{T,bb}$ (left) and the geometric separation of the $b$-jets, $R_{bb}$ (right). We only consider events inside the mass window $m_{bb} = 114 - 138$ GeV. The significance is computed for an integrated luminosity of $50~\ifb$.[]{data-label="fig:zh_slices"}](zh-significance-r_bb "fig:"){width="31.00000%"}\
As an estimate of the maximum significance over the entire phase space we obtain $2.7 \pm 0.3~\sigma$ for an integrated luminosity of $50~\ifb$. The error bar is given by a $\pm 20\%$ variation in the normalization of the signal rate. The total maximum significance at higher integrated luminosities can be approximately computed by a Gaussian scaling. In the left panel of Fig. \[fig:zh\_slices\] we show how this maximum significance is shared between slices of $p_{T,bb}$. The phase space regime best suited to distinguish signal from background events at this limited luminosity is $$\begin{aligned}
{5}
p_{T,H} = 50 - 100~\gev \; .\end{aligned}$$ Boosted Higgs analyses in the $ZH$ channel indeed significantly reduce the QCD continuum background, but Higgs taggers should be optimized for as low $p_{T,H}$ values as possible. The reason for this finite transverse momentum range is on the one hand that for the background there does not exist a large mass scale in the process, which means that a sizeable $b\bar{b}$ invariant mass has to be generated through a geometric separation, namely $m_{bb}^2 \simeq 2E_1
E_2 (1 - \cos \theta)$. In contrast, if we require the $H \to
b\bar{b}$ decay to be boosted, the $b$-jets in the signal will move closer together, making it harder for the QCD background to fake the Higgs signal. On the other hand, while we would naively expect higher transverse momenta to carry more and more weight in the analysis, the number of signal events in this range is strongly limited. Asking for $p_{T,bb} > 150$ already drives us into a strongly statistics limited phase space regime.
In the right panel of Fig. \[fig:zh\_slices\] we show how the maximum significance is composed by slices in the geometric separation $R_{bb}$. This example illustrates how <span style="font-variant:small-caps;">MadMax</span> can be used to analyze the maximum significance distribution in terms of any phase space observable. The separation of the two $b$-tagged jets is crucial for the definition of the fat jet as the starting point of any Higgs tagger. From Fig. \[fig:zh\_production\] it is clear that hardly any signal events lie in the range $R_{bb} \lesssim 1.0$. To optimize a boosted Higgs analysis it appears to be beneficial to extend the fat jet size towards $R_{bb} \sim 2.0$.
Additional effects can still modify the outcome of our study. First, we assume that the detector performance does not depend on the boost of the $b\bar{b}$ system. This is clearly only true up to a certain $p_{T,bb}$ range, where the two bottom jets start overlapping. Second, we assume theory uncertainties to be independent of $p_{T,bb}$. It is not clear if this statement holds for QCD effects, and it is clearly not true in the electroweak sector, once we include electroweak Sudakov logarithms. Finally, we did not actually study dangerous observables, for example, defined in the pre-jet stage of the analysis or at odds with the assumed factorization properties of the signal and background predictions. In that sense <span style="font-variant:small-caps;">MadMax</span> clearly does not deliver a realistic estimate of systematic and theoretical uncertainties. It is merely a first step which allows us to study phase space patterns easily and reliably. For example the question to what degree the estimated uncertainties are realistic and what the effect of shape uncertainties in the background might be will be left to possible further studies.
Boosted $\mathbf{t\bar{t}H}$ production {#sec:tth}
=======================================
The second process for which we want to ask the question where the main distinguishing phase space features exist is $t\bar{t}H$ production with a Higgs decay to $b$-quarks. Two studies indicate that boosted top and Higgs configurations might be promising to extract the signal from the background: for purely hadronic events the buckets methods successfully targets moderate transverse momenta around $p_{T,t} \gtrsim 150$ GeV [@buckets]. For semileptonic top pairs the <span style="font-variant:small-caps;">HEPTopTagger</span> study shows that slightly larger boosts $p_{T,t} \gtrsim 200 - 250$ GeV can be successfully probed [@tth]. Purely leptonic top pairs have recently been shown to lead to promising results based on a <span style="font-variant:small-caps;">MadWeight</span> study [@tth_lep]. In the semi–leptonic and hadronic cases it is obvious that the boosted kinematics is an excellent way to resolve combinatorial issues [@tth; @tth_comb]. The open question is to what degree the arguments presented in the last section also point to a boosted $t\bar{t}H$ search in terms of the signal and background matrix elements.
To answer this question we will first assume that the continuum $t\bar{t}b\bar{b}$ background is the most relevant issue. For the semileptonic analysis this has been shown, once we require at least three $b$-tags [@tth]. For the purely hadronic channel the removal of the QCD backgrounds is considerably more tedious, but appears to be possible [@buckets]. We follow this study and as a first step require four $b$-tagged jets with an efficiency of 60% each. In addition, we assume a set of global cuts or other ways to reduce the multi–jet backgrounds with a conservatively estimated efficiency around 10% for the $t\bar{t}H$ signal and the irreducible $t\bar{t}b\bar{b}$ background.
To study the combinatorics of the $b$-jets identified as Higgs decay we would have to simulate top decays. However, this is precisely the issue which we want to separate from our significance study over phase space. Therefore, we can assume the top quarks to be fully reconstructed. For hadronic top decays a major part of the buckets [@buckets] and <span style="font-variant:small-caps;">HEPTopTagger</span> studies [@heptop] have been devoted to quantifying the quality of this momentum reconstruction. Moreover, in the absence of additional missing energy from the hard process one hadronic top tag can be combined with an efficient approximate reconstruction of the boosted leptonic top decay [@leptonic_top]. The efficiency of actually reconstructing a hadronic top decay using a top tagger is strongly dependent on the transverse momentum. We roughly estimate it to 33% per top quark, in addition to the branching ratio of 68%. This way we should obtain at least a semi-realistic number for the integrated maximum significance. As mentioned before, such global efficiencies will not affect the main outcome of our study, the distribution of the maximum significance over the transverse momentum range of the heavy particles.
For the Higgs decay to bottoms we again include a branching ratio of 58% [@hdecay]. To approximately include detector effects we evaluate the Higgs propagator with a Gaussian width of $\pm
12$ GeV. The mass window of $m_{bb} = 114-138$ GeV with a signal efficiency of 68% we carry through the entire analysis, both for the signal and for the background. As discussed in the previous section this ensures that we avoid gluon splitting issues in the background simulation and at the same time allows for side bands in the obvious $m_{bb}$ distribution [@tth].
![Left: $p_{T,bb}$ for signal and backgrounds inside the mass window $m_{bb} = 114 - 138$ GeV. Center and right: transverse momenta of the two tops. For the signal and the QCD background the solid lines approximate the merged results by using the $p_T$-dependent correction factors of Eqs. and . The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:tth_pt"}](tth-pt_bb "fig:"){height="31.00000%"} ![Left: $p_{T,bb}$ for signal and backgrounds inside the mass window $m_{bb} = 114 - 138$ GeV. Center and right: transverse momenta of the two tops. For the signal and the QCD background the solid lines approximate the merged results by using the $p_T$-dependent correction factors of Eqs. and . The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:tth_pt"}](tth-pt_t1 "fig:"){height="31.00000%"} ![Left: $p_{T,bb}$ for signal and backgrounds inside the mass window $m_{bb} = 114 - 138$ GeV. Center and right: transverse momenta of the two tops. For the signal and the QCD background the solid lines approximate the merged results by using the $p_T$-dependent correction factors of Eqs. and . The merged multi–jet simulation are shown as dotted lines.[]{data-label="fig:tth_pt"}](tth-pt_t2 "fig:"){height="31.00000%"}\
For the $t\bar{t}H$ signal the next-to-leading order corrections are known [@tth_nlo]. Since our study is focused on the phase space structure, we approximately include them by correcting the $p_{T,H}$ distribution to agree with a matched $t\bar{t}H$ plus zero and one jet simulation in the <span style="font-variant:small-caps;">Mlm</span> scheme [@mlm] the same way as we do with the $Zb\bar{b}$ background in Eq.. We find a correction factor for the $t\bar{t}H$ signal, $$\begin{aligned}
{5}
\log \frac{d\sigma_\text{ME+PS}}{d\sigma_\text{LO}} =
0.53 - 2.5\times 10^{-3}\,p_{T,H}
+ 2.0\times 10^{-5}\,p_{T,H}^2
- 3.9\times 10^{-8}\,p_{T,H}^3 \; .
\label{eq:tth_kfac_s}\end{aligned}$$ For a $t\bar{t}b\bar{b}$ background study in the boosted Higgs regime it is absolutely crucial that we correctly simulate the transverse momentum of the $b\bar{b}$ system. While the events entering the <span style="font-variant:small-caps;">MadMax</span> analysis are generated as $t\bar{t}b\bar{b}$ production in <span style="font-variant:small-caps;">Madgraph</span>, we can correct the $p_{T,bb}$ distribution to agree with matched $t\bar{t}b\bar{b}$ plus zero and one jet simulation using $$\begin{aligned}
{5}
\log \frac{d\sigma_\text{ME+PS}}{d\sigma_\text{LO}} =
0.98 - 6.7\times 10^{-3}\,p_{T,bb}
+ 3.8\times 10^{-5}\,p_{T,bb}^2
- 7.6\times 10^{-8}\,p_{T,bb}^3 \; .
\label{eq:tth_kfac_b}\end{aligned}$$ As for the $ZH$ case we cut off both correction factors using a constant value above $p_{T,bb} = 350$ GeV. This reweighting of the differential cross section should account for the leading logarithmic higher–order corrections [@tth_nlo; @ttbb_nlo], in particular linked to different partons in the initial state. In the left panel of Fig. \[fig:tth\_pt\] we show the corresponding distributions, after including the two $p_T$-dependent correction factors.
![Maximum significance for the $t\bar{t}H$ signal for slices in the reconstructed $p_{T,bb}$ (left), the geometric separation of the two $b$-jets, $R_{bb}$ (center), and the leading top transverse momentum $p_{T,t^\text{leading}}$ (right). We only consider events inside the mass window $m_{bb} = 114 - 138$ GeV. The significance is computed for an integrated luminosity of $50~\ifb$.[]{data-label="fig:tth_slices"}](tth-significance-p_bb "fig:"){width="31.00000%"} ![Maximum significance for the $t\bar{t}H$ signal for slices in the reconstructed $p_{T,bb}$ (left), the geometric separation of the two $b$-jets, $R_{bb}$ (center), and the leading top transverse momentum $p_{T,t^\text{leading}}$ (right). We only consider events inside the mass window $m_{bb} = 114 - 138$ GeV. The significance is computed for an integrated luminosity of $50~\ifb$.[]{data-label="fig:tth_slices"}](tth-significance-r_bb "fig:"){width="31.00000%"} ![Maximum significance for the $t\bar{t}H$ signal for slices in the reconstructed $p_{T,bb}$ (left), the geometric separation of the two $b$-jets, $R_{bb}$ (center), and the leading top transverse momentum $p_{T,t^\text{leading}}$ (right). We only consider events inside the mass window $m_{bb} = 114 - 138$ GeV. The significance is computed for an integrated luminosity of $50~\ifb$.[]{data-label="fig:tth_slices"}](tth-significance-pt_t1 "fig:"){width="31.00000%"}\
For the $t\bar{t}H$ analysis not only the boost of the Higgs candidate, but also the boost of each top is relevant for the analyses [@tth; @buckets]. Therefore, we need to test to what degree the reweighting in Eqs. and affects the top kinematics. In the center and right panels of Fig. \[fig:tth\_pt\] we compare the background $p_{T,t}$ distributions for the fully merged event sample with the $p_{T,bb}$-reweighted $t\bar{t}b\bar{b}$ sample. We see that the only phase space region not perfectly described by the $p_{T,bb}$-dependent correction factor is the low-$p_T$ range of the two tops. However, the difference is a mere 5%, covered by the theory uncertainty, in a phase space region which will turn out relatively unimportant.
For the set of approximate efficiencies listed above we estimate the maximum significance of the $t\bar{t}H$ to $5.3 \pm 0.5~\sigma$ for an integrated luminosity of $50~\ifb$ at a collider energy of 14 TeV. The quoted error bar is again defined by a $\pm 20\%$ variation of the signal rate. Again, values for higher luminosities can be inferred by Gaussian scaling. In Fig.\[fig:tth\_slices\] we show this maximum significance in slices of the transverse momentum of the reconstructed Higgs system, the separation of the Higgs decay jets, and the transverse momentum of the leading top. For the Higgs boost we find a similar range as for the $ZH$ channel, while the most promising transverse momentum range for the heavier top roughly scales with $m_t/m_H$, $$\begin{aligned}
{5}
p_{T,H} &= \; 50 - 100~\gev \notag \\
p_{T,t} &= 100 - 250~\gev \; .\end{aligned}$$ This result indicates that the requirements for the Higgs tagger [@bdrs] in the $ZH$ and $t\bar{t}H$ analyses are very similar. The only difference between these two channels is that for the $t\bar{t}H$ process the Higgs tagger has to be adapted to higher jet multiplicities, as done in Ref. [@tth]. For the top tagger it is crucial that we gain access to transverse momenta well below $p_{T,t} = 300$ GeV, ruling out most of the currently available analysis tools optimized for heavy $t\bar{t}$ resonance searches [@top_tagger].
Conclusions {#sec:conclusions}
===========
In this paper we have introduced <span style="font-variant:small-caps;">MadMax</span> as a novel approach to studying the composition of a signal significance in terms of the signal and background phase space. It allows us to determine those phase regions which are best suited for the extraction of a signal process from irreducible backgrounds in an efficient and mathematically well–defined manner [@orig].
Relying on <span style="font-variant:small-caps;">MadMax</span> we have studied the two Higgs search channels involving a hadronic $H \to b\bar{b}$ decay, namely associated $ZH$ and $t\bar{t}H$ production. In both cases the central question is to what degree boosted Higgs configurations benefit the signal extraction and what range in transverse Higgs momenta we should target. Unlike in the original study [@orig], we specifically did not focus on predicting the integrated maximum significance for each of these processes, so detector effects as well as global efficiencies are only adjusted to obtain a semi-realistic number.
For the $ZH$ channel we find that for integrated luminosities in the $50-100~\ifb$ range the most promising phase space regime is around or below $p_{T,H} =100$ GeV, challenging the development of Higgs taggers. For the $t\bar{t}H$ analysis the first reason to rely on boosted top and Higgs decay topologies is the otherwise overwhelming combinatorial background. In this study we estimated the additional motivation for using these topologies based on the matrix element structure of the signal and the irreducible background. The most promising range in the transverse momentum of the reconstructed Higgs boson again came out as $p_{T,H} = 50 - 100$ GeV, indicating that the same Higgs tagger should suit both analyses. To include as many of the relevant signal events as possible the size of the fat jet could be extended towards $R_{bb} \sim 2$, if possible. For the transverse momentum of the leading top quark the significance is mostly collected for $p_{T,t} = 100 - 250$ GeV, seriously challenging the development of hadronic top taggers.
We note that <span style="font-variant:small-caps;">MadMax</span> is an automized tool which can be used in the <span style="font-variant:small-caps;">Madgraph5</span> framework to produce multi–dimensional differential distributions adding to the maximum significance in any kinematic observable. Some of its current limitations, like the focus on irreducible backgrounds can be overcome easily. Likewise, simple transfer functions can be included in a straightforward manner. Once these effects are taken into account, we should be able to also give a more reliable estimate of the integrated maximum significance for a signal–background combination.
[**Acknowledgments**]{}
First of all, we would like to thank Kyle Cranmer for supporting this extension of the original work of Ref.[@orig]. PS would like to thank the IMPRS [*Precision studies of fundamental symmetries*]{} for their never–ending support. TP would like to thank the CCPP at New York University for their hospitality in a crucial phase of this paper. DW is grateful to the University of Pittsburgh and to Ayres Freitas for supporting him while writing this paper, as well as to Jan Pawlowski for agreeing to co-referee his rather technical thesis. Finally, all of us would like to thank Fabio Maltoni for suggesting the name <span style="font-variant:small-caps;">MadMax</span>.
MadMax in Madgraph {#sec:appendix}
==================
To understand how <span style="font-variant:small-caps;">MadMax</span> works together with <span style="font-variant:small-caps;">Madgraph5</span> [@madgraph] we first describe some of the main <span style="font-variant:small-caps;">Madgraph</span> features. To construct the log–likelihood map of Eq. we need the squared matrix elements for the signal and background, which <span style="font-variant:small-caps;">Madgraph5</span> computes using a modified version of <span style="font-variant:small-caps;">Helas</span> [@helas]. These matrix elements are integrated using the so-called single diagram enhanced method [@madevent] to account for the propagator structure. A good example process is $pp \to \mu^+ \mu^-$, where we compute the cross–section via $$\begin{aligned}
{5}
\sigma_\text{tot} = \int dx_1 \, dx_2 \, d\text{LIPS} \;
f_p(x_1,\mu_F) \, f_p(x_2,\mu_F) \;
\left| \sum \limits_n \mathcal{M}_n \right|^2 \; .\end{aligned}$$ The phase space is described by a random number vector $r$ and transforms the integral into a sum over phase space cells $\Delta
r$. The parton densities are evaluated together with the matrix elements, so in the following we implicitly include them in $\mathcal{M}$. We can improve the convergence if we know the leading behavior of the individual matrix elements, $$\begin{aligned}
{5}
\sigma_\text{tot}
= \sum \limits_r \; \Delta r \;
\left| \sum \limits_n \mathcal{M}_n (r) \right|^2
= \sum \limits_{r,i} \Delta r \,
\frac{ \left| \mathcal{M}_i (r) \right|^2 }{ \sum \limits_n
\left| \mathcal{M}_n (r) \right|^2 } \times \left| \sum
\limits_n \mathcal{M}_n (r) \right|^2 \; .\end{aligned}$$ The optimized phase space mapping is different for each diagram $i$. For our example we optimize for a $1/s$ scaling in the photon exchange and for a Breit–Wigner propagator in the $Z$ exchange. In addition, the sum over diagrams is decomposed into incoherent partial sums for different incoming partons, so an additional weight accounts for the parton densities, $$\begin{aligned}
{5}
\sigma_\text{tot} =
\sum \limits_{r,i,p} \Delta r \,
\frac{\left| \mathcal{M}_{i_p} (r) \right|^2}
{\sum \limits_{n_p} \left| \mathcal{M}_{n_p} (r) \right|^2}
\times \left| \sum \limits_{n_p} \mathcal{M}_{n_p} (r) \right|^2
\times \omega_p^\text{pdf} (r) \; .\end{aligned}$$
In this framework <span style="font-variant:small-caps;">MadMax</span> computes the maximum significance using a log–likelihood ratio integration. It starts from the single event likelihood for signal and background. For each phase space point we need to know simultaneously $$\begin{aligned}
{5}
d\sigma_s(r) &=
\left| \sum \limits_{n_s} \mathcal{M}_{n_s}(r) \right|^2
\qqqquad
d\sigma_b (r) =
\left| \sum \limits_{n_b} \mathcal{M}_{n_b}(r) \right|^2
\notag \\
q(r) &= -\sigma_{\text{tot,s}} \, \mathcal{L}
+ \log \left( 1 + \frac{d\sigma_s(r)}{d\sigma_b(r)}
\right) \; ,\end{aligned}$$ as quoted in Eq.. From the logarithm in the log–likelihood ratio it is clear that the we cannot use the single enhanced diagram method. Instead, we will use a modified version closer to the original proposal [@madevent]: let us assume our signal process consists of $a_s \in [1, \ldots, n_s]$ sub-processes, with different partons in the initial and final state. Each sub-process will be computed using $i_{a_s} \in [1 , \ldots, n_{a_s}
]$ matrix elements. The same is true for the background, namely $a_b
\in [1, \ldots, n_b]$ subprocesses with $i_{a_b} \in [1 , \ldots,
n_{a_b}]$ matrix elements. Any function $f(r)$ we can construct using the basis elements $$\begin{aligned}
{5}
\frac{ \left| \mathcal{M}_i (r) \right|^2 }{
\sum \limits_{\substack{a_s,i_{a_s}\\a_b,i_{a_b}}} \left|
\mathcal{M}_n (r) \right|^2 } \times f(r) \; .\end{aligned}$$ In this basis we can integrate all signal and background rates. Furthermore, the points $q(r)$ defining the single event probability have the correct weight as well. The only remaining issue is that events in <span style="font-variant:small-caps;">Madgraph5</span> are usually computed with a dynamical factorization and renormalization scale choice. In <span style="font-variant:small-caps;">MadMax</span> we add, divide, and combine matrix elements from signal as well as background processes. In the current implementation of <span style="font-variant:small-caps;">MadMax</span> we use fixed factorization and renormalization scales, to simplify the interface to <span style="font-variant:small-caps;">Madgraph5</span>. The implementation of a general scale choice would be straightforward.
![Example <span style="font-variant:small-caps;">MadMax</span> output for the $ZH$ channel. We give the computed signal and background cross–sections as well as the corresponding standard <span style="font-variant:small-caps;">Madgraph</span> results. In addition, we give all constant efficiencies and $K$-factors included in the statistical analysis.[]{data-label="fig:example_results"}](example_result){width="85.00000%"}
We test and validate our implementation in two ways. First, we have compared our findings for weak–boson–fusion Higgs production with a Higgs decay to muons with the original results [@orig]. Second, for each set of signal and background processes we can compare our signal and background cross–sections with the <span style="font-variant:small-caps;">Madgraph5</span> results obtained in parallel, an example output is shown in Fig. \[fig:example\_results\]. The numbers always agree within the numerical uncertainties. An extensive validation study as well as more information on the implementation of MadMax can be found in Ref. [@daniel].
[99]{}
K. Cranmer and T. Plehn, Eur. Phys. J. C [**51**]{}, 415 (2007). M. Klute, R. Lafaye, T. Plehn, M. Rauch and D. Zerwas, Europhys. Lett. [**101**]{}, 51001 (2013). J. M. Butterworth, A. R. Davison, M. Rubin, G. P. Salam, Phys. Rev. Lett. [**100**]{}, 242001 (2008); ATLAS note, ATL-PHYS-PUB-2009-088.
T. Plehn, G. P. Salam and M. Spannowsky, Phys. Rev. Lett. [**104**]{}, 111801 (2010). M. R. Buckley, T. Plehn and M. Takeuchi, JHEP [**1308**]{}, 086 (2013); M. R. Buckley, T. Plehn, T. Schell and M. Takeuchi, arXiv:1310.6034 \[hep-ph\]. for recent overviews see A. Altheimer, S. Arora, L. Asquith, G. Brooijmans, J. Butterworth, M. Campanelli, B. Chapleau and A. E. Cholakian [*et al.*]{}, J. Phys. G [**39**]{}, 063001 (2012); T. Plehn and M. Spannowsky, J. Phys. G [**39**]{}, 083001 (2012). D. E. Soper and M. Spannowsky, Phys. Rev. D [**84**]{}, 074002 (2011); D. E. Soper and M. Spannowsky, Phys. Rev. D [**87**]{}, no. 5, 054012 (2013). J. Alwall, P. Demin, S. de Visscher, R. Frederix, M. Herquet, F. Maltoni, T. Plehn and D. L. Rainwater [*et al.*]{}, JHEP [**0709**]{}, 028 (2007); J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer and T. Stelzer, JHEP [**1106**]{}, 128 (2011). for a proof and corresponding definitions, see J. Stuart, A. Ord and S. Arnold, [*Kendall’s Advanced Theory of Statistics, Vol 2A (6th Ed.)*]{} (Oxford University Press, New York, 1994); for a pedagogical introduction in the context of high-energy physics, see A. Read, in ‘1st Workshop on Confidence Limits’, CERN Report No. CERN-2000-005 (2000).
for some early examples see K. Kondo, J. Phys. Soc. Jap. [**57**]{}, 4126 (1988); D. Atwood and A. Soni, Phys. Rev. D [**45**]{}, 2405 (1992); M. Diehl and O. Nachtmann, Eur. Phys. J. C [**1**]{}, 177 (1998). see some early examples see V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Lett. B [**617**]{}, 1 (2005); A. Abulencia [*et al.*]{} \[CDF Collaboration\], Phys. Rev. D[**73**]{}, 092002 (2006); V.M Abazov [*et. al.*]{} \[D0 Collaboration\], Nature 429, 638 (2004).
P. Artoisenet, V. Lemaitre, F. Maltoni and O. Mattelaer, JHEP [**1012**]{}, 068 (2010). H. Hu and J. Nielsen, in ‘1st Workshop on Confidence Limits’, CERN 2000-005 (2000) \[arXiv:physics/9906010\].
K. Cranmer, LEPStats4LHC, <https://plone4.fnal.gov/P0/phystat/packages/0703002>
D. Lopez-Val, T. Plehn and M. Rauch, JHEP [**1310**]{}, 134 (2013). for a pedagogical introduction see T. Plehn, Lect. Notes Phys. [**844**]{}, 1 (2012); \[arXiv:0910.4182 \[hep-ph\]\]. T. Plehn, M. Spannowsky, M. Takeuchi and D. Zerwas, JHEP [**1010**]{}, 078 (2010); T. Plehn, M. Spannowsky and M. Takeuchi, Phys. Rev. D [**85**]{}, 034029 (2012); G. Aad [*et al.*]{} \[ATLAS Collaboration\], JHEP [**1301**]{}, 116 (2013); <http://www.physi.uni-heidelberg.de//Publications/Kasieczka-Doktor.pdf>
A. Djouadi, J. Kalinowski and M. Spira, Comput. Phys. Commun. [**108**]{}, 56 (1998). O. Brein, A. Djouadi and R. Harlander, Phys. Lett. B [**579**]{}, 149 (2004); S. Dawson, T. Han, W. K. Lai, A. K. Leibovich and I. Lewis, Phys. Rev. D [**86**]{}, 074007 (2012). M. L. Mangano, M. Moretti, F. Piccinini and M. Treccani, JHEP [**0701**]{}, 013 (2007). J. M. Campbell and R. K. Ellis, Phys. Rev. D [**62**]{}, 114012 (2000); F. Febres Cordero, L. Reina and D. Wackeroth, Phys. Rev. D [**80**]{}, 034015 (2009). P. Artoisenet, P. de Aquino, F. Maltoni and O. Mattelaer, arXiv:1304.6414 \[hep-ph\]. J. Cammin and M. Schumacher, ATL-PHYS-2003-024; for an experimental update see S. Allwood-Spiers, ATL-PHYS-PROC-2011-280. J. Thaler and L. -T. Wang, JHEP [**0807**]{}, 092 (2008); K. Rehermann and B. Tweedie, JHEP [**1103**]{}, 059 (2011); T. Plehn, M. Spannowsky and M. Takeuchi, JHEP [**1105**]{}, 135 (2011). W. Beenakker, S. Dittmaier, M. Krämer, B. Plümper, M. Spira and P. M. Zerwas, Nucl. Phys. B [**653**]{}, 151 (2003); S. Dawson, C. Jackson, L. H. Orr, L. Reina and D. Wackeroth, Phys. Rev. D [**68**]{}, 034022 (2003). A. Bredenstein, A. Denner, S. Dittmaier and S. Pozzorini, Phys. Rev. Lett. [**103**]{}, 012002 (2009); A. Bredenstein, A. Denner, S. Dittmaier and S. Pozzorini, JHEP [**1003**]{}, 021 (2010); G. Bevilacqua, M. Czakon, C. G. Papadopoulos, R. Pittau and M. Worek, JHEP [**0909**]{}, 109 (2009). H. Murayama, I. Watanabe and K. Hagiwara, KEK-91-11, <http://madgraph.kek.jp/~kanzaki/Tutorial/helas.pdf>
F. Maltoni and T. Stelzer, JHEP [**0302**]{}, 027 (2003). D. Wiegand, master thesis (2013), [http://www.thphys.uni-heidelberg.de/\\textasciitilde
plehn/includes/theses/wiegand.pdf](http://www.thphys.uni-heidelberg.de/\textasciitilde
plehn/includes/theses/wiegand.pdf)
[^1]: It is well known that control regions where the effective hard process is $Zg^* \to Zb\bar{b}$ production should not be used to probe QCD features of continuum $Zb\bar{b}$ production.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the quantum dynamics of a two-level system interacting with a quantized harmonic oscillator in the [*deep strong coupling regime*]{} (DSC) of the Jaynes-Cummings model, that is, when the coupling strength $g$ is comparable or larger than the oscillator frequency $\omega$ ($g / \omega \gtrsim 1$). In this case, the rotating-wave approximation cannot be applied or treated perturbatively in general. We propose an intuitive and predictive physical frame to describe the DSC regime where [*photon number wavepackets*]{} bounce back and forth along [*parity chains*]{} of the Hilbert space, while producing [*collapse and revivals*]{} of the initial population. We exemplify our physical frame with numerical and analytical considerations in the qubit population, photon statistics, and Wigner phase space.'
author:
- 'J. Casanova'
- 'G. Romero'
- 'I. Lizuain'
- 'J. J. García-Ripoll'
- 'E. Solano'
title: 'Deep Strong Coupling Regime of the Jaynes-Cummings model'
---
The interaction between a two-level system and a harmonic oscillator is ubiquitous in different physical setups, ranging from quantum optics to condensed matter and applications to quantum information. Typically, due to the parameter accessibility of most experiments, the rotating-wave approximation (RWA) can be applied producing a solvable dynamics called the Jaynes-Cummings (JC) model [@jaynescummings63]. In this case, Rabi oscillations inside the JC doublets or collapses and revivals of the system populations [@Eberly80] are paradigmatic examples of the intuitive physics behind the JC dynamics. To achieve these and other phenomena in the lab, the strong coupling (SC) regime is required, that is, the qubit-oscillator coupling has to be comparable or larger than all decoherence rates. This model accurately describes the dynamics of cavity QED [@raimond01; @walther06], trapped ion experiments [@leibfried03], and several setups in mesoscopic physics, where the qubit-oscillator model is essential in modeling superconducting qubits [@clarke08] with either coplanar transmission lines [@wallraff04; @blais04; @chiorescu04; @hofheinz09] or nanomechanical resonators [@cleland04; @lahaye09]. Nowadays, solid-state semiconductor [@gunter09] or superconductor systems [@sornborger04; @abdumalikov08; @bourassa09; @borja09; @Lizuain10; @gross10; @paul10] have allowed the advent of the ultrastrong coupling (USC) regime, where the coupling strength is comparable or larger than appreciable fractions of the mode frequency: $g / \omega \gtrsim 0.1$. In this regime, the RWA breaks down and the model becomes analytically unsolvable, although some limits can be explored [@irish07; @chen08; @hwang10; @nori10; @chen10]. Confident of the impressive fast development of current technology, one could explore further regimes where the rate between the coupling strength and oscillator frequency could reach $g/\omega \gtrsim 1$, here called deep SC (DSC) regime. This unusual regime, yet to be experimentally explored, is the focus of our current efforts. In this letter, we introduce a rigorous and intuitive description of the DSC regime of the JC model, providing an insightful picture where photon number wavepackets propagate coherently along two independent parity chains of states. In this way, the Hilbert space splits in two independent chains, exhibiting a comprehensible collapse-revival pattern of the system populations.
We consider the Jaynes-Cummings Hamiltonian without the RWA, also called Rabi Hamiltonian, describing a two-level system coupled to a single mode harmonic oscillator $$H=\frac{\hbar}{2} \omega_0 \sigma_z + \hbar \omega a^{\dag} a + \hbar g (\sigma^{+} + \sigma^{-})(a + a^{\dag}).
\label{tHamiltonian}$$ Here, $a$ and $a^{\dag}$ are the annihilation and creation operators of the mode with frequency $\omega,$ while $\sigma_z$ and $\sigma^{\pm}$ are Pauli operators associated to a qubit with ground state $| \rm g \rangle$, excited state $| \rm e \rangle$, and transition frequency $\omega_0$. We concentrate in the study of the DSC regime, $g / \omega \gtrsim 1$, with no particular relation between $\omega$ and $\omega_0$. We do not refer to any particular system because several of them in quantum optics and condensed matter may profit from the physical insight developed here [@comment1]. We start by observing that the parity operator [@shore73] $$\label{parity}
\Pi=-\sigma_z (-1)^{n_a} = - (| {\rm e} \rangle \langle { \rm e} | - | { \rm g } \rangle \langle {\rm g} |) (-1)^{a^{\dag} a} ,$$ with $\Pi | p \rangle = p | p \rangle$ and $p = \pm 1$, is a key element associated to the Hamiltonian in Eq. (\[tHamiltonian\]). It is instrumental to understand how the system dynamics moves inside the Hilbert space split in two unconnected subspaces or parity chains, $$\begin{aligned}
| {\rm g} 0_a \rangle \leftrightarrow | {\rm e} 1_a \rangle \leftrightarrow | {\rm g} 2_a \rangle \leftrightarrow | {\rm e} 3_a \rangle \leftrightarrow \ldots (p = +1), \nonumber \\
| {\rm e} 0_a \rangle \leftrightarrow | {\rm g} 1_a \rangle \leftrightarrow | {\rm e} 2_a \rangle \leftrightarrow | {\rm g} 3_a \rangle \leftrightarrow \ldots ( p = -1).
\label{chains}\end{aligned}$$ Neighboring states within each parity chain may be connected via either rotating or counter-rotating terms. For example, in the parity chain with $p = +1$, the counter-rotating term $\sigma^{+} a^{\dagger}$ induces the transition $| {\rm g} 2_a \rangle \rightarrow | {\rm e} 3_a \rangle$, while the rotating term $\sigma^{+} a$ induces $| {\rm e} 1_a \rangle \leftarrow | {\rm g} 2_a \rangle$. When going back from DSC$\rightarrow$USC$\rightarrow$SC, the parity chains break into the known Jaynes-Cummings doublets $\{ | {\rm g} , n _a+ 1 \rangle , | {\rm e} , n_a \rangle \}$ because we enter into the domain of applicability of the RWA.
![(Color online) (a)-(b) Round trip of a photon number wavepacket and collapse-revivals due to DSC dynamics with initial state $| + , 0_b \rangle = | {\rm g} , 0_a \rangle$. c) Collapse-revivals with secondary peaks due to counterpropagating photon number wavepackets starting in initial state $| + , 2_b \rangle = | {\rm g} , 2_a \rangle$. For all cases, $\omega_0 = 0$ and $g / \omega = 2$. []{data-label="Fig1"}](fig_1){width="\linewidth"}
We introduce the parity basis $| p ,n_b\rangle,$ where $b^{\dagger}b | n_b \rangle = n_b | n_b \rangle$, and $b = \sigma_x a$ such that $\, b | p ,n_b \rangle = \sqrt{n_b} | p ,n_b - 1 \rangle$. Using this basis, the Hamiltonian in Eq. (\[tHamiltonian\]) can be written as $$H = \hbar \omega b^{\dag} b+\hbar g (b+b^{\dag})-\hbar\frac{\omega_0}{2}(-1)^{b^{\dag}b}\Pi .
\label{pHamiltonian}$$ This Hamiltonian commutes with the parity operator $\Pi$, and for each parity chain ($p = \pm 1$) there is an independent Hamiltonian describing a perturbed harmonic oscillator. Note that the term $-\hbar \omega_0(-1)^{b^{\dag}b}\Pi/2$ behaves as an energy shift proportional to $\omega_0$. In the DSC regime, we can get rid of the term $\hbar g (b+b^\dagger)$ in Eq. (\[pHamiltonian\]) by changing to the basis $D(-\beta_0)|p , n_b \rangle$, with $D(\beta_0)=e^{\beta_0 b^\dagger - \beta_0^*b}$ and $\beta_0=g/\omega$. The eigenenergies and eigenfunctions can be approximated as a series in $\omega_0/\omega$ $$\begin{aligned}
E^{\beta_0}_{p,n_b} /\hbar &\approx& \omega n_b - g^2/\omega -
\frac{\omega_0}2 p (-1)^{n_b} \Delta_{n_b n_b} + \label{eq:approx}\\
&+&\sum_{m_b \neq n_b} \frac{\omega_0^2}{4\omega(n_b - m_b)}|\Delta_{n_b m_b}|^2 + \mathcal{O}(\omega_0^3/\omega^3).\nonumber\end{aligned}$$ Alternative approximations can be found in the literature [@irish07]. To first order we get a displacement in the energy levels due to the coupling $\Delta_{n_b n_b} = \langle n_b | D(2\beta_0)| n_b \rangle,$ a correction which is much smaller than one, $|\Delta_{n_b m_b}| \ll 2^{-(n_b + m_b)}.$ Note that this formalism is rigorously valid in the DSC regime.
We study now the DSC dynamics with the initial state $|\psi(0)\rangle = | + , 0_b \rangle=| {\rm g} , 0_a \rangle$, as we activate the interaction in Eq. (\[pHamiltonian\]). We observe that the photon statistics, $P_{n_b} (t)$, will spread independently along each parity chain, eventually reaching an energy barrier and bouncing repeatedly. Remarkably, an intuitive picture can be found, as displayed in Figs. \[Fig1\] and \[Fig2\] , that provides physical insight in a problem that is, in general, analytically intractable. Note that, in Figs. \[Fig1\]a and \[Fig1\]b, the round trip of the initial photon number wave packet induces collapse-revivals that are not reminiscent of the SC regime of the JC model [@Eberly80], where initial large coherent states are required. In the DSC limit, with $\omega_0 = 0$, this intuitive picture can be rigorously confirmed integrating the evolution $$\begin{aligned}
|\psi(t)\rangle &=& D^{\dagger} (\beta_0)e^{-i(\omega b^\dagger b - g^2/\omega) t} D(\beta_0)
| + , 0_b \rangle \label{eq:U0}\\
&=& U(t,\omega_0=0)|\psi(0)\rangle = e^{i \frac{g^2}{\omega} t} e^{- i (\frac{g}{\omega})^2 \sin(\omega t)} | + , \beta(t) \rangle , \nonumber\end{aligned}$$ where $\beta(t)=\beta_0(e^{-i\omega t}-1)$ is the amplitude of a coherent state. The revival probability of the initial state reads $$P_{+0_b}(t) = |\langle \psi(0)|\psi(t)\rangle|^2 = e^{-|\beta(t)|^2} ,$$ exhibiting periodic collapses and full revivals [@comment2]. When the initial state is $| + , 2_b \rangle = | {\rm g} , 2_a \rangle$, as in Fig. \[Fig1\]c, the DSC dynamics generates counterpropagating photon number wavepackets in both directions that bounce back and forth producing interference secondary peaks. Similar intuition follows when considering initial superposition states, e. g. $( | + , 0_b \rangle + | + , 2_b \rangle) / \sqrt{2}$, as long as the state components belong to the same parity chain, otherwise no secondary peaks appear. When we break the qubit degeneracy, $\omega_0 \ne 0,$ the intuitive picture remains but we lose the integrability of the problem. Probability still spreads along each parity chain, as seen in Fig. \[Fig2\], but now the photon number wavepacket suffers self-interference, it distorts and its center no longer follows the periodic orbits of $\omega_0=0.$ The result are full collapses and partial revivals where probability $P_{+0_b}$ is not completely restored, and whose maximum value deteriorate as time passes.
![(Color online) (a) Photon statistics at different times of the evolution with $\omega_0=0.5~\omega$. (b) Comparison of probability $P_{+,0_b}(t)$ calculated for $\omega_0=0$ (solid line) and $\omega_0=0.5~\omega$ (dashed line). In all simulations the initial state is $| + , 0_b \rangle$ and $g/\omega=2$.[]{data-label="Fig2"}](fig_2){width="8cm"}
![(Color online) Modulus of the Wigner function, $|W(x,p)|,$ and trajectory $(x , p),$ for Eq. (\[pHamiltonian\]). (a) For $\omega_0=0$ the Wigner function is a symmetric Gaussian (a coherent state) and moves clockwise in phase space along a circle. (b)-(c)-(d) For $\omega_0=0.5~\omega$ the Wigner function loses its symmetry and moves clockwise along a spiral trajectory as shown in the sequence corresponding to $t = 0.5, 1, 5$. In both simulations the initial state is $| + , 0_b \rangle$ and $g/\omega=2$.[]{data-label="Fig3"}](fig_3){width="7.5cm"}
The collapses and revivals have also interesting consequences in phase space, which we have analyzed using the Wigner function and phase space trajectories, $(\bar{x}(t),\bar{p}(t))=\langle (b + b^\dagger, i b^\dagger - i b)/\sqrt{2}\rangle.$ In the integrable case, $\omega_0=0$, the Wigner function of the state is a Gaussian centered on a point $x+ip=\beta(t)$ which draws periodic circular orbits on the plane. As soon as we switch on the term proportional to $\omega_0,$ the wavepacket suffers two distortions, see Fig. \[Fig3\]. The first one is a squeezing tangential to the orbit, shown in Fig. \[Fig3\]c. Accompanied by the difussion and interference in the Wigner wavepacket, the orbits also distort, becoming spirals that relax towards the center of the original orbits, $-\beta_0.$
The phenomenon of collapses and revivals for the nonintegrable case, $\omega_0 \neq 0$, even if partial, reveal a structure in the Hamiltonian spectrum, which is approximately equispaced. We can write the revival probability $$P_{+0_b}(t) = {\Large|} \sum_{\ell} | \langle \psi(0)|\phi_{\ell} \rangle |^2 e^{-i E_{\ell} t/\hbar} {\Large|}^2,
\label{eq:revivals}$$ as a function of the overlap of the initial state with the eigenstates of the full model, $(H - E_{\ell})| \phi_{\ell} \rangle=0.$ When $\omega_0 = 0,$ the eigenenergies are regularly spaced, $E_{\ell} = \hbar \omega \ell$ and the function becomes periodic with period $2\pi/\omega.$ This causes an initial Gaussian wavepacket in phase space to get reconstructed at the same position for $t=2\pi,4\pi,6\pi,\ldots.$ In the DSC case with $\omega_0\neq 0,$ the energy levels deviate very slightly from this regular distribution, $E_{\ell} = \hbar\omega {\ell} - \hbar \delta_{\ell} \omega_0,$ where the correction $\delta_{\ell} \omega_0$ is less than $10\%$ for the examples considered in this work. The reconstruction of the wavepackets is incomplete and different partial waves may get delayed or accelerated with respect to the original orbit, $\beta(t)$. This causes the squeezing of the Wigner function and the self-interference in the photon number wavepacket, as displayed in Fig. \[Fig2\]a.
![(Color online) Collapses and revivals of $P_{+,0_b}(t)$ for $g/\omega=2$ and (a) $\omega_0=0.3\omega$ and (b) $\omega_0=0.5\omega.$ We plot the exact numerical solution (area) at first order in $\omega_0/\omega$ (solid) and a two-mode approximation (red, dashed). (c) Distribution of probability of the different detunings $\delta_{n_b} = (\hbar\omega n_b - E_{n_b}) / \hbar\omega_0,$ weighted by their contribution to the wavefuncion given in Eq. (\[eq:revivals\]). We have marked the level $N^r_b$ which is used for the two-mode approximation in curves (a) and (b).[]{data-label="Fig4"}](fig_4){width="7.5cm"}
We have also found that the overall dynamics is very accurately captured by the first order correction to the eigenenergies, shown in Eq. (\[eq:approx\]). If we use $\delta_{n_b} = (-1)^{n_b} p \Delta_{n_b n_b}/2$ in Eq. (\[eq:revivals\]), together with the initial condition $\langle \psi(0) | \phi_{n_b}\rangle = \langle \psi(0) | D(-\beta_0) |p {n_b} \rangle = \delta_{+,p} \exp(-|\beta|^2/2)\beta^{n_b}/{n_b}!$, we obtain curves that approximate very well the exact result. This is shown in Figs. \[Fig4\]a and \[Fig4\]b for values of $\omega_0=0.3\omega$ and $0.5\omega$ in the DSC regime, $g/\omega=2$. In both cases the revivals happen close to $t = k \,(2\pi / \omega)$, with integer $k$, but decreasing in intensity and with a large fraction of the curve moving with a slower speed, reconstructing itself at later times. The effect of this is more evident in Fig. \[Fig4\]b, and also in the photon number wavepacket plot in Fig. \[Fig3\]a, where one appreciates two waves with slightly different periods interfering with each other. The existence of this delayed revivals is due to the structure of the wavefunction $\psi(t)$ that, as shown in Fig. \[Fig4\]c, is composed of many contributions close to zero detuning, $\delta_{n_b}\simeq 0,$ and a few large contributions with $\delta=-0.116,-0.223.$ The former constitute the main revivals, while the second ones make the revivals at slightly longer periods, $2\pi\omega-0.223\omega_0,$ forming the second wavefront in Fig. \[Fig2\]a and Fig. \[Fig4\]a-b.
Based on our previous results, we have developed a heuristic approximation that allows us to reproduce the main revivals. Our method recognizes that if we start the dynamics with state $|\psi(0)\rangle=|+,N_b\rangle,$ the main contribution to the wavefunction is around a level $N^r_b=[(g/\omega)^2+N_b],$ where $[\cdot]$ denotes the closest integer. This is indeed the case in the considered examples, as shown in Fig. \[Fig4\]c. We will only consider the energy correction for this level, $\delta_{N^r_b},$ and neglect the dephasing of all other *off-resonant* terms. Under this criteria, we approximate the system state, up to normalization, as $$\begin{aligned}
| \psi(t) \rangle \approx && U(t,\omega_0=0) | \psi(0) \rangle + \psi_{N^r_b}(e^{-i(\omega N_b \nonumber - \omega_0\Delta_{N^r_b N^r_b} /2)t} \\ && - e^{-i\omega N_b t}) D(-\beta_0) |+,N^r_b\rangle,
\label{jcstate1} \end{aligned}$$ where $U(t,\omega_0=0)$ is the evolution operator at $\omega_0=0$ introduced in Eq. (\[eq:U0\]), and $\psi_{N^r_b}=\langle \phi_{N^r_b}|\psi(0)\rangle$. This form of the state is motivated by the behavior of the Wigner function, as seen in Fig. \[Fig3\](b): a central core approximated by the solution of $\omega_0=0$ plus a delayed correction capturing the effects of $\omega_0,$ forming the squeezed tail. Considering the simplest case of $N_b = 0$, from state in Eq. (\[jcstate1\]), it is straightforward to obtain a simple analytical expression for the revival probability $$\begin{aligned}
\label{prob}
P_{+0_b}(t)
\approx & 2 & e^{-|\beta(t)|^2/2-\beta_0^2} \frac{\beta_0^{2N^r_b}}{N^r_b!}\left[ \cos (\omega_0 \delta_{N^r_b} t/2) -1 \right] \nonumber\\
& + & e^{-|\beta(t)|^2},
\label{jcapprox}\end{aligned}$$ where small terms are neglected. This expression has been compared with the exact solution, as shown in Fig. \[Fig4\]a and \[Fig4\]b, giving a good estimate of the height of the partial revivals as a function of time and $\omega_0.$ The other features, such as the delayed front is not reproduced because this approximation does not contain the contributions with $\delta=-0.223,$ but this can be improved by including more resonant levels.
![(Color online) Parity chains for the two-qubit DSC regime.[]{data-label="Fig5"}](fig_5){width="7.5cm"}
[*Two qubits and a mode*]{}$.-$ It is also possible to give a qualitative and quantitative description of the DSC considering the case of two qubits. Here, the $2\otimes 2 \otimes N$-dimensional Hilbert space bifurcates into two independent parity chains of tetrahedra, see Fig. \[Fig5\], where each vertex is connected to their neighbors via rotating or counter-rotating terms. The same dynamical properties of probability collapses and revivals can be found, as well as interesting entanglement properties.
[*Conclusions*]{}$.-$ The SC regime of the JC model is considered nowadays an intuitive and comprehensive field. The USC regime is described by the SC regime plus RWA and higher-order corrections. In this work, we have aimed at developing an insightful description of the DSC regime of the JC model. The transition between the SC and DSC regimes remains a rather diffuse crossover with well understood frontiers.
We acknowledge funding from Basque Government grants BFI08.211 and IT472-10; Spanish MICINN projects FIS2009-12773-C02-01, FIS2009-10061, and Juan de la Cierva Program; QUITEMAD, and SOLID European project.
[99]{}
E. T. Jaynes and F. W. Cummings, Proc. IEEE **51**, 89 (1963).
J. H. Eberly, N. B. Narozhny, and J. J. Sanchez-Mondragon, Phys. Rev. Lett. [**44**]{}, 1323 (1980).
J. M. Raimond, M. Brune, and S. Haroche, Rev. Mod. Phys. [**73**]{}, 565 (2001).
H. Walther, B. T. H. Varcoe, B. G. Englert, and T. Becker, [*Rep. Prog. Phys.*]{} **69**, 1325 (2006).
D. Leibfried, R. Blatt, C. Monroe, and D. Wineland, Rev. Mod. Phys. [**75**]{}, 281 (2003).
J. Clarke and F. K. Wilhelm, Nature **453**, 1031 (2008).
A. Wallraff [*et al.*]{}, Nature **431**, 162 (2004).
A. Blais [*et al.*]{}, Phys. Rev. A **69**, 062320 (2004).
I. [Chiorescu]{} [*et al.*]{}, Nature **431**, 159 (2004).
M. Hofheinz [*et al.*]{}, Nature [**459**]{}, 546 (2009).
A. N. Cleland and M. R. Geller, Phys. Rev. Lett. **93**, 070501 (2004).
M. D. LaHaye [*et al.*]{}, Nature **459**, 960 (2009).
G. Günter [*et al.*]{}, Nature. **458**, 178 (2009).
A. T. Sornborger, A. N. Cleland, and M. R. Geller, Phys. Rev. A **70**, 052315 (2004).
A. A. Abdumalikov, O. Astafiev, Y. Nakamura, Y. A. Pashkin, and J. Tsai, Phys. Rev. B **78**, 180502 (2008).
J. Bourassa [*et al.*]{}, Phys. Rev. A **80**, 032109 (2009).
B. Peropadre, P. Forn-Díaz, E. Solano, and J. J. García-Ripoll, Phys. Rev. Lett. [**105**]{}, 023601 (2010).
I. Lizuain, J. Casanova, J. J. García-Ripoll, J. G. Muga, and E. Solano, Phys. Rev. A **81**, 062131 (2010).
T. Niemczyk, [*et al.*]{}, Nature Phys. [**6**]{}, 772 (2010).
P. Forn-Díaz, [*et al.*]{}, Phys. Rev. Lett. [**105**]{}, 237001 (2010).
E. K. Irish, Phys. Rev. Lett. [**99**]{}, 173601 (2007).
S. Ashhab and F. Nori, Phys. Rev. A [**81**]{}, 042311 (2010).
Q.-H. Chen, Y.-Y. Zhang, T. Liu, and K.-L. Wang, Phys. Rev. A [**78**]{}, 051801 (2008).
M.-J. Hwang and M.-S. Choi, Phys. Rev. A [**82**]{}, 025802 (2010).
Q.-H. Chen, L. Li, T. Liu, and K.-L. Wang, arXiv: 1007.1747.
For the strong coupling of the Jaynes-Cummings model in ion traps, see: J. F. Poyatos [*et al.*]{}, Phys. Rev. A [**54**]{}, 1532 (1996), and W. C. Campbell [*et al.*]{}, Phys. Rev. Lett. [**105**]{}, 090502 (2010).
Various parity operators have been previously used, see H. B. Shore and L. M. Sander, Phys. Rev. B [**7**]{}, 4537 (1973).
The case of $\omega_0 = 0$ in Eq. (\[tHamiltonian\]) is related to studies in bichromatically excited trapped ions: K. M[ø]{}lmer and A. S[ø]{}rensen, Phys. Rev. Lett. [**82**]{}, 135 (1999), and E. Solano, R. L. de Matos Filho, and N. Zagury, Phys. Rev. Lett. [**87**]{}, 060402 (2001).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Controlling complex networked systems to a desired state is a key research goal in contemporary science. Despite recent advances in studying the impact of network topology on controllability, a comprehensive understanding of the synergistic effect of network topology and individual dynamics on controllability is still lacking. Here we offer a theoretical study with particular interest in the diversity of dynamic units characterized by different types of individual dynamics. Interestingly, we find a global symmetry accounting for the invariance of controllability with respect to exchanging the densities of any two different types of dynamic units, irrespective of the network topology. The highest controllability arises at the global symmetry point, at which different types of dynamic units are of the same density. The lowest controllability occurs when all self-loops are either completely absent or present with identical weights. These findings further improve our understanding of network controllability and have implications for devising the optimal control of complex networked systems in a wide range of fields.'
author:
- Chen Zhao
- 'Wen-Xu Wang'
- 'Yang-Yu Liu'
- 'Jean-Jacques Slotine'
title: Individual dynamics induces symmetry in network controllability
---
As a key notion in control theory, controllability denotes our ability to drive a dynamic system from any initial state to any desired final state in finite time [@Kalman-JSIAM-63; @Luenberger-Book-79]. For the canonical linear time-invariant (LTI) system $\dot{\bf x}= A
{\bf x} + B {\bf u}$ with state vector ${\bf x} \in
\mathbb{R}^N$, state matrix $A \in \mathbb{R}^{N\times N}$ and control matrix $B \in \mathbb{R}^{N\times M}$, Kalman’s rank condition $\text{rank}[B,AB,\cdots, A^{N-1}B] = N$ is sufficient and necessary to assure controllability. Yet, in many cases system parameters are not exactly known, rendering classical controllability tests impossible. By assuming that system parameters are either fixed zeros or freely independent, structural control theory (SCT) helps us overcome this difficulty for linear time-invariant systems [@Lin:1974; @Shields-IEEE-76; @Hosoe-IEEE-80; @SContrl1; @Dion-Automatica-03]. Quite recently, many research activities have been devoted to study the structural controllability of systems with complex network structure, where system parameters (e.g., the elements in $A$, representing link weights or interaction strengths between nodes) are typically not precisely known, only the zero-nonzero pattern of $A$ is known [@LSB:2011; @NV:2012; @LSB:pOne; @WNLG:2012; @Kurth:2012; @WGG:2012; @PLSB:2013; @bimodality:2013]. Network controllability problem can be typically posed as a combinatorial optimization problem, i.e., identify a minimum set of driver nodes, with size denoted by $N_\mathrm{D}$, whose control is sufficient to fully control the system’s dynamics [@LSB:2011]. Other controllability related issues, e.g., energy cost, have also been extensively studied for complex networked systems [@Yan:2012; @GSZPB:2012; @LSB:observability; @SM:PRL]. While the intrinsic individual dynamics can be incorporated in the network model, it would be more natural and fruitful to consider their effect separately. Hence, most of the previous studies focused on the impact of network topology, rather than the individual dynamics of nodes, on network controllability [@LSB:2011; @WNLG:2012].
If one explores the impact of individual dynamics on network controllability in the SCT framework, a specious result would be obtained — a single control input can make an arbitrarily large linear system controllable. Although this result as a special case of the minimum inputs theorem can be proved [@LSB:2011] and its implication was further emphasized in [@Cowan:2012], this result is inconsistent with empirical situations, implying that the SCT is inapplicable in studying network controllability, if individual dynamics of nodes are imperative to be incorporated to capture the collective dynamic behavior of a networked system. To overcome this difficulty, and more importantly, to understand the impact of individual dynamics on network controllability, we revisit the key assumption of SCT, i.e., the independency of system parameters. We anticipate that major new insights can be obtained by relaxing this assumption, e.g., considering the natural diversity and similarity of individual dynamics. This also offers a more realistic characterization of many real-world networked systems where not all the system parameters are completely independent.
To solve the network controllability problem with dependent system parameters, we rely on the recently developed exact controllability theory (ECT) [@exact_control]. ECT enables us to systematically explore the role of individual dynamics in controlling complex systems with arbitrary network topology. In particular, we consider prototypical linear forms of individual dynamics (from first-order to high-orders) that can be incorporated within the network representation of the whole system in a unified matrix form. This paradigm leads to the discovery of a striking symmetry in network controllability: if we exchange the fractions of any two types of dynamic units, the system’s controllability (quantified by $N_\mathrm{D}$) remains the same. This exchange-invariant property gives rise to a global symmetry point, at which the highest controllability (i.e., lowest number of driver nodes) emerges. This symmetry-induced optimal controllability holds for any network topology and various categories of individual dynamics. We substantiate these findings numerically in a variety of network models.\
Exact controllability theory (ECT) [@exact_control] claims that for arbitrary network topology and link weights characterized by the state matrix $A$ in the LTI system $\dot{\mathbf{x}}= A\mathbf{x}+B\mathbf{u} $, the minimum number of driver nodes $N_\text{D}$ required to be controlled by imposing independent signals to fully control the system is given by the maximum geometric multiplicity $\max_i\{\mu(\lambda_i)\}$ of $A$’s eigenvalues $\{\lambda_i\}$ [@multi1; @multi2; @multi3; @multi4; @multi5]. Here $\mu(\lambda_i) \equiv N-
\text{rank}(\lambda_i I_N-A)$ is the geometric multiplicity of the eigenvalue $\lambda_i$ and $I_N$ is the identity matrix. Calculating all the eigenvalues of $A$ and subsequently counting their geometric multiplicities are generally applicable but computationally prohibitive for large networks. If $A$ is symmetric, e.g., in undirected networks, $N_\text{D}$ is simply given by the maximum algebraic multiplicity $\max_i\{\delta(\lambda_i)\}$, where $\delta(\lambda_i)$ denotes the degeneracy of eigenvalue $\lambda_i$. Calculating $N_\text{D}$ in the case of symmetric $A$ is more computationally affordable than in the asymmetric case. Note that for structured systems where the elements in $A$ are either fixed zeros or free independent parameters, ECT offers the same results as that of the SCT [@exact_control].\
We first study the simplest case of first-order individual dynamics $\dot{x}_i=a_0x_i$. The dynamical equations of a linear time-invariant control system associated with first-order individual dynamics [@Slotine:book] can be written as $$\label{eq:1st_order_node}
\dot{\mathbf{x}}=\Lambda\mathbf{x}+ A\mathbf{x}+B\mathbf{u} = \Phi\mathbf{x}+B\mathbf{u},$$ where the vector $\mathbf{x} = (x_1,\cdots,x_N)^\text{T}$ captures the states of $N$ nodes, $\Lambda\in \mathbb{R}^{N\times N}$ is a diagonal matrix representing intrinsic individual dynamics of each node, $A\in \mathbb{R}^{N\times N}$ denotes the coupling matrix or the weighted wiring diagram of the networked system, in which $a_{ij}$ represents the weight of a directed link from node $j$ to $i$ (for undirected networks, $a_{ij}=a_{ji}$). $\mathbf{u}=(u_1,u_2,\cdots,u_M)^\text{T}$ is the input vector of $M$ independent signals, $B \in \mathbb{R}^{N\times M}$ is the control matrix, and $\Phi \equiv \Lambda + A$ is the state matrix. Without loss of generality, we assume $\Lambda$ is a “constant" matrix over the field $\mathbb{Q}$ (rational numbers), and $A$ is a structured matrix over the field $\mathbb{R}$ (real numbers). In other words, we assume all the entries in $\Phi$ have been rescaled by the individual dynamics parameters. The resulting state matrix $\Phi$ is usually called a *mixed matrix* with respect to $(\mathbb{Q}, \mathbb{R})$ [@mixed_matrix]. The first-order individual dynamics in $\Phi$ is captured by self-loops in the network representation of $\Phi$ (see Fig. \[fig:illus\_motif\]a). $N_\text{D}$ can then be determined by calculating the maximum geometric multiplicity $\max_i\{ \mu(\lambda_i)\}$ of $\Phi$’s eigenvalues.
We study two canonical network models (Erdös-Rényi and Scale-free) with random edge weights and a $\rho_s$ fraction of nodes associated with identical individual dynamics (i.e., self-loops of identical weights). As shown in Fig. \[fig:1st\_order\]a,b, the fraction of driver nodes $n_\text{D}\equiv N_\text{D}/N$ is symmetric about $\rho_\text{s}=0.5$, regardless of the network topology. Note that the symmetry cannot be predicted by SCT in the sense that in case of completely independent self-loop weights $n_\text{D}$ will monotonically decrease to $1/N$ as $\rho_\text{s}$ increases to 1, implying that a single driver node can fully control the whole network [@Cowan:2012]. The symmetry can be theoretically predicted (see SM Sec.2.2). An immediate but counterintuitive result from the symmetry is that $n_\text{D}$ in the absence of self-loops is exactly the same as the case that each node has a self-loop with identical weight. This is a direct consequence of Kalman’s rank condition for controllability [@Kalman-JSIAM-63]:
$$\text{rank}[B,AB,\cdots, A^{N-1}B] =
\text{rank}[B,(A+w_\text{s}I_N)B,\cdots, (A+w_\text{s}I_N)^{N-1}B]
\label{eq:kalman_equal}$$
where the left and the right hand sides are the rank of controllability matrix in the absence and full of identical self-loops, respectively (see SM Sec.1 for proof).
The presence of two types of nonzero self-loops $s_2$ and $s_3$ leads to even richer behavior of controllability. If the three types of self-loops (including self-loops of zero weights) are randomly distributed at nodes, the impact of their fractions on $n_\text{D}$ can be visualized by mapping the three fractions into a 2D triangle (or 2-simplex), as shown in Fig. \[fig:1st\_order\]c,d. We see that $n_\text{D}$ exhibits symmetry in the triangle and the minimum $n_\text{D}$ occurs at the center that represents identical fractions of the three different self-loop types. The symmetry-induced highest controllability can be generalized to arbitrary number of self-loops. Assume there exist $n$ types of self-loops $s_1,\cdots ,s_n$ with weights $w_\text{s}^{(1)},\cdots ,w_\text{s}^{(n)}$, respectively, we have $$\begin{aligned}
N_\text{D}= N- \min_i \bigg\{ \text{rank}\big(\Phi-w_\text{s}^{(i)} I_N \big) \bigg\}
\label{eq:unify_1st}\end{aligned}$$ for sparse networks with random weights (see SM Sec. 2 for detailed derivation and the formula of dense networks). An immediate prediction of Eq. (\[eq:unify\_1st\]) is that $N_\text{D}$ is primarily determined by the self-loop with the highest density, simplifying Eq. (\[eq:unify\_1st\]) to be $N_\text{D} = N- \text{rank}(\Phi-w_\text{s}^\text{max} I_N )$, where $w_\text{s}^\text{max}$ is the weight of the prevailing self-loop (see SM Sec. 2). Using Eq. (\[eq:unify\_1st\]) and the fact that $\Phi$ is a mixed matrix, we can predict that $N_\text{D}$ remains unchanged if we exchange the densities of any two types of self-loops (see SM Sec. 2), accounting for the symmetry of $N_\text{D}$ for arbitrary types of self-loops. Due to the dominance of $N_\text{D}$ by the self-loop with the highest density and the exchange-invariance of $N_\text{D}$, the highest controllability with the lowest value of $N_\text{D}$ emerges when distinct self-loops are of the same density.
To validate the symmetry-induced highest controllability predicted by our theory, we quantify the density heterogeneity of self-loops as follows: $$\Delta \equiv \sum_{i=1}^{N_\text{s}}\left| \rho_\text{s}^{(i)}-\frac{1}{N_\text{s}} \right|,$$ where $N_\text{s}$ is the number of different types of self-loops (or the diversity of self-loops). Note that $\Delta=0$ if and only if all different types of self-loops have the same density, i.e., $\rho_\text{s}^{(1)}=
\rho_\text{s}^{(2)}=\cdots \rho_\text{s}^{(N_\text{s})}=\frac{1}{N_\text{s}}$, and the larger value of $\Delta$ corresponds to more diverse case. Figure \[fig:multi\_dynamics\]a,b shows that $n_\text{D}$ monotonically increases with $\Delta$ and the highest controllability (lowest $n_\text{D}$) arises at $\Delta=0$, in exact agreement with our theoretical prediction. Figure \[fig:multi\_dynamics\]c,d display $n_\text{D}$ as a function of $N_\text{s}$. We see that $n_\text{D}$ decreases as $N_\text{s}$ increases, suggesting that the diversity of individual dynamics facilitates the control of a networked system. When $N_\text{s}=N$ (i.e., all the self-loops are independent), $n_\text{D}=1/N$, which is also consistent with the prediction of SCT [@LSB:2011; @Cowan:2012].\
In some real networked systems, dynamic units are captured by high-order individual dynamics, prompting us to check if the symmetry-induced highest controllability still holds for higher-order individual dynamics. The graph representation of dynamic units with 2nd-order dynamics is illustrated in Fig. \[fig:illus\_motif\]b. In this case, the eigenvalues of the dynamic unit’s state matrix $\left(
\begin{array}{cc}
0 & 1 \\
a_0 & a_1 \\
\end{array}
\right)$ play a dominant role in determining $N_\text{D}$. For two different units as distinguished by distinct ($a_0$ $a_1$) one can show that their state matrices almost always have different eigenvalues, except for some pathological cases of zero measure that occur when the parameters satisfy certain accidental constraints. The eigenvalues of the state matrix of dynamic units take over the roles of self-loops in the 1st-order dynamics, accounting for the following formulas for sparse networks $$N_\text{D} = 2N- \min_i\bigg\{ \text{rank}(\Phi - \lambda^{(i)} I_{2N}) \bigg\},
\label{eq:2nd_nodal_formu}$$ where $\lambda^{(i)}$ is either one of the two eigenvalues of type-$i$ dynamic unit’s state matrix. The formula implies that $N_\text{D}$ is exclusively determined by the prevailing dynamic unit, (see SM Sec. 2). The symmetry of $N_\text{D}$, i.e., exchanging the densities of any types of dynamic units, does not alter $N_\text{D}$ (see SM Sec. 2), and the emergence of highest controllability at the global symmetry point can be similarly proved as we did in the case of 1st-order individual dynamics.
The 3rd-order individual dynamics are graphically characterized by a dynamic unit composed of three nodes (Fig. \[fig:illus\_motif\]c), leading to a $3N\times 3N$ state matrix (Fig. \[fig:illus\_motif\]c). We can generalize Eq. (\[eq:2nd\_nodal\_formu\]) to arbitrary order of individual dynamics: $$N_\text{D} = dN- \min_i\bigg\{ \text{rank}(\Phi - \lambda^{(i)}_d I_{dN}) \bigg\},
\label{eq:unify_any_order}$$ where $d$ is the order of the dynamic unit, $\lambda^{(i)}_d$ is any one of the $d$ eigenvalues of type-$i$ dynamic units and $I_{dN}$ is the identity matrix of dimension $dN$. In analogy with the simplified formula for the 1st-order dynamics, insofar as a type of individual dynamics prevails in the system, Eq (\[eq:unify\_any\_order\]) is reduced to $N_\text{D} = dN- \text{rank}(\Phi - \lambda^\text{max}_d I_{dN})$, where $\lambda^\text{max}_d$ is one of the eigenvalues of the prevailing dynamic unit’s state matrix. Similar to the case of 1st-order individual dynamics, the global symmetry of controllability and the highest controllability occurs at the global symmetry point can be proved for individual dynamics of any order and arbitrary network topology (see SM Sec.2 and 3 for theoretical derivations and see SM Sec. 4 for numerical and analytical results of high-order individual dynamics).
In summary, we map individual dynamics into dynamic units that can be integrated into the matrix representation of the system, offering a general paradigm to explore the joint effect of individual dynamics and network topology on the system’s controllability. The paradigm leads to a striking discovery: the universal symmetry of controllability as reflected by the invariance of controllability with respect to exchanging the fractions of any two different types of individual dynamics, and the emergence of highest controllability at the global symmetry point. These findings generally hold for arbitrary networks and individual dynamics of any order. The symmetry-induced highest controllability has immediate implications for devising and optimizing the control of complex systems by for example, perturbing individual dynamics to approach the symmetry point without the need to adjust network structure.
The theoretical paradigm and tools developed here also allow us to address a number of questions, answers to which could offer further insights into the control of complex networked systems. For example, we may consider the impact of general parameter dependency (e.g., link weight similarity), instead of focusing on self-loops or individual dynamics. Our preliminary results show that introducing more identical link weights will not affect the network controllability too much, unless the network is very dense and almost all link weights are identical (see SM Sec.5). We still lack a comprehensive understanding of the impact of parameter dependency on structural controllability for arbitrary complex networks. Moreover, at the present we are incapable of tackling general nonlinear dynamical systems in the framework of ECT, which is extremely challenging for both physicists and control theorists. Nevertheless, we hope our approach could inspire further research interests towards achieving ultimate control of complex networked systems.
[99]{}
R. E. Kalman, J. Soc. Indus. and Appl. Math Ser. A [**1**]{}, 152 (1963).
D. G. Luenberger, [*Introduction to Dynamic Systems: Theory, Models, & Applications*]{} (John Wiley & Sons, New York, 1979).
C. T. Lin, IEEE Trans. Autom. Control [**19**]{}, 201 (1974).
R. W. Shields, and J. B. Pearson, IEEE Trans. Autom. Control [**21**]{}, 203 (1976).
S. Hosoe, IEEE Trans. Autom. Control [**25**]{}, 1192 (1980).
C. Commault, J.-M. Dion, and J. W. van der Woude, Kybernetika [**38**]{}, 503 (2002).
J.-M. Dion, C. Commault, and J. van der Woude, Automatica [**39**]{}, 1125 (2003).
Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási, Nature [**473**]{}, 167 (2011). T. Nepusz, and T. Vicsek, Nature Phys. [**8**]{}, 568-573 (2012).
Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási, PLoS ONE [**7**]{}, e44459 (2012).
W.-X. Wang, X. Ni, Y.-C. Lai, and C. Grebogi, Phys. Rev. E [**85**]{}, 026115 (2012).
Y. Tang, H. Gao, W. Zou, and J. Kurths, PLoS ONE [**7**]{}, e41375 (2012).
B. Wang, L. Gao, and Y. Gao, J. Stat. Mech. : Theor. Exp. [**2012**]{}, P04011 (2012).
M. Pósfai, Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási, Sci. Rep. [**3**]{}, 1067 (2013).
T. Jia, Y.-Y. Liu, E. Csóka, M. Pósfai, J.-J. Slotine, and A.-L. Barabási, Nature Commun. [**4**]{}, 2002 (2013).
G. Yan, J. Ren, Y.-C. Lai, C.-H. Lai, and B. Li, Phys. Rev. Lett. [**108**]{}, 218703 (2012).
R. Gutiérez, I. Sendiña-Nadal, M. Zanin, D. Papo, and S. Boccaletti, Sci. Rep. [**2**]{}, 396 (2012).
Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási, Proc. Natl Acad. Sci. [**110**]{}, 2460-2465 (2013).
J. Sun, and A. E. Motter, Phys. Rev. Lett. [**110**]{}, 208701 (2013).
N. J. Cowan, E. J. Chastain, D. A. Vilhena, J. S. Freudenberg, and C. T. Bergstrom, PLoS ONE [**7**]{}, e38398 (2012).
Z. Yuan, C. Zhao, Z. Di, W.-X. Wang, and Y.-C. Lai, Nature Commun. [**4**]{}, 2447 (2013).
E. D. Sontag, [*Mathematical Control Theory: Deterministic Finite Dimensional Systems*]{} (Springer, 2nd edn, 1998).
P. J. Antsaklis, and A. N. Michel, [*Linear Systems*]{} (McGraw-Hill, 1997).
N. Cai, J.-X. Xi, Y.-S. Zhong, and H.-Y. Ma, Int. J. Innov. Comput. I [**8**]{}, 3315 (2012).
G. Parlangeli, and G. Notarstefano, IEEE Trans. Autom. Control [**57**]{}, 743 (2012).
G. Notarstefano, and G. Parlangeli, IEEE Trans. Autom. Control [**58**]{}, 1719 (2013).
J.-J. Slotine, and W. Li, [*Applied Nonlinear Control*]{} (Prentice-Hall, 1991).
K. Murota, [*Matrices and Matroids for Systems Analysis*]{} (Springer Heidelberg Dordrecht London, New York, USA, 2010).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that Lagrangian submanifolds in six-dimensional nearly Kähler (non Kähler) manifolds and in twistor spaces $Z^{4n+2}$ over quaternionic Kähler manifolds $Q^{4n}$ are minimal. Moreover, we will prove that any Lagrangian submanifold $L$ in a nearly Kähler manifold $M$ splits into a product of two Lagrangian submanifolds for which one factor is Lagrangian in the strict nearly Kähler part of $M$ and the second factor is Lagrangian in the Kähler part of $M$. Using this splitting theorem we then describe Lagrangian submanifolds in nearly Kähler manifolds of dimensions six, eight and ten.'
address: 'Leibniz Universität Hannover, Institut für Differentialgeometrie, Welfengarten 1, 30167 Hannover, Germany'
author:
- Lars Schäfer
- Knut Smoczyk
title: Decomposition and minimality of Lagrangian submanifolds in nearly Kähler manifolds
---
[^1]
Introduction
============
Suppose $(M,\omega)$ is a symplectic manifold. Then a submanifold $L\subset M$ is called [*Lagrangian*]{}, if $\omega_{|TL}=0$ and $2\dim(L)=\dim M$. If $L$ can be embedded (immersed) as a Lagrangian submanifold in $\complex{n}$, then by Darboux’s theorem $L$ can also be embedded (immersed) into any other symplectic manifold $(M,\omega)$. One of the most interesting examples for symplectic manifolds are Kähler manifolds $(M,J,g)$, where the symplectic form $\omega$ is given by $\omega(X,Y)=g(JX,Y)$.
In this paper we will consider Lagrangian submanifolds in nearly Kähler manifolds. An almost Hermitian manifold $(M,J,g)$ is called [*nearly Kähler*]{}, if its almost complex structure $J$ satisfies $$\label{nk1}
\nabla_X(J)X=0\,,\quad\forall\,X\in TM\,.$$ A nearly Kähler manifold $(M,J,g)$ is Kähler, iff $\nabla J=0$. $(M,J,g)$ is called strict nearly Kähler, if $\nabla_XJ\neq 0$ for all $X\in TM$, $X\neq 0$.
In contrast to the Kähler case, nearly Kähler (non Kähler) manifolds are not symplectic manifolds (at least not with their characteristic $2$-form $\omega$).
Anyway, also in this case, it is common to say that $L\subset M$ is Lagrangian, if $\omega_{|TL}=0$ and if $L$ has half the dimension of $M$. Under the weaker condition that $(M,J,g)$ is nearly Kähler, it may not be possible to find Lagrangian submanifolds of a given topological type $L$ in $M$, even if $L$ is a Lagrangian submanifold in $\complex{n}$. This is because there is no analogue of Darboux’s theorem in the nearly Kähler case. Nevertheless, there exist numerous examples for nontrivial Lagrangian submanifolds in the nearly Kähler six-sphere $S^6$ [@Vrancken-2003].
Nearly Kähler geometry was first studied in the 1970s by Gray (see [@Gray-1970],[@Gray-1976]) in the context of weak holonomy. The most prominent example of a nearly Kähler manifold is $S^6$ with its standard almost complex structure and Riemannian metric. We shortly resume the state of the classification of nearly Kähler manifolds $M$. The reader is invited to consult section 5 of [@Butruille-2006] for a short survey with detailed information and more references. In the rest of this paragraph we consider $M$ to be complete and simply connected. After splitting of the Kähler factor $M$ can supposed to be strict nearly Kähler. Nagy [@Nagy-2002a] has reduced the classification of strict nearly Kähler manifolds using previous work of Cleyton and Swann [@Cleyton-Swann] to almost hermitian products of:
- (naturally reductive) three-symmetric spaces [@Gray-1972; @Gray-Wolf],
- twistor spaces of non locally symmetric positive quaternionic Kähler manifolds endowed with their non-integrable complex structure,
- irreducible six-dimensional nearly Kähler manifolds.
The class of examples from three-symmetric spaces can be divided into three types. One type are nearly Kähler structures on twistor spaces over symmetric spaces. In this case the isotropy representation is complex reducible. Another type are those with irreducible isotropy representation and for the third type the isotropy representation is real reducible. In the case of a homogenous nearly Kähler six-manifold Butruille [@Butruille-2005] showed that there are only the previous known examples [@Gray-1972; @Gray-Wolf] $$\begin{aligned}
S^6&=&G_2/SU_3, S^3 \times S^3=SU_2\times SU_2\times SU_2/SU_2,\\ \mathbb
CP^3&=&SO_5/U_2\times S^1 \mbox{ and } F(1,2)=SU_3/T^2,\end{aligned}$$ which are all three-symmetric spaces with the metric defined by the Killing form. In dimension eight strict nearly Kähler manifolds are almost hermitian products of a six dimensional nearly Kähler factor and a Riemannian surface. In dimension ten strict nearly Kähler manifolds are either almost hermitian products of a six dimensional nearly Kähler factor and a complex surface or they come from twistor spaces. Therefore we later study twistor spaces and three-symmetric spaces in more detail.
Since there is no Darboux theorem, the existence of Lagrangian submanifolds in nearly Kähler manifolds, even locally, is not unobstructed. Indeed, in this paper we will show that Lagrangian submanifolds tend to be minimal in the strict nearly Kähler directions of $M$, a fact which was previously known only for the nearly Kähler six-sphere $S^6$ [@Ejiri-1981]. To be precise, among other things we will prove the following three theorems:
[**Theorem A:**]{} [*If $L$ is a Lagrangian submanifold in a strict nearly Kähler six-manifold, then $L$ is orientable and minimal.*]{}
[**Theorem B:**]{} [*If $L\subset Z^{4n+2}$ is a Lagrangian submanifold in a twistor space over a positive quaternionic Kähler manifold $M^{4n}$, then $L$ is minimal. Moreover if $n>1$, then $TL=\mathcal{D}^\perp\oplus\mathcal{D}$ decomposes into a $2n$-dimensional horizontal distribution $\mathcal{D}^\perp$ and a vertical line bundle $\mathcal{D}$ and the second fundamental form $II(\mathcal{D},\cdot)$ of the vertical part vanishes identically.* ]{}
[**Theorem C:**]{} [*Let $M$ be a nearly Kähler manifold and $L\subset M$ be Lagrangian. Then $M$ and $L$ decompose locally into products $M=M_K\times M_{SNK}$, $L=L_K\times L_{SNK}$, where $M_K$ is Kähler, $M_{SNK}$ is strict nearly Kähler and $L_K\subset M_K$, $L_{SNK}\subset M_{SNK}$ are both Lagrangian. The dimension of $L_K$ is given by $$\dim L_K=\frac{1}{2}\dim\operatorname{ker}(r)\,,$$ where $r=\operatorname{Ric}-\operatorname{Ric}^*$ is defined as in (\[op r\]). Moreover, if the splitting of $M$ is global and $L$ is simply connected, then $L$ decomposes globally as well.* ]{}
The organization of the paper is as follows: In section \[sec 1\] we will recall the basic geometric quantities related to nearly Kähler geometry. Section \[sec 2\] concerns the geometry of Lagrangian submanifolds $L$ in nearly Kähler manifolds. Here we will define some new tensors, that are couplings of the second fundamental form $II$ of $L$ with the torsion of the canonical connection $\nb$. Then we will prove some important identities related to these quantities. In section \[sec 3\] we will consider the case of Lagrangian submanifolds in six-dimensional nearly Kähler manifolds and we will prove our main theorem A. Section \[sec 5\] is devoted to the decomposition of Lagrangian submanifolds $L$ in nearly Kähler manifolds $M$ and we will show that the Lagrangian condition is compatible with the splitting of the nearly Kähler manifold $M$ induced by the operator $r=\operatorname{Ric}-\operatorname{Ric}^*$. This will be theorem C. We will need this result to prove theorem B in the following section, where we analyse Lagrangian submanifolds in twistor spaces over positive quaternionic Kähler manifolds in more detail. In the remainder we will prove that left invariant Lagrangian submanifolds in three-symmetric spaces are totally geodesic (Section \[sec 7\]) and then we will analyse in Section \[sec 8\] the space of deformations of a Lagrangian submanifold in strict nearly Kähler six-manifolds in terms of coclosed eigenforms of the Hodge-Laplacian.
The geometry of nearly Kähler manifolds {#sec 1}
=======================================
Let $(M,J,g)$ be a smooth nearly Kähler manifold with Riemannian metric $\langle\cdot,\cdot\rangle:=g(\cdot,\cdot)$, Levi-Civita connection $\nabla$ and almost complex structure $J\in\operatorname{End}(TM)=\Gamma(TM\otimes T^*M)$. Then the nearly Kähler condition is $$\begin{aligned}
\langle JX,JY\rangle&=&\langle X,Y\rangle\,,\quad\forall\, X,\,Y\in TM\,,\nonumber\\
J^2&=&-\operatorname{Id}\,,\nonumber\\
\nabla_X(J)Y&=&-\nabla_Y(J)X\,,\quad\forall\, X,\,Y\in TM\,.\nonumber\end{aligned}$$ This gives $$\label{nk4}
\langle\nabla_X(J)Y,Z\rangle=-\langle\nabla_Y(J)X,Z\rangle=-\langle\nabla_X(J)Z,Y\rangle$$ and $$\label{domega}
d\omega=3\nabla\omega\,.$$ Let $$R(W,X,Y,Z)=\langle\nabla_{[W,X]}Y,Z\rangle-\langle[\nabla_W,\nabla_X]Y,Z\rangle$$ denote the value of the curvature tensor of $M$ on vector fields $W,X,Y,Z\in\mathscr{X}(M)$. The following is well known (see [@Gray-1970],[@Gray-1976]): $$\begin{aligned}
\nabla_{JX}(J)Y&=&\nabla_{X}(J)JY=-J\nabla_X(J)Y\,.\label{nk5}\end{aligned}$$ Let $\{e_1,\dots,e_{2n}\}$ be a local orthonormal frame field. Then the Ricci and Ricci$\phantom{}^*$ curvatures are $$\operatorname{Ric}(X,Y)=\sum_{i=1}^{2n}R(X,e_i,Y,e_i)\,,$$ $$\operatorname{Ric}^*(X,Y)=\frac{1}{2}\sum_{i=1}^{2n}R(X,JY,e_i,Je_i)\,.$$ We define the endomorphism $$\label{op r}
r\in\operatorname{End}(TM)\,,\quad\langle rX,Y\rangle:=\operatorname{Ric}(X,Y)-\operatorname{Ric}^*(X,Y)\,.$$ Then $r$ satisfies the following identities (see [@Koto-1960]): $$\begin{aligned}
[r,J]&=&0\,,\label{koto 1}\\
\langle rX,Y\rangle&=&\sum_{i=1}^{2n}\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle
=\sum_{i=1}^{2n}\langle\nabla_{e_i}(J)X,\nabla_{e_i}(J)Y\rangle\,,\label{koto 2}\\
r&=&-\sum_{i=1}^{2n}\nabla^2_{e_ie_i}(J)J\,.\label{koto 3}\end{aligned}$$ On a nearly Kähler manifold we can define another connection $\nb$ by $$\nb_XY:=\nabla_XY-\frac{1}{2}J\nabla_X(J)Y\,.$$ As has been shown in [@Friedrich-Ivanov], $\nb$ is the uniquely defined connection on $M$ with $$\begin{aligned}
\nb g&=&0\,,\label{nb g}\\
\nb J&=&0\,\label{nb J}\end{aligned}$$ such that $$\label{tau}
\tau(X,Y,Z):=g(T(X,Y),Z)$$ is completely skew symmetric, where $T(X,Y)$ denotes the torsion of $\nb$. This can be compared to the Bismut connection in complex geometry. In our case the torsion $T(X,Y)$ now becomes $$T(X,Y)=-J\nabla_X(J)Y\,.$$ By the nearly Kähler condition, the torsion satisfies $$\label{tau2}
T(JX,Y)=-JT(X,Y)=T(X,JY)\,.$$
It has been shown in [@Belgun-Moroianu] that $\nabla(J)$ and the torsion $T$ are $\nb$-parallel. Nagy proved the following theorem in [@Nagy-2002]:
[**Theorem:** ]{}(Nagy) [*Any complete, strict nearly Kähler manifold $M$ has positive Ricci curvature and hence is compact with finite fundamental group.*]{}
The operator $r$ is also parallel w.r.t. $\nb$, i.e. $\nb r=0$ (cf. [@Nagy-2002]). In fact in this reference it is further proven that even the Ricci and Ricci$\phantom{}^*$ curvatures are $\nb$-parallel. Since $r$ is parallel w.r.t. a metric connection $\nb$, the spectrum of $r$ does not depend on $p\in M$. Moreover, $r$ is selfadjoint and positive semidefinite. Note, that $(M,J,g)$ is strict nearly Kähler, iff $r$ is positive definite, which is equivalent to $\operatorname{ker}(r)=\{0\}$. In general, since $r$ commutes with $J$, all eigenspaces $E_p(\lambda)\subset T_pM$ of $r$ are complex subspaces of $T_pM$ and in particular they are even dimensional.
Lagrangian submanifolds {#sec 2}
=======================
[F]{}or the rest of this section let us assume that $L\subset M$ is a Lagrangian submanifold of a nearly Kähler manifold $(M^{2n},J,g)$. Then since $n=\dim(L)=\frac{1}{2}\dim(M)$ we have $$\langle JX,Y\rangle=0\,,\quad\forall\,X,\,Y\in TL\quad\Leftrightarrow\quad
J:TL\to T^\perp L\quad\text{is an isomorphism}.$$ [F]{}rom $\omega_{|TL}=0$ we deduce $d\omega_{|TL}=0$. On the other hand (\[domega\]) implies $$d\omega(X,Y,Z)=3\langle\nabla_X(J)Y,Z\rangle\,.$$ [F]{}rom this and (\[nk4\]) the following Lemma easily follows (see also [@Hou]).
\[lemma 1\] Suppose $L\subset M$ is a Lagrangian submanifold in a nearly Kähler manifold $(M,J,g)$. Then $$\begin{aligned}
&&\nabla_X(J)Y\in T^\perp L\,,\quad\forall\,X,\,Y\in TL\,,\label{lag 1}\\
&&\nabla_X(J)Y\in T^\perp L\,,\quad\forall\,X,\,Y\in T^\perp L\,,\label{lag 2}\\
&&\nabla_X(J)Y\in TL\,,\text{ if }\,X\in TL, Y\in T^\perp L\,\text{ or if }\,X\in T^\perp L, Y\in TL\,.\label{lag 3}\end{aligned}$$
Denote by $II$ the second fundamental form of the Lagrangian immersion $ L \subset M^{2n}$ into a nearly Kähler manifold $M.$
\[2nd\_fund\_info\] [F]{}or a Lagrangian submanifold $ L \subset M^{2n}$ in a nearly Kähler manifold we have the following information.
- The second fundamental form is given by $\langle II(X,Y),U \rangle = \langle
\nb_XY,U \rangle$ for $X,Y \in TL$ and $U \in
T^\perp L.$
- The tensor $C(X,Y,Z):=\langle II(X,Y),JZ \rangle =\omega(II(X,Y),Z)$, $\forall\,X,Y,Z \in TL$ is totally symmetric.
[F]{}rom Lemma \[lemma 1\] we compute for $X,Y \in TL$ and $U \in
T^\perp L$ the second fundamental form $II$ $$\langle II(X,Y),U \rangle = \langle \nabla_XY,U \rangle = \langle \nb_X Y -\frac{1}{2} J\nabla_X (J)Y,U \rangle = \langle \nb_XY,U \rangle\,.$$ This yields part (i). Next we prove (ii): First we observe for $X,Y,Z \in TL$ $$\begin{aligned}
C(X,Y,Z)&=&\langle II(X,Y),JZ \rangle = \langle \nb_XY,JZ \rangle = - \langle Y,\nb_X(JZ) \rangle\\ &=& -\langle Y,J\nb_X
Z \rangle=\langle \nb_XZ,JY \rangle =C(X,Z,Y).\end{aligned}$$ Since the second fundamental form is symmetric, it follows that $C$ is totally symmetric.
Next we generalize an identity of [@Ejiri-1981] to nearly Kähler manifolds of arbitrary dimension. This and the next lemma will be crucial to prove that Lagrangian submanifolds in strict nearly Kähler six-manifolds and in twistor spaces $Z^{4n+2}$ over quaternionic Kähler manifolds with their canonical nearly Kähler structure are minimal.
\[Lemma cyclic id\] The second fundamental form $II$ of a Lagrangian immersion $L \subset M^{2n}$ into a nearly Kähler manifold and the tensor $\nabla(J)$ satisfy the following identity $$\quad \quad \langle II( X,J\nabla_Y(J)Z),U \rangle = \langle J\nabla_{
II(X,Y)}(J)Z,U \rangle + \langle J\nabla_Y(J) II(X,Z),U \rangle$$ with $X,Y \in TL$ and $U \in T^\perp L.$
The proof of this identity uses $\nb (J)=0,$ $\nb (\nabla
J)=0$ and Lemma \[lemma 1\]. With $X,Y,Z \in TL$ and $U \in T^\perp L$ we obtain $$\begin{aligned}
\langle II(X,J\nabla_Y(J)Z),U \rangle &=& \langle \nb_X(J\nabla_Y(J)Z),U
\rangle \overset{\nb (J)=0}=
\langle J\nb_X(\nabla_Y(J)Z),U \rangle\\
&\overset{\nb (\nabla (J))=0}=&\langle J \left[ \nabla_{\nb_XY}(J)Z+\nabla_Y(J)\nb_X Z \right],U \rangle \\
&=&\langle J[ -\nabla_Z(J)\nb_X Y + \nabla_Y(J) \nb_X Z ],U \rangle\\
&=& -\langle {\nb _XY}, {\nabla_Z(J) J
U}\rangle + \langle {\nb_XZ},{\nabla_Y(J)JU}\rangle\\
&=& -\langle {II(X,Y)}, {\nabla_Z(J) J
U}\rangle + \langle {II(X,Z)},{\nabla_Y(J)JU}\rangle\\
&=& -\langle J\nabla_Z(J) II(X,Y),U\rangle+ \langle J\nabla_Y(J) II(X,Z),U\rangle \\
&=& \langle J\nabla_{ II(X,Y)}(J)Z,U\rangle + \langle J\nabla_Y(J) II(X,Z),U\rangle.\end{aligned}$$ This is exactly the claim of the Lemma.
Given the tensor $C(X,Y,T(Z,V))$ we define the following traces $$\begin{aligned}
\alpha(X,Y)&:=&\sum_{i=1}^nC(e_i,X,T(e_i,Y))\,,\nonumber\\
\beta(X,Y,Z)&:=&\sum_{i=1}^nC(T(e_i,X),Y,T(e_i,Z))\,,\nonumber\end{aligned}$$ where $\{e_1,\dots,e_n\}$ is a local orthonormal frame of $TL$.
\[lemma cyc\] [F]{}or a Lagrangian immersion in a nearly Kähler manifold and any $X,Y,Z,V\in TL$ holds: $$\begin{aligned}
&&C(X,Y,T(Z,V))+C(X,Z,T(V,Y))+C(X,V,T(Y,Z))=0\,,\label{cyc 2}\\
&&\alpha(X,Y)-\alpha(Y,X)=\langle\overrightarrow H,JT(X,Y)\rangle\,,\label{cyc mean}\\
&&\beta(X,Y,Z)=\beta(Z,Y,X)=\beta(Y,X,Z)+\alpha(T(Y,X),Z)\,,\label{cyc beta}\\
&&\alpha(T(X,Y),Z)+\alpha(T(Y,Z),X)+\alpha(T(Z,X),Y)=0\,.\label{cyc beta2}\end{aligned}$$ Here $\overrightarrow H$ denotes the mean curvature vector of $L$.
Let us first rewrite the identity in Lemma \[Lemma cyclic id\] in terms of the tensor $C$ and the torsion $T(X,Y)=-J\nabla_X(J)Y$. Let $X,Y,Z,V\in TL$ be arbitrary. Then Lemma \[Lemma cyclic id\] gives $$\begin{aligned}
C(X,T(Z,Y),V)&=&\langle II( X,J\nabla_Y(J)Z),JV \rangle\nonumber\\
&=& \langle J\nabla_{II(X,Y)}(J)Z,JV \rangle + \langle J\nabla_Y(J) II(X,Z),JV \rangle\nonumber\\
&=&\langle J\nabla_Z(J)(JV),II(X,Y)\rangle-\langle J\nabla_Y(J)(JV),II(X,Z)\rangle\nonumber\\
&=&-\langle J^2\nabla_Z(J)V,II(X,Y)\rangle+\langle J^2\nabla_Y(J)V,II(X,Z)\rangle\nonumber\\
&=&\langle JT(Z,V),II(X,Y)\rangle-\langle JT(Y,V),II(X,Z)\rangle\nonumber\\
&=&C(X,Y,T(Z,V))-C(X,Z,T(Y,V))\,.\nonumber\end{aligned}$$ This is (\[cyc 2\]). Taking a trace gives $$\begin{aligned}
\alpha(X,Y)&=&\sum_{i=1}^nC(e_i,X,T(e_i,Y))\nonumber\\
&\overset{(\ref{cyc 2})}{=}&\sum_{i=1}^nC(e_i,e_i,T(X,Y))+\sum_{i=1}^nC(e_i,Y,T(e_i,X))\nonumber\\
&=&\langle\overrightarrow H,JT(X,Y)\rangle+\alpha(Y,X)\,,\nonumber\end{aligned}$$ which is (\[cyc mean\]). The first identity in (\[cyc beta\]) is clear since $C$ is fully symmetric. If we apply (\[cyc 2\]) to $\beta(X,Y,Z)$, then we get $$\begin{aligned}
\beta(X,Y,Z)
&=&\sum_{i=1}^nC(T(e_i,X),Y,T(e_i,Z))\nonumber\\
&=&-\sum_{i=1}^nC(T(X,Y),e_i,T(e_i,Z))-\sum_{i=1}^nC(T(Y,e_i),X,T(e_i,Z))\nonumber\\
&=&\sum_{i=1}^nC(e_i,T(Y,X),T(e_i,Z))+\sum_{i=1}^nC(T(e_i,Y),X,T(e_i,Z))\nonumber\\
&=&\alpha(T(Y,X),Z)+\beta(Y,X,Z)\,.\nonumber\end{aligned}$$ This is the second identity in (\[cyc beta\]). In view of this we also get $$\begin{aligned}
&&\alpha(T(X,Y),Z)+\alpha(T(Y,Z),X)+\alpha(T(Z,X),Y)\nonumber\\
&=&\beta(Y,X,Z)-\beta(X,Y,Z)+\beta(Z,Y,X)\nonumber\\
&&-\beta(Y,Z,X)+\beta(X,Z,Y)-\beta(Z,X,Y)\nonumber\\
&=&0\nonumber\end{aligned}$$ and this is (\[cyc beta2\]).
Lagrangian submanifolds in nearly Kähler six-manifolds {#sec 3}
======================================================
By a well known theorem of Ejiri [@Ejiri-1981] Lagrangian submanifolds of $S^6$ are minimal. In this section we will see that this is a special case of a much more general theorem. As a consequence of Lemma \[lemma cyc\] we will observe that Lagrangian submanifolds in strict nearly Kähler manifolds of dimension six are minimal. As the next theorem shows, the situation in three dimensions is quite special.
\[Minimal\_th\_six\] Let $L^3$ be a Lagrangian immersion in a strict nearly Kähler six-manifold $M^6$. Then we have $$\begin{aligned}
\alpha&=&0\label{3 alpha}\,,\\
\overrightarrow H&=&0\,.\label{3 mean}\end{aligned}$$ In particular, any Lagrangian immersion in a strict nearly Kähler six-manifold is orientable and minimal.
Let $\{e_1,e_2,e_3\}$ be an orthonormal basis of $T_pL$ for a fixed point $p\in L$. [F]{}rom the skew-symmetry of $\langle T(X,Y),Z\rangle$ we see that there exists a (nonzero) constant $a$ such that $T(e_1,e_2)=ae_3$. Then we also have $T(e_2,e_3)=ae_1,\, T(e_3,e_1)=ae_2$. The symmetry of $C$ implies $$\begin{aligned}
\alpha(e_1,e_1)
&=&C(e_1,e_1,T(e_1,e_1))+C(e_2,e_1,T(e_2,e_1))+C(e_3,e_1,T(e_3,e_1))\nonumber\\
&=&0-aC(e_2,e_1,e_3)+aC(e_3,e_1,e_2)=0\nonumber\end{aligned}$$ and $$\begin{aligned}
\alpha(e_1,e_2)
&=&C(e_1,e_1,T(e_1,e_2))+C(e_2,e_1,T(e_2,e_2))+C(e_3,e_1,T(e_3,e_2))\nonumber\\
&=&aC(e_1,e_1,e_3)+0-aC(e_3,e_1,e_1)=0\,.\nonumber\end{aligned}$$ Similarly we prove that $\alpha(e_i,e_j)=0$ for all $i,j=1,\dots,3$. This shows $\alpha=0$. But then (\[cyc mean\]) also implies $\overrightarrow H=0$. The observation, that the frame $\{ e_1,e_2,e_3\}$ defines an orientation on $L,$ finishes the proof.
\[bem type\] The constant $a$ in the formula $T(e_1,e_2)=ae_3$ from above is related to the type constant $\alpha$ of the nearly Kähler manifold $M$ by the formula $$a^2=\alpha\,.$$ This follows directly from the defining equation for the type constant which is $$\langle\nabla_X(J)Y,\nabla_X(J)Y\rangle
=\alpha\left(|X|^2\cdot|Y|^2-\langle X,Y\rangle^2-\langle X,JY\rangle^2\right)\,.$$ In addition, recall that the type constant is related to the Ricci curvature (of the Levi-Civita connection) of $M$ by $$\operatorname{Ric}=5\alpha g\,,$$ so that the scalar curvature $s$ is $s=30\alpha$. In particular, strict six-dimensional nearly Kähler manifolds are Einstein with $s>0$.
The splitting theorem {#sec 5}
=====================
The following example shows that theorem \[Minimal\_th\_six\] does not extend to eight dimensions:
Let $L'\subset M^6_{SNK}$ be a (minimal) Lagrangian submanifold in a strict nearly Kähler manifold $M^6_{SNK}$ and suppose $\gamma\subset\Sigma$ is a curve on a Riemann surface $\Sigma$. Then the Lagrangian submanifold $L:=L'\times\gamma\subset M$ in the nearly Kähler manifold $M:=\Sigma\times M_{SNK} $ is minimal, if and only if $\gamma$ is a geodesic in $\Sigma$.
In this section we will see that this is basically the only counterexample to theorem \[Minimal\_th\_six\] that occurs in dimension eight.
Nearly Kähler manifolds $(M,J,g)$ split locally into a Kähler factor and a strict nearly Kähler factor and under the assumption, that $M$ is complete and simply connected this splitting is global [@Nagy-2002]. The natural question answered in the following theorem is in which way Lagrangian submanifolds lie in this decomposition.
\[split\_th\] Let $(M,J,g)$ be a nearly Kähler manifold and $L\subset M$ a Lagrangian submanifold. Then $M$ and $L$ split (locally) into $M=M_K\times M_{SNK}$, $L=L_K\times L_{SNK}$, where $M_K$ is Kähler, $M_{SNK}$ is strict nearly Kähler and $L_K\subset M_K$, $L_{SNK}\subset M_{SNK}$ are both Lagrangian. Moreover, the dimension of $L_K$ is given by $$\operatorname{dim}(L_K)=\frac{1}{2}\operatorname{dim}(\operatorname{ker}(r))\,.$$ If the splitting of $M$ is global and $L$ is simply connected, then the splitting of $L$ is global as well.
i) \[p1\] We define $$K_p:=\{X\in T_pM:rX=0\}\,,\quad K_p^\perp:=\{Y\in T_pM:\langle X,Y\rangle=0\,,\forall\, X\in K_p\}\,.$$ Because of $\nb r=0, \nb g=0$ this defines two orthogonal smooth distributions $$\mathscr{D}_K:=\bigcup_{p\in M}K_p\,,\quad \mathscr{D}_{SNK}:=\bigcup_{p\in M}K_p^\perp$$ on $M$.
ii) \[p2\] Since $$\langle rX,Y\rangle=\sum_{i=1}^{2n}\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle
=\langle\nabla_X(J),\nabla_Y(J)\rangle_{\operatorname{End}(TM)}$$ $r$ is a selfadjoint and positive semidefinite operator. Then $\langle rX,X\rangle=||\nabla_X(J)||^2$ implies $$X\in \operatorname{ker}(r)\quad\Leftrightarrow\quad \nabla_X(J)=0\,.$$
iii) \[p3\] Let $K\in\Gamma(\mathscr{D}_K)$, $X\in\mathscr{X}(M)$ be smooth vector fields. Then $\nb r=0$ and $r(K)=0$ imply $$X(r(K))=0=\nb_X(r)K+r(\nb_XK)=r(\nb_XK)\,.$$ Hence $\nb_XK\in\Gamma(\mathscr{D}_K)$ as well.
iv) Suppose $K_1,K_2\in\Gamma(\mathscr{D}_K)$. Then $$\begin{aligned}
[K_1,K_2]
&=&\nabla_{K_1}K_2-\nabla_{K_2}K_1\nonumber\\
&=&\nb_{K_1}K_2+\frac{1}{2}J\nabla_{K_1}(J)K_2-\nb_{K_2}K_1-\frac{1}{2}J\nabla_{K_2}(J)K_1\nonumber\\
&\overset{(\ref{p2})}{=}&\nb_{K_1}K_2-\nb_{K_2}K_1\overset{(\ref{p3})}{\in}\Gamma(\mathscr{D}_K)\,.\nonumber\end{aligned}$$ Hence $\mathscr{D}_K$ is an integrable distribution.
v) The orthogonal distribution $\mathscr{D}_{SNK}$ is also integrable. For $X\in\mathscr{X}(M)$, $S\in\Gamma(\mathscr{D}_{SNK})$, $K\in\Gamma(\mathscr{D}_K)$ we compute $$\begin{aligned}
X\langle K,S\rangle&=&0\nonumber\\
&=&\langle\nabla_XK,S\rangle+\langle K,\nabla_XS\rangle\nonumber\\
&=&\langle\nb_XK+\frac{1}{2}J\nabla_X(J)K,S\rangle+\langle K,\nabla_XS\rangle\nonumber\\
&\overset{(\ref{p3})}{=}&\frac{1}{2}\langle J\nabla_X(J)K,S\rangle+\langle K,\nabla_XS\rangle\nonumber\\
&\overset{(\ref{nk4})}{=}&-\frac{1}{2}\langle J\nabla_K(J)X,S\rangle+\langle K,\nabla_XS\rangle\nonumber\\
&\overset{(\ref{p2})}{=}&\langle K,\nabla_XS\rangle\,,\nonumber\end{aligned}$$ so that $\nabla_XS\in\Gamma(\mathscr{D}_{SNK})$. Then again $$[S_1,S_2]=\nabla_{S_1}S_2-\nabla_{S_2}S_1\in\Gamma(\mathscr{D}_{SNK})\,.$$
vi) The splitting theorem of de Rham can now be applied to the distributions $\mathscr{D}_K$ and $\mathscr{D}_{SNK}$ and the nearly Kähler manifold $(M,J,g)$ splits (locally) into a Riemannian product $$(M,J,g)=(M_K,J_K,g_K)\times (M_{SNK},J_{SNK},g_{SNK})\,,$$ where $TM_K=\mathscr{D}_K$, $TM_{SNK}=\mathscr{D}_{SNK}$ and where $(M_K,J_K,g_K)$ is Kähler and $(M_{SNK},J_{SNK},g_{SNK})$ is strict nearly Kähler.
vii) Now let $L\subset M=M_K\times M_{SNK}$ be Lagrangian. We prove that $r$ leaves tangent and normal spaces of $L$ invariant. To see this, we fix an adapted local orthonormal frame field $\{e_1,\dots,e_{2n}\}$ of $M$ such that $e_1,\dots,e_n$ are tangent to $L$ and $e_{n+1}=Je_1,\dots,e_{2n}=Je_n$ are normal to $L$. Since for any three vectors $X,Y,Z$ we have $$\langle\nabla_X(J)JZ,\nabla_Y(J)JZ\rangle=\langle J\nabla_X(J)Z,J\nabla_Y(J)Z\rangle
=\langle\nabla_X(J)Z,\nabla_Y(J)Z\rangle\,,$$ we obtain $$\begin{aligned}
\langle rX,Y\rangle
&=&\sum_{i=1}^{2n}\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle\nonumber\\
&=&\sum_{i=1}^{n}\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle
+\sum_{i=1}^{n}\langle\nabla_X(J)Je_i,\nabla_Y(J)Je_i\rangle\nonumber\\
&=&2\sum_{i=1}^{n}\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle.\nonumber\end{aligned}$$ Now, if $X\in TL$, $Y\in T^\perp L$, then by Lemma \[lemma 1\] we have $$\nabla_X(J)e_i\in T^\perp L\,,\quad\nabla_Y(J)e_i\in TL\,,$$ so that $$\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle=0\,,\quad\forall\, i=1,\dots,n\,.$$ [F]{}urther it follows $$\langle rX,Y\rangle=2\sum_{i=1}^n\langle\nabla_X(J)e_i,\nabla_Y(J)e_i\rangle=0\,.$$ Since this works for any $X\in TL, Y\in T^\perp L$ and since $r$ is selfadjoint we conclude $$r(TL)\subset TL\,,\quad r(T^\perp L)\subset T^\perp L\,.$$ At a given point $p\in L$ we may now choose an orthonormal basis $\{f_1,\dots,f_n\}$ of $T_pL$ that consists of eigenvectors of $r_{|TL}$ considered as an endomorphism of $TL$. Since $[r,J]=0$ and $L$ is Lagrangian, the set $\{f_1,\dots,f_n,Jf_1,\dots,Jf_n\}$ then also determines an orthonormal eigenbasis of $r\in\operatorname{End}(TM)$. In particular, since $J$ leaves the eigenspaces invariant, $K_p=\operatorname{ker}(r(p))$ and $T_pL$ intersect in a subspace $K^L_p$ of dimension $\frac{1}{2}\dim(K_p)=\frac{1}{2}\dim(M_K)$. [F]{}or the same reason $K_p^\perp\cap T_pL$ gives an $\frac{1}{2}\dim(M_{SNK})$-dimensional subspace. The corresponding distributions, denoted by $\mathscr{D}_K^L$ and $\mathscr{D}_{SNK}^L$ are orthogonal and both integrable, since in view of $$\mathscr{D}_K^L=\mathscr{D}_K\cap TL\,,\quad\mathscr{D}_{SNK}^L=\mathscr{D}_{SNK}\cap TL$$ they are given by intersections of integrable distributions. We may now apply again the splitting theorem of de Rham to the Lagrangian submanifold. This completes the proof.
A more detailed analysis of the proof of the last theorem shows
\[cor r\] If $L\subset M$ is Lagrangian and $p\in L$ a fixed point, then to each eigenvalue $\lambda$ of the operator $r$ at $p$ there exists a basis $e_1,\dots,e_k,f_1,\dots,f_k$ of eigenvectors of $\operatorname{Eig}(\lambda)$ such that $e_1,\dots,e_k\in T_pL$, $f_1,\dots,f_k\in T_p^\perp L$. Here, $2k$ denotes the multiplicity of $\lambda$.
Let us recall the situation in dimension eight and ten [@Gray-1976; @Nagy-2002].
\[NK\_dim8\_10\]
(i) Let $M^8$ be a simply connected complete nearly Kähler manifold of dimension eight. Then $M^8$ is a Riemannian product $M^8= \Sigma \times M_{SNK}^6$ of a Riemannian surface $\Sigma$ and a six-dimensional strict nearly Kähler manifold $M_{SNK}^6.$
(ii) Let $M^{10 }$ be a simply connected complete nearly Kähler manifold of dimension ten. Then $M^{10}$ is either the product $M_K^4 \times M_{SNK}^6$ of a Kähler surface $M_K^4$ and a six-dimensional strict nearly Kähler manifold $M_{SNK}^6$ or $M$ is a twistor space over a positive, eight dimensional quaternionic Kähler manifold.
Note, that any complete, simply connected eight dimensional quaternionic manifold equals one of the following three spaces: $\mathbb{HP}^2, \mathbb{G}r_2(\complex{2}), G_2/SO(4)$
In the next theorem, part (i) and (ii) collect the information on Lagrangian submanifolds in nearly Kähler manifolds of dimension eight and ten.
\[theoC\]
(i) Let $L$ be a Lagrangian submanifold in a simply connected nearly Kähler manifold $M^8$. Then $M^8=\Sigma \times M^6_{SNK}$, where $\Sigma$ is a Riemann surface, $M^6_{SNK}$ is strict nearly Kähler and $L=\gamma\times L'$ is a product of a (real) curve $\gamma\subset\Sigma$ and a minimal Lagrangian submanifold $L' \subset M^6_{SNK}.$
(ii) \[theoC2\] Let $L$ be a Lagrangian submanifold in a simply connected complete nearly Kähler manifold $M^{10}$, then either
(a) $M^{10}=M_K^4 \times M^6_{SNK}$ and the manifold $L= S \times L'$ is a product of a Lagrangian (real) surface $S\subset
M_K^4$ and a minimal Lagrangian submanifold $L' \subset M^6_{SNK}$ or
(b) \[theoC2b\] the manifold $L$ is a Lagrangian submanifold in a twistor space over a positive, eight dimensional quaternionic Kähler manifold.
(iii) Let $M_1$, $M_2$ be two nearly Kähler manifolds. Denote the operator $r$ on $M_i$ by $r_i$, $i=1,2$. If $\operatorname{Spec}(r_1)\cap\operatorname{Spec}(r_2)=\emptyset$ and $L\subset M_1\times M_2$ is Lagrangian, then $L$ splits (locally) into $L=L_1\times L_2$, where $L_i\subset M_i$, $i=1,2$ are Lagrangian. If $L$ is simply connected, then the decomposition is global.
This is a combination of the results of Theorem \[split\_th\], Corollary \[cor r\] and Proposition \[NK\_dim8\_10\].
Theorem \[theoC\] (\[theoC2\]), part (b) motivates the discussion of Lagrangian submanifolds in twistor spaces in the subsequent section. Indeed, the results derived in the next section imply that Lagrangian submanifolds in twistor spaces are, regardless their dimension, always minimal.
Lagrangian submanifolds in twistor spaces {#twistor_sec}
=========================================
As mentioned in the introduction an important class of examples for nearly Kähler manifolds is given by twistor spaces $Z^{4n+2}$ over positive quaternionic Kähler manifolds $N^{4n}.$ We recall that the twistor space is the bundle of almost complex structures over $N.$ It can be endowed with a Kähler structure $(Z,J^Z,g^Z),$ such that the projection $\pi \,:\, Z \rightarrow N$ is a Riemannian submersion with totally geodesic fibers $S^2.$ Denote by $\mathcal H$ and $\mathcal V$ the horizontal and the vertical distributions of the submersion $\pi.$ Then the direct decomposition $$TZ = \mathcal H \oplus \mathcal V \label{TM_dec_tw}$$ is orthogonal and compatible with the complex structure $J^Z.$ Let us consider now a second almost hermitian structure $(J,g)$ on $Z$ which is defined by $$g:=
\begin{cases}
g^Z(X,Y), \mbox{ for } X,Y \in \mathcal{H}, \\
\frac{1}{2} g^Z(V,W), \mbox{ for } V,W \in \mathcal{V},\\
g^Z(V,X)=0, \mbox{ for } V \in \mathcal{V},
X \in \mathcal{H}
\end{cases}$$ and $$J:=
\begin{cases}
J^Z \mbox{ on } \mathcal{H}, \\
-J^Z \mbox{ on } \mathcal{V}.
\end{cases}$$
Note, that in view of (\[TM\_dec\_tw\]), the decomposition $TZ=\mathcal H\oplus\mathcal V$ is also compatible w.r.t. $J$ and orthogonal w.r.t. $g$. Then the following result is well-known.
\[can\_NK\_tw\] The manifold $(Z,J,g)$ is a strict nearly Kähler manifold and the distributions $\mathcal V$ and $\mathcal H$ are parallel w.r.t. the connection $\nb.$ The projection $\pi$ is also a Riemannian submersion with totally geodesic fibers for the metric $g.$
We summarize some information which will be useful in the later text.
\[info\_lemma\] In the situation of Proposition \[can\_NK\_tw\] we have
(a) The torsion $T =-
J\nabla (J)$ of the canonical connection satisfies $$\begin{aligned}
T(X,Y) \in \mathcal V,&& \mbox{ for } X,Y \in \mathcal H, \label{Tor_1_equ}\\
T(X,V) \in \mathcal H,&& \mbox{ for } X \in \mathcal H, V \in \mathcal V,\label{Tor_2_equ}\\
T(U,V) =0, && \mbox{ for } U,V \in \mathcal V. \label{Tor_3_equ}\end{aligned}$$
(b) The association $$\label{ass 1}
\mathcal H \ni X \mapsto T(Y,X)\in\mathcal V$$ is surjective for $0\ne Y \in \mathcal H$ and the map $$\label{ass 2}
\Phi^V\,:\, \mathcal H \ni X \mapsto T(V,X)$$ with $0\ne V \in \mathcal V$ is invertible and squares to $-\kappa^2 Id_{\mathcal H}$ for some $\kappa \in \mathbb{R}.$
(c) The operator $r$ has eigenvalues $\lambda_{\mathcal{H}}=4\kappa^2$, $\lambda_{\mathcal V}=4n\kappa^2$. If $n>1$, then the eigenbundle of $\lambda_{\mathcal H}$ is $\mathcal{H}$ and $\mathcal{V}$ is the eigenbundle of $\lambda_{\mathcal V}$.
Property follows from Propostion 4.1 (ii) of [@Nagy-2002]. The second property follows, since $g(T(\cdot,\cdot),\cdot)$ is a three-form and follows by a direct computation from the nearly Kähler condition. The second part of (b) is a consequence of Proposition 4.1 (i) of [@Nagy-2002]. We claim that for a fixed $Y \in
\mathcal H$ there exists $X \in \mathcal H,$ such that $\mathcal V \ni V=T(Y,X) \ne 0.$ This yields $JV=-T(Y,JX) \ne 0$ and proves the first part of (b). It remains to prove the claim: $T(Y,X)=0$ for all $X \in \mathcal H$ implies $g(T(Y,X),V)=-
g(J\nabla_Y(J)X,V)=g(\nabla_Y(J)X,JV)=-g(\nabla_Y(J)V,JX) =0$ for all $X \in \mathcal H$ and $V \in \mathcal V.$ In other words $\nabla_Y(J) =0$ in contradiction to strict nearly Kähler. Part (c) can be found in Corollary 4.1 (i) of [@Nagy-2002].
In the rest of this section we consider a nearly Kähler manifold $(M=Z,J,g)$ of this type and study Lagrangian submanifolds $L\subset M.$
\[2nd\_fund\_tw\] Let $L^{2n+1}\subset M^{4n+2}$ be a Lagrangian submanifold in a twistor space as in Proposition \[can\_NK\_tw\]. Then the second fundamental form $II$ satisfies $$\begin{aligned}
&II(X,Y) \in \pi^{\mathcal H}(T^\perp L),& \mbox{ for } X,Y \in \pi^{\mathcal H}(TL), \\
&II(X,Y) \in \pi^{\mathcal V}(T^\perp L),& \mbox{ for } X,Y \in \pi^{\mathcal V}(TL),\\
&II(X,Y) =0,& \mbox{ for } X \in \pi^{\mathcal H}(TL), Y \in \pi^{\mathcal V}(TL).\end{aligned}$$ Here $\pi^{\mathcal H}(TL)$ and $\pi^{\mathcal V}(TL)$ are the orthogonal projections of $TL$ to $\mathcal H$ resp. $\mathcal V$.
The second fundamental form is given by $C(X,Y,Z)=\langle \nb_X Y, JZ \rangle$ for $X,Y,Z \in TL.$ The lemma follows since the decomposition is $\nb$-parallel, orthogonal and $J$-invariant and as the tensor $C$ is completely symmetric.
As will be shown in the next theorem, for $n>1$ we have $$\pi^{\mathcal H}(TL)=\mathcal H\cap TL\,,\quad\pi^{\mathcal V}(TL)=\mathcal V\cap TL\,.$$
\[min\_\_Lgra\_tw\] Let $L^{2n+1}\subset M^{4n+2}$ be a Lagrangian submanifold in a twistor space as in Proposition \[can\_NK\_tw\]. Then $L$ is minimal. If $n>1$, then the tangent space of $L$ splits into a one-dimensional vertical part and a $2n$-dimensional horizontal part. Moreover, the second fundamental form $II$ of the vertical normal direction vanishes completely if $n>1$.
i) By theorem \[Minimal\_th\_six\] it suffices to consider the case $n>1$.
ii) \[Lemma\_distr\_D\] Let $L\subset M$ be a Lagrangian submanifold. Since $n>1$, the two eigenvalues $\lambda_{\mathcal H},\,\lambda_{\mathcal V}$ of $r$ are distinct and the eigenspace $\mathcal{V}$ of $\lambda_{\mathcal V}$ is two-dimensional. By Corollary \[cor r\] this induces a one-dimensional vertical tangential distribution $\mathcal D$ on $L$. Then, by the Lagrangian condition, $\mathcal D := \pi^{\mathcal V}(TL) $.
iii) Denote by $\mathcal D^\perp$ the orthogonal complement of $\mathcal D$ in $TL.$ The trace of the second fundamental form $II$ of $L$ restricted to $\mathcal D^\perp$ is zero.
[F]{}irst we observe that by Lemma \[2nd\_fund\_tw\] we can restrict the second fundamental form $II$ to $\mathcal D^\perp =\pi^{\mathcal H}(TL).$ We fix $U \in \mathcal D$ of unit length. Using Lemma \[lemma 1\] and Lemma \[info\_lemma\] we observe, that $\Phi(X):= \frac{1}{\kappa} J\nabla_U(J)X$ defines an (almost) complex structure on $\mathcal D^\perp$ which is compatible with the metric. [F]{}irst we observe with $X \in \mathcal D^\perp.$ $$\begin{aligned}
II(X,X)&=&-II(X,\Phi(\Phi(X)))=-\frac{1}{\kappa}II(X,J\nabla_U(J)\,\Phi(X))\\
&=&-\frac{1}{\kappa}J\left[\nabla_{II(X,U)}(J)\Phi(X)
+\nabla_U(J)II(X,\Phi(X)) \right]\\
&=&-\frac{1}{\kappa}J\nabla_U(J)\, II(X,\Phi(X))=-\Phi II(X,\Phi(X))\,. \end{aligned}$$ After polarizing we obtain $$\label{pluriminimal}
II(\Phi X,\Phi Y)=-II(X,Y)\,,\quad\forall\, X,\,Y\in\mathcal D^\perp\,.$$ In particular, taking a trace over (\[pluriminimal\]) we get $$\begin{aligned}
\operatorname{tr}^{\mathcal D^\perp}II &=& 0\,.\end{aligned}$$
iv) We have $$\alpha(X,Y)=0\,,\quad\forall\, X,\,Y\in\mathcal D^\perp\,.$$
By (ii) we may choose an orthonormal frame $\{e_1,\dots,e_{2n+1}\}$ of $TL$ such that $e_1,\dots,e_{2n}\in\mathcal D^\perp$ and $e_{2n+1}\in\mathcal D$. Since $$\alpha(X,Y)=\sum_{i=1}^{2n+1}C(e_i,X,T(e_i,Y))$$ and the tensor $C$ is fully symmetric we see that by Lemma \[2nd\_fund\_tw\] all terms on the RHS vanish since either $e_i\in \mathcal D=\pi^{\mathcal V}(TL)$, $X\in\mathcal D^\perp=\pi^{\mathcal H}(TL)$ or $e_i, X\in\mathcal D^\perp$ and $T(e_i,Y)\in\mathcal D$ (cf. Lemma \[info\_lemma\]).
v) By (iii) the mean curvature vector $\overrightarrow H$ satisfies $$\overrightarrow H\in\mathcal D\,.$$ [F]{}rom (\[cyc mean\]) and (iv) we get $$\label{mean 1}
\langle J\overrightarrow H, T(X,Y)\rangle=0\,,\quad\forall\,X,\,Y\in\mathcal D^\perp\,.$$ Since $J$ maps $\mathcal V$ to itself, we also have $J\overrightarrow H\in\mathcal D\subset\mathcal V$. Now we choose $X\in \mathcal{D}^\perp$ and $\tilde Y\in\mathcal H$ with $$T(X,\tilde Y)=J\overrightarrow H\,.$$ This is possible since the map $\mathcal H\ni \tilde Y\mapsto T(X,\tilde Y)\in\mathcal V$ is surjective by (\[ass 1\]). Let $\tilde Y=Y+Y^\perp$ be the orthogonal decomposition of $\tilde Y$ into the tangent and normal parts of $\tilde Y$. Note that $Y,Y^\perp$ are both horizontal. We have $$T(X,Y^\perp)=0$$ since $T(X,\cdot)$ maps tangent to tangent and normal to normal vectors and $$T(X,\tilde Y)=T(X,Y)+T(X,Y^\perp)=J\overrightarrow H\in TL\,.$$ Therefore there exist two tangent vectors $X,Y\in\mathcal{D}^\perp$ with $$T(X,Y)=J\overrightarrow H\,.$$ This implies $$|\overrightarrow H|^2=\langle J\overrightarrow H,T(X,Y)\rangle\overset{(\ref{mean 1})}{=}0$$ which proves that the mean curvature vector vanishes. [F]{}rom this, the fact that $\mathcal D$ is one-dimensional and from Lemma \[2nd\_fund\_tw\] it follows that $II(V,\cdot)=0$ for any $V\in\mathcal D$.
Let $L\subset M$ be a Lagrangian submanifold in a twistor space $M^{4n+2}$ as above with $n>1$. Then the integral manifolds $c$ of the distribution $\mathcal D$ are geodesics (hence locally great circles) in the totally geodesic fibers $S^2.$
The last theorem implies that the geodesic curvature vanishes and that in consequence an integral manifold $c$ of $\mathcal D$ is totally geodesic in the fibers.
Invariant Lagrangian submanifolds of three-symmetric spaces {#sec 7}
===========================================================
The idea of a three-symmetric space is to replace the symmetry of order two as in the case of a symmetric space by a symmetry of order three. Nearly Kähler geometry on such spaces was first studied in [@Gray-1972; @Gray-Wolf].\
Like symmetric spaces three-symmetric spaces have a homogenous model, which we shortly resume. Let $G$ be a connected Lie group and $s$ an automorphism of order $3$ and let $G^s_0 \subset H \subset G^s$ be a subgroup contained in the fix-point set $G^s$ of $s.$ The differential $s_*$ decomposes $$\mathfrak g \otimes \mathbb C = \mathfrak h \otimes \mathbb C \oplus
\mathfrak m^+ \oplus \mathfrak m^-$$ into the eigenspaces of $s_*$ with eigenvalues $1$ and $\frac{1}{2}(-1 \pm
\sqrt{-3}).$ With the definition $\mathfrak m := (\mathfrak m^+ \oplus
\mathfrak m^-)\cap \mathfrak g$ the decomposition $$\label{3-sym_dec} \mathfrak g=\mathfrak h \oplus \mathfrak m$$ is reductive and $G/H$ is a reductive homogenous space. The canonical complex structure is then defined by $$s_*| \mathfrak m= -\frac{1}{2} Id + \frac{\sqrt{3}}{2} J.$$ The choice of an $Ad(H)$-invariant and $s_*$-invariant (pseudo-)metric $B$ on $\mathfrak m$ makes $G/H$ into a (pseudo-)Riemannian three-symmetric space, such that $B$ is almost hermitian with respect to $J.$\
By Proposition 5.6 of [@Gray-1972] the data $(M=G/H,J,B)$ defines a nearly Kähler manifold if and only if the decomposition is naturally reductive, i.e. $$B([X,Y]_{\mathfrak m},Z)=B(X,[Y,Z]_{\mathfrak m})\mbox{ for } X,Y,Z \in \mathfrak m.$$ Since $M$ is homogenous and naturally reductive the Levi-Civita connection in the canonical base point $p$ (after identifying $\mathfrak m$ and the tangent space) is given by $$\label{LC_3_sym_NK} 2B(\nabla_XY,Z) = B([X,Y],Z) \mbox{
for } X,Y,Z \in \mathfrak m.$$ Moreover it is $B(\nabla_X(J)Y,JZ)= B(X,[Y,Z]) \mbox{
for } X,Y,Z \in \mathfrak m$ which implies $$J \nabla_X(J)Y =
[X,Y]_{\mathfrak m}\mbox{ for } X,Y \in \mathfrak m$$ in $p.$ With this information the canonical connection $\nb$ and the Levi-Civita connection $\nabla$ are related in $p$ by $$\label{nb_3_sym_NK}
\nb_XY - \nabla_XY = -\frac{1}{2} J \nabla_X(J)Y= -\frac{1}{2}[X,Y]_{\mathfrak m}\mbox{
for } X,Y,Z \in \mathfrak m.$$
Let $L$ be an invariant Lagrangian submanifold in $M=G/H$ with the above given nearly Kähler structure. Then $L$ is totally geodesic. In particular $L$ is minimal.
By Proposition \[2nd\_fund\_info\] the second fundamental form is given by $$B(II(X,Y),JZ)= B(\nabla_XY,JZ) = B(\nb_XY,JZ)\,.$$ This shows with equation $$0=B(\nb_XY-\nabla_XY,JZ)= -\frac{1}{2}B([X,Y]_{\mathfrak m},J Z) \mbox{
for } X,Y,Z \in \mathfrak m\cap T_pL.$$ Using equation we obtain $$B(II(X,Y),JZ) = B(\nabla_XY,JZ)= \frac{1}{2} B( [X,Y]_{\mathfrak
m},JZ),$$ which is zero by the previous equation. This shows that $II=0,$ i.e. $L$ is totally geodesic and consequently minimal.
Examples on $S^3\times S^3$
---------------------------
A convenient way to introduce the nearly Kähler structure on $M=S^3\times S^3$ is to identify $S^3$ with $SU(2)$ and $M$ with $SU(2) \times
SU(2).$ Let us denote by $\mathcal E =(e_1,e_2,e_3)$ a frame of left invariant one-forms on the first factor and by $\mathcal F= (f_1,f_2,f_3)$ a frame left invariant one-forms on the second factor. One can choose the frames $\mathcal E$ and $\mathcal F$ (see for example [@Butruille-2005]) in such a way that the fundamental two-form $\omega$ is given by $$\omega = c\sum_{i=1}^3 e_i \wedge f_i \mbox{ with } c\in \mathbb{R}.$$ With this preparation we characterize left invariant Lagrangian submanifolds in the following theorem.
Let $L\subset M=S^3\times S^3$ be an invariant Lagrangian submanifold. Then $L$ is $S^3$ after inclusion in the first or second factor or an integral manifold of the left invariant distributions $\mathcal D= \{ X \oplus DX \,|\,
X \in \mathfrak{su}(2)\} \subset\mathfrak{su}(2) \oplus \mathfrak{su}(2) $ with $D\in \{\text{diag}(1,1,1), \text{diag}(1,1,-1), \text{diag}(1,-1,-1), \text{diag}(-1,-1,-1)\}.$
After the action of an element of $SU(2) \times SU(2)$ we suppose, that the neutral element $e$ lies in $L$ and consider $\mathcal L=T_eL \subset T_e(SU(2) \times SU(2)) = \mathfrak{su}(2) \oplus \mathfrak{su}(2).$ It is not difficult to see that inclusion of $S^3$ in the first or second factor yields a Lagrangian submanifold. Therefore we suppose that $T_eL$ is not one of the factors. Since $\mathcal L$ is Lagrangian we can express it as a graph, i.e. $\mathcal L= X \oplus AX$ for $X \in \mathfrak{su}(2)$ and a linear map $A\,:\, \mathfrak{su}(2) \rightarrow \mathfrak{su}(2).$ The integrability of $\mathcal L$ yields that $[X \oplus AX,Y \oplus AY]= [X,Y] \oplus [AX,AY]$ equals $[X,Y] \oplus A[X,Y].$ This is equivalent to $A[X,Y]=[AX,AY].$ If we identify $\mathfrak{su}(2)$ with $\mathbb{R}^3$ endowed with the cross product, then the non-trivial solutions of this constraint are $A^tA=1.$ The condition $\omega_{| \mathcal L}=0$ is equivalent to $A^t=A.$ This means we end up with a symmetric $3\times 3$ matrix satisfying $A^2=1.$ Using Sylvester’s theorem the solutions are $SO(3)$-orbits of one of diagonal matrices $D\in \{ \text{diag}(1,1,1), \text{diag}(1,1,-1), \text{diag}(1,-1,-1), \text{diag}(-1,-1,-1) \}.$ The action of $SO(3)$ corresponds to the choice of some basis of $ \mathfrak{su}(2)$ w.r.t. to some basis left invariant one-forms.
Deformations of Lagrangian submanifolds in nearly Kähler manifolds {#sec 8}
==================================================================
Our aim in this section is to study the space of deformations of a given Lagrangian (and hence minimal Lagrangian) submanifold $L$ in a strict six-dimensional nearly Kähler manifold $M^6$. In an article by Moroianu, Nagy and Semmelmann [@MNS-2008] the deformation space of nearly Kähler structures on six-dimensional nearly Kähler manifolds has been related to the space of coclosed eigenforms of the Hodge-Laplacian. As we will show below, a similar statement holds for the deformation of Lagrangian submanifolds in strict nearly Kähler six-manifolds.
To this end we assume that $$F:L\times(-\epsilon,\epsilon)\to M$$ is a smooth variation of Lagrangian immersions $F_t:=F(\cdot,t):L\to M$, $t\in(-\epsilon,\epsilon)$ into a nearly Kähler manifold $M$. Let $$V:=\frac{d}{dt} F_t$$ denote the variation vector field. Since tangential deformations correspond to diffeomorphisms acting on $L$, we may assume w.l.o.g. that $V\in\Gamma(T^\perp L)$ is a normal vector field. The Cartan formula and $F^*_t\omega=0$ for all $t$ then implies $$0=d(i_V\omega)+i_Vd\omega$$ holds everywhere on $L$. By the nearly Kähler condition this is equivalent to $$\label{def 1}
d(V\cont\omega)+3V\cont\nabla\omega=0$$ on $L$. Let us define the variation $1$-form $\theta\in\Omega^1(L)$ by $$\theta:=V\cont\omega\,.$$
Let $F_t:L\to M$ be a variation of Lagrangian immersions in a six-dimensional nearly Kähler manifold $M$. Then the variation $1$-form $\theta$ is a coclosed eigenform of the Hodge-Laplacian, where the eigenvalue $\lambda$ satisfies $\lambda=9\alpha$ with the type constant $\alpha$ of $M$.
[F]{}or $X,Y\in TL$ and $V\in T^\perp L$ we compute $$\begin{aligned}
(V\cont\nabla\omega)(X,Y)
&=&\nabla\omega(V,X,Y)\nonumber\\
&=&\langle \nabla_X(J)Y,V\rangle\nonumber\\
&=&\langle J\nabla_X(J)Y,JV\rangle\nonumber\\
&=&-\langle JV,T(X,Y)\rangle\,.\nonumber\end{aligned}$$ Since $T$ induces an orientation on the Lagrangian submanifold by the three-form $$\tau(X,Y,Z):=\langle T(X,Y),Z\rangle\,,$$ we obtain a naturally defined $*$-operator $*:\Omega^p(L)\to\Omega^{3-p}(L)$ which for $1$-forms is given by $$*\phi:=\frac{1}{\sqrt{\alpha}}\,\phi\circ T\,.$$ Here, $\alpha$ is the type constant of $M$ (cf. Remark \[bem type\]). This implies that equation (\[def 1\]) can be rewritten in the form $$\label{def 2}
d\theta=3\sqrt{\alpha}\,{*\theta}\,.$$ Consequently $$\delta\theta=*d{*\theta}=0$$ and $$\begin{aligned}
\delta d\theta
&=&3\sqrt{\alpha}\,{*d{**}}\theta\nonumber\\
&\overset{(*^2=\operatorname{Id})}{=}&3\sqrt{\alpha}\,{*d}\theta\nonumber\\
&\overset{(\ref{def 2})}{=}&9\alpha**\theta\nonumber\\
&=&9\alpha\theta\,.\nonumber\end{aligned}$$ In total $$\Delta_{\operatorname{Hodge}}\theta=(\delta d+d\delta)\theta=9\alpha\theta\,.$$ This proves the theorem. By remark \[bem type\] this is equivalent to $$\Delta_{\operatorname{Hodge}}\theta=\frac{3}{10}s\,\theta\,,$$ where $s$ is the scalar curvature of $M$.
[10000]{}
: [*Nearly Kähler 6-manifolds with reduced holonomy*]{}, [Ann. Global Anal. Geom.]{}, [19/4]{}, [(2001)]{}, [307–319]{}
: [*Second order families of special Lagrangian 3-folds*]{}, [Perspectives in Riemannian geometry, CRM Proc. Lecture Notes]{}, [40]{}, [AMS (2006)]{}, [63–98]{}
: [*Classification des variétés approximativement kähleriennes homogènes*]{}, [Ann. Global Anal. Geom.]{}, [27/3]{}, [(2005)]{},[201–225]{}
: [*Homogeneous nearly Kähler manifolds*]{}, [to appear in Handbook of pseudo-Riemannian geometry and supersymmetry]{}, (2006)
: [*Einstein metrics via intrinsic or parallel torsion*]{}, [Math. Z.]{}, [247/3]{},[(2004)]{}, [513–528]{}
: [*Totally real submanifolds in a $6$-sphere*]{}, [Proc. Amer. Math. Soc.]{}, [83/4]{}, (1981), [759–763]{}
: [*Parallel spinors and connections with skew-symmetric torsion in string theory*]{}, [Asian J. Math.]{}, [6/2]{}, [(2002)]{}, [303–335]{}
: [*Nearly Kähler manifolds*]{}, [J. Differential Geometry]{}, [4]{}, [(1970)]{}, [283–309]{}
: [*Riemannian manifolds with geodesic symmetries of order $3$*]{}, [J. Differential Geometry]{}, [7]{}, [(1972)]{}, [343–369]{}
: [*The structure of nearly Kähler manifolds*]{}, [Math. Ann.]{}, [223/3]{}, [(1976)]{}, [233–248]{}
: [*Homogeneous spaces defined by Lie group automorphisms. I,II*]{}, [J. Differential Geometry]{}, [2]{}, [(1968)]{}, [77–159]{}
: [*On totally real submanifolds in a nearly Kähler manifold*]{}, [Port. Math. (N.S.)]{}, [58/2]{}, [(2001)]{}, [219–231]{}
: [*Some theorems on almost Kählerian spaces*]{}, [J. Math. Soc. Japan]{}, [12]{}, [(1960)]{}, [422–433]{}
: [*Deformations of nearly Kähler structures*]{}, [Pacific J. Math.]{}, [235/1]{}, [(2008)]{}, [57–72]{}
: [*Nearly Kähler geometry and Riemannian foliations*]{}, [Asian J. Math.]{}, [6/3]{}, [(2002)]{}, [481–504]{}
: [*On nearly-Kähler geometry*]{}, [Ann. Global Anal. Geom.]{}, [22/2]{}, [(2002)]{}, [167–178]{}
: [*Special Lagrangian submanifolds of the nearly Kaehler 6-sphere*]{}, [Glasg. Math. J.]{}, [45/3]{}, [(2003)]{}, [415–426]{}
[^1]: Supported by DFG, priority program SPP 1154, SM 78/4-2
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider diffeomorphisms of compact Riemmanian manifolds which have a Gibbs-Markov-Young structure, consisting of a reference set $\Lambda$ with a hyperbolic product structure and a countable Markov partition. We assume polynomial contraction on stable leaves, polynomial backward contraction on unstable leaves, a bounded distortion property and a certain regularity of the stable foliation. We establish a control on the decay of correlations and large deviations of the SRB measure of the dynamical system, based on a polynomial control on the Lebesgue measure of the tail of return times. Finally, we present an example of a dynamical system defined on the torus and prove that it verifies the properties of the Gibbs-Markov-Young structure that we considered.'
address:
- |
José F. Alves\
Departamento de Matemática, Faculdade de Ciências da Universidade do Porto\
Rua do Campo Alegre 687, 4169-007 Porto, Portugal
- |
Davide Azevedo\
Departamento de Matemática, Faculdade de Ciências da Universidade do Porto\
Rua do Campo Alegre 687, 4169-007 Porto, Portugal
author:
- 'José F. Alves'
- Davide Azevedo
title: |
Statistical properties of diffeomorfisms with\
weak invariant manifolds
---
Introduction
============
In this work we consider discrete-time dynamical systems $f:M\to M$, where $f$ is a diffeomorphism of a compact Riemannian manifold $M$. We are interested in the study of the statistical behavior of typical orbits when $f$ has non-uniformly hyperbolic behavior; more specifically, to study the large deviations and decay of correlations with respect to the SRB measure for diffeomorphisms having some [*Gibbs-Markov-Young structure*]{}.
Bowen, Ruelle and Sinai [@s72; @b75; @r76] obtained exponential decay of correlations for uniformly hyperbolic diffeomorphisms with respect to some measures which describe the statistics of a large set of initial states in the phase space, the so-called Sinai-Ruelle-Bowen (SRB) measures. Later, some classes of non-uniformly hyperbolic diffeomorphisms were considered. First, Young [@y98] proved an exponential rate for the decay of correlations assuming there exists a reference set $\Lambda \subseteq M$ with a [*hyperbolic product structure*]{} and, among other properties, exponential contraction along stable leaves and exponential backward contraction on unstable leaves. Alves and Pinheiro in [@ap08] weakened these assumptions, removing the backward contraction but still imposing an exponential contraction along stable leaves. In that paper, they proved exponential or polynomial decay of correlations, depending on different hypothesis that we will explain later. In addition, Young also obtained, in [@y99], a control on the rate of decay of correlations for non-invertible dynamical systems and, together with Benedicks in [@by00], for Hénon maps.
Many authors have proved results on large deviations for uniformly hyperbolic dynamical systems, some of which can be found in [@op88; @k90; @l90; @y90; @w96]. Later, Araújo and Pacifico, in [@ap06], studied large deviations for certain classes of non-uniformly expanding maps and partially hiperbolic non-uniformly expanding diffeomorphisms. In [@a07], Araújo extended these results to a more general case. Melbourne and Nicol, in [@mn08], obtained a control on large deviations for non-uniformly hyperbolic systems that verify certain properties, including exponential contraction on stable leaves and exponential backward contraction on unstable leaves. In [@m09], Melbourne obtained a slightly better result for large deviations.
Using a tower structure, introduced by Young in [@y98], it is possible, under certain conditions, to obtain a relation between the measure of the tail of the return time function and both the decay of correlations and large deviations. Young, in [@y98], for systems with exponential behaviour in stable and unstable leaves, proved exponential decay of correlations when the measure of the tail of the return time decreases exponentially. In [@y99], Young also proved, for non-invertible systems, both polynomial and exponential decay of correlations based, respectively, on polynomial and exponential control on the tail of the return time. Alves and Pinheiro, in [@ap08], extended the result of [@y98] to a more general case, obtaining, in addition to the exponential decay of correlations, a polynomial decay of correlations assuming a polynomial return time. As for the large deviations, Melbourne and Nicol, in [@mn08], also obtained exponential and polynomial control of large deviations, with the corresponding hypothesis on the tail of the return time.
Overview
--------
The main goal of Section \[amr\] is to define Gibbs-Markov-Young structures and present the two main results of this paper, which give a polynomial control on both the decay of correlations and large deviations, from a polynomial control of the tail of the return time associated to such a structure. In Subsection \[ae\] we present an example on the torus having a Gibbs-Markov-Young structure and the return time function with polynomial tail. In Section \[tm\] we introduce the Young tower and quotient tower, and present some auxiliary results. Sections \[pa\] and \[pb\] are devoted to the proof of the two main theorems. In the Appendix we prove a result in the quotient tower that is important for the control of large deviations.
Statement of results {#amr}
====================
Let $M$ be a finite dimensional Riemannian compact manifold, $d$ be the distance in $M$ and ${\operatorname{Leb}}$ be the Lebesgue measure on the Borel sets of $M$. Given a submanifold $\gamma$ of $M$, let ${\operatorname{Leb}}_{\gamma}$ denote the measure on $\gamma$ induced by the restriction of the Riemannian structure to $\gamma$ and $d_\gamma$ the distance induced in $\gamma$. Consider a diffeomorphism $f: M\rightarrow M$.
Gibbs-Markov-Young structures {#gmys}
-----------------------------
An embedded disk $\gamma \subseteq M$ is called an [*unstable manifold*]{} if, for every $x,y \in \gamma$, $$d (f^{-n}(x),f^{-n}(y)) \underset{n}{\rightarrow} 0.$$ Analogously, an embedded disk $\gamma \subseteq M$ is called a [*stable manifold*]{} if, for every $x,y \in \gamma$, $$d (f^{n}(x),f^{n}(y)) \underset{n}{\rightarrow} 0.$$ We say that $\Gamma^u=\{\gamma^u\}$ is a [*continuous family of $C^1$ unstable manifolds*]{} if there is a compact set $K^s$, a unit disk $D^u$ of some ${\mathbb{R}}^n$ and a map $\phi^u:K^s\times D^u \rightarrow M$ such that:
- $\gamma^u=\phi^u(\{x\} \times D^u)$ is an unstable manifold;
- $\phi^u$ maps $K^s \times D^u$ homeomorphically onto its image;
- $x \rightarrow \phi^u|_{\{x\} \times D^u}$ defines a continuous map from $K^s$ to ${\operatorname{Emb}}^1(D^u,M)$, where ${\operatorname{Emb}}^1(D^u,M)$ denotes the space of $C^1$ embeddings from $D^u$ into $M$.
[*Continuous families of $C^1$ stable manifolds*]{} are defined similarly. We say that $\Lambda \subseteq M$ has a [*hyperbolic product structure*]{} if there exists a continuous family of stable manifolds $\Gamma^s=\{\gamma^s\}$ and a continuous family of unstable manifolds $\Gamma^u=\{\gamma^u\}$ such that:
- $\Lambda=(\bigcup\gamma^s)\bigcap(\bigcup\gamma^u)$;
- $\dim \gamma^s + \dim \gamma^u=\dim M$;
- each $\gamma^s$ intersects each $\gamma^u$ in exactly one point;
- stable and unstable manifolds are transversal with angles bounded away from $0$.
A subset $\Lambda_1 \subseteq \Lambda$ is called an [*$s$-subset*]{} if $\Lambda_1$ also has a hyperbolic product structure and its defining families $\Gamma^s_1$ and $\Gamma^u_1$ can be chosen with $\Gamma^s_1 \subseteq \Gamma^s$ and $\Gamma^u_1 = \Gamma^u$. A subset $\Lambda_2 \subseteq \Lambda$ is called a [*$u$-subset*]{} if $\Lambda_2$ also has a hyperbolic product structure and its defining families $\Gamma^s_2$ and $\Gamma^u_2$ can be chosen with $\Gamma^s_2 = \Gamma^s$ and $\Gamma^u_2 \subseteq \Gamma^u$.
Given $x\in \Lambda$, denote by $\gamma^*(x)$ the element of $\Gamma^*$ containing $x$, for $*\in\{s,u\}$. For each $n\geq 1$ denote by $(f^n)^u$ the restriction of the map $f^n$ to $\gamma^u$-disks, and by $\det D(f^n)^u$ the Jacobian of $(f^n)^u$.
From now on we consider $\Lambda \subseteq M$ having a hyperbolic product structure, with $\Gamma^s$ and $\Gamma^u$ as their defining families. We say that $\Lambda$ has a [*Gibbs-Markov-Young (GMY) structure*]{} if the properties (P$_{0}$)-(P$_{5}$) listed bellow hold.
- [*Lebesgue detectable:*]{} there exists $\gamma\in\Gamma^u$ such that ${\operatorname{Leb}}_\gamma(\Lambda\cap\gamma)>0$.
- [*Markovian:*]{} there are pairwise disjoint $s$-subsets $\Lambda_1, \Lambda_2,...\subseteq \Lambda$ such that:
- ${\operatorname{Leb}}_\gamma \big((\Lambda \backslash \bigcup_{i=1}^\infty \Lambda_i)\bigcap \gamma \big)=0$ on each $\gamma \in \Gamma^u$;
- for each $i \in {\mathbb{N}}$ there exists a $R_i \in {\mathbb{N}}$ such that $f^{R_i} (\Lambda_i)$ is an $u$-subset and, for all $x \in \Lambda_i$, $$f^{R_i} (\gamma^s(x)) \subseteq \gamma^s (f^{R_i}(x)) \quad\mbox{and} \quad f^{R_i} (\gamma^u(x)) \supseteq \gamma^u (f^{R_i}(x)).$$
We introduce a [*return time function*]{} $R: \Lambda \rightarrow {\mathbb{N}}$ and a [*return function*]{} $f^R: \Lambda \rightarrow \Lambda$ defined for each $i \in {\mathbb{N}}$ as $$R|_{\Lambda_i}=R_i \quad\mbox{and}\quad f^R|_{\Lambda_i}=f^{R_i}|_{\Lambda_i}.$$ For $x,y \in \Lambda$, let the [*separation time*]{} $s(x,y)$ be defined as $$s(x,y)=\min\left\{n\in {\mathbb{N}}_0: (f^R)^n(x) \mbox{ and } (f^R)^n(y) \mbox{ are in distinct } \Lambda_i\right\}.$$ For the remaining properties we assume that $C>0$, $\alpha>1$ and $0<\beta<1$ are constants depending only on $f$ and $\Lambda$.
- [*Polynomial contraction on stable leaves:*]{}
$\displaystyle\forall\, \gamma^s\in \Gamma^s \ \forall x,y\in\gamma^s \ \forall n\in{\mathbb{N}}\quad d(f^n(x),f^n(y))\leq \frac{C}{n^\alpha}$.
- [*Backward polynomial contraction on stable leaves:*]{}
$\displaystyle\forall\, \gamma^u\in \Gamma^u \ \forall x,y\in\gamma^u \ \forall n\in{\mathbb{N}}\quad d(f^{-n}(x),f^{-n}(y))\leq \frac{C}{n^\alpha}$.
- [*Bounded distortion:*]{}
For $\gamma \in \Gamma^u$ and $x,y\in \Lambda \cap \gamma$ $$\log \frac{\det D(f^R)^u(x)}{\det D(f^R)^u(y)}\leq C\beta^{s(f^R(x),f^R(y))}.$$
- [*Regularity of the stable foliation:*]{}
For each $\gamma, \gamma' \in \Gamma^u$, defining $$\begin{array}{rccc}
\Theta_{\gamma', \gamma} :&\gamma' \cap \Lambda& \rightarrow& \gamma \cap \Lambda\\
& x& \mapsto& \gamma^s(x) \cap \gamma,
\end{array}$$ then
- $\Theta$ is absolutely continuous and $$\frac{d(\Theta_* {\operatorname{Leb}}_{\gamma'})}{d{\operatorname{Leb}}_{\gamma}}(x)=\prod_{n=0}^\infty \frac{\det\, Df^u(f^n(x))}{\det\,Df^u(f^n(\Theta^{-1}(x))};$$
- denoting $$u(x)=\frac{d(\Theta_* {\operatorname{Leb}}_{\gamma'})}{d{\operatorname{Leb}}_{\gamma}}(x),$$ we have $$\forall\, x,y\in \gamma'\cap \Lambda \quad \log \frac{u(x)}{u(y)}\leq C\beta^{s(x,y)}.$$
The properties of $f$ that we present here are related to similar properties defined in [@y98] and [@ap08]. The main difference here is that we only assume polynomial contraction on stable leaves, in opposition to the exponential contraction in [@y98] and [@ap08]. Our Markovian property (P$_{1}$) is the same as in [@ap08] and weaker than the Markov property in [@y98]; see [@y98 Remark 2.3]. Properties (P$_{2}$) and (P$_{3}$), about polynomial contraction on stable leaves and backward polynomial contraction on unstable leaves, are an improvement over [@y98], where exponential contraction is assumed. In [@ap08], there is no backward contraction assumed on unstable leaves, however, exponential contraction is assumed on the stable ones. Properties (P$_{4}$) and (P$_{5}$) coincide with properties (P$_{4}$) and (P$_{3}$) in [@ap08], respectively. Our properties (P$_{4}$) and (P$_{5}$) are weaker than the corresponding ones in [@y98]; see [@y98 Remark 2.4].
Main results {#mt}
------------
An $f$-invariant probability measure $\mu$ is called a [*Sinai-Ruelle-Bowen (SRB) measure*]{} if all the Lyapunov exponents of $f$ are nonzero $\mu$ almost everywhere and the conditional measures on local unstable manifolds are absolutely continuous with respect to the Lebesgue measures on these manifolds. It was proved in [@y98 Theorem 1] that if $f$ has a hyperbolic structure $\Lambda$ such that the return time $R$ is integrable with respect to ${\operatorname{Leb}}_\gamma$, for some $\gamma\in \Gamma^u$, then $f$ has some SRB measure $\mu$. Given $0<\eta\leq 1$, we define the [*space of $\eta$-Hölder continuous functions*]{} $$H_\eta=\big\{\varphi:M\rightarrow {\mathbb{R}}:\, \exists\, C>0 \ \forall\, x,y\in M \quad |\varphi(x)-\varphi(y)|\leq Cd(x,y)^\eta\big\}$$ with the seminorm $$|\varphi|_\eta= \inf\big\{C>0: \, \forall\, x,y\in M \quad |\varphi(x)-\varphi(y)|\leq Cd(x,y)^\eta\big\}.$$ We denote the *correlation of observables* $\varphi, \psi \in H_\eta$ by $$\mathcal C_n(\varphi,\psi,\mu)=\Big|\int (\varphi {\mbox{\,\small $\circ$}}f^n) \psi\, d\mu - \int \varphi \,d\mu \int \psi \, d\mu\Big|.$$
\[TheoremA\] Suppose that $f$ admits a GMY structure $\Lambda$ with gcd$\{R_i\}=1$ for which there are $\gamma\in \Gamma^u$, $\zeta>1$ and $C_1>0$ such that $${\operatorname{Leb}}_\gamma\{R>n\}\leq\frac{C_1}{n^\zeta}.$$ Then, given $\varphi,\psi\in H_\eta$, there exists $C_2>0$ such that for every $n\ge 1$ $$\mathcal C_n(\varphi,\psi,\mu)\leq C_2\max\Big\{\frac{1}{n^{\zeta-1}},\frac1{n^{\alpha\eta}}\Big\},$$ where $\alpha>0$ is the constant in (P$_{2}$) and (P$_{3}$).
The proof of this theorem will be given in Section \[pa\]. If $\mu$ is an ergodic probability measure and $\varepsilon>0$, the [*large deviation*]{} at time $n$ of the time average of the observable $\phi$ from its spatial average is given by $$LD(\phi,\varepsilon,n,\mu)=\mu\left\{\left|\frac{1}{n}\sum_{i=1}^{n-1} \phi{\mbox{\,\small $\circ$}}f^i-\int \phi\, d\mu\right|>\varepsilon\right\}.$$
\[TheoremB\] Suppose that $f$ admits a GMY structure $\Lambda$ with gcd$\{R_i\}=1$ for which there are $\gamma\in \Gamma^u$, $\zeta>1$ and $C_1>0$ such that $${\operatorname{Leb}}_\gamma\{R>n\}\leq\frac{C_1}{n^\zeta}.$$ Then there are $\eta_0>0$ and $\zeta_0=\zeta_0(\eta_0)>1$ such that for all $\eta>\eta_0$, $1<\zeta<\zeta_0$, $\varepsilon>0$, $p>\max\{1,\zeta-1\}$ and $\phi\in \mathcal H_\eta$, there exists $C_2>0$ such that for every $n\ge 1$ $$LD(\phi,\varepsilon,n,\mu)\leq\frac{C_2}{\varepsilon^{2p}}\frac{1}{n^{\zeta-1}}.$$
This theorem will be proved in Section \[pb\]. We note that, in both theorems, the conclusions remain valid for a power of $f$ when the assumption gcd$\{R_i\}=1$ is removed.
An example {#ae}
----------
Here we give an example of a diffeomorphism $f$ of the two-dimensional torus $\mathbb{T}^2={\mathbb{R}}^2/\mathbb{Z}^2$ with a GMY structure $\Lambda$ having polynomial decay for the Lebesgue measure of the tail of the return time. As a consequence, we deduce that $f$ satisfies the results on polynomial decay of correlations and large deviations from Section \[mt\].
We start with an orientation preserving $C^2$ Anosov diffeomorphism $ f_0$ of $\mathbb{T}^2$ and we consider a finite Markov partition $W_0,\ldots,W_d$ for $f_0$ such that the fixed point $(0,0)$ belongs to the interior of $W_0$. Considering the hyperbolic decomposition into stable and unstable sub-bundles $TM=E^s\oplus E^u$, we assume that there is $0<\lambda<1$ such that $$\|Df|_{E^s}\|<\lambda \quad\text{and}\quad \|Df^{-1}|_{E^u}\|<\lambda.$$ We assume moreover that the transition matrix $A$ of $f_0$ is [*aperiodic*]{}, i.e. some power of $A$ having all entries strictly positive. By a suitable change of coordinates we can suppose that $f_0(a,b)=\big(\phi_0(a),\psi_0(b)\big)$ for all $(a,b)\in W_0$, the local stable manifold of $(0,0)$ is $\{a=0\}$, the local unstable manifold of $(0,0)$ is $\{b=0\}$ and both $\phi_0$ and $\psi_0$ are orientation preserving.
Now we consider $f:\mathbb{T}^2\to \mathbb{T}^2$, a perturbation of $f_0$ that coincides with $f_0$ out of $W_0$ and $f(a,b)=\big(\phi(a),\psi(b)\big)$ for $(a,b)\in W_0$. For definiteness we assume that $W_0=[a_0',a_0]\times [b_0',b_0]$ and consider $V_0=[\phi^{-1}_0(a'_0),\phi^{-1}_0(a_0)]\times [\psi_0(b'_0),\psi_0(b_0)]$. Observe that $V_0$ is a neighborhood of $(0,0)$ strictly contained in $W_0$. We assume that for some $0<\theta<1$ we have $$\phi(a)=a(1+a^\theta)\quad\text{and}\quad\psi(b)=\phi^{-1}(b)\quad\forall(a,b)\in V_0,$$ and assume that $\phi$ and $\psi$ coincide respectively with $\phi_0$ and $\psi_0$ in $W_0\setminus V_0$. Note that $\phi$ is the so-called intermittent map of the type considered in [@y99 Section 6.2] and $(0,0)$ is a fixed point of $f$ with $\phi'(0)=1=\psi'(0)$.
The following theorems, on the decay of correlations and large deviations for the diffeomorphism $f$, will be proved in Section \[theae\].
\[th.ex1\] Let $f$ be as above and take $\varphi,\psi\in H_\eta$. Then, $f$ has a physical measure $\mu$ and there exists $C_2>0$ such that for every $n\ge 1$
1. if $ \eta>\frac{1}{\theta+1},$ then $\displaystyle \mathcal C_n(\varphi,\psi,\mu)\leq \frac{C_2}{n^{1/\theta}}$;
2. if $\eta\leq\frac{1}{\theta+1}$, then $\displaystyle \mathcal C_n(\varphi,\psi,\mu)\leq \frac{C_2}{n^{(1+1/\theta)\eta}}$.
\[th.ex2\] Let $f$ be as above. There are $\eta_0>0$ and $\zeta_0=\zeta_0(\eta_0)>1$ such that for all $\eta>\eta_0$, $1<\zeta<\zeta_0$, $\varepsilon>0$, $p>1/\theta$ and $\phi\in \mathcal H_\eta$, there exists $C_2>0$ such that for every $n\ge 1$ we have $$LD(\phi,\varepsilon,n,\mu)\leq\frac{C_2}{\varepsilon^{2p}}\frac{1}{n^{{1/\theta}}}.$$
Tower structures {#tm}
================
In this section we are going to define a tower structure and a quotient tower, originally introduced by Young in [@y98]. We assume that $f:M\to M$ has a GMY structure $\Lambda$ with return time function $R: \Lambda\to\mathbb N$.
Tower maps {#ts}
----------
We define a [*tower*]{} $$\Delta=\{(x,l): x \in \Lambda \mbox{ and } 0 \leq l<R(x)\}$$ and a [*tower map*]{} $F:\Delta \rightarrow \Delta$ as $$F(x,l)=\left\{ \begin{array} {ll}
(x,l+1) & \mbox{ if } l+1<R(x),\\
(f^R(x),0) & \mbox{ if } l+1=R(x).
\end{array}\right.$$ The [*$l$-th level of the tower*]{} is defined as $$\Delta_l=\{(x,l) \in \Delta \}.$$ There is a natural identification between $\Delta_0$, the $0$-th level of the tower, and $\Lambda$. So, we will make no distinction between them. Under this identification we easily conclude from the definitions that $F^R=f^R$ for each $x \in \Delta_0$. The $l$-th level of the tower is a copy of the set $\{R>l\} \subseteq \Delta_0$. We define a projection map $$\label{pi}
\begin{array} {rccl}
\pi:&\Delta & \rightarrow & \displaystyle{\bigcup_{n=0}^\infty f^n (\Delta_0)}\\
&(x,l) & \mapsto & f^l(x)
\end{array}$$ and observe that $f{\mbox{\,\small $\circ$}}\, \pi=\pi {\mbox{\,\small $\circ$}}F$.
Let $\mathcal{P}$ be a partition of $\Delta_0$ into subsets $\Delta_{0,i}$ with $\Delta_{0,i}=\Lambda_i$ for $i\in {\mathbb{N}}$. We can now define a partition on each level of the tower, $\Delta_l$, by defining its elements as $$\Delta_{l,i}=\{(x,l) \in \Delta_l: x \in \Delta_{0,i} \}.$$ So, the set $\mathcal{Q}={\{\Delta_{l,i}\}}_{l,i}$ is a partition of $\Delta$. We introduce a sequence of partitions $(\mathcal{Q}_n)_{n\geq 0}$ of $\Delta$ defined for $n\in{\mathbb{N}}$ as follows $$\label{partition}
\mathcal{Q}_0=\mathcal{Q}\quad \mbox{ and }\quad \mathcal{Q}_n=\bigvee_{i=0}^n F^{-i} \mathcal{Q} .$$ For each point $x\in \Delta$, let $Q_n(x)$ be the element of $\mathcal{Q}_n$ that contains that point.
Next, we establish a polynomial upper bound on the diameter of the elements of the tower partition.
\[diam\] There exists $C>0$ such that, for all $k\in {\mathbb{N}}$ and $Q\in \mathcal{Q}_{2k}$, $${\operatorname{diam}}(\pi F^k(Q))\leq \frac{C}{k^\alpha}.$$
Take $k>0$ and $Q\in \mathcal{Q}_{2k}$. Fixing $x,y \in Q$, there exists $z=\gamma^u(x)\cap \gamma^s(y)$. Choosing $l$ such that $Q \subseteq \Delta_l$, then $y_0= F^{-l}(y)$ and $z_0= F^{-l}(z)$ are both in $\Delta_0$ and are in the same stable leaf. So, using [(P$_{2}$)]{}, $$\begin{aligned}
\label{dstable}
\nonumber d(\pi F^k(y),\pi F^k(z))&=d(\pi F^{k+l}(y_0),\pi F^{k+l}(z_0))=d(f^{k+l}(\pi y_0), f^{k+l}(\pi z_0)) \\
&\leq \frac{C}{(k+l)^\alpha}d(\pi y_0,\pi z_0) \leq \frac{C_1}{k^\alpha},\end{aligned}$$ because $M$ is compact.
The points $x_0= F^{-l}(x)$ and $z_0= F^{-l}(z)$ are both in $\Delta_0$ and are in the same unstable leaf. So, as above, $$d(\pi F^k(x),\pi F^k(z))=d(f^{k+l}(\pi x_0),f^{k+l}(\pi z_0)).$$ Since $x,z \in Q\cap\Delta_l$ and $Q\in\mathcal{Q}_{2k}$, each pair of points $F^{-i}(x)$ and $F^{-i}(z)$, for $i=0,\dots, l$, belongs to the same element of $\mathcal{Q}$. Then $x_0,z_0 \in Q'$, for some $Q' \in \mathcal{Q}_{2k+l}$, which implies that $F^{2k+l}(x_0),F^{2k+l}(z_0)\in \Delta_{l',i'}$, for some $l',i' \in {\mathbb{N}}$. Therefore, there exists $j\in{\mathbb{N}}_0$ such that $F^{2k+l+j}(x_0),F^{2k+l+j}(z_0)\in \Delta_0$ and so $f^{2k+l+j}(\pi x_0),f^{2k+l+j}(\pi z_0)\in \Lambda$. Then, using [(P$_{3}$)]{} and the compactness of $M$, $$\begin{aligned}
\label{dunstable}
\nonumber d(f^{k+l}(\pi x_0),f^{k+l}(\pi z_0))&=d\big(f^{-k-j}(f^{2k+l+j}(\pi x_0)),f^{-k-j}(f^{2k+l+j}(\pi z_0))\big)\\
&\leq \frac{C}{(k+j)^\alpha}d\big(f^{2k+l+j}(\pi x_0),f^{2k+l+j}(\pi z_0)\big) \leq \frac{C_1}{k^\alpha}.\end{aligned}$$ From and , the conclusion follows.
Quotient towers {#qd}
---------------
We now introduce a quotient tower, obtained from the tower by identifying points in the same stable leaf. Let $\sim$ be the equivalence relation defined on $\Lambda$ by $x\sim y$ if $y\in \gamma^s(x)$. Consider $\bar\Lambda=\Lambda/\!\!\sim$ and the [*quotient tower*]{} $\bar\Delta$, whose levels are $\bar\Delta_l=\Delta_l/\!\!\sim$ and set $\bar\Delta_{l,i}=\Delta_{l,i}/\!\!\sim$. Since the tower map $F$ takes $\gamma^s$-leaves to $\gamma^s$-leaves, we can define $\bar F: \bar\Delta\rightarrow \bar\Delta$ as the function obtained from $F$ by this identification. We introduce a partition of $\bar\Delta$, $\mathcal{\bar Q}={\{\bar\Delta_{l,i}\}}_{l,i}$ and a sequence of partitions $(\mathcal{\bar Q}_n)_{n\in{\mathbb{N}}_0}$ of $\bar \Delta$, defined analogously to , as follows $$\mathcal{\bar Q}_0=\mathcal{\bar Q}\quad \mbox{ and }\quad \mathcal{\bar Q}_n=\bigvee_{i=0}^n \bar F^{-i} \mathcal{\bar Q} \ \mbox{ for } n\in{\mathbb{N}}.$$ Since $R$ is constant on each stable leaf and $f^R$ takes $\gamma^s$-leaves to $\gamma^s$-leaves, then the definitions of the [*return time*]{} $\bar R:\bar\Delta_0\rightarrow{\mathbb{N}}$ and the separation time $\bar s:\bar\Delta_0\times\bar\Delta_0\rightarrow{\mathbb{N}}$ are naturally induced by the corresponding definitions in $\Delta_0$. We extend the [*separation time*]{} $\bar s$ to $\bar\Delta\times\bar\Delta$ in the following way: if $x$ and $y$ belong to the same $\bar\Delta_{l,i}$, take $\bar s(x,y)=\bar s(x_0,y_0)$, where $x_0,y_0$ are the corresponding elements of $\bar\Delta_{0,i}$; otherwise, take $\bar s(x,y)=0$.
We now present an auxiliary result whose proof can be found in [@ap08 Lemma 3.4].
\[J\] There exists a constant $C_F>0$ such that, given $k\in {\mathbb{N}}$ and $x,y\in\bar\Delta$ belonging to the same element of $\mathcal{\bar Q}_{k-1}$, we have $$\left|\frac{J\bar F^k(x)}{J\bar F^k(y)}-1\right|\leq C_F\beta^{\bar s(\bar F^k(x),\bar F^k(y))}.$$
Using property (P$_{5}$) we will be able to define a reference measure $\bar m$ on the quotient tower $\bar\Delta$. We start by defining measures $m_\gamma$ on each $\gamma\cap\Lambda$, $\gamma\in \Gamma^u$. Fix $\widehat\gamma\in \Gamma^u$ and, for any given $\gamma\in\Gamma^u$ and $x\in\gamma\cap\Lambda$, let $\widehat x$ be the point in $\gamma^s(x)\cap\widehat\gamma$. Define $$\widehat u(x)=\prod_{n=0}^\infty \frac{\det\, Df^u(f^n(x))}{\det\, Df^u(f^n(\widehat x))}$$ and note that $\widehat u$ satisfies (P$_{5}$)-(b). For each $\gamma\in\Gamma^u$, define $m_\gamma$ as the measure in $\gamma$ such that $$\frac{dm_\gamma}{d{\operatorname{Leb}}_\gamma}=\widehat u 1_{\gamma\cap\Lambda}.$$ We are going to see that, if $\Theta=\Theta_{\gamma,\gamma'}$ is as defined in (P$_{5}$), then $$\label{mgamma}
\Theta_* m_\gamma=m_{\gamma'}.$$ To show this it is enough to verify that the density of both measures with respect to ${\operatorname{Leb}}_{\gamma'}$ coincide. Indeed, from (P$_{5}$)-(a) we have $$\frac{\widehat u(x')}{\widehat u(x)}=
\prod_{n=0}^\infty \frac{\det\, Df^u(f^n(x'))}{\det\,Df^u(f^n(\widehat x))} \cdot\frac{\det\, Df^u(f^n(\widehat x))}{\det\,Df^u(f^n(x))}=
\frac{d\Theta_* {\operatorname{Leb}}_\gamma}{d{\operatorname{Leb}}_{\gamma'}}(x'),$$ and so, $$\frac{d\Theta_* m_\gamma}{d{\operatorname{Leb}}_{\gamma'}}(x')=\widehat u(x)\frac{d\Theta_* {\operatorname{Leb}}_\gamma}{d{\operatorname{Leb}}_{\gamma'}}(x')=\widehat u(x')=
\frac{d\Theta_* m_{\gamma'}}{d{\operatorname{Leb}}_{\gamma'}}(x').$$
Now define $m$ as the measure on $\Lambda$ whose conditional measures on $\gamma\cap\Lambda$ for $\gamma\in\Gamma^u$ are the measures $m_\gamma$. Take a measure in $\Delta$, also denoted by $m$, by letting $m_{|\Delta_l}$ be induced by the natural identification of $\Delta_l$ and a subset of $\Lambda$. Finally, since holds, we can define a measure $\bar m$ on $\bar \Delta$ whose representative on each $\gamma\in\Gamma^u$ is the measure $m_\gamma$.
Given $0<\beta<1$, we define the functional spaces $$\begin{aligned}
&{\mathcal F}_\beta=\big\{\varphi:\bar\Delta\rightarrow {\mathbb{R}}:\, \exists\, C_\varphi>0 \ \forall\, x, y\in \bar\Delta
\quad |\varphi( x)-\varphi( y)|\leq C_\varphi \beta^{ \bar s( x, y)}\big\},\\
&{\mathcal F}_\beta^{\,+}=\big\{\varphi\in{\mathcal F}_\beta :\, \exists\, C_\varphi>0\text{ such that on each }\bar\Delta_{l,i}, \text{ either }
\varphi\equiv 0\text{ or }\\
&\hspace{14mm}\varphi>0\text{ and for all } x, y\in \bar\Delta_{l,i}\quad
\Big|\frac{\varphi( x)}{\varphi( y)}-1\Big|\leq C_\varphi \beta^{\bar s( x, y)}\big\}.\end{aligned}$$ From now on, we denote by $C_\varphi$ both the infimum of the constant in the definition of ${\mathcal F}_\beta$ and of ${\mathcal F}_\beta^+$ with respect to $\varphi$. We also denote by ${\mathcal F}_\beta$ and ${\mathcal F}_\beta^+$ the analogous sets defined for functions with domain $M$ or $\Delta$. The proof of the following result can be found in [@y99 Lemma 2] and [@y98 Theorem 1].
\[nu\] Assume that $\bar R$ is integrable with respect to $\bar m$. Then
1. $\bar F$ has a unique invariant probability measure $\bar\nu$ equivalent to $\bar m$;
2. $d\bar\nu/d\bar m\in{\mathcal F}_\beta^{\,+}$ and is bounded from below by a positive constant;
3. $(\bar F,\bar\nu)$ is mixing.
The next theorem plays a key role in the proof of Theorem \[TheoremA\] and is an improved version of [@y99 Theorem 2] given in [@ap08 Theorem 3.6] whose proof can be found in [@ap08 Appendix A].
\[3.6\] For $\varphi\in {\mathcal F}_\beta^{\,+}$ let $\bar\lambda$ be the measure whose density with respect to $\bar m$ is $\varphi$.
1. If ${\operatorname{Leb}}\{\bar R>n\}\leq C n^{-\zeta}$, for some $C>0$ and $\zeta>1$, then there is $C'>0$ such that $$\left|\bar F^n_*\bar\lambda-\bar\nu\right|\leq C' n^{-\zeta+1}.$$
2. If ${\operatorname{Leb}}\{\bar R>n\}\leq C e^{-cn^\eta}$, for some $C,c>0$ and $0<\eta\leq 1$, then there is $C',c'>0$ such that $$\left|\bar F^n_*\bar\lambda-\bar\nu\right|\leq C' e^{-c'n^\eta}.$$
Moreover, $c'$ does not depend on $\varphi$ and $C'$ depends only on $C_\varphi$.
To prove Thoerem \[TheoremB\] we need to consider more general functional spaces, due to the fact that the polynomial estimates in the stable manifolds interfere with the regularity of the function $\psi$ that will appear in Proposition \[metric\]. Given $\theta>0$, we define $$\begin{aligned}
&\mathcal G_\theta=\left\{\varphi:\bar{\Delta}\rightarrow {\mathbb{R}}:
\exists\, D_\varphi>0\ \forall x,y\in \bar{\Delta}\quad\ |\varphi(x)-\varphi(y)|\leq \frac{D_\varphi}{\max\{\bar s(x,y),1\}^\theta}\right\},\\
&\mathcal G_\theta^+=\Bigg\{\varphi\in \mathcal G_\theta:
\exists\, D_\varphi>0 \text{ such that on each } \bar{\Delta}_{l,i}, \text{ either } \varphi\equiv 0 \text{ or } \\
&\hspace{14mm} \varphi>0 \text{ and for all } x,y\in \bar{\Delta}_{l,i} \quad \left|\frac{\varphi(x)}{\varphi(y)}-1\right|\leq \frac{D_\varphi}{\max\{\bar s(x,y),1\}^\theta}\Bigg\}.\end{aligned}$$ As above, we denote by $D_\varphi$ both the infimum of the constant in the definition of ${\mathcal G}_\theta$ and of ${\mathcal G}_\theta^+$ with respect to $\varphi$. The sets ${\mathcal G}_\theta$ and ${\mathcal G}_\theta^+$ also represent the analogous sets defined for functions with domain $M$ or $\Delta$.
\[Theorem C\] Assume that there is $C>0$ such that $$m\{\bar R>n\}\leq \frac{C}{n^\zeta} .$$ Then there are $\theta_0>1$ and $1<\zeta_0=\zeta_0(\theta)$ such that for all $\theta\ge \theta_0$ and $1<\zeta<\zeta_0$, given $\varphi\in \mathcal G_\theta^+$ there exists $C'>0$, depending only on $D_\varphi$, such that $$\left|\bar F^n_*\bar\lambda-\bar\nu\right|\leq \frac{C'}{n^{\zeta-1}},$$ where $\bar\lambda$ is the measure whose density with respect to $\bar m$ is $\varphi$.
This theorem is similar to [@y99 Theorem 3] and will be proved in Appendix \[appendix\]. Note that we are only assuming $\varphi\in \mathcal G_\theta^+$, instead of $\mathcal F_\beta^+$, which forces us to impose some extra assumptions. In practice, we only need the following consequence of Theorem \[Theorem C\].
\[Corollary\] Assume that there is $C>0$ such that $$m\{\bar R>n\}\leq \frac{C}{n^\zeta} .$$ Then there are $\theta_0>1$ and $1<\zeta_0=\zeta_0(\theta)$ such that for all $\theta\ge \theta_0$ and $1<\zeta<\zeta_0$, given $\varphi\in \mathcal G_\theta$ and $\psi\in L^\infty$ there exists $C'>0$, depending only on $D_\varphi$ and $\|\psi\|_\infty$, such that $$\mathcal C_n(\psi,\varphi,\bar\nu)\leq \frac{C'}{n^{\zeta-1}}.$$
Let $\rho=\frac{d\bar \nu}{d{\operatorname{Leb}}}$ and take $\widetilde\varphi=b(\varphi+a)$, where $a\geq 0$ is such that $\widetilde\varphi$ is bounded from below by a strictly positive constant and $b>0$ is such that $ \int \widetilde\varphi\rho\, d{\operatorname{Leb}}=1$. Note that, since $\varphi\in \mathcal G_\theta$, then $\widetilde\varphi\in \mathcal G_\theta^+$. In addition, as $\rho\in \mathcal F_\beta^+$ by Theorem \[nu\] and $\mathcal F_\beta^+\subseteq \mathcal G_\theta^+$, then $\widetilde\varphi\rho\in \mathcal G_\theta^+$.
Let $P:L^2(\bar\Delta)\rightarrow L^2(\bar\Delta)$ be the Perron-Frobenius operator associated with $\bar F$, defined as follows: $$\forall\,v,w\in L^2(\bar\Delta) \quad \int_{\bar\Delta} P(v)\,w\,d\bar\nu=
\int_{\bar\Delta} v\,w{\mbox{\,\small $\circ$}}\bar F\,d\bar\nu.$$ This means that if $\mu$ is a signed measure and $\phi=\frac{d\mu}{d{\operatorname{Leb}}}$, then $P(\phi)=\frac{d(F_*\mu)}{d{\operatorname{Leb}}}$. So, if $\lambda$ is the measure such that $\frac{d\lambda}{d{\operatorname{Leb}}}=\widetilde\varphi\rho$, we have $$\begin{aligned}
\mathcal C_n(\psi,\varphi,\bar\nu)&=\frac{1}{b}\Big|\int (\psi {\mbox{\,\small $\circ$}}\bar F^n) (\widetilde\varphi\rho)\, d{\operatorname{Leb}}- \int \psi\rho \,d{\operatorname{Leb}}\int \widetilde\varphi\rho \, d{\operatorname{Leb}}\Big|\\
&=\frac{1}{b}\Big|\int \psi P^n(\widetilde\varphi\rho)\, d{\operatorname{Leb}}- \int \psi\rho \,d{\operatorname{Leb}}\Big|\\
&\leq\frac{1}{b}\int |\psi| \big|P^n(\widetilde\varphi\rho) -\rho\big| \,d{\operatorname{Leb}}\leq\frac{1}{b} \|\psi\|_\infty \big|\bar F_*^n\lambda -\bar\nu\big|.\end{aligned}$$ Since we have $\frac{d\lambda}{d{\operatorname{Leb}}}=\widetilde\varphi\rho\in\mathcal G_\theta^+$, the conclusion follows from Theorem \[Theorem C\].
Decay of correlations {#pa}
=====================
This section is dedicated to the proof of Theorem \[TheoremA\], adapting the approach of [@y98] and [@ap08] to our more general conditions on the definition of GMY structure. First of all we note that it was shown in [@y99 Sections 2 and 4] that there exists a measure $\nu$ on the tower $\Delta$ such that $\mu=\pi_* \nu$ and $\bar{\nu}=\bar{\pi}_* \nu$. Fixing $\varphi, \psi \in H_\eta$, let $$\label{eq.til}
\widetilde{\varphi}=\varphi {\mbox{\,\small $\circ$}}\, \pi\quad\text{and}\quad \widetilde{\psi}=\psi {\mbox{\,\small $\circ$}}\, \pi.$$ Observe that $$\int (\varphi {\mbox{\,\small $\circ$}}f^n) \psi\, d\mu=\int (\varphi {\mbox{\,\small $\circ$}}f^n) \psi\, d(\pi_*\nu)= \int (\varphi {\mbox{\,\small $\circ$}}\pi{\mbox{\,\small $\circ$}}F^n)\widetilde\psi\, d\nu=\int (\widetilde\varphi {\mbox{\,\small $\circ$}}F^n)\widetilde\psi\, d\nu,$$ and, arguing as above, $$\int \varphi\,d\mu \int \psi \ d\mu = \int \widetilde{\varphi}\, d\nu \int \widetilde{\psi} \, d\nu.$$ Hence we have $\mathcal C_n(\varphi,\psi,\mu)=\mathcal C_n(\widetilde\varphi,\widetilde\psi,\nu)$. Given $n\in {\mathbb{N}}$, we fix a positive integer $k<n/2$ and define the discretization $\bar{\varphi}_k$ of $\widetilde \varphi$ on the tower $\Delta$ as $$\bar{\varphi}_k|_Q=\inf\{\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k(x):x \in Q\},\quad \mbox{ for } Q \in \mathcal{Q}_{2k}.$$
There exists $C_2>0$ depending only on $|\varphi|_\eta$ and on $\|\psi\|_\infty$ such that $$|\mathcal C_n(\widetilde\varphi,\widetilde\psi,\nu)-\mathcal C_{n-k}(\bar \varphi_k,\widetilde\psi,\nu)| \leq \frac{C_2}{k^{\alpha \eta}}.$$
Notice that as $\nu$ is $F$-invariant $$\label{Cn-k}
\mathcal C_{n-k}(\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k,\widetilde{\psi},\nu)=
\Big|\int (\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k {\mbox{\,\small $\circ$}}F^{n-k}) \widetilde{\psi} d\nu - \int \widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k d\nu \int \widetilde{\psi} \ d\nu\Big|=\mathcal C_n(\widetilde{\varphi},\widetilde{\psi},\nu).$$ Using the fact that $\varphi$ is H" older continuous and Lemma \[diam\], we observe that for $Q\in \mathcal Q_{2k}$ and all $x,y \in Q$, $$\begin{aligned}
|\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k(x)-\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k(y)|&=|\varphi {\mbox{\,\small $\circ$}}\, \pi {\mbox{\,\small $\circ$}}F^k(x)-\varphi {\mbox{\,\small $\circ$}}\, \pi {\mbox{\,\small $\circ$}}F^k(y)|\\
& \leq |\varphi|_\eta d(\pi F^k(x),\pi F^k(y))^\eta \leq |\varphi|_\eta \Big(\frac{C}{k^\alpha} \Big)^\eta,\end{aligned}$$ which implies that, for any $x\in Q$, $$\label{phik1}|
\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k(x)-\bar{\varphi}_k(x)| \leq |\varphi|_\eta \Big(\frac{C}{k^\alpha} \Big)^\eta.$$ Applying , and the $F$-invariance of $\nu$ we obtain $$\begin{aligned}
\big|\mathcal C_n(\widetilde{\varphi},\widetilde{\psi},\nu)&-\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi},\nu)\big|=
\big|\mathcal C_{n-k}(\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k,\widetilde{\psi},\nu)-\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi},\nu)\big|\\
&\leq \Big|\int (\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k-\bar{\varphi}_k) {\mbox{\,\small $\circ$}}F^{n-k} \widetilde{\psi} d\nu\Big| + \Big|\int (\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k-\bar{\varphi}_k) d\nu \int \widetilde{\psi} d\nu\Big|\\
&\leq \|\psi\|_\infty \Big(\int \Big|(\widetilde{\varphi} {\mbox{\,\small $\circ$}}F^k-\bar{\varphi}_k) {\mbox{\,\small $\circ$}}F^{n-k}\big| d\nu + \int \big|\widetilde{\varphi}
{\mbox{\,\small $\circ$}}F^k-\bar{\varphi}_k\Big| d\nu \Big) \\
& \leq 2\|\psi\|_\infty |\varphi|_\eta \Big(\frac{C}{k^\alpha} \Big)^\eta.\end{aligned}$$ We only need to take $C_2=2\|\psi\|_\infty |\varphi|_\eta C^\eta$.
Now we define $\bar{\psi}_k$ in a similar way to $\bar{\varphi}_k$. Denote by $\bar{\psi}_k \nu$ the signed measure whose density with respect to $\nu$ is $\bar{\psi}_k$ and by $\widetilde{\psi}_k$ the density of $F^k_*\bar{\psi}_k \nu$ with respect to $\nu$. Let $|\nu|$ denote the total variation of a signed measure $\nu$.
There exists $C_3>0$ depending only on $|\psi|_\eta$ and on $\|\varphi\|_\infty$ such that $$\big|\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi},\nu)-\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi}_k,\nu)\big| \leq \frac{C_3}{k^{\alpha \eta}}.$$
Observe that, since $\|\bar\varphi_k\|_\infty\leq \|\varphi\|_\infty$, we have $$\begin{aligned}
\label{c-c}
\nonumber\big|\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi},\nu)-\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi}_k,\nu)\big| &
\leq \int \big|\bar{\varphi}_k {\mbox{\,\small $\circ$}}F^{n-k}\big|\, \big|\widetilde{\psi}-\widetilde{\psi}_k\big| d\nu + \int \big|\bar{\varphi}_k\big| d\nu \int \big|\widetilde{\psi}-\widetilde{\psi}_k\big| d\nu\\
& \leq 2\|\varphi\|_\infty \int \big|\widetilde{\psi}-\widetilde{\psi}_k\big| d\nu.\end{aligned}$$ Note also that $$F^k_*((\widetilde{\psi} {\mbox{\,\small $\circ$}}F^k) \nu)=\widetilde{\psi} \nu$$ and so, by the definition of $\widetilde\psi_k$, we have $$\label{F*}
\int \big|\widetilde{\psi}-\widetilde{\psi}_k\big| d\nu= \big|F^k_*((\widetilde{\psi} {\mbox{\,\small $\circ$}}F^k) \nu)-F^k_*(\bar{\psi}_k \nu)\big|
\leq\big|(\widetilde{\psi} {\mbox{\,\small $\circ$}}F^k-\bar{\psi}_k)\nu \big|=\int \big|\widetilde{\psi} {\mbox{\,\small $\circ$}}F^k-\bar{\psi}_k \big| d\nu$$ Using Lemma \[diam\], , and the same argument as in we get $$|\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi},\nu)-\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi}_k,\nu)|\leq 2\|\varphi\|_\infty \int |\widetilde{\psi} {\mbox{\,\small $\circ$}}F^k-\bar{\psi}_k| d\nu \leq 2\|\varphi\|_\infty |\psi|_\eta \Big(\frac{C_1}{k^\alpha} \Big)^\eta.$$ To conclude, we just need to take $C_3=2\|\varphi\|_\infty|\psi|_\eta C_1^\eta$.
$\mathcal C_{n-k}(\bar{\varphi}_k,\widetilde{\psi}_k,\nu)=\mathcal C_n(\bar{\varphi}_k,\bar{\psi}_k,\bar\nu).$
By definition of $\widetilde{\psi}_k$ we have $$\int (\bar{\varphi}_k {\mbox{\,\small $\circ$}}F^{n-k}) \widetilde{\psi}_k d\nu=\int \bar{\varphi}_k d(F_*^{n-k} (\widetilde{\psi}_k \nu))=\int \bar{\varphi}_k d(F_*^n (\bar{\psi}_k \nu)).$$ Since $\bar{\varphi}_k$ is constant on $\gamma^s$ leaves and $F$ and $\bar{F}$ are semi-conjugated by $\bar{\pi}$, then $$\int \bar{\varphi}_k d(F_*^n (\bar{\psi}_k \nu))=\int \bar{\varphi}_k d(\bar{\pi}F_*^n (\bar{\psi}_k \nu)))
=\int \bar{\varphi}_k d(\bar{F}_*^n (\bar{\psi}_k \bar{\nu}))=\int (\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n) \bar{\psi}_k d\bar{\nu}.$$ So, we have proved that $$\label{part1}
\int (\bar{\varphi}_k {\mbox{\,\small $\circ$}}F^{n-k}) \widetilde{\psi}_k d\nu=\int (\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}_*^n) \bar{\psi}_k d\bar{\nu}.$$ Additionally, as $\bar{\varphi}_k$ is constant on $\gamma^s$ leaves, and using the definition of $\widetilde{\psi}_k$ and the $\bar F$-invariance of $\bar\nu$, we may write $$\begin{aligned}
\label{part2}
\int \bar{\varphi}_k d\nu \int \widetilde{\psi}_k d\nu&=\int \bar{\varphi}_k d\bar{\nu} \int d(F^k_*(\bar{\psi}_k \nu))
=\int \bar{\varphi}_k d\bar{\nu} \int \bar{\psi}_k d\bar{\nu}.\end{aligned}$$ Gathering and , we obtain the conclusion.
Without loss of generality we may assume that $\bar{\psi}_k$ is not the null function. Defining $$b_k=\left(\int (\bar{\psi}_k+2\|\bar{\psi}_k\|_\infty) d\bar{\nu}\right)^{-1} \quad \mbox{and} \quad \widehat{\psi}_k=b_k(\bar{\psi}_k+2\|\bar{\psi}_k\|_\infty),$$ we obtain $$\frac{1}{3\|\bar{\psi}_k\|_\infty} \leq b_k \leq \frac{1}{\|\bar{\psi}_k\|_\infty}.$$ Defining $\displaystyle \bar\rho=\frac{d\bar\nu}{d\bar m}$, it follows from the definition of $b_k$ that $$\int \widehat{\psi}_k \,\bar{\rho}\, d\bar{m}=1.$$ Since $\bar{\psi}_k$ is constant on elements of $\mathcal{Q}_{2k}$, the same holds for $\widehat{\psi}_k$. Let $\widehat{\lambda}_k$ be the probability measure on $\bar{\Delta}$ whose density with respect to $\bar{m}$ is $\widehat{\psi}_k \bar{\rho}$.
There exists $C_4>0$ depending only on $\|\varphi\|_\infty$ and on $\|\psi\|_\infty$ such that $$\mathcal C_n(\bar{\varphi}_k,\bar{\psi}_k,\bar\nu) \leq C_4 |\bar{F}^{n-2k}_* \bar{\lambda}_k-\nu|.$$
Notice that, by the definition of $\widehat\psi_k$ and the $\bar F$-invariance of $\bar\nu$, $$\begin{aligned}
\int(\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n)\,\bar{\psi}_k \,d\bar{\nu}&=\int(\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n)\,\Big(\frac{1}{b_k} \widehat\psi_k-2\|\bar\psi_k\|_\infty \Big) \,d\bar{\nu}\\
& = \frac{1}{b_k}\int(\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n)\, \widehat\psi_k\,d\bar\nu- 2\|\bar\psi_k\|_\infty \int\bar{\varphi}_k \,d\bar{\nu}\end{aligned}$$ and, similarly, $$\int \bar{\varphi}_k \,d\bar{\nu}\int \bar{\psi}_k \,d\bar{\nu}=\frac{1}{b_k}\int \bar{\varphi}_k \,d\bar{\nu}\int \widehat{\psi}_k \,d\bar{\nu}-2\|\bar\psi_k\|_\infty \int\bar{\varphi}_k \,d\bar{\nu}.$$ Then, using the last two equalities and the definitions of $\bar\rho$ and $\widehat\lambda_k$, we obtain $$\begin{aligned}
\label{Cn}
\nonumber \mathcal C_n(\bar{\varphi}_k,\bar{\psi}_k,\bar\nu)&=\frac{1}{b_k} \Big|\int(\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n) \,\widehat{\psi}_k \,d\bar{\nu}-\int \bar{\varphi}_k \,d\bar{\nu} \int \widehat{\psi}_k \,d\bar{\nu} \Big|\\
\nonumber &=\frac{1}{b_k} \Big|\int(\bar{\varphi}_k {\mbox{\,\small $\circ$}}\bar{F}^n) \,\widehat{\psi}_k \bar\rho\,d\bar{m}-\int \bar{\varphi}_k \bar\rho\,d\bar{m} \int \widehat{\psi}_k \bar\rho\,d\bar{m} \Big|\\
&\leq \frac{1}{b_k} \int |\bar{\varphi}_k|\,\Big|\frac{d(\bar{F}^n_* \widehat{\lambda}_k)}{d\bar{m}}-\bar{\rho} \Big|\,d\bar{m}.\end{aligned}$$ Setting $\bar{\lambda}_k=\bar{F}^{2k}_* \widehat{\lambda}_k$ and since $k<n/2$, we have $$\frac{d}{d\bar{m}} \bar{F}^n_* \widehat{\lambda}_k=\frac{d}{d\bar{m}} \bar{F}^{n-2k}_* \bar{\lambda}_k,$$ and so, using and since $\frac{1}{b_k} \leq 3\|\bar{\psi}_k\|_\infty$, we have $$\begin{aligned}
\mathcal C_n(\bar{\varphi}_k,\bar{\psi}_k,\bar\nu) &\leq \frac{1}{b_k} \int |\bar{\varphi}_k| \Big|\frac{d(\bar{F}^n_* \widehat{\lambda}_k)}{d\bar{m}}-\bar{\rho} \Big|d\bar{m}\\&= \frac{1}{b_k} \int |\bar{\varphi}_k| \,\Big|\frac{d(\bar{F}^{n-2k}_* \bar{\lambda}_k)}{d\bar{m}}-\frac{d\bar{\nu}}{d\bar{m}} \Big|d\bar{m}\\
& \leq \frac{1}{b_k} \|\bar{\varphi}_k\|_\infty |\bar{F}^{n-2k}_* \bar{\lambda}_k-\bar\nu|\\
&\leq 3\|\bar{\psi}_k\|_\infty \|\bar{\varphi}_k\|_\infty |\bar{F}^{n-2k}_* \bar{\lambda}_k-\nu|\\
& \leq 3\|\psi\|_\infty \|\varphi\|_\infty |\bar{F}^{n-2k}_* \bar{\lambda}_k-\nu|.\end{aligned}$$ We just need to take $C_4=3\|\psi\|_\infty \|\varphi\|_\infty$.
Gathering everything that was proved in the previous lemmas, we get $$\begin{aligned}
\mathcal C_n(\varphi,\psi,\mu)&=\mathcal C_n(\widetilde\varphi,\widetilde\psi,\mu)\nonumber\\
&\leq \mathcal C_{n-k}(\bar\varphi_k,\widetilde\psi,\mu)+\frac{C_2}{k^{\alpha\eta}}\nonumber\\
&\leq \mathcal C_{n-k}(\bar\varphi_k,\widetilde\psi_k,\nu)+\frac{C_3+C_2}{k^{\alpha\eta}}\nonumber\\
&\leq C_4 |\bar{F}^{n-2k}_* \bar{\lambda}_k-\nu| +\frac{C_3+C_2}{k^{\alpha\eta}}.\label{ineq.nova}\end{aligned}$$ Let $\phi_k$ be the density of the measure $\bar{\lambda}_k$ with respect to $\bar{m}$. The next lemma, whose proof is given in [@ap08 Lemma 4.1], gives that $\phi_k\in {\mathcal F}^+_\beta$.
\[phik\] There is $C>0$, not depending on $\phi_k$, such that $$|\phi_k(\bar x)-\phi_k(\bar y)|\leq C\beta^{\bar s(\bar x,\bar y)},\qquad \forall\,\bar x,\bar y\in\bar\Delta.$$
Now Lemma \[phik\] together with allow us to use Theorem \[3.6\] and obtain $$\mathcal C_n(\varphi,\psi,\mu)\leq C_4\,\frac{C'}{(n-2k)^{\zeta-1}}+\frac{C_3+C_2}{k^{\alpha\eta}}\leq C\max\Big\{
\frac{1}{n^{\zeta-1}},\frac1{n^{\alpha\eta}}\Big\}.$$ This concludes the proof of Theorem \[TheoremA\].
Large deviations {#pb}
================
In this section we prove Theorem \[TheoremB\]. Though our assumptions are different from [@mn08] and [@m09], we will follow the approach in these papers. The proof of Theorem \[TheoremB\] uses the construction of a function $\psi\in\mathcal G_\theta(\bar\Delta)$, which will be done in Proposition \[metric\], for $\theta=\alpha\eta-1$.
\[2b\] There exists $C_3>0$ such that, for all $x,y\in\gamma^u$ with $s(x,y)\neq 0$ and $0\leq k<R$ we have $$d(f^kx,f^ky)\le \frac{C_3}{s(x,y)^\alpha}.$$
Let $n\in{\mathbb{N}}$ be such that $s(x,y)=n$. Using (P$_{3}$), we get $$d(f^kx,f^ky)=d(f^{k-R_n}(f^{R_n}x), f^{k-R_n}(f^{R_n}y))\leq \frac{C'}{(R_n-k)^\alpha} d(f^{R_n}x, f^{R_n}y)\leq \frac{C_3}{(R_n-k)^\alpha}.$$ Since $R-k\geq1$, then $R_n-k\geq n$, and so $$d(f^kx,f^ky)\leq \frac{C_3}{(s(x,y))^\alpha}.$$
We say that $\psi:\Delta\rightarrow {\mathbb{R}}$ *depends only on future coordinates* if, given $x,y\in \Delta$ with $y\in \gamma_s(x)$, then $\psi(x)=\psi(y)$. In particular, a function $\psi:\Delta\rightarrow {\mathbb{R}}$ depending only on future coordinates can be interpreted as defined in the quotient $\bar\Delta$. The following result is an adaptation of [@mn05 Lemma 3.2] to the polynomial case.
\[metric\] Let $f$ have a GMY structure $\Lambda$ and $\phi:M\rightarrow{\mathbb{R}}$ be a function in $\mathcal{H}_\eta$ with $\eta> 1/\alpha$. Then there exist functions $\chi,\psi:\Delta\rightarrow{\mathbb{R}}$ such that:
1. $\chi\in L^\infty(\Delta)$ and $\|\chi\|_\infty$ depends only on $|\phi|_\eta$;
2. $\phi\circ\pi=\psi+\chi-\chi\circ F$;
3. $\psi$ depends only on future coordinates;
4. the function $\psi:\bar\Delta\rightarrow{\mathbb{R}}$ belongs to $\mathcal G_\theta$, for $\theta=\alpha\,\eta-1$.
Let us fix $\gamma^u\in \Gamma^u$. Given $p=(x,l)\in\Delta$, let $\widehat p=(\widehat x,l)$, where $\widehat x$ is the unique point in $\gamma^s(x)\cap \gamma^u$ and define $$\chi(p)=\sum_{j=0}^\infty\big(\phi \pi F^j(p)-\phi \pi F^j(\widehat p)\big).$$ Observing that $\pi\circ F^j(p)=f^j\circ\pi(p)=f^{j+l}(x)$ and using [(P$_2)$]{}, we have $$\begin{aligned}
|\chi(p)|\le\sum_{j=0}^\infty\Big|\phi \pi F^j (p) -\phi \pi F^j (\widehat{Fp}) \Big|&\le \sum_{j=0}^\infty |\phi|_\eta\, d(f^{j+l}(x),f^{j+l}(\widehat x))^\eta\\
&\le |\phi|_\eta\, C^\eta\sum_{j=0}^\infty\frac1{j^{\alpha\eta}}=C'\, |\phi|_\eta,\end{aligned}$$ since $\alpha\eta>1$. So, the first item holds. Defining $\psi=\phi\circ\pi-\chi+\chi\circ F$, the second item is verified and, as $$\begin{aligned}
\psi(p)&=\phi\pi (p)-\sum_{j=0}^\infty \phi\pi F^j (p)+\sum_{j=0}^\infty \phi\pi F^j (\widehat p)+\sum_{j=0}^\infty \phi\pi F^{j+1} (p)
-\sum_{j=0}^\infty\phi \pi F^j (\widehat{Fp})\\
&=\sum_{j=0}^\infty\big(\phi\pi F^j (\widehat p)-\phi \pi F^j (\widehat{Fp})\big),\end{aligned}$$ $\psi$ depends only on future coordinates. So, the third item is proved. We are left to prove the last item. Let $n\in{\mathbb{N}}$ and $p,q\in\Delta$. Then $$\begin{gathered}
\label{psi}
|\psi(p)-\psi(q)|\le\sum_{j=0}^n\big|\phi \pi F^j (\widehat p)-\phi \pi F^j (\widehat q)\big|
+\sum_{j=0}^{n-1}\big|\phi \pi F^j (\widehat{Fp})-\phi \pi F^j (\widehat{Fq})\big|\\
+\sum_{j=n+1}^\infty\big|\phi \pi F^j (\widehat p)-\phi \pi F^{j-1} (\widehat{Fp})\big|
+\sum_{j=n+1}^\infty\big|\phi \pi F^j (\widehat{q})-\phi \pi F^{j-1} (\widehat{Fq})\big|.\end{gathered}$$ Since the choice of $n$ is arbitrary we can assume that $s(p,q)\approx 2n$. This means that there will be no separation during the calculations of the first two terms. We will consider separately each term of the right-hand side of . We start with the third term. When $p\neq(x,R(x)-1)$ then $F\widehat p=\widehat{Fp}$. If $p=(x,R(x)-1)$ then $F\widehat p=(f^{R}\widehat x,0)$ and $\widehat{Fp}=(\widehat{f^{R}x},0)$ and so $\pi F^j\widehat p= f^{j-1} f^{R}\widehat x$ and $\pi F^{j-1}\widehat{Fp}= f^{j-1} \widehat{f^{R}x}$. But $f^{R}\widehat x$ and $\widehat{f^{R}x}$ belong to the same stable leaf, and then, using [(P$_2)$]{}, $$\begin{aligned}
\big|\phi \pi F^j (\widehat{p})-\phi \pi F^{j-1} (\widehat{Fp})\big|&=\big|\phi f^{j-1} f^{R}(\widehat x)-\phi f^{j-1}(\widehat{f^{R}x})\big|\\
&\le |\phi|_\eta {d(f^{j-1} f^{R}(\widehat x),f^{j-1}(\widehat{f^{R}x}))}^\eta\le |\phi|_\eta\frac1{(j-1)^{\theta+1}}\end{aligned}$$ and then, recalling that $s(p,q)\approx 2n$, $$\label{cima}
\sum_{j=n+1}^\infty\big|\phi \pi F^j \widehat{p}-\phi \pi F^{j-1} \widehat{Fp}\big|\le |\phi|_\eta\sum_{j=n+1}^\infty\frac1{(j-1)^{\theta+1}}
\le C'|\phi|_\eta\frac1{n^{\theta}} \approx 2^\theta C'|\phi|_\eta\frac1{s(p,q)^{\theta}}.$$ The calculations for the fourth term of the right-hand side of are similar.
Consider now the first term and take $p=(x,l)$ and $q=(y,l)$. Then $$\pi F^j(\widehat p)=f^{j+l}(\widehat x)=f^L {f^{R(x)}}^J(\widehat x),\quad\text{
where }\quad J\le j\ \text{ and }\ L<R\big((f^R)^J(\widehat x)\big),$$ and analogously for $\pi F^j(\widehat q)$. Then, since $\phi\in \mathcal{H}_\eta$ and using the calculations above and Lemma \[2b\], $$\begin{aligned}
\big|\phi \pi F^j (\widehat{p})-\phi \pi F^{j} (\widehat{q})\big|&\le |\phi|_\eta {d(\pi F^j (\widehat{p}),\pi F^j (\widehat{q}))}^\eta=
|\phi|_\eta {d(f^L {f^{R(x)}}^J (\widehat x), f^L {f^{R(x)}}^J (\widehat y))}^\eta\\
&\le C_3|\phi|_\eta \frac1{{s({f^{R(x)}}^J (\widehat x), {f^{R(x)}}^J (\widehat y))}^{\alpha\eta}}=C_3|\phi|_\eta \frac1{\big(s(\widehat x,\widehat y)-J\big)^{\theta+1}}\\
&\le C_3|\phi|_\eta \frac1{\big(s(\widehat x,\widehat y)-j\big)^{\theta+1}}\approx C_3|\phi|_\eta \frac1{(2n-j)^{\theta+1}}.\end{aligned}$$ So, we have $$\label{baixo}
\sum_{j=0}^n\big|\phi \pi F^j (\widehat{p})-\phi \pi F^{j} (\widehat{q})\big|\le C_3|\phi|_\eta \sum_{j=0}^n\frac1{(2n-j)^{\theta+1}} \\
\le C''|\phi|_\eta \frac1{n^\theta} \approx 2^\theta C''|\phi|_\eta \frac1{s(p,q)^{\theta}}.$$ The calculations for the second term of the right-hand side of are analogous. From and , we obtain $$|\psi(p)-\psi(q)|\le \frac{D_\psi}{s(p,q)^{\theta}},$$ where $D_\psi$ depends only on $|\phi|_\eta$.
We now present an auxiliary result presented in [@r00 Theorem 2.5].
\[martingales\] Let $\{X_i\}$ be a sequence of $L^2$ random variables with filtration $\mathcal{G}_i$. Let $p \geq 1$ and define $$\label{bin}
b_{i,n}=\max_{i\leq k \leq n}\left\|X_i \sum_{j=i}^k E(X_j|\mathcal{G}_i)\right\|_p.$$ Then $$E|X_1+\cdots+X_n|^{2p}\leq\left(4p\sum_{i=1}^n b_{i,n}\right)^p.$$
Given $\psi:\bar\Delta \rightarrow {\mathbb{R}}$, we define $$\displaystyle \psi_n=\sum_{i=1}^{n-1} \psi {\mbox{\,\small $\circ$}}\bar F^i.$$ In the next proposition we prove that, in the quotient tower, a control on the decay of correlations implies a control on large deviations. This proof is based on [@m09 Theorem 1.2].
\[previous\] Let $\zeta>0$ and $\psi\in \mathcal G_\theta(\bar\Delta)$, for some $\theta>0$. Suppose there exists $C_4>0$ such that, for all $w\in L^\infty (\bar\Delta)$ and all $n\geq n_0$ we have $$\mathcal C_n(w,\psi,\bar\nu)\leq\frac{C_4}{n^\zeta},$$ where $C_4$ depends only on $D_\psi$ and $\|w\|_\infty$. Then, for $\varepsilon>0$ and $p>\max\{1,\zeta\}$, $$LD(\psi,\varepsilon,n,\bar\nu)\leq \frac{C_5}{\varepsilon^{2p}n^\zeta},$$ where $C_5>0$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$.
We may assume, without loss of generality, that $\displaystyle\int\psi\,d\bar\nu=0$. By Markov’s Inequality, we have $$\bar\nu\Big\{\Big|\frac{1}{n}\psi_n\Big|>\varepsilon\Big\}= \bar\nu\Big\{\Big|\frac{1}{n}\psi_n\Big|^{2p}>\varepsilon^{2p}\Big\}
\leq \frac{1}{\varepsilon^{2p}}\int_{\bar\Delta}\Big|\frac{1}{n}\psi_n\Big|^{2p}\,d\bar\nu=\| \psi_n\|_{2p}^{2p}\frac{1}{\varepsilon^{2p}}
\frac{1}{n^{2p}},$$ and so we only need to prove that $$\label{objective}
\| \psi_n\|_{2p}^{2p}\leq C_5\,n^{2p-\zeta},$$ where $C_5>0$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$. By the definition of the Perron-Frobenius operator and the hypothesis, we have for all $w\in L^\infty(\bar\Delta)$ $$\label{pn}
\Big|\int_{\bar\Delta} P^n(\psi)\,w\,d\bar\nu\Big|=
\Big|\int_{\bar\Delta} \psi\,w{\mbox{\,\small $\circ$}}\bar F^n\,d\bar\nu\Big|=\mathcal C_n(w,\psi,\bar\nu) \leq \frac{C_4}{n^\zeta}.$$ Choosing $w=\operatorname*{sgn}P^n(\psi)$ in we get $${\|P^n(\psi)\|}_1=\int_{\bar\Delta} P^n(\psi)\,\operatorname*{sgn}(P^n\psi)\,d\bar\nu\leq \frac{C_4}{n^\zeta}.$$ Note that $C_4$ depends only on $D_\psi$ as $\|\operatorname*{sgn}P^n(\psi)\|_\infty=1$. Since $\|P^n(\psi)\|_\infty \leq \|\psi\|_\infty$ we have $$\label{pkk}
\nonumber\|P^n(\psi)\|_p= \leq {\|P^n(\psi)\|}_\infty^{1-\frac{1}{p}}{\|P^n(\psi)\|}_1^\frac{1}{p}\leq \|\psi\|_\infty^{1-\frac{1}{p}}\frac{(C_4)^\frac{1}{p}}{n^\frac{\zeta}{p}}=\frac{C'}{n^\frac{\zeta}{p}},$$ where $C'$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$. Define $$\label{chik}
\displaystyle \chi_k=\sum_{n=1}^k P^n(\psi) \quad\text{and} \quad\varphi_k=\psi-\chi_k{\mbox{\,\small $\circ$}}F+\chi_k-P^k(\psi).$$ Observe that $\chi_k,\varphi_k\in L^p(\bar\Delta)$, $$\label{chikk}
\|\chi_k\|_p\leq \sum_{n=1}^k \|P^n(\psi)\|_p\leq C' \sum_{n=1}^k \frac{1}{n^\frac{\zeta}{p}}\leq
C' \frac{k^{1-\frac{\zeta}{p}}}{1-\frac{\zeta}{p}}$$ and $$\label{varphikk}
\|\varphi_k\|_p\leq\|\psi\|_p+2\|\chi_k\|_p+\|P^k(\psi)\|_p\leq 4C' \frac{k^{1-\frac{\zeta}{p}}}{1-\frac{\zeta}{p}}$$ for $k$ sufficiently large. Now we are going to prove that $P(\varphi_k)=0$. In fact, given $w\in L^2 (\bar\Delta)$, we have, since $\bar\nu$ is $F$-invariant, $$\begin{aligned}
\int_{\bar\Delta} P(\chi_k) w\, d\bar\nu& -\int_{\bar\Delta} P(\chi_k{\mbox{\,\small $\circ$}}F) w\, d\bar\nu =
\int_{\bar\Delta} \chi_k w{\mbox{\,\small $\circ$}}F\, d\bar\nu-\int_{\bar\Delta} \chi_k{\mbox{\,\small $\circ$}}F w{\mbox{\,\small $\circ$}}F\, d\bar\nu\\
&=\int_{\bar\Delta} \chi_k\, w{\mbox{\,\small $\circ$}}F\, d\bar\nu-\int_{\bar\Delta} \chi_k\, w\, d\bar\nu =\sum_{n=1}^k \int_{\bar\Delta} P^n( \psi)\, w{\mbox{\,\small $\circ$}}F\, d\bar\nu-\sum_{n=1}^k \int_{\bar\Delta} P^n( \psi)\, w\, d\bar\nu\\
&=\sum_{n=1}^k \int_{\bar\Delta} \psi (w{\mbox{\,\small $\circ$}}F^{n+1}-w{\mbox{\,\small $\circ$}}F^n)\, d\bar\nu =\int_{\bar\Delta} \psi(w{\mbox{\,\small $\circ$}}F^{k+1}-w{\mbox{\,\small $\circ$}}F)\, d\nu.\end{aligned}$$ On the other hand, $$\int_{\bar\Delta} P(\psi)w-P^{k+1}(\psi)w\, d\nu=\int_{\bar\Delta} \psi(w{\mbox{\,\small $\circ$}}F-w{\mbox{\,\small $\circ$}}F^{k+1})\, d\nu.$$ So, $P(\varphi_k)=P(\psi)-P(\chi_k{\mbox{\,\small $\circ$}}F)+P(\chi_k)-P^{k+1}(\psi)=0$.
The operator $P$ is the adjoint operator of $U:L^2(\bar\Delta, \bar\nu) \rightarrow L^2(\bar\Delta, \bar\nu)$ defined by $U(v)=v{\mbox{\,\small $\circ$}}F$. Besides, $P{\mbox{\,\small $\circ$}}\,U=I$, where $I$ is the identity operator, and $U{\mbox{\,\small $\circ$}}P=E(\cdot|F^{-1}\mathcal{M})$, where $\mathcal{M}$ is the underlying $\sigma$-algebra. So, $E(\varphi_k|F^{-1}\mathcal{M})=0$ and $E(\varphi_k{\mbox{\,\small $\circ$}}F^j|F^{-(n+1)}\mathcal{M})=0$. Then, $\{\varphi_k{\mbox{\,\small $\circ$}}F^n:n\in {\mathbb{N}}_0\}$ is a sequence of reverse martingale differences. Passing to the natural extension (see [@r00]), $\{\varphi_k{\mbox{\,\small $\circ$}}F^n:n\in {\mathbb{N}}_0\}$ is a sequence of martingale differences with respect to a filtration $\{\mathcal{G}_n:n\in N_0\}$. Defining $X_j=\psi{\mbox{\,\small $\circ$}}F^j$ in Theorem \[martingales\], we have $$b_{i,n}=\max_{i\leq l \leq n} \big\|\psi{\mbox{\,\small $\circ$}}F^i \sum_{j=i}^l E(\psi{\mbox{\,\small $\circ$}}F^j|\mathcal{G}_i)\big\|_p\leq
\|\psi\|_\infty\max_{i\leq l \leq n}\big\|\sum_{j=i}^l E(\psi{\mbox{\,\small $\circ$}}F^j|\mathcal{G}_i)\big\|_p$$ and, by that theorem, we obtain $$\label{psin}
\|\psi_n\|_{2p}^{2p}\leq \big(4p\sum_{i=1}^n b_{i,n}\big)^p.$$ Recalling the definition of $\varphi_k$ in , we have $$\sum_{j=i}^l E(\psi{\mbox{\,\small $\circ$}}F^l|\mathcal{G}_i)=\varphi_k{\mbox{\,\small $\circ$}}F^i+E(\chi_k{\mbox{\,\small $\circ$}}F^{l+1}|\mathcal{G}_i)-E(\chi_k{\mbox{\,\small $\circ$}}F^i|\mathcal{G}_i)+
\sum_{j=i}^l E(P^k(\psi){\mbox{\,\small $\circ$}}F^l|\mathcal{G}_i),$$ and so, using , and , we obtain $$\Big\|\sum_{j=i}^l E(\psi{\mbox{\,\small $\circ$}}F^l|\mathcal{G}_i)\Big\|_p \leq \|\varphi_k\|_p+2\|\chi_k\|_p+n\|P^k(\psi)\|_p\leq
C'\left(6\frac{k^{1-\frac{\zeta}{p}}}{1-\frac{\zeta}{p}}+\frac{n}{k^{\frac{\zeta}{p}}}\right).$$ Then, $$b_{i,n}\leq C''\left(10\frac{k^{1-\frac{\zeta}{p}}}{1-\frac{\zeta}{p}}+\frac{n}{k^{\frac{\zeta}{p}}}\right),$$ where $C''$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$. Then, recalling and choosing $k=n$, we conclude that $$\|\psi_n\|_{2p}^{2p}\leq \big(4p\sum_{i=1}^n b_{i,n}\big)^p\leq C_5 n^{2p-\zeta},$$ where $C_5$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$.
\[Melbourne\] Suppose that $f$ has a GMY structure $\Lambda$ and $\phi:M\rightarrow{\mathbb{R}}$ is a function belonging to $\mathcal{H}_\eta$. Assume that there exist $\psi\in\mathcal G_\theta$, for some $\theta>0$, and $\chi\in L^\infty (\Delta)$ such that $\phi\circ\pi=\psi+\chi-\chi\circ \bar F$ where $\psi$ depends only on future coordinates. Fixing $\zeta>0$, assume that, for all $w\in L^\infty (\bar\Delta)$ and all $n\geq n_0$ there exists $C_4>0$, depending only on $D_\psi$ and $\|w\|_\infty$, such that $$C_n(w,\psi,\bar\nu)\leq\frac{C_4}{n^\zeta}.$$ Then, for $\varepsilon>0$ and $p>\max\{1,\zeta\}$, $$LD(\phi,\varepsilon,n,\mu)\leq\frac{C}{\varepsilon^{2p}}\frac{1}{n^{\zeta}},$$ where $C>0$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$.
We may assume, without loss of generality, that $ \int\phi\,d\mu=0$. By assumption, we can write $\phi\circ\pi=\psi+\chi-\chi\circ \bar F$ where $\psi\in\mathcal G_\theta$, for some $\theta>0$, $\chi\in L^\infty(\Delta)$ and $\psi$ depends only on future coordinates. By Proposition \[previous\] we have $$\label{nupsi}
\bar\nu \Big\{\Big|\frac{1}{n}\psi_n \Big|>\varepsilon \Big\}\leq \frac{C_5}{\varepsilon^{2p}n^{\zeta}},$$ where $C_5>0$ depends only on $p, D_\psi$ and $\|\psi\|_\infty$. Note that $$\label{munu}
\mu\Big\{\Big|\frac{1}{n}\phi_n(x)\Big|>\varepsilon\Big\}=\nu\Big\{\Big|\frac{1}{n}\phi_n(\pi y)\Big|>\varepsilon\Big\}$$ and $$\phi_n{\mbox{\,\small $\circ$}}\pi= \sum_{k=0}^{n-1} \phi{\mbox{\,\small $\circ$}}f^k{\mbox{\,\small $\circ$}}\pi
=\sum_{k=0}^{n-1} \psi{\mbox{\,\small $\circ$}}F^k+\sum_{k=0}^{n-1} \chi{\mbox{\,\small $\circ$}}F^k-\sum_{k=0}^{n-1} \chi{\mbox{\,\small $\circ$}}F^{k+1}=\psi_n+\chi-\chi{\mbox{\,\small $\circ$}}F^n.$$ Let $y\in\Delta$ be such that $\tfrac{1}{n}\big|\psi_n(y)+\chi(y)-\chi(F^n y)\big|>\varepsilon$. Then $\tfrac{1}{n}\big|\psi_n(y)\big|+\tfrac{2}{n}\|\chi\|_\infty>\varepsilon$ and so $$\Big\{\frac{1}{n}\big|\phi_n{\mbox{\,\small $\circ$}}\pi\big|>\varepsilon\Big\}
\subseteq \Big\{\frac{1}{n}\big|\psi_n\big|>\varepsilon-\frac{2}{n}\|\chi\|_\infty\Big\}.$$ From , and the last inclusion we get, for a sufficiently large $n_0$ and $n\geq n_0$, $$\label{constants}
\mu\Big\{\Big|\frac{1}{n}\phi_n(x)\Big|>\varepsilon\Big\}\leq\bar\nu\Big\{\frac{1}{n}\big|\psi_n\big|>\varepsilon-\frac{2}{n}\|\chi\|_\infty\Big\}
\leq \frac{C}{\varepsilon^{2p}} \frac{1}{n^{\zeta}},$$ where $C>0$ depends only on $p, D_\psi$ and $\|\psi\|_\infty$.
Note that to obtain the conclusion we only need to verify the assumptions of Proposition \[Melbourne\]. Taking $\eta_0=1/\alpha$, under the hypothesis of Theorem \[TheoremB\] we can use Proposition \[metric\] to conclude that there exist $\psi\in\mathcal G_\theta(\Delta)$, for $\theta=\alpha\eta-1$, and $\chi\in L^\infty (\Delta)$ such that $\phi\circ\pi=\psi+\chi-\chi\circ \bar F$, where $\psi$ depends only on future coordinates and $D_\psi$ depends only on $|\phi|_\eta$. So, we may apply Corollary \[Corollary\], obtaining for all $w\in L^\infty(\bar\Delta)$ $$C_n(w,\psi,\nu)\leq \frac{C_4}{n^{\zeta-1}},$$ where $C_4$ depends only on $D_\psi$ and $\|w\|_\infty$. Consequently, using Proposition \[Melbourne\], $$LD(\phi,\varepsilon,n,\mu)\leq\frac{C}{\varepsilon^{2p}}\frac{1}{n^{\zeta-1}}$$ where $C>0$ depends only on $p$, $D_\psi$ and $\|\psi\|_\infty$ and so, only on $p$, $|\phi|_\eta$ and $\|\psi\|_\infty$. As $\|\psi\|_\infty\leq \|\phi\|_\infty+2\|\chi\|_\infty$ and, by Proposition \[metric\], $\|\chi\|_\infty$ depends only on $D_\phi$, we conclude that $C$ depends only on $p$, $|\phi|_\eta$ and $D_\phi$.
The example {#theae}
===========
Let $f$ be the diffeomorphism of the two-dimensional torus $\mathbb{T}^2={\mathbb{R}}^2/\mathbb{Z}^2$ introduced in Subsection \[ae\]. As described in Subsection \[ae\], $f$ coincides with an Anosov diffeomorphism $f_0$ in all rectangles $W_1,\dots,W_d$ of a Markov partition of $f_0$ but one, $W_0$. Recall that we have taken $W_0=[a_0',a_0]\times [b_0',b_0]$ a neighborhood of $(0,0)$ and $f_0(a,b)=\big(\phi_0(a),\psi_0(b)\big)$ for all $(a,b)\in W_0$. Moreover, $f(a,b)=\big(\phi(a),\psi(b)\big)$ for each $(a,b)\in W_0$, where $$\phi(a)=a(1+a^\theta)\quad\text{and}\quad\psi(b)=\phi^{-1}(b),\quad\forall(a,b)\in V_0,$$ for some $0<\theta<1$ and $V_0$ the neighborhood $[\phi^{-1}_0(a'_0),\phi^{-1}_0(a_0)]\times [\psi_0(b'_0),\psi_0(b_0)]$ of $(0,0)$ contained in $W_0$ .
Observe that as we have not modified the geometric structure of $f_0$, then the set $W_1$ is completely foliated by a set $\Gamma^s$ of stable leaves and a set $\Gamma^u$ of unstable leaves. To obtain the conclusions of Theorems \[th.ex1\] and \[th.ex2\] we shall prove that $f$ satisfies the properties (P$_{1}$)-(P$_{5}$) on the set $\Lambda=W_1$ (any other $W_i\neq W_0$ would be fine) and that we have recurrence times with polynomial decay to some unstable leaf on $W_1$, thus being in the conditions of Theorems \[TheoremA\] and \[TheoremB\].
We consider the sequences $(a_n)_n$ and $(a_n')_n$ defined recursively for $n\ge 1$ as $$a_n=\phi^{-1}(a_{n-1})\quad\text{and}\quad a_n'=\phi^{-1}(a_{n-1}').$$ For all n$\geq 0$, set $$J_n=[a_{n+1},a_n]\times [b_0',b_0]\quad\text{and}\quad J_n'=[a_n',a_{n+1}'] \times [b_0',b_0] .$$ Observe that these sets form a (lebesgue mod 0) partition of $W_0$. Setting for $i=1,\dots, k$ and $n\ge 0$ $$\widehat R|_{W_i}=1,\quad \widehat R|_{J_n}=n+1\quad\text{and}\quad \widehat R|_{J'_n}=n+1,$$ define $${\widehat R}_1={\widehat R}-1+n_0, \quad
{\widehat R}_i={\widehat R}_{i-1}+({\widehat R}-1){\mbox{\,\small $\circ$}}f^{{\widehat R}_{i-1}}+n_0 \quad \text{for } i \geq 2$$ and, for $x\in W_1$, let $R(x)$ be equal to the smallest $\widehat R_i$ such that $f^{{\widehat R}_i}(x)\in W_1$. Note that as we are assuming the transition matrix of $f_0$ (and thus of $f$) with respect to the partition $W_0,\dots,W_k$ to be aperiodic, then $R$ is well defined.
Invariant manifolds
-------------------
Here we prove that the manifolds in $\Gamma^s$ and $\Gamma^u$ satisfy (P$_{2}$) and (P$_{3}$). We start by proving some useful estimates about the map $\phi$. It follows from the results in the beginning of [@y99 Section 6.2] that $(a_n)_n$ and $(a_n')_n$ have the same asymptotics of the sequence $1/n^{1/\theta}$. In particular, there is $C>0$ such that for all $n\ge1$ we have $$\label{deltaan}
\Delta a_n:=a_n-a_{n+1}\le \frac{C}{n^{1+1/\theta}},$$ and a similar estimate holds for $(a_n')_n$. For the sake of notational simplicity we shall consider $\tau=1/\theta$.
\[phitau\] There exists $C>0$, such that for all $n\geq 0$ and all $x\in [a_{n+1},a_n]$, we have $$|(\phi^n)'(x)|\geq Cn^{\tau+1}.$$
By the definition of $a_n$, we have $$|\phi^n(a_n)-\phi^n(a_{n+1})|=|a_0-a_1|$$ and so, using the Mean Value Theorem and , we get, for some $\xi\in [a_{n+1},a_n]$, $$|(\phi_u^n)'(\xi_k)|=\frac{\Delta a_0}{\Delta a_n} \geq C n^{\tau+1}.$$ Using the previous lemma for $a=\xi$ and any $b\in [a_{n+1},a_n]$, we obtain the same conclusion for any point in $[a_{n+1},a_n]$, concluding the proof.
To simplify notation, we write $f_u'$ to mean the derivative of $f$ in the unstable direction.
\[contractex\] There exists $C>0$ such that
- for all $n\in{\mathbb{N}}$ and $x,y\in\gamma^u\in \Gamma^u$ we have $$d(f^{-n}(x),f^{-n}(y))\leq \frac{C}{n^{\tau+1}}d(x,y);$$
- for all $n\in{\mathbb{N}}$ and $x,y\in\gamma^s\in \Gamma^s$ we have $$d(f^{n}(x),f^{n}(y))\leq \frac{C}{n^{\tau+1}}d(x,y).$$
We shall prove (a). The proof of (b) follows similar arguments.
Consider $x,y\in\gamma^u$ and let $n\in {\mathbb{N}}$. We first assume that the orbits of $x$ and $y$ visit $W_0$ exactly at the same moments. Then, it is enough to prove that there is $C>0$ such that for a given point $z\in T^2$ we have $$|(f_u^{-n})'(z)|\le \frac{C}{n^{\tau +1}}.$$ Let $K=\{j\in{\mathbb{N}},\,0\le j\le n:\ f^{-j} (x)\in W_0\}$ and $M=\{0,\ldots,n\}\setminus K$. The set $K$ can be written as $K=\cup_{i=1}^k K_i$, where $$K_i=\{-k_i,\ldots, -k_i+p_i\,|\, -k_i-1\not\in K, \quad -k_i+p_i+1\not\in K\}.$$ Analogously, we write $M=\cup_{i=1}^m M_i$, where $$M_i=\{-m_i,\ldots, -m_i+q_i\,|\, -m_i-1\not\in M, \quad -m_i+q_i+1\not\in M\}.$$ Considering $P=\sum_{i=1}^k p_i$ and $Q=\sum_{i=1}^k q_i$, we have $P+Q=n$.
Note that, since $k_i+p_i\in K$ and $k_i+p_i+1\not\in K$ , then $f^{k_i+p_i}(x) \in J_0$. Since we assumed that the orbits of $x$ and $y$ visit $W_0$ exactly at the same moments, then $f^{k_i+p_i}(x) \in J_0$. Observe that $f$ coincides with $\phi$ in $K\cap \gamma^u$. Using the Mean Value Theorem and Lemma \[phitau\] we get, for some $\xi\in J_0$, $$\label{insideW0}
d(f^{-k_i}(x),f^{-k_i}(y))\leq (\phi^{-p_i})'(\xi) d(f^{-k_i+p_i}(x),f^{-k_i+p_i}(y)) \leq \frac{C}{p_i^{\tau+1}} d(f^{-k_i+p_i}(x),f^{-k_i+p_i}(y)).$$
For the iterates $m\in M$ we have $f^{-m}(x)\notin W_0$, and so the behavior of $(f_u)'$ is the same of the unperturbed Anosov case. In particular, there is exponential backward contraction: there is $\lambda>1$ such that $$\label{outsideW0}d(f^{-m}(x),f^{-m}(y))\leq \lambda d(f^{-(m-1)}(x),f^{-(m-1)}(y)).$$ Gathering and , we obtain, for any $n\in{\mathbb{N}}$, $$d(f^{-n}(x),f^{-n}(y))\leq \lambda^{Q} \prod_{i=1}^k \frac{C}{p_i^{\tau+1}}d(x,y).$$ Now it is enough to prove that $$\label{eq.CP}
\prod_{i=1}^k \frac{C}{p_i^{\tau+1}}\le \frac{C}{P^{\tau+1}}.$$ We have for each $i$ $$\label{eq.CP2}
\frac{C}{p_i^{\tau+1}}=\left(\frac{p_i}{C^{\frac{1}{\tau+1}}}\right)^{-\tau-1}.$$ With no loss of generality, we may assume that each ${p_i}/C^{\frac1{\tau+1}}\ge 2$. Actually, if this were not the case we would have the $p_i$’s uniformly bounded, meaning that the corresponding $p_i$ iterates would be uniformly bounded away from the stable leaf of $(0,0)$. In particular, there would be some $0<\lambda_0<1$ such that $|(f_u^{-1})'|\le \lambda_0$ and this case could be treated as the case of the previous case with $\lambda_0$ playing the role of $\lambda$.
Let us now prove under the assumption that ${p_i}/C^{\frac1{\tau+1}}\ge 2$ for each $1\le i\le k$. This in particular implies that $$\prod_{i=1}^k \frac{p_i}{C^{\frac{1}{\tau+1}}} \ge \sum_{i=1}^k \frac{p_i}{C^{\frac{1}{\tau+1}}}.$$ Using this we get $$\label{eq.CP2}
\prod_{i=1}^k \frac{C}{p_i^{\tau+1}}= \left(\prod_{i=1}^k \frac{C^{\frac{1}{\tau+1}}}{p_i}\right)^{\tau+1}\le \left(\sum_{i=1}^k \frac{C^{\frac{1}{\tau+1}}}{p_i}\right)^{\tau+1}=\left( \frac{C^{\frac{1}{\tau+1}}}{P}\right)^{\tau+1}=\frac{C}{P^{\tau+1}},$$ thus proving .
Let us finally consider the case where the orbits of $x$ and $y$ do not visit $W_0$ at the same moments. Assume that there is $j\le n$ such that $f^j(x)\in J\cup J'$ and $f^j(y)\notin J\cup J'$. Choosing the size of the rectangle $W_1$ sufficiently small (and thus the length of $\gamma^u(x)$), we may assure that we necessarily have $f^j(x)$ (uniformly) bounded away from $\gamma^s(0,0)$. In particular, there is some $\lambda_0$ such that $|f_u'|\ge \lambda_0$, and so we may repeat the calculations above we $\lambda_0$ playing the role of $\lambda$.
Bounded distortion
------------------
Here we prove the bounded distortion property (P$_{4}$).
The following lemma is proved in [@y99 Lemma 5].
\[yd\] There exists $C>0$ such that, for all $i,n\in{\mathbb{N}}$ with $i\leq n$, and for all $a,b\in [a_{n+1},a_n]$, $$\log \frac{(\phi^i)'(a)}{(\phi^i)'(b)}\leq
C\frac{\left|\phi^i(a)-\phi^i(b)\right|}{\Delta a_{n-i}}\le C.$$
\[beta\] There exists $C>0$ and $0<\beta<1$ such that for all $x,y\in \gamma^u\in\Gamma^u$ we have $ d(x,y)\leq C\beta^{s(x,y)}.$
We will start by showing that there exists $0<\beta<1$ such that, for $x,y\in \Lambda\cap\gamma^u$ with $s(x,y)\neq 0$, we have $d(x,y)\leq \beta\, d(f^R(x),f^R(y))$. In fact, since $f^R(x),f^R(y)\not\in W_0$ and $f$ behaves like an Anosov diffeomorphism outside $W_0$, then $$d(f^{R-1}(x),f^{R-1}(y))\leq \beta\, d(f^R(x),f^R(y))\quad\mbox{for some } 0<\beta<1$$ and so $d(x,y)\leq \beta\,d(f^R(x),f^R(y)) $.
Applying this inequality successively, we obtain $$d(x,y)\leq \beta\,d(f^R(x),f^R(y))\leq\cdots\leq \beta^s\, d(\left(f^R\right)^s(x),\left(f^R\right)^s(y))\leq C\beta^{s(x,y)},$$ where $C$ is the diameter of $M$.
\[prop:bound:dist\] For $\gamma \in \Gamma^u$ and $x,y\in \Lambda \cap \gamma$, $$\left|\log\frac{\big(f_u^R\big)'(x)}{\big(f_u^R\big)'(y)}\right|\leq C\beta^{s(f^R(x),f^R(y))}.$$
Let $\gamma \in \Gamma^u$ and $x,y\in \Lambda \cap \gamma$. We have $$\label{dl}
\left|\log\frac{(f_u^R)'(x)}{(f_u^R))'(y)}\right|\leq \sum_{j=0}^{R-1}|\log f_u'(f^j x)-\log f_u'(f^j y)|.$$ As in Proposition \[contractex\], without loss of generality we may assume that the orbits of $x$ and $y$ visit $W_0$ exactly at the same moments. Let $K=\{j\in{\mathbb{N}}_0,\,0\le j\le R-1:\ f^j (x)\in W_0\}$ and $M=\{0,\ldots,R\}\setminus K$. The set $K$ can be written as $K=\cup_{i=1}^k K_i$, where $$K_i=\{k_i,\ldots, k_i+p_i\,|\, k_i-1\not\in K, \quad k_i+p_i+1\not\in K\}.$$ Analogously we can write $M=\cup_{i=1}^m M_i$, where $$M_i=\{m_i,\ldots, m_i+q_i\,|\, m_i-1\not\in M, \quad m_i+q_i+1\not\in M\}.$$ We will consider now the terms of the right hand side of which belong to some $K_i$. Note that, since $k_i+p_i\in K$ and $k_i+p_i+1\not\in K$, then $f^{k_i+p_i}\in J_0$. From Lemma \[yd\], $$\begin{aligned}
\sum_{j=0}^{p_i-1}|\log f_u'(f^{k_i+j}x)-\log f_u'(f^{k_i+j}y)|&\leq
C_1 d(f^{k_i+p_i}x,f^{k_i+p_i}y).\end{aligned}$$ So, adding the term $j=p_i$, we obtain $$\sum_{j=0}^{p_i}|\log f_u'(f^{k_i+j}x)-\log f_u'(f^{k_i+j}y)|\leq C d(f^{k_i+p_i}x,f^{k_i+p_i}y),$$ because there exists $\xi\in J_0$ such that $$|\log f_u'(f^{k_i+p_i}x)-\log f_u'(f^{k_i+p_i}y)|= \left|\frac{f_u''(\xi)}{f_u'(\xi)}\right| d(f^{k_i+p_i}x,f^{k_i+p_i}y).$$ Let us now consider the terms belonging to some $M_i$. Since $f$ is of class $C^2$ outside $W_0$, using the Mean Value Theorem and the fact that $f$ is uniformly expanding on unstable leaves, we have, for $x\in M_i$, $$\begin{gathered}
\sum_{j=0}^{q_i}|\log f_u'(f^{m_i+j}x)-\log f_u'(f^{m_i+j}y)|\leq C_3\sum_{j=0}^{q_i}d(f^{m_i+j}x,f^{m_i+j}y)\\
\leq C_3\sum_{j=0}^{q_i}\beta^{q_i-j+1}d(f^{m_i+q_i}x,f^{m_i+q_i}y)\leq C d(f^{m_i+q_i}x,f^{m_i+q_i}y).\end{gathered}$$ Gathering the conclusions we obtained for $K$ and $M$, using Proposition \[contractex\]-(a), $$\begin{aligned}
\sum_{p\in K\cup L\cup M}|\log f_u'(f^px)-\log f_u'(f^py)|\leq & C\Big(\sum_{i=0}^kd(f^{k_i+p_i}x,f^{k_i+p_i}y)
+\sum_{i=0}^md(f^{m_i+q_i}x,f^{m_i+q_i}y)\Big)\\
\leq &C \Big(\sum_{i=0}^k\tfrac1{(R-k_i-p_i)^{\tau+1}} +
\sum_{i=0}^m \tfrac1{(R-m_i-q_i)^{\tau+1}} \Big)d(f^Rx,f^Ry)\\
\leq &C'd((f^R)x,(f^R)y),\end{aligned}$$ and so, $$\label{distort}
\left|\log\frac{\big(f_u^R\big)'(x)}{\big(f_u^R\big)'(y)}\right|\leq C'd((f^R)x,(f^R)y).$$ Applying Lemma \[beta\], we have $d((f^R)x,(f^R)y)\leq C_2 \beta^{s(f^R(x),f^R(y))}$ for some $C_2>0$, thus concluding the proof.
Regularity of the stable foliation
----------------------------------
To prove property [(P$_{5}$)]{}-(a), we follow the ideas in [@ap10 Section 3.5]. The proof of the following lemma can be found in [@m87 Theorem 3.3.].
\[unif\] Let $N$ and $P$ be manifolds, where $P$ has finite volume, and, for every $n\in {\mathbb{N}}$, let $\Theta_n:N\rightarrow P$ be an absolutely continuous map with Jacobian $J_n$. If we assume that
- $\Theta_n$ converges uniformly to an injective continous map $\Theta:N\rightarrow P$,
- $J_n$ converges uniformly to an integrable continous map $J:N\rightarrow {\mathbb{R}}$,
then $\Theta$ is absolutely continuous with Jacobian $J$.
Until the end of this section we denote $\Theta=\Theta_{\gamma',\gamma}(x)$ to simplify the notation. The next lemma can be found in [@ap10 Lemma 3.11] and it is a consequence of [@m87 Lemma 3.8].
\[Gn\] Given $\gamma,\gamma' \in\Gamma^u$ and $\Theta:\gamma'\rightarrow\gamma$, then, for every $n\in{\mathbb{N}}$, there exists an absolutely continous function $\pi_n:f^n(\gamma')\rightarrow f^n(\gamma)$ with Jacobian $G_n$ such that
- $\displaystyle \lim_{n\rightarrow\infty} \sup_{x\in\gamma}\, \big\{d_{f^n(\gamma')} \big(f^n(x),f^n(\Theta(x))\big)\big\}=0$;
- $\displaystyle \lim_{n\rightarrow\infty} \sup_{x\in f^n(\gamma)} \big\{\big|1-G_n(x)\big|\big\}=0$.
\[prod\] There exists $C>0$ such that for all $x,y\in \gamma^s\in \Gamma^s$ and $n\in{\mathbb{N}}$ we have $$\log\prod_{i=n}^\infty \frac{\det\, Df(f^i(x))}{\det\, Df(f^i(y))}\leq \frac{C}{n^{\tau}}.$$
Note that $$\log\prod_{i=n}^\infty \frac{\det\, Df(f^i(x))}{\det\, Df(f^i(y))}\leq\sum_{i=n}^\infty \big|\log\big(\det\, Df(f^i(x))\big)-\log\big(\det\, Df(f^i(y))\big)\big|.$$ We now need to control each term of the above sum. We divide this in three cases.
Assume first that $f^i(x),f^i(y)\in W_0$. Since $f^i(y)\in \gamma^s(f^i(x))$ and $f$ has a product form in $W_0$, then $\big|\log\big(\det\, Df(f^i(x))\big)-\log\big(\det\, Df(f^i(y))\big)\big|=0$.
Assume now that $f^i(x),f^i(y) \not\in W_0$. As $f$ behaves like an Anosov diffeomorphism outside $W_0$, then $\log\det Df$ is Lipschitz. So, using the polynomial contraction on stable leaves, we get $$\big|\log\big(\det\, Df(f^i(x))\big)-\log\big(\det\, Df(f^i(y))\big)\big|\leq C_1 d(f^i(x),f^i(y))\leq \frac{C_2}{i^{\tau+1}}.$$
Finally, for $f^i(x)\in W_0$ and $f^i(y)\not\in W_0$, choose the point $z$ in the same stable leaf as $f^i(x)$ such that $z$ is in the boundary of $W_0$ and between $f^i(x)$ and $f^i(y)$. Then, applying the first case to $f^i(x)$ and $z$, and the second case to $z$ and $f^i(y)$, we obtain the conclusion.
Adding all the terms, we conclude that $$\sum_{i=n}^\infty \big|\log\big(\det\, Df(f^i(x))\big)-\log\big(\det\, Df(f^i(y))\big)\big|\leq C_3 \sum_{i=n}^\infty \frac{1}{i^{\tau+1}}
\leq \frac{C}{n^{\tau}}.$$
We define, for $n\in{\mathbb{N}}$, the map $\Theta_n:\gamma'\rightarrow \gamma$ as $\Theta_n=f^{-R_n}{\mbox{\,\small $\circ$}}\pi_{R_n}{\mbox{\,\small $\circ$}}f^{R_n}$. Note that $\Theta_n$ is absolutely continuous, its Jacobian is $$J_n(x)=\frac{|\det(Df^{R_n})(x)|}{|\det(Df^{R_n}) \Theta_n(x)|}G_{R_n}(f^{R_n}(x))$$ and the Jacobian of $\Theta$ is given by $$J(x)=\frac{d(\Theta_* {\operatorname{Leb}}_\gamma)}{d {\operatorname{Leb}}_{\gamma'}}.$$
For $\gamma', \gamma\in\Gamma^u$, the function $\Theta$ is absolutely continuous and its Jacobian is given by $$J(x)=\prod_{i=0}^\infty \frac{\det\, Df(f^i(x))}{\det\, Df(f^i(\Theta(x)))}.$$
Note that Lemma \[prod\] implies that the product in the above proposition is finite. The proof of this proposition is a direct consequence of the following lemma together with Lemma \[unif\].
The functions $\Theta_n$ converge uniformly to $\Theta$ and their Jacobians $J_n$ converge uniformly to $J$.
Using [(P$_{3}$)]{}, we have, for $x\in\gamma$, $$d_\gamma (\Theta_n(x), \Theta(x)) =d_\gamma(f^{-R_n} \pi_{R_n}f^{R_n}(x), f^{-R_n} f^{R_n}\Theta(x)) \leq \tfrac{C}{(R_n)^{\tau+1}} d_{f^{R_n}(\gamma)} (\pi_{R_n}f^{R_n}(x), f^{R_n}\Theta(x))$$ and, since $R_n\underset{n}{\rightarrow}\infty$ and $d_{f^{R_n}(\gamma)} (\pi_{R_n}f^{R_n}(x), f^{R_n}\Theta(x))$ is bounded, then the uniform convergence follows.
We write $$J_n(x)=\frac{|\det(Df^{R_n})(x)|}{|\det(Df^{R_n}) \Theta(x)|}\ \frac{|\det(Df^{R_n}) \Theta(x)|}{|\det(Df^{R_n}) \Theta_n(x)|}
\ G_{R_n}(f^{R_n}(x)),$$ By Lemma \[Gn\], $G_{R_n}(f^{R_n}(x))$ converges uniformly to one. To control the second factor note that, by applied to the point $\Theta(x)$ and $\Theta_n(x)$, we have $$\left|\log \frac{|\det(Df^{R_n}) \Theta(x)|}{|\det(Df^{R_n}) \Theta_n(x)|}\right| \leq Cd_{f^{R_n}(\gamma')}(f^{R_n}(\Theta(x),f^{R_n}(\Theta_n(x)).$$ So, $$\frac{|\det(Df^{R_n}) \Theta(x)|}{|\det(Df^{R_n}) \Theta_n(x)|}\underset{n}{\rightarrow} 1.$$ We are left to prove that the first factor converges uniformly to $J$. Notice that $$\log\frac{|\det(Df^{R_n})(x)|}{|\det(Df^{R_n}) \Theta(x)|}=\sum_{i=0}^{R_n} \log\frac{\det\, Df(f^i(x))}{\det\, Df(f^i(\Theta(x)))}$$ and so $$\log J(x)-\log\frac{|\det(Df^{R_n})(x)|}{|\det(Df^{R_n}) \Theta(x)|}=\sum_{R_n+1}^{\infty} \log\frac{\det\, Df(f^i(x))}{\det\, Df(f^i(\Theta(x)))},$$ which converges uniformly to zero, by Lemma \[prod\].
The next proposition proves [(P$_{5}$)]{}-(b).
For each $\gamma, \gamma' \in \Gamma^u$, the map $\Theta$ is absolutely continuous and denoting $$u(x)=\frac{d(\Theta_* {\operatorname{Leb}}_{\gamma'})}{d{\operatorname{Leb}}_{\gamma}}(x),$$ we have $$\log \frac{u(x)}{u(y)}\leq C\beta^{s(x,y)},\quad \forall\, x,y\in \gamma'\cap \Lambda.$$
It is known that [(P$_{5}$)]{}-(b) is satisfied by Anosov diffeomorphisms. But $f$ is topologically conjugate to the Anosov diffeomorphism $f_0$. Since the separation time is invariant by topological conjugacy, then so is [(P$_{5}$)]{}-(b).
Recurrence times
----------------
Our goal in this subsection is to prove that there exists $C>0$ such that for all $\gamma\in \Gamma^u$ and $n\in{\mathbb{N}}$ we have $${\operatorname{Leb}}_\gamma\{R>n\}\leq \frac{C}{n^{\tau+1}}.$$ Since we have assumed the transition matrix of the initial Markov partition aperiodic, then there is $n_0\in{\mathbb{N}}$ such that $f^n(W_j)$ intersects $W_k$, for all $j,k$ and all $n\geq n_0$.
\[n0delta0\] For $L\in\{W_1,\ldots,W_d,J_0,J_0'\}$, there exists $n_0\in{\mathbb{N}}$ and $\delta_0>0$ such that, for all $n\geq n_0$ and $j\in\{1,\ldots,d\}$, we have $${\operatorname{Leb}}_\gamma\big(f^{-n}(W_j)\cap L\big)\geq \delta_0.$$
Choosing $n_0$ as in above, we know that, for all $n\geq n_0$, we have $f^n(L)$ intersects $W_k$, for all $k$. Since, in addition, $f^n(L)$ must cross the entire length of the unstable direction of any $W_k$ it intersects, then $f^n(L)$ crosses the entire length of the unstable direction of every $W_k$. Then $$\label{int}
\frac{{\operatorname{Leb}}_\gamma(f^{-n}(W_j)\cap L)}{{\operatorname{Leb}}_\gamma(L)}=\frac{\displaystyle \int_{f^n(f^{-n}(W_j)\cap L)} (f_u^{-n})'d{\operatorname{Leb}}_\gamma}{\displaystyle \int_{f^n(L)} (f_u^{-n})'d{\operatorname{Leb}}_\gamma}=\frac{\displaystyle\int_{W_j} (f_u^{-n})'d{\operatorname{Leb}}_\gamma}{\displaystyle\int_{\cup W_k} (f_u^{-n})'d{\operatorname{Leb}}_\gamma}.$$ Let $R_0=0$. Choosing $k\in{\mathbb{N}}_0$ such that $R_k\leq n<R_{k+1}$, note that $(f_u^{n-R_k})'(x)\geq 1$ and so, $$\big(f_u^n\big)'(x)= (f_u^{R_k})'(f^{n-R_k}(x))(f_u^{n-R_k})'(x)\geq (f_u^{R_k})'(f^{n-R_k}(x)).$$ Analogously, since $(f_u^{n-R_{k+1}})'(x)\leq 1$, then $$\big(f_u^n\big)'(x)=(f_u^{R_{k+1}})'(f^{n-R_{k+1}}(x))(f_u^{n-R_{k+1}})'(x)\leq (f_u^{R_{k+1}})'(f^{n-R_{k+1}}(x)).$$ Consequently, $$\label{1sobre}
\frac{1}{(f_u^{R_{k+1}})'(f^{n-R_{k+1}}(y))}\leq \big(f_u^{-n}\big)'(y) \leq \frac{1}{(f_u^{R_k})'(f^{n-R_k}(y))}.$$ Applying (P$_{4}$), there exists $C>0$ such that, for all $m\in{\mathbb{N}}$, $$e^{1/C}\leq \frac{\big(f_u^{R_m}\big)'(z)}{\big(f_u^{R_m}\big)'(w)}\leq e^C.$$ Fixing $w_0$ we have $$(f_u^{R_k})'(f^{n-R_k}(y))\geq e^{1/C} \big(f_u^{R_k}\big)'(w_0) \quad \text{and} \quad
(f_u^{R_{k+1}})'(f^{n-R_{k+1}}(y))\leq e^C \big(f_u^{R_{k+1}}\big)'(w_0).$$ Then, from , and the previous inequalities, we obtain $$\begin{aligned}
\frac{{\operatorname{Leb}}_\gamma(f^{-n}(W_j)\cap L)}{{\operatorname{Leb}}_\gamma(L)}& \geq\frac{\displaystyle\int_{W_j} \frac{1}{(f_u^{R_{k+1}})'(f^{n-R_{k+1}})}d{\operatorname{Leb}}_\gamma}{\displaystyle\int_{\cup W_k}
\frac{1}{(f_u^{R_k})'(f^{n-R_k})}d{\operatorname{Leb}}_\gamma}\geq \frac{\displaystyle\frac{1}{e^C \big(f_u^{R_{k+1}}\big)'(w_0)}{\operatorname{Leb}}_\gamma(W_j)}{\displaystyle\frac{1}{e^{1/C} \big(f_u^{R_k}\big)'(w_0)}{\operatorname{Leb}}_\gamma(\cup W_k)}\\
&=\frac{e^{1/C}}{e^C\big(f_u^R\big)'(w_0)}\frac{\big(f_u^{R_k}\big)'(w_0)}{\big(f_u^{R_k}\big)'(f^R(w_0))}\frac{{\operatorname{Leb}}_\gamma(W_j)}{{\operatorname{Leb}}_\gamma(\cup W_k)}\geq \frac{e^{2/C}}{e^C\big(f_u^R\big)'(w_0)}\frac{{\operatorname{Leb}}_\gamma(W_j)}{{\operatorname{Leb}}_\gamma(\cup W_k)},\end{aligned}$$ using (P$_{4}$) in the last step. Finaly, $${\operatorname{Leb}}_\gamma(f^{-n}(W_j)\cap L)\geq \frac{e^{2/C}{\operatorname{Leb}}_\gamma(L)}{e^C\big(f_u^R\big)'(w_0){\operatorname{Leb}}_\gamma(\cup W_k)}\ \min_{j=1,\ldots,d}\{{\operatorname{Leb}}_\gamma(W_j)\}=\delta_0.$$
Define the $\sigma$-algebra $${\mathcal B}_i=\bigvee_{j=0}^{\widehat R_{i-1}}f^{-j}{\mathcal A}.$$
\[1\] There exists ${\varepsilon}_0>0$ such that, for all $i\in{\mathbb{N}}$ and all $\omega\in{\mathcal B}_i$ with $R_{|\omega}> \widehat R_{i-1}$, $${\operatorname{Leb}}_\gamma\big\{R=\widehat R_i\,|\,\omega\big\}\geq{\varepsilon}_0.$$
Fix $i\in {\mathbb{N}}$ and let $\omega\in{\mathcal B}_i$ be such that $R_{|\omega}> \widehat R_{i-1}$. It follows from the definition of ${\mathcal B}_i$ that $f^{\widehat R_{i-1}}\omega\in{\mathcal A}$. Set $n=\widehat R_{i-1}+(\widehat R-1){\mbox{\,\small $\circ$}}f^{\widehat R_{i-1}}$. If $f^{\widehat R_{i-1}}\omega=W_l$, for some $l\neq 0$, since $(\widehat R-1)W_l=W_l$, then $f^nW_l=W_l$. If $f^{\widehat R_{i-1}}\omega=J_l$, for some $l\in{\mathbb{N}}_0$, since $(\widehat R-1)J_l=J_0$, then $f^nJ_l=J_0$ (analogously, $f^nJ'_l=J'_0$). So, we proved that $f^n\omega=L\in\big\{W_1,\ldots,W_d, J_0,J'_0\big\}$.
Calling $A=L\cap f^{-n_0}W_k$ and noting that $\widehat R_i(x)=n+n_0$, we have $$\begin{aligned}
B=\big\{x\in\omega:R(x)=\widehat R_i(x)\big\}=&\{x\in\omega:f^{n+n_0}(x)\in W_k\big\}\\
=&\{x\in f^{-n}L:x\in f^{-(n+n_0)}W_k\big\}=f^{-n}(A).\end{aligned}$$ From Lemma \[n0delta0\], we know that ${\operatorname{Leb}}_\gamma(A)\geq\delta_0>0$. We are left to prove that ${\operatorname{Leb}}_\gamma(f^{-n}(A))\geq{\varepsilon}_0$. But, if we prove that $\big(f_u^{-n}\big)'_{|_A}\geq \delta_1>0$, then we get $${\operatorname{Leb}}_\gamma(f^{-n}(A))=\int_A\big(f_u^{-n}\big)'d{\operatorname{Leb}}_\gamma\geq \delta_1\,\delta_0={\varepsilon}_0.$$
To prove that $\big(f_u^{-n}\big)'_{|_A}\geq \delta_1>0$, we only need to find an upper bound for $(f_u^{n})'$ in $B$. If $z\in A$ then $z=f^n(x)$, for some $x\in B$ and $R(x)=\widehat R_i(x)=n+n_0$. So, $$(f_u^{n})'(x)=\big(f^{-n_0}{\mbox{\,\small $\circ$}}f_u^{R(x)}\big)'(x)=\big(f_u^{-n_0}\big)'(f^R(x))\,\big(f_u^R\big)'(x).$$ Since $n_0$ is fixed and $\big(f_u^{-n_0}\big)'$ is a continuous function with a compact domain, then $\big(f_u^{-n_0}\big)'$ has an upper bound. So, we only need to control $\big(f_u^R\big)'$ in $B$. Using , there exists a constant $C>0$ such that, for $x,y\in L$, $$\left|\log\frac{\big(f_u^R\big)'(x)}{\big(f_u^R\big)'(y)}\right|\leq Cd(f^R(x),f^R(y))$$ and so $$\left|\frac{\big(f_u^R\big)'(x)}{\big(f_u^R\big)'(y)}\right|\leq e^{C\,\text{diam}(M)}.$$ Fixing $y_0\in L$, we get $$\big|\big(f_u^R\big)'(x)\big|\leq e^{C\,\text{diam}(M)}\big|\big(f_u^R\big)'(y_0)\big|=C_1,$$ concluding the proof.
\[2\] For all $i,n\in{\mathbb{N}}$ and all $\omega\in{\mathcal B}_i$, $${\operatorname{Leb}}_\gamma\big\{\widehat R_{i+1}-\widehat R_i>n_0+n\,|\,\omega\big\}\leq {\operatorname{Leb}}_\gamma\big\{\widehat R>n\big\}.$$
Let $A=\big\{x\in\omega:\widehat R_{i+1}(x)-\widehat R_i(x)>n_0+n\big\}$. For $x\in A$ we have $(\widehat R-1)(f^{\widehat R_i}(x))=\widehat R_{i+1}(x)-\widehat R_i(x)-n_0>n$. Then $f^{\widehat R_i}(A)\subseteq \bigcup_{k\geq n+2}(J_k\cup J'_k)$. So $$A\subseteq f^{-\widehat R_i}\Big(\bigcup_{k\geq n+2}(J_k\cup J'_k)\Big)\subseteq f^{-\widehat R_i}\big\{\widehat R>n\big\}$$ and, as $(f_u^{-\widehat R_i})'\leq 1$, then $${\operatorname{Leb}}_\gamma(A)\leq {\operatorname{Leb}}_\gamma\big(f^{-\widehat R_i}\big\{\widehat R>n\big\}\big)=\int_{\{\widehat R>n\}}\big(f_u^{-\widehat R_i}\big)'d{\operatorname{Leb}}_\gamma\leq {\operatorname{Leb}}_\gamma\{\widehat R>n\}.$$
In the proof of the next result we use ideas from [@y99 Section 4.1] and [@ap08 Section A.2.1].
There exists $C>0$ such that, for sufficiently large $n$, $${\operatorname{Leb}}_\gamma\{R>n\}\leq \frac{C}{n^{\tau+1}}.$$
We start by noting that $$\label{Rchapeu}
{\operatorname{Leb}}_\gamma\{\widehat{R}>n\}={\operatorname{Leb}}_\gamma\big( \bigcup_{i\geq n}(J_i\cup J'_i)\big)={\operatorname{Leb}}_\gamma([0,a_n] \cup[a'_n,0])\leq \frac{C}{n^{\tau+1}}.$$ Defining $\widehat R_0=0$, observe that ${\operatorname{Leb}}_\gamma\{R>n\}=\text{(I)}+\text{(II)}$, where $$\begin{aligned}
\text{(I)} & = \sum_{i\leq \frac{1}{2}\left[\frac{n}{n_0}\right]}{\operatorname{Leb}}_\gamma\{R>n; \widehat{R}_{i-1}\leq n<\widehat{R}_i\},\\
\text{(II)} & = {\operatorname{Leb}}_\gamma\big\{R>n; n\geq \widehat{R}_{\frac{1}{2}\left[\frac{n}{n_0}\right]}\Big\}.\end{aligned}$$ First we will see that there exists $\varepsilon_0>0$ and $C>0$, a constant depending on $f$, but not on $n$, such that $$\text{(II)} \leq C(1-\varepsilon_0)^{\frac{1}{2}\left[\frac{n}{n_0}\right]}.$$ In fact, taking $n\geq 4n_0$, and so $\frac{1}{2}\big[\frac{n}{n_0}\big]\geq 2$, we have $$\begin{aligned}
\label{II}
\nonumber\text{(II)} & = {\operatorname{Leb}}_\gamma\big\{R>n; n\geq \widehat{R}_{\frac{1}{2}\left[\frac{n}{n_0}\right]}\Big\}
\leq {\operatorname{Leb}}_\gamma\big\{R\geq \widehat{R}_{\frac{1}{2}\left[\frac{n}{n_0}\right]}\Big\}\\
\nonumber & = {\operatorname{Leb}}_\gamma\{R>\widehat{R}_2\}\,{\operatorname{Leb}}_\gamma\{R>\widehat{R}_3\mid R>\widehat{R}_2\}\ \cdots\ {\operatorname{Leb}}_\gamma\big\{R>\widehat{R}_{\frac{1}{2}
\left [\frac{n}{n_0}\right]}\mid R>\widehat{R}_{\frac{1}{2}\left[\frac{n}{n_0}\right]-1}\Big\}\\
& \leq C(1-\varepsilon_0)^{\frac{1}{2}\left[\frac{n}{n_0}\right]},\quad\mbox{applying Lemma \ref{1} to each factor}.\end{aligned}$$
We will now focus on (I). Let $k\geq 2n_0$. By , $$\label{4}
{\operatorname{Leb}}_\gamma\big\{\widehat{R}>\tfrac{n}{i}-n_0\big\} \leq \frac{C}{(\frac{n}{i}-n_0)^{\tau+1}} \leq C_1\left(\frac{i}{n_0}\right)^{\tau+1},\quad \forall i\leq \tfrac{1}{2}\big[\tfrac{n}{n_0}\big].$$ Fixing $i$, we have $$\begin{aligned}
\label{eq3}
{\operatorname{Leb}}_\gamma\left\{R>n\mid \widehat{R}_{i-1}\leq n<\widehat{R}_i\right\} & \leq {\operatorname{Leb}}_\gamma\left\{R> \widehat{R}_{i-1}; n< \widehat{R}_i\right\}\nonumber\\
& \leq \sum_{j=1}^i {\operatorname{Leb}}_\gamma\left\{R> \widehat{R}_{i-1};\widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i}\right\}.\end{aligned}$$ The last inequality is true because there exists $j\leq i$ such that $\widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i}$. In fact, if we assume the opposite, then $\frac{n}{i}\,i\geq \sum_{j=1}^i\big(\widehat{R}_j-\widehat{R}_{j-1}\big)=\widehat{R}_i$, which contradicts the assumption.
We will now prove that each term of the sum is less then or equal to $C(1-\varepsilon_0)^i\,\frac{i^{\tau+1}}{n^{\tau+1}}$. Considering first the case $i,j\geq 2$, define $$\begin{aligned}
&a= {\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_2\big\}\,{\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_3 \mid R> \widehat{R}_2\big\}\ \cdots\ {\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_{j-2} \mid R> \widehat{R}_{j-3}\big\},\\
&b={\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_{j-1}; \widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i} \mid R>\widehat{R}_{j-2}\big\},\\
&c={\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_j \mid R>\widehat{R}_{j-1}; \widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i} \big\} \cdots m
\big\{R> \widehat{R}_{i-1} \mid R>\widehat{R}_{i-2}; \widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i} \big\},\end{aligned}$$ where if $j=2$ or $j=3$ we take $a=1$ and if $j=i$ we take $c=1$. Note that $${\operatorname{Leb}}_\gamma\left\{R> \widehat{R}_{i-1};\widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i}\right\}=a\cdot b\cdot c.$$ Applying Lemma \[1\] to each factor in $a$, we get $a\leq (1-\varepsilon_0)^{j-1}$. Each factor in $c$ is of the form $ {\operatorname{Leb}}_\gamma\big\{R> \widehat{R}_k \mid R>\widehat{R}_{k-1}; \widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i} \big\}$ with $j\leq k<i$. Using again Lemma \[1\], we conclude that $c\leq (1-\varepsilon_0)^{i-j}$. Using Lemma \[2\] and , we get $$b \leq {\operatorname{Leb}}_\gamma\big\{\widehat{R}_j-\widehat{R}_{j-1}>\tfrac{n}{i}| R>\widehat{R}_{j-2}\big\} \leq \, {\operatorname{Leb}}_\gamma\big\{\widehat{R}>\tfrac{n}{i}-n_0\big\}\leq C\Big(\frac{i}{n}\Big)^{\tau+1}.$$ Gathering all the estimates above we get $$\label{I}
\text{(I)} \leq {\displaystyle\sum_{i\leq \frac{1}{2}\left[\frac{n}{n_0}\right]}} a\cdot b\cdot c \leq
= C{\displaystyle\sum_{i\leq \frac{1}{2}\left[\frac{n}{n_0}\right]}}(1-\varepsilon_0)^{i-1}\Big(\frac{i}{n}\Big)^{\tau+1}\leq
\frac{C}{n^{\tau+1}}{\displaystyle\sum_{i=1}^\infty}(1-\varepsilon_0)^{i-1}i^{\tau+1}=\frac{C_1}{n^{\tau+1}}.$$ For the term $i=1$ of (I), we have, by the definition of $\widehat R_1$, $$\begin{aligned}
{\operatorname{Leb}}_\gamma\{R>\widehat R_0; \widehat R_0<n<\widehat R_1\}\leq& {\operatorname{Leb}}_\gamma\{\widehat R_1>n\}={\operatorname{Leb}}_\gamma\{\widehat{R}>n-n_0+1\}\\
=&{\operatorname{Leb}}_\gamma\Big( \bigcup_{k\geq n-n_0+1}(J_k\cup J'_k)\Big)={\operatorname{Leb}}_\gamma([0,a_{n-n_0+1}] \cup[a'_{n-n_0+1},0])\\
\leq& \frac{C}{(n-n_0+1)^{\tau+1}}\leq\frac{C_1}{n^{\tau+1}},\end{aligned}$$ for any $n\geq n_1$, with $n_1$ sufficiently large. For $i\geq 2$ and $j=1$, considering each term of the sum in , $$\begin{aligned}
{\operatorname{Leb}}_\gamma\big\{R>&\widehat R_{i-1}; \widehat R_1-\widehat R_0>\tfrac{n}i\big\}\leq {\operatorname{Leb}}_\gamma\big\{\widehat R_1-\widehat R_0>\frac{n}i\big\}\,
{\operatorname{Leb}}_\gamma\big\{R>\widehat R_1|\widehat R_1-\widehat R_0>\frac{n}i\big\}\cdot\\
& \cdot
{\operatorname{Leb}}_\gamma\big\{R>\widehat R_2|R>\widehat R_1;\widehat R_1-\widehat R_0>\frac{n}i\big\}\cdots
{\operatorname{Leb}}_\gamma\big\{R>\widehat R_{i-1}|R>\widehat R_{i-2};\widehat R_1-\widehat R_0>\frac{n}i\big\}\\
&\leq C(1-{\varepsilon}_0)^{i-1},\end{aligned}$$ arguing as we did to estimate $c$ in the general case. Finally, from , and the calculations for the small terms, we have, for sufficiently large $n$, $$\text{(I)}+ \text{(II)}\leq \frac{C_1}{n^{\tau+1}}+C(1-\varepsilon_0)^{\frac{1}{2}\left[\frac{n}{n_0}\right]}\leq \frac{C_2}{n^{\tau+1}}.$$
Coupling measures {#appendix}
=================
In this appendix we prove Theorem \[Theorem C\]. To simplify notation, we shall remove all bars. Though the proof follows the same steps of [@y99 Theorem 3], we have decided to include it here, as our polynomial assumptions imply some changes in the estimates.
Assume that there is $C>0$ such that $$m\{ R>n\}\leq \frac{C}{n^\zeta} .$$ Let $\lambda$ and $\lambda'$ be probability measures in $\Delta$ whose densities with respect to $m$ are in the space $\mathcal G_\theta^+$ and denote $$\varphi=\frac{d\lambda}{dm} \quad\text{and}\quad \varphi'=\frac{d\lambda'}{dm}.$$ Consider the function $$\begin{array}[t]{rccc}
F\times F: & \Delta\times\Delta & \longrightarrow & \Delta\times\Delta\\
& (x,y) & \longmapsto & (F(x),F(y)),
\end{array}$$ the measure $P=\lambda\times\lambda'$ in $\Delta\times\Delta$ and let $\pi,\pi':\Delta\times\Delta\rightarrow \Delta$ be the projections on the first and second coordinates, respectively. Note that $F^n\circ \pi=\pi\circ(F\times F)^n$, for all $n\in{\mathbb{N}}$.
Consider the partition $\mathcal Q=\{\Delta_{l,i}\}$ of $\Delta$ introduced in Section \[qd\] and the partition $\mathcal Q\times \mathcal Q$ of $\Delta\times\Delta$. Observe that each element of $\mathcal Q\times \mathcal Q$ is sent bijectively by $F\times F$ onto a union of elements of $\mathcal Q\times \mathcal Q$. For $n\in{\mathbb{N}}$, we define $$(\mathcal Q\times \mathcal Q)_n=\bigvee_{i=0}^{n-1}(F\times F)^{-i}(\mathcal Q\times \mathcal Q)$$ and denote by $(\mathcal Q\times \mathcal Q)_n(x,x')$ the element of $(\mathcal Q\times \mathcal Q)_n$ that contains the pair $(x,x')$ of $\Delta\times\Delta$. Define $\widehat{R}:\Delta\rightarrow {\mathbb{N}}$ as $$\widehat{R}(x)=\min\{n\in{\mathbb{N}}_0:F^n(x)\in \Delta_0\}.$$ Note that $\widehat{R}_{|\Delta_0}=R_{|\Delta_0}$.
As $(F,\nu)$ is mixing and $\frac{d\nu}{dm}\in L^\infty$, there exists $n_0\in{\mathbb{N}}$ and $\delta_0>0$ such that, for all $n\geq n_0$, we have $m(F^{-n}(\Delta_0)\cap\Delta_0)\geq\delta_0$. Consider the sequence of [*stopping times*]{} $0\equiv \tau_0<\tau_1<\cdots$, defined in $\Delta\times\Delta$, as $$\begin{aligned}
\tau_1(x,x') & = n_0+\widehat{R}(F^{n_0}x)\\
\tau_2(x,x') & = \tau_1+n_0+\widehat{R}(F^{\tau_1}x')\\
\tau_3(x,x') & = \tau_2+n_0+\widehat{R}(F^{\tau_2}x)\\
\tau_4(x,x') & = \tau_3+n_0+\widehat{R}(F^{\tau_3}x')\\
& \vdots &\end{aligned}$$ Observe that $\tau_{i+1}-\tau_i\geq n_0$ for all $i\in{\mathbb{N}}$. We introduce now the [*simultaneous return time*]{} $T:\Delta\times\Delta\rightarrow {\mathbb{N}}$ as $$T(x,x')=\min_{i\geq 2}\big\{\tau_i:(F^{\tau_i}x,F^{\tau_i}x')\in\Delta_0\times\Delta_0\big\}.$$ Note that, as $(F,\nu)$ is mixing, then $(F\times F,\nu\times\nu)$ is ergodic. So $T$ is well defined $m\times m$ a.e.. We define a sequence of partitions of $\Delta\times\Delta$, $\xi_1<\xi_2<\cdots$ as follows:
- $\xi_1(x,x')=\left(F^{-\tau_1(x)+1}\mathcal Q\right)(x)\times\Delta$. The elements of $\xi_1$ are of the form $\Gamma=A\times\Delta$, where ${\tau_1}_{|A\times\Delta}$ is constant and $A$ is sent bijectively to $\Delta_0$ by $F^{\tau_1}$;
- for $i$ even, $\xi_i$ is the refinement of $\xi_{i-1}$ obtained by partitioning $\Gamma\in\xi_{i-1}$ in the $x'$ direction into sets $\widetilde{\Gamma}$ such that ${\tau_i}_{|\widetilde{\Gamma}}$ is constant and $\pi'(\widetilde{\Gamma})$ is sent bijectively to $\Delta_0$ by $F^{\tau_i}$;
- for $i$ odd, $i>1$, we do the same as in the previous point replacing the $x'$ direction by the $x$ direction and $\pi'$ by $\pi$.
For convenience we define $\xi_0=\{\Delta\times\Delta\}$. Note that, for all $i\in{\mathbb{N}}$, $\{T=\tau_i\}$ and $\{T>\tau_i\}$ are $\xi_{i+1}$-measurable and, for all $n\leq i$, $\tau_n$ is $\xi_i$-measurable. Define a sequence of [*stopping times*]{} in $\Delta\times\Delta$, $0\equiv T_0<T_1<\cdots$, as $$T_1=T \quad\text{and}\quad T_n=T_{n-1}+T{\mbox{\,\small $\circ$}}(F\times F)^{T_{n-1}}, \mbox{ for } n\geq 2.$$ We consider the dynamical system $\widehat{F}=(F\times F)^T:\Delta\times\Delta\rightarrow \Delta\times \Delta$. Observe that, for all $n\in{\mathbb{N}}$, $\widehat{F}^n=(F\times F)^{T_n}.$
Let $\widehat{\xi}_1$ be a partition of $\Delta\times\Delta$ composed by rectangles $\widehat{\Gamma}$ such that $T_{|\widehat{\Gamma}}$ is constant and $\widehat{F}: \widehat{\Gamma}\rightarrow \Delta_0\times\Delta_0$ is bijective. Define a sequence of partitions, $\widehat{\xi}_2$, $\widehat{\xi}_3, \ldots$, by $\widehat{\xi}_n=\widehat{F}^{-(n-1)}\widehat{\xi}_1$, for $n\geq 2$. Note that $T_n$ is constant on each element of $\widehat{\xi}_n$ and $\widehat{F}_n$ maps each element of $\widehat{\xi}_n$ bijectively to $\Delta_0\times\Delta_0$.
Consider the measure $m\times m$ for the dynamical system $\widehat{F}$ and the Jacobian, $J\widehat F$, of $\widehat{F}$ with respect to $m\times m$. Define a [*separation time*]{} $\widehat{s}:(\Delta\times\Delta)\times (\Delta\times\Delta)\rightarrow {\mathbb{N}}_0$ as $$\widehat{s}(z,w)=\min\big\{n\in{\mathbb{N}}_0: \text{$\widehat{F}z$ and $\widehat{F}w$ belong to different elements of $\widehat{\xi}_1$}\big\}.$$ Denoting $$\Phi=\frac{dP}{d(m\times m)},$$ we observe that $\Phi(x,x')=\varphi(x)\varphi'(x')$. We may assume without loss of generality that $\varphi>0$ and $\varphi'>0$. The proof of the following lemma can be found in [@y99 Sublemma 3].
\[A4\] For $z,w\in\Delta\times\Delta$ such that $\widehat s(z,w)\geq n$, for some $n\in{\mathbb{N}}$, we have $$\left|\log\frac{J\widehat F^n z}{J\widehat F^n w}\right|\leq2 C_{ F} \beta^{\widehat s(\widehat F^n z, \widehat F^n w)}.$$
\[A5\] For all $z,w\in\Delta\times\Delta$, we have $$\left|\log\frac{\Phi(z)}{\Phi(w)}\right|\leq \frac{D_\Phi}{\widehat s(z,w)^\theta},$$ where $D_\Phi=D_\varphi+D_{\varphi'}$.
Let $z=(x,x')$ and $w=(y,y')$. Then, since $\log x\leq x-1$ for $x\in{\mathbb{R}}^+$ and $\varphi,\varphi'\in \mathcal G_\theta^+$, $$\begin{aligned}
\left|\log\frac{\Phi(z)}{\Phi(w)}\right|&\leq \left|\log\frac{\varphi(x)}{\varphi(y)}\right|+\left|\log\frac{\varphi'(x')}{\varphi'(y')}\right|\leq\Big|\frac{\varphi(x)}{\varphi(y)}-1\Big|+\Big|\frac{\varphi'(x')}{\varphi'(y')}-1\Big|\vspace{3mm}\\
& \leq D_\varphi\frac{1}{s(x,y)^\theta}+D_{\varphi'}\frac{1}{s(x',y')^\theta}\leq \frac{D_\varphi+D_{\varphi'}}{\widehat s(z,w)^\theta}.\end{aligned}$$
\[A6\] There exists a constant $C>0$ depending only on $D_\varphi$ and $D_{\varphi'}$, such that, for all $i\in {\mathbb{N}}$, $\Gamma\in\widehat \xi_i$, $z,w\in\Delta_0\times\Delta_0$ and $Q=\widehat F_*^i(P|\Gamma)$, we have $$\left|\frac{dQ}{d{\operatorname{Leb}}}(z)\Big/\,\frac{dQ}{d{\operatorname{Leb}}}(w)\right| \leq C.$$
Take $z_0,w_0\in\Gamma$ such that $\widehat F^i(z_0)=z$ and $\widehat F^i(w_0)=w$. As $\widehat s(z_0,w_0)\geq i$, using Lemma \[A4\] and Lemma \[A5\], we get $$\left|\frac{dQ}{d{\operatorname{Leb}}}(z)\Big/\frac{dQ}{d{\operatorname{Leb}}}(w)\right|=\frac{\Phi(z_0)}{\Phi(w_0)}\left|\frac{J\widehat F^i(w_0)}{J\widehat F^i(z_0)}\right|\leq e^{D_\Phi}e^{C_{\widehat F}}.$$
Recalling Lemma \[A4\] we define $C_{\widehat F}=2C_F$. We take $$\label{c1}
K>C_{\widehat F}+\frac{C_{\widehat F}}{1-\beta}$$ and $\widehat C=K-C_{ \widehat F}$. Observe that $$\widehat C>\frac{C_{\widehat F}}{1-\beta} .$$ From here on we assume that $\theta>2e^{K}.$
\[Proposition A\] There exists $\varepsilon_0>0$ such that, for all $i\geq 2$ and $\Gamma\in\xi_i$ with $T_{|\Gamma}>\tau_{i-1}$, we have $$P\{T=\tau_i\,|\,\Gamma\}\geq \varepsilon_0.$$ The constant $\varepsilon_0$ depends only on $D_\varphi$, $D_{\varphi'}$ and, if we choose $i\geq i_0(D_\varphi, D_{\varphi'})$, the dependence can be removed.
\[Proposition B\] There exists $k_0>0$ such that, for all $i\in{\mathbb{N}}_0$, $\Gamma\in\xi_i$ and $n\in{\mathbb{N}}_0$, $$P\{\tau_{i+1}-\tau_i>n_0+n | \Gamma\}\leq k_0\,{\operatorname{Leb}}\{\widehat{R}>n\}.$$ The constant $k_0$ depends only on $D_\varphi$, $D_{\varphi'}$ and, if we choose $i\geq i_0(D_\varphi,D_{\varphi'})$, the dependence can be removed.
The proofs of these two propositions follow the same steps of the proofs of (E1) and (E2) in [@ap08 Subsections A.3.1 and A.3.2]. We only need to adapt the proof of [@ap08 Lemma A.2] to our case, which we do next.
\[A2\] There exists $C_0=C_0(\varphi)>0$ such that, for all $k\in{\mathbb{N}}$, ${\displaystyle A\in \bigvee_{i=0}^{k-1}F^{-i}(\mathcal Q)}$ with $F^k(A)=\Delta_0$, $\mu=F_*^k(\lambda|A)$ and $x,y\in\Delta_0$, we have $$\left|\frac{d\mu}{d{\operatorname{Leb}}}(x)\Big/\frac{d\mu}{d{\operatorname{Leb}}}(y)\right|\leq C_0.$$ The dependence of $C_0$ on $D_\varphi$ may be removed if we assume that the number of visits $j\leq k$ of $A$ to $\Delta_0$ is bigger then a certain $j_0=j_0(D_\varphi)$.
Given $x_0,y_0\in A$ such that $F^k(x_0)=x$ and $F^k(y_0)=y$ then, as $\varphi\in \mathcal G_\theta^+$ and using Lemma \[J\], $$\begin{aligned}
\left|\frac{d\mu}{d{\operatorname{Leb}}}(x)\Big/\frac{d\mu}{d{\operatorname{Leb}}}(y)\right| =\frac{\varphi(x_0)}{\varphi(y_0)}\left|\frac{JF^k(y_0)}{JF^k(x_0)}\right|
& \leq \Big(1+\frac{D_\varphi}{s(x_0,y_0)^\theta}\Big)\big(1+C_F\beta^{s(F^k(x_0),F^k(y_0))}\big) \\
& \leq \Big(1+\frac{D_\varphi}{j^\theta}\Big)(1+C_F)=C_0.\end{aligned}$$
The following proposition, whose proof can be found in [@ap08 Subsection A.2.1], follows from Propositions \[Proposition A\] and \[Proposition B\].
\[poldecay\] Let $C>0$ and $\zeta>1$ be such that ${\operatorname{Leb}}\{R>n\}\leq Cn^{-\zeta}$. Then, there exists $C'>0$ such that $$P\{T>n\}\leq \frac{C'}{n^{\zeta-1}}.$$
We want to define a sequence of densities $(\widehat\Phi_i)$ in $\Delta\times\Delta$ such that $\widehat\Phi_0\geq \widehat\Phi_1\geq\cdots$ and for all $i\in{\mathbb{N}}$ and $\widehat\Gamma\in\widehat\xi_i$, $$\label{condition}
\pi_* \widehat F_*^i \big((\widehat\Phi_{i-1}-\widehat\Phi_i)((m\times m)|\widehat\Gamma)\big)=
\pi'_* \widehat F_*^i \big((\widehat\Phi_{i-1}-\widehat\Phi_i)((m\times m)|\widehat\Gamma)\big).$$ Take $\zeta$ as in Theorem \[Theorem C\]. Noting that $1<\zeta<\frac{\theta}{e^{K}}-1$, we fix $\rho$ such that $$\label{gamma}
\zeta+1<\rho<\frac{\theta}{e^{K}},$$ Take $$\varepsilon_i=e^{K}\Big(1-\Big(\frac{i-1}{i}\Big)^\rho\Big),$$ for $i\geq i_0$, where $i_0$ is such that $\varepsilon_{i_0}<1$. Further restrictions on $i_0$ will be imposed during the proof of Lemma \[steps\]. Define $\widehat \Phi_i\equiv \Phi$ for $i<i_0$, and for $i\geq i_0$ $$\label{phii}
\widehat \Phi_i(z)=\left(\frac{\widehat \Phi_{i-1}(z)}{J\widehat F^i(z)}-\varepsilon_i\min_{w\in\widehat\xi_i(z)} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}\right) J\widehat F^i(z),$$ where $\widehat\xi_i(z)$ is the element of $\widehat\xi_i$ which contains $z$. It is easy to verify that the sequence $\big(\Phi_i(z)\big)$ satisfies condition . For $z\in\Delta\times\Delta$, let $$\widetilde\Psi_{i_0-1}(z)=\frac{\Phi}{J\widehat F^{i_0-1}(z)}$$ and for $i\geq i_0$ $$\Psi_i(z) =\frac{\widetilde\Psi_{i-1}(z)}{J\widehat F(\widehat F^{i-1})(z)},$$ $$\varepsilon_{i,z}=\varepsilon_i\min_{w\in\widehat\xi(z)}\Psi_i(w)$$ and $$\widetilde\Psi_i(z)=\Psi_i(z)-\varepsilon_{i,z}.$$
\[steps\] There exists $i_0\in{\mathbb{N}}$ such that, for $i\geq i_0$ and for all $z,w\in \Delta\times\Delta$ with $w\in\widehat\xi_i(z)$, we have $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\right|\leq \widehat C.$$
We divide this proof into several steps.
[*Step 1:*]{} By the definition of $\Psi_i$ and Lemma \[A4\], $$\begin{aligned}
\left|\log\frac{\Psi_i(z)}{\Psi_i(w)}\right| &\leq \left|\log\frac{\widetilde\Psi_{i-1}(z)}{\widetilde\Psi_{i-1}(w)}\right|+\left|\log\frac{J\widehat F(\widehat F^{i-1}w)}{J\widehat F(\widehat F^{i-1}z)}\right|\\
&\leq \left|\log\frac{\widetilde\Psi_{i-1}(z)}{\widetilde\Psi_{i-1}(w)}\right|+C_{\widehat F}\beta^{\widehat s(\widehat F^iz,\widehat F^iw)}.\end{aligned}$$
[*Step 2:*]{} Setting $\widehat\varepsilon_i=\varepsilon_{i,z}=\varepsilon_{i,w}$, we get $$\begin{aligned}
\label{s2.1}
\nonumber\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}-\log\frac{\Psi_i(z)}{\Psi_i(w)}\right| &= \left|\log\left(\frac{\Psi_i(z)-\widehat\varepsilon_i}{\Psi_i(z)}\,\frac{\Psi_i(w)}{\Psi_i(w)-\widehat\varepsilon_i}\right)\right|\\
&=\left|\log\left(1+\frac{\frac{\widehat\varepsilon_i}{\Psi_i(w)}-\frac{\widehat\varepsilon_i}{\Psi_i(z)}}{1-\frac{\widehat\varepsilon_i}{\Psi_i(w)}}\right)\right|.\end{aligned}$$ We may assume that $\Psi_i(w)\leq \Psi_i(z)$. Otherwise, we can swap the positions of $z$ and $w$. We can easily verify that, for all $0<a\leq b<1$, we have $$\log\Big(1+\frac{b-a}{1-b}\Big)\leq \frac{b}{1-b}\log \frac{b}{a}.$$ Taking $a=\frac{\widehat \varepsilon_i}{\Psi_i(z)}$ and $b=\frac{\widehat \varepsilon_i}{\Psi_i(w)}$ and recalling the definition of $\widehat\varepsilon_i$, we obtain $$\label{s2.2}
\left|\log\left(1+\frac{\frac{\widehat\varepsilon_i}{\Psi_i(w)}-\frac{\widehat\varepsilon_i}{\Psi_i(z)}}{1-\frac{\widehat\varepsilon_i}{\Psi_i(w)}}\right)\right|
\leq \frac{\frac{\widehat\varepsilon_i}{\Psi_i(w)}}{1-\frac{\widehat\varepsilon_i}{\Psi_i(w)}}\left|\log \frac{\frac{\widehat\varepsilon_i}{\Psi_i(w)}}{\frac{\widehat\varepsilon_i}{\Psi_i(z)}}\right|
\leq \frac{\varepsilon_i}{1-\varepsilon_i}\left|\log\frac{\Psi_i(z)}{\Psi_i(w)}\right|.$$ Gathering the expressions and , we obtain $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}-\log\frac{\Psi_i(z)}{\Psi_i(w)}\right|\leq\,\frac{\varepsilon_i}{1-\varepsilon_i}\,\left|\log\frac{\Psi_i(z)}{\Psi_i(w)}\right|.$$ Denoting $\varepsilon'_i=\frac{\varepsilon_i}{1-\varepsilon_i}$, we conclude that $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\right|\leq (1+\varepsilon'_i) \left|\log\frac{\Psi_i(z)}{\Psi_i(w)}\right|.$$
[*Step 3:*]{} Note that $$\Psi_{i_0}(z)=\frac{\widetilde\Psi_{i_0-1}(z)}{J\widehat F(\widehat F^{i_0-1}(z))}=
\frac{\Phi(z)}{J\widehat F^{i_0-1}(z)J\widehat F(\widehat F^{i_0-1}(z))}=\frac{\Phi(z)}{J\widehat F^{i_0}(z)},$$ and so, using step 2, Lemma \[A4\] and Lemma \[A5\], $$\begin{aligned}
\label{Psitilde}
\nonumber\left|\log\frac{\widetilde\Psi_{i_0}(z)}{\widetilde\Psi_{i_0}(w)}\right| & \leq (1+\varepsilon'_{i_0}) \left(\left|\log\frac{\Phi(z)}{\Phi(w)}\right|+\left|\log\frac{J\widehat F^{i_0}(w)}{J\widehat F^{i_0}(z)}\right|\right)\\
\nonumber&\leq (1+\varepsilon'_{i_0}) \left(\frac{D_\Phi}{\widehat s(z,w)^\theta}+C_{\widehat F}\beta^{\widehat s(\widehat F^{i_0} z,\widehat F^{i_0} w)}\right)\\
&= (1+\varepsilon'_{i_0}) \left(\frac{D_\Phi}{\big(\widehat s(\widehat F^{i_0} z,\widehat F^{i_0} w)+i_0\big)^\theta}+C_{\widehat F}\beta^{\widehat s(\widehat F^{i_0} z,\widehat F^{i_0} w)}\right).\end{aligned}$$ Then $$\left|\log\frac{\widetilde\Psi_{i_0}(z)}{\widetilde\Psi_{i_0}(w)}\right| \leq (1+\varepsilon'_{i_0}) \left(\frac{D_\Phi}{i_0^\theta}+C_{\widehat F}\right) \underset{i_0\rightarrow\infty}{\rightarrow}
C_{\widehat F}<\widehat C$$ and so we can choose $i_0$ sufficiently large such that $$\left|\log\frac{\widetilde\Psi_{i_0}(z)}{\widetilde\Psi_{i_0}(w)}\right|\leq \widehat C,$$ obtaining the conclusion of the Lemma for $i=i_0$.
[*Step 4:*]{} Using steps 2 and 1, we obtain $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\right|\leq (1+\varepsilon'_i) \left(\left|\log\frac{\widetilde\Psi_{i-1}(z)}{\widetilde\Psi_{i-1}(w)}\right|+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}\right)$$
[*Step 5:*]{} Using the equality $\widehat s(\widehat F^{i-j} z,\widehat F^{i-j} w)=\widehat s(\widehat F^i z,\widehat F^i w)+j$ and the inequalities in steps 3 and 4, we get, for $i\geq i_0+1$, $$\begin{aligned}
\Bigg|&\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\Bigg|\\
&\leq (1+\varepsilon'_i) \left((1+\varepsilon'_{i-1}) \left(\left|\log\frac{\widetilde\Psi_{i-2}(z)}{\widetilde\Psi_{i-2}(w)}\right|+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)+1}\right)+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}\right)\\
&\leq \left|\log\frac{\widetilde\Psi_{i_0}(z)}{\widetilde\Psi_{i_0}(w)}\right|\prod_{j=i_0+1}^i (1+\varepsilon'_j)+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}\Big(\beta^{i-i_0-1}\prod_{j=i_0+1}^i (1+\varepsilon'_j)\\
& \quad+\cdots+ \beta(1+\varepsilon'_i)(1+\varepsilon'_{i-1})+(1+\varepsilon'_i)\Big)\\
& \leq \Big(\frac{D_\Phi}{\widehat s(z,w)^\rho}+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)+i-i_0}\Big)\prod_{j=i_0}^i (1+\varepsilon'_j)\\
& \quad+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}\Big(\beta^{i-i_0-1}\prod_{j=i_0+1}^i (1+\varepsilon'_j) +\cdots+(1+\varepsilon'_i)\Big)\\
&= \frac{D_\Phi}{\big(\widehat s(\widehat F^i z,\widehat F^i w)+i\big)^\theta}\prod_{j=i_0}^i (1+\varepsilon'_j)+C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}(1+\varepsilon'_i)\sum_{k=i_0}^i \Big(\prod_{j=k}^i (1+\varepsilon'_j)\beta\Big).\end{aligned}$$
In the next two steps we will control the two terms of the previous expression.
[*Step 6:*]{} Recalling that $\varepsilon'_i=\frac{\varepsilon_i}{1-\varepsilon_i}$ and $\varepsilon_i=e^{K}\big(1-\big(\frac{i-1}{i}\big)^\rho\big)$, it is easy to check that $$\lim_{i\to\infty} \frac{\varepsilon'_i}{ {1}/{i}}=e^{K}\,\rho.$$ Remember that, in , we chose $\rho$ such that $\theta>e^{K}\rho$. So, for $i_0$ sufficiently large and $i\geq i_0$, we have $\varepsilon'_i< \frac{\theta}{i}$. As $\log(1+x)\leq x$ for $x>0$, then $$\log\prod_{j=i_0}^i (1+\varepsilon'_j)=\sum_{j=i_0}^i \log(1+\varepsilon'_j)\leq \sum_{j=i_0}^i \varepsilon'_j \leq \theta \sum_{j=i_0}^i \frac{1}{j} \leq \theta \log \frac{i}{i_0-1}.$$ So, $$\prod_{j=i_0}^i (1+\varepsilon'_j) \leq \Big(\frac{i}{i_0-1}\Big)^\theta$$ and $$\frac{D_\Phi}{\big(\widehat s(\widehat F^i z,\widehat F^i w)+i\big)^\theta}\prod_{j=i_0}^i (1+\varepsilon'_j)\leq
\frac{D_\Phi}{i^\theta}\Big(\frac{i}{i_0-1}\Big)^\theta=\frac{D_\Phi}{(i_0-1)^\theta}.$$
[*Step 7:*]{} We may choose $i_0$ sufficiently large such that $(1+\varepsilon'_{i_0})\beta<1$. Note that we will later impose additional restrictions on $i_0$. So, recalling that $(\varepsilon'_i)$ is a decreasing sequence converging to zero, then, for all $i\geq i_0$, $$\begin{aligned}
C_{\widehat F}\beta^{\widehat s(\widehat F^i z,\widehat F^i w)}(1+\varepsilon'_i)\sum_{k=i_0}^i \Big(\prod_{j=k}^i (1+\varepsilon'_j)\beta\Big) &\leq C_{\widehat F}(1+\varepsilon'_{i_0})\sum_{k=0}^\infty \big((1+\varepsilon'_{i_0})\beta\big)^k\\
&= \frac{C_{\widehat F}(1+\varepsilon'_{i_0})}{1-(1+\varepsilon'_{i_0})\beta}.\end{aligned}$$
[*Step 8:*]{} Replacing the conclusions of steps 6 and 7 on the expression in step 5, we obtain, for $i\geq i_0+1$, $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\right|\leq \frac{D_\Phi}{(i_0-1)^\theta}+ \frac{C_{\widehat F}(1+\varepsilon'_{i_0})}{1-(1+\varepsilon'_{i_0})\beta}.$$ As $\varepsilon'_{i_0} \underset{i_0\rightarrow \infty}{\rightarrow} 0$, then $$\frac{D_\Phi}{(i_0-1)^\theta}+ \frac{C_{\widehat F}(1+\varepsilon'_{i_0})}{1-(1+\varepsilon'_{i_0})\beta} \underset{i_0\rightarrow \infty}{\rightarrow} \frac{C_{\widehat F}}{1-\beta}.$$ Observing that we chose $\widehat C>\frac{C_{\widehat F}}{1-\beta}$, then there exists $i_0$ large enough such that, for $i\geq i_0+1$, $$\left|\log\frac{\widetilde\Psi_i(z)}{\widetilde\Psi_i(w)}\right|\leq \widehat C.$$ Recalling that we proved the same result for $i=i_0$ in step 3, this concludes the proof.
\[max\] There exists $i_0\in{\mathbb{N}}$ such that, for all $i\geq i_0$ and $\widehat\Gamma\in \widehat\xi_i$, $${\max_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\Bigg/\!{\min_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\leq e^{K}.$$
Notice that, by the definitions, we have, for $i\geq i_0$, $$\label{Phifrac}
\frac{\widehat\Phi_i}{J\widehat F^i(z)}=\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^i(z)}-\varepsilon_i \min_{w\in\widehat\xi_i(z)}\frac{\widehat\Phi_{i-1}(w)}{J\widehat F^i(w)}$$ and $$\label{Psifrac}
\widetilde\Psi_i(z)=\frac{\widetilde\Psi_{i-1}(z)}{J\widehat F(\widehat F^{i-1}(z))}-\varepsilon_i \min_{w\in\widehat\xi_i(z)}\frac{\widetilde\Psi_{i-1}(w)}{J\widehat F(\widehat F^{i-1}(w))}.$$ We will prove by induction that for all $z\in\Delta\times\Delta$ and all $i\geq i_0$ we have $$\label{cima1}
\widetilde\Psi_{i-1}(z)=\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^{i-1}(z)},$$ which, since $J\widehat F^i(z)=J\widehat F(\widehat F^{i-1}(z))J\widehat F^{i-1}(z)$, is equivalent to $$\label{baixo1}
\frac{\widetilde\Psi_{i-1}(z)}{J\widehat F(\widehat F^{i-1}(z))}=\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^i(z)}.$$ If $i=i_0$, then is true by definition. Supposing now, by induction, that is true, we will prove that it is also true replacing $i-1$ by $i$. In fact, using in and remembering , we obtain $$\widetilde\Psi_i(z)=\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^i(z)}-\varepsilon_i \min_{w\in\widehat\xi_i(z)}\frac{\widehat\Phi_{i-1}(w)}{J\widehat F^i(w)}=\frac{\widehat\Phi_i}{J\widehat F^i(z)},$$ which concludes the proof of . Using , we have $$\frac{\widetilde\Psi_{i-1}(z)}{\widetilde\Psi_{i-1}(w)}
=\frac{\displaystyle\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^{i}(z)}\, J\widehat F(\widehat F^{i-1}(z))}{\displaystyle\frac{\widehat\Phi_{i-1}(w)}{J\widehat F^{i}(w)}\,{J\widehat F(\widehat F^{i-1}(w))}}$$ and so $$\frac{\displaystyle\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^{i}(z)}}{\displaystyle\frac{\widehat\Phi_{i-1}(w)}{J\widehat F^{i}(w)}}=\frac{\widetilde\Psi_{i-1}(z)}{\widetilde{\Psi}_{i-1}(w)}\,\frac{J\widehat F(\widehat F^{i-1}(w))}{J\widehat F(\widehat F^{i-1}(z))}.$$ Since, by Lemma \[steps\], $$\frac{\widetilde\Psi_{i-1}(z)}{\widetilde\Psi_{i-1}(w)}\leq e^{\widehat C},$$ and by Lemma \[A4\] $$\frac{J\widehat F(\widehat F^{i-1}(w))}{J\widehat F(\widehat F^{i-1}(z))}\leq e^{C_{\widehat F}},$$ then $${\max_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\Bigg/\!{\min_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\leq e^{\widehat C+C_{\widehat F}}=e^{K}.$$
\[3\] There exists $i_0\in{\mathbb{N}}$ such that, for $i\geq i_0$, we have $$\widehat \Phi_i\leq \Big(\frac{i-1}{i}\Big)^\rho \widehat \Phi_{i-1} \quad \text{in} \quad \Delta\times\Delta.$$
Observe that, for $i\geq i_0$ and $z\in\Delta\times\Delta$, $$\begin{aligned}
\widehat \Phi_i(z)\leq \Big(\frac{i-1}{i}\Big)^\rho \widehat \Phi_{i-1}(z)\
&\Leftrightarrow \varepsilon_i\min_{w\in\widehat\xi_i(z)} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}
\geq \Big(1-\Big(\frac{i-1}{i}\Big)^\rho\Big) \frac{\widehat\Phi_{i-1}(z)}{J\widehat F^i(z)}\\
&\Leftrightarrow \varepsilon_i \geq \Big(1-\Big(\frac{i-1}{i}\Big)^\rho\Big)
\frac{\frac{\widehat\Phi_{i-1}(z)}{J\widehat F^i(z)}}{\min_{w\in\widehat\xi_i(z)} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)} }.\end{aligned}$$ Since, by Lemma \[max\], for all $\widehat\Gamma\in \widehat\xi_i$, $$\label{maxmin}
{\max_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\Bigg/\!{\min_{w\in\widehat\Gamma} \frac{\widehat \Phi_{i-1}(w)}{J\widehat F^i(w)}}\leq e^{K}$$ the conclusion follows from our choice of $\varepsilon_i$.
The proof of the following result is an adaptation of the proofs of [@y99 Lemma 4] and (E3) in [@ap08 Subsection A.3.3]. The first part of the proof is the same, but we present it for the sake of completeness.
\[Proposition C\] There exists a constant $k_1>0$ such that, for all $n\in {\mathbb{N}}$, $$|F^n_*\lambda-F^n_*\lambda'|\leq 2P\{T>n\}+k_1\sum_{i=1}^\infty \frac{1}{i^\rho}\,P\{T_i\leq n<T_{i+1}\}.$$ The constant $k_1$ depends only on $D_\varphi$ and $D_{\varphi'}$.
Given $n \in{\mathbb{N}}_0$, $z\in\Delta\times\Delta$ and recalling the definition of $\widehat \Phi_i$ given in , let $\Phi_0,\Phi_1,\ldots$ be defined as follows: $$\label{Phin}
\Phi_n(z)=\widehat{\Phi}_i(z)\quad\text{for} \quad T_i(z)\leq n<T_{i+1}(z).$$ We will prove that, for all $n\in{\mathbb{N}}$, $$\label{Flambda}
\left|F^n_*\lambda-F^n_*\lambda'\right|\leq 2\int\Phi_n\, d(m\times m).$$ In fact, observing that ${\displaystyle \Phi=\Phi_n+\sum_{k=1}^n\left(\Phi_{k-1}-\Phi_k\right)}$, we have $$\begin{aligned}
\left|F^n_*\lambda-F^n_*\lambda'\right| &=\left|\pi_*(F\times F)^n_*\left(\Phi(m\times m)\right)- \pi'_*(F\times F)^n_*\left(\Phi(m\times m)\right)\right|\\
&=\left|\pi_*(F\times F)^n_*\left(\Phi_n(m\times m)\right)- \pi'_*(F\times F)^n_*\left(\Phi_n(m\times m)\right)\right|\\
& +\sum_{k=1}^n\big|(\pi-\pi')_*\big((F\times F)^n_*(\Phi_{k-1}-\Phi_k)(m\times m)\big)\big|.\end{aligned}$$ The first term in the last expression is bounded as follows $$\left|\pi_*(F\times F)^n_*\left(\Phi_n(m\times m)\right)- \pi'_*(F\times F)^n_*\left(\Phi_n(m\times m)\right)\right|\leq 2\int \Phi_n\,d(m\times m).$$ We will now verify that the other terms vanish. Let $A_{k,i}=\{z\in\Delta\times\Delta: k=T_i(z)\}$ and $A_k=\bigcup A_{k,i}$. Note that each of the sets $A_{k,i}$ is a union of elements of $\Gamma\in\widehat{\xi}_i$ and $A_{k,i}\neq A_{k,j}$ for $i\neq j$. By we have $\Phi_{k-1}-\Phi_k=\widehat{\Phi}_{i-1}-\widehat{\Phi}_i$ on $\Gamma\in\widehat{\xi}_i|A_{k,i}$ and $\Phi_{k-1}=\Phi_k$ on $\Delta\times\Delta \setminus A_k$. Given $k\in{\mathbb{N}}$ and remembering that, from , $$\pi_* \widehat F_*^i \big((\widehat\Phi_{i-1}-\widehat\Phi_i)((m\times m)|\widehat\Gamma)\big)=
\pi'_* \widehat F_*^i \big((\widehat\Phi_{i-1}-\widehat\Phi_i)((m\times m)|\widehat\Gamma)\big),$$ we have $$\begin{aligned}
\pi_*(F\times F)^n_* &(\Phi_{k-1}-\Phi_k)(m\times m) \\
&=\sum_i\sum_{\Gamma\subseteq A_{k,i}}F^{n-k}_*\pi_*(F\times F)^{T_i}_*\left((\widehat{\Phi}_{i-1}-\widehat{\Phi}_i)(m\times m)|\Gamma\right)\\
&=\sum_i\sum_{\Gamma\subseteq A_{k,i}}F^{n-k}_*\pi'_*(F\times F)^{T_i}_*\left((\widehat{\Phi}_{i-1}-\widehat{\Phi}_i)(m\times m)|\Gamma\right)\\
&= \pi'_*(F\times F)^n_*\left(\Phi_{k-1}-\Phi_k)(m\times m)\right).\end{aligned}$$ This completes the proof of . As a consequence, we have $$\begin{aligned}
\label{Fnlambda}
\nonumber |F_*^n\lambda-F_*^n\lambda'| &\leq 2\int \Phi_n\, d(m\times m)\\
&=2\int_{\{T_{i_0}>n\}}\Phi_n\,d(m\times m)+2\sum_{i=i_0}^\infty\int_{\{T_i\leq n<T_{i+1}\}}\Phi_n\,d(m\times m).\end{aligned}$$ For the first term of this expression we have $$\int_{\{T_{i_0}>n\}}\Phi_nd(m\times m)=\int_{\{T_{i_0}>n\}}\Phi\,d(m\times m)=P\{T_{i_0}>n\}$$ and for each of the others, using Lemma \[3\], we obtain $$\begin{aligned}
\int_{\{T_i\leq n<T_{i+1}\}}\Phi_n\,d(m\times m) &=\int_{\{T_i\leq n<T_{i+1}\}}\widehat\Phi_i\,d(m\times m)\\
&\leq\int_{\{T_i\leq n<T_{i+1}\}}\Big(\frac{i_0-1}{i}\Big)^\rho\Phi\,d(m\times m)\\
&=\Big(\frac{i_0-1}{i}\Big)^\rho P\{T_{i}\leq n<T_{i+1}\}.\end{aligned}$$ So, replacing the previous two expressions in , we get $$\left|F^n_*\lambda-F^n_*\lambda'\right| \leq 2P\{T_{i_0}>n\}+2(i_0-1)^\rho\sum_{i=i_0}^\infty \frac{1}{i^\rho}\, P\{T_{i}\leq n<T_{i+1}\}.$$ On the other hand, $$\begin{aligned}
P\{T_{i_0}>n\} &= P\{T>n\}+(i_0-1)^\rho\sum_{i=1}^{i_0-1} \frac{1}{(i_0-1)^\rho}\, P\{T_{i}\leq n<T_{i+1}\}\\
& \leq P\{T>n\}+(i_0-1)^\rho\sum_{i=1}^{i_0-1} \frac{1}{i^\rho}\, P\{T_{i}\leq n<T_{i+1}\}.
\end{aligned}$$ Gathering the last two inequalities we conclude that $$\left|F^n_*\lambda-F^n_*\lambda'\right|\leq 2P\{T>n\}+k_1\sum_{i=1}^\infty \frac{1}{i^\rho}\,P\{T_{i}\leq n<T_{i+1}\},$$ where $k_1$ depends only on $i_0$. Fixing $i_0$ sufficiently large, from Lemma \[3\] we obtain the dependence of $k_1$ on $\varphi$ and $\varphi'$.
The proof of the following proposition can be found in [@ap08 Subsection A.3.4]. We remark that though it uses [@ap08 Lemma A.6], whose proof does not necessarily follow for functions in $\mathcal G_\theta^+$, we obtained the same conclusion in Lemma \[A6\].
\[Proposition D\] There exists a constant $k_2>0$ such that, for $n\in{\mathbb{N}}$ and $i\in{\mathbb{N}}_0$, $$P\{T_{i+1}-T_i>n\}\leq k_2(m\times m)\{T>n\}.$$ The constant $k_2$ depends only on $D_\varphi$ and $D_{\varphi'}$.
We are now ready to prove Theorem \[Theorem C\].
[*Proof of Theorem \[Theorem C\].*]{} Given $i\in{\mathbb{N}}$, we have $$\label{pti}
P\{T_i\leq n < T_{i+1}\} \leq \sum_{j=0}^i P\Big\{T_{j+1} -T_j >\frac{n}{i+1}\Big\}.$$ The last inequality is true because there exists $j\leq i$ such that $T_{j+1}-T_j>\tfrac{n}{i+1}$. In fact, if we assume the opposite, then $\frac{n}{i}\,i\geq \sum_{j=1}^i\big(T_{j+1}-T_j\big)=T_{i+1}$, which contradicts the assumption.
It follows, respectively from Proposition \[Proposition C\], and Proposition \[Proposition D\] that $$\begin{aligned}
\left|F^n_*\lambda-F^n_*\lambda'\right| & \leq 2P\{T>n\}+k_1\sum_{i=1}^\infty \frac{1}{i^\rho}\, P\{T_{i}\leq n<T_{i+1}\}\\
& \leq 2P\{T>n\}+k_1\sum_{i=1}^\infty \frac{1}{i^\rho}\sum_{j=0}^iP\Big\{T_{j+1}-T_j>\frac{n}{i+1}\Big\}\\
& \leq 2P\{T>n\}+k_1k_2\sum_{i=1}^\infty \frac{1}{i^\rho}(i+1)\,(m\times m)\Big\{T>\frac{n}{i+1}\Big\}.
\end{aligned}$$ We know from Proposition \[poldecay\] that $P\{T >n\}\leq C/n^{\zeta-1}$. So, taking $P = m\times m$ we obtain $$\begin{aligned}
2P\{T>n\} +k_1k_2\sum_{i=1}^\infty \frac{1}{i^\rho}(i+1)\,(m\times m)\Big\{T>\frac{n}{i+1}\Big\}\leq \Big(2C+k_1k_2\sum_{i=1}^\infty \frac{(i+1)^\zeta}{i^\rho}\Big)\frac{1}{n^{\zeta-1}}.
\end{aligned}$$ Since, in , we chose $\rho>\zeta+1$, we obtain $$\left|F^n_*\lambda-F^n_*\lambda'\right|\leq C'\frac{1}{n^{\zeta-1}}.$$
[10]{} J. F. Alves and V. Pinheiro, *Slow rates of mixing for dynamical systems with hyperbolic structures*, J. Stat. Phys. **131** (2008), 505–534.
J. F. Alves and V. Pinheiro, *Gibbs-Markov structures and limit laws for partially hyperbolic attractors with mostly expanding central direction*, Advances in Mathematics **223** (2010), 1706–1730.
V. Ara' ujo, *Large deviations for semiflows over a non-uniformly expanding base*, Bull. Braz. Math. Soc. (N.S.) **38** (2007), 335–376.
V. Ara' ujo and M. J. Pacifico *Large deviations for non-uniformly expanding maps*, J. Stat. Phys. **125** (2006), 411–453.
R. Bowen, *Equilibrium states and the ergodic theory of Anosov diffeomorphisms*, Lecture Notes in Mathematics, vol. 470, Springer, New York (1975).
M. Benedicks and L.-S. Young *Markov extensions and decay of correlations for certain H' enon maps*, Ast' erisque **261** (2000), 13–56.
Y. Kifer, *Large deviations in dynamical systems and stochastic processes*, Trans. Amer. Math. Soc., **321** (1990), 505–524.
A. Lopes, *Entropy and large deviations*, Nonlinearity, **3** (1990), 527–546.
R. Mañé, *Ergodic theory and differentiable dynamics*, Springer-Verlag, Berlin (1987).
I. Melbourne, *Large and moderate deviations for slowly mixing dynamical systems*, Proc. Amer. Math. Soc. **137**, no.5 (2009), 1735–1741.
I. Melbourne and M. Nicol, *Almost sure invariance principle for nonuniformly hyperbolic systems*, Commun. Math. Phys. **260** (2005), 131–146.
I. Melbourne and M. Nicol, *Large deviations for nonuniformly hyperbolic systems*, Transactions of the American Mathematical Society **360** (2000), 6661–6676.
S. Orei and S. Pelikan, *Large deviations principles for stationary processes*, Ann. Probab., **16** (1988), 1481–1495.
D. Ruelle, *A measure associated with Axiom A attractors*, Am. J. Math. **98** (1976), 619–654.
E. Rio, *Th' eorie asymptotique des processus al' eatoires faiblement d' ependants*, Math' ematiques & Applications(Berlin) \[Mathematics and Applications\] **31**, Springer-Verlag, Berlin (2000).
Ya. Sinai, *Gibbs measure in ergodic theory*, Russ. Math. Surv. **27** (1972), 21–69.
S. Waddington, *Large deviations asymptotics for Anosov flows*, Ann. Inst. H. Poincar' e Anal. Non Linéaire, **13** (1996), 445–484.
L.-S. Young, *Large deviations in dynamical systems*, Trans. Amer. Math. Soc. **318** (1990), 525–543.
L.-S. Young, *Statistical properties of dynamical systems with some hyperbolicity*, Ann. Math. **147** (1998), 585–650.
L.-S. Young, *Recurrence times and rates of mixing*, Israel J. Math. **110** (1999), 153–188.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'With the development of affordable aberration-correctors, analytical scanning transmission electron microscopy (STEM) studies of complex interfaces can now be conducted at high spatial resolution at laboratories worldwide. Energy-dispersive X-ray spectroscopy (STEM-EDS) in particular has grown in popularity, since it enables elemental mapping over a wide range of ionization energies. However, the interpretation of atomically-resolved data is greatly complicated by beam-sample interactions that are often overlooked by novice users. Here we describe the practical factors—namely, sample thickness and the choice of ionization edge—that affect the quantification of a model perovskite oxide interface. Our measurements of the same sample in regions of different thickness indicate that interface profiles can vary by as much as 2–5 unit cells, depending on the spectral feature. This finding is supported by multislice simulations, which reveal that on-axis maps of even perfectly abrupt interfaces exhibit significant delocalization. Quantification of thicker samples is further complicated by channeling to heavier sites across the interface, as well as an increased signal background. We show that extreme care must be taken to prepare samples to minimize channeling effects and argue that it may not be possible to extract atomically-resolved information from many chemical maps.'
author:
- 'Steven R. Spurgeon'
- Yingge Du
- 'Scott A. Chambers'
bibliography:
- 'References.bib'
title: 'Measurement Error in Atomic-Scale STEM-EDS Mapping of a Model Oxide Interface'
---
[^1]
Introduction {#introduction .unnumbered}
============
Interfaces control the behavior of a variety of emergent properties in oxides, ranging from electron gas formation [@Chakhalian2014] to ferroelectricity [@Mannhart2010]. While great strides have been made in the precision synthesis of atomically-sharp thin film heterostructures [@Martin2010], interface charge [@Nakagawa2006], strain [@SankaraRamaKrishnan2014], and entropy can all drive film-substrate intermixing that can greatly affect properties. With the development of commercial spherical aberration (C$_\textrm{s}$) corrected microscopes, analytical scanning transmission electron microscopy (STEM) is now increasingly used to characterize nanoscale interfaces [@Krivanek2008]. C$_\textrm{s}$-correction has enabled the large convergence angles, small probe sizes, and high probe currents needed for efficient chemical mapping of individual atomic columns [@Lu2013; @Allen2012; @DAlfonso2010; @Muller2008]. Complementary analysis using energy-dispersive X-ray spectroscopy (STEM-EDS) and electron energy loss spectroscopy (STEM-EELS) now permits measurements of composition and chemistry at the Ångström scale. Moreover, cost reductions have now made aberration-correctors more attainable than ever, leading to their widespread adoption by universities and laboratories. While the site-specific nature of these techniques can offer rich insight into local interface environments, the interpretation of the resulting data is far from simple and still poorly understood, as discussed in a recent case study of the perovskite SrTiO$_3$ (STO) [@Kothleitner2014]. A complex array of physical processes, including beam broadening and channeling effects [@Oxley2007], can lead to serious misinterpretations of chemical maps. Channeling is particularly problematic, since it tends to occur when imaging along low-order zone axes commonly used for atomic-scale imaging; in this case, the strong Coulombic interaction between the electron probe and the atoms in the crystal focuses the probe intensity along columns, complicating the analysis of ionization signals [@Lugg2014].
The newly developed technique of atomic-column STEM-EDS mapping exemplifies the challenges associated with the quantification of high-resolution analysis of crystalline materials. In one of the first demonstrations of atomic-scale STEM-EDS mapping of STO [@DAlfonso2010], D’Alfonso *et al.* argued that the localization of X-ray scattering potentials can lead to a directly interpretable chemical map analogous to high-angle annular dark field (STEM-HAADF) imaging [@Allen2012]. However, subsequent work has shown that the contribution of thermally scattered electrons can affect contrast in both EDS and EELS maps [@Forbes2012], while studies of interfaces have revealed that the apparent atomic column size is affected by probe channeling, which directly depends on thickness [@Lu2014; @Lu2013]. Similar difficulties are encountered in STEM-EELS mapping, where image contrast reversals have been observed that are attributed to off-column channeling of the probe [@Wang2008; @Oxley2007]. Probe broadening due to a finite sample thickness can also complicate the analysis of diffuse interfaces. The few quantitative studies conducted to date have shown that the characterization of unknown structures is difficult [@Kotula2012], generally requiring extensive modeling [@Neish2015], and *a priori* sample information [@Kothleitner2014]. Promising recent work by Chen *et al.* has shown that it may be possible to quantify EDS maps on an absolute scale by combining thin samples with rigorous simulations [@Chen2016; @Chen2015].
A study of several key parameters in STEM-EDS measurements is needed to inform the growing community and raise awareness of potential sources of error. Here we consider how sample thickness and the choice of X-ray ionization edge influence measurements of interface composition and mixing in a model perovskite oxide interface. We find that measurements of the interface width can vary widely from 2–5 unit cells and that spectral components can exhibit different thickness dependencies, leading to variations in measurement error for each species. Moreover, channeling of the delocalized signal to heavier sites across the interface can complicate direct quantification of peak areas. Our study highlights important practical considerations for atomic-scale STEM-EDS mapping of interfaces, emphasizing the need for both extremely thin samples and supporting simulations to accurately interpret experimental data. We caution against direct interpretation of chemical maps absent these qualifications.
Materials and Methods {#materials-and-methods .unnumbered}
=====================
We have selected a model system consisting of a 30 nm-thick La$_{0.88}$Sr$_{0.12}$CrO$_3$ ($p$-type LSCO) thin film deposited onto a 0.1 Wt% Nb:SrTiO$_3$ (001) ($n$-type Nb:STO) substrate using molecular beam epitaxy [@Zhang2015], as illustrated in Figure \[eds\_map\](a). This system is intriguing for its potential use as an all-perovskite transparent $p-n$ junction, as will be reported elsewhere. Predictive control of device performance depends on a clear understanding of the interface structure to preserve the band offset between the two layers. TEM samples were prepared using a standard lift out method on an FEI Helios DualBeam focused ion beam (FIB) microscope. To generate a range of sample thicknesses, a wedge-shaped sample was prepared at 4–7$^{\circ}$ incidence angle, using ion beam energies of 2–30 keV. STEM measurements were performed at 200 keV using an aberration-corrected JEOL ARM-200CF microscope equipped with a JEOL Centurio silicon drift detector (quoted solid angle of 0.98 sr) for EDS analysis. All images were acquired along the \[100\] zone-axis with a $\sim 1$ Å probe size and a 27.5 mrad convergence semi-angle, yielding an approximate probe current of $\sim 130$ pA. STEM-HAADF images were acquired with 90–370 mrad inner-outer collection angles, respectively, while STEM-EDS maps were acquired in multiple regions using the Thermo Noran System 7 software, with an approximate instantaneous pixel dwell time of $50$ s px$^{-1}$, an effective total dwell time of 9–27 ms px$^{-1}$, and a total collection time of 5–17 min. Maps were processed for net counts, with a background removal and multiple linear least squares fit of reference spectra to deconvolve overlapping peaks. STEM-EELS zero-loss peak thickness measurements indicate that the measured regions are approximately 28, 33, 50, 66, 70, and 75 nm-thick.
Multislice simulations were conducted using the quantum excitation of phonons (QEP) model [@Forbes2010], which accounts for beam channeling effects and allows for separation of elastic and thermal electrons, in contrast to the frozen phonon model. Our simulations used a supercell consisting of six unit cells of TiO$_2$-terminated SrTiO$_3$ (STO) interfaced with six unit cells of La$_{0.88}$Sr$_{0.12}$CrO$_3$ (LSCO). STEM-EDS ionization maps were simulated using the STEM v4.5 software package [@Allen2015] with a $10 \times 1$ supercell tiling and $2500 \times 2160$ px grid sampling for 10, 28, 50, and 100 nm-thick crystals. The total simulation time was 25 days. Using these parameters we are able account for electrons scattered to a maximum angle of approximately 300 mrad. X-ray absorption has not been included in these models, which may impact the results for thicker samples. Crystal models were converted into the appropriate input format using the XTL-Converter program [@SpurgeonNov2015]. Chen *et al.* have shown that the size of the necessary Gaussian finite source correction can be broadened by multiple-frame averaging [@Chen2016]; therefore, to determine the most appropriate source size we calculated a range of corrections and compared them to the experimental data. This procedure yielded a FWHM = 0.19 nm as the best fit.
Results and Discussion {#results-and-discussion .unnumbered}
======================
Figures \[eds\_map\](b–c) show composite STEM-EDS maps and corresponding averaged line profiles for Sr $L_\alpha$ (1.806 kV), La $L_\alpha$ (4.650 kV), Ti $K_\alpha$ (4.508 kV), and Cr $K_\alpha$ (5.412 kV) taken from $\sim$33 and $\sim$75 nm-thick regions, respectively. The profiles were extracted from the net X-ray counts maps, averaged in the film plane, and smoothed using an adjacent averaging filter. For comparison to simulation, we propose the use of a logistic fit to the extrema on either side of the interface; we define the interface width ($\delta$) as the difference of 90% and 10% of the signal maxima. Using this procedure we measure the interface width as a function of sample thickness, as shown in Figure \[eds\_map\](d). Even though all measurements are conducted on the same crystal, we find that there is a clear increase in the apparent interface width in thicker mapping regions. $\delta_\textrm{La}$ by far exhibits the largest change in interface width, increasing nearly fourfold by $\sim 1.8$ nm moving from 28 to 75 nm-thick regions. $\delta_\textrm{Sr}$ and $\delta_\textrm{Cr}$ increase by $\sim 0.86$ and $\sim 0.96$ nm, respectively, while $\delta_\textrm{Ti}$ increases by $\sim 0.87$ nm. Depending on the species, these values correspond to nearly 2–5 unit cells of measurement error. We find that the absolute value of $\delta_\textrm{Ti}$ is generally larger than the other species in the thinnest regions; significant Ti interdiffusion is not unexpected in this system and may actually promote a more stable interface [@Colby2013]. We note that atomic force microscope (AFM) maps of the substrate surface show that at most one step can be contained in the foil along the beam direction, which could also affect the absolute value of the interface width in any MBE-grown samples. In addition to the broadening of interface profiles, we also find that the peak-to-background ratios of the profiles significantly decrease with increasing thickness, as evidenced by comparing Figures \[eds\_map\](b–c). These results are troubling and suggest that the accuracy of interface measurements, as well as the quantification of resulting peak areas, can greatly depend on sample thickness.
![(a) Illustration of the STO / LSCO crystal structure. (b–c) Unfiltered composite STEM-EDS maps and corresponding $A$- and $B$-site net X-ray count line profiles for $\sim$33 and $\sim$75 nm-thick STO / LSCO interfaces, respectively. The line profiles have been averaged in the plane of the maps. (d) Interface width as a function of sample thickness for each edge. \[eds\_map\]](eds.png){width="\textwidth"}
To gain insight into the effects of delocalization and channeling on the resulting ionization maps, we have performed multislice simulations, systematically varying the thickness of a model interface. We consider an ideal abrupt STO / La$_{0.88}$Sr$_{0.12}$CrO$_3$ interface, thereby avoiding the complications of modeling different interface geometries. We have simulated ionization maps for 10, 28, 50, and 100 nm-thick samples; selected ionization maps and line profiles for the 10 and 50 nm-thick samples are shown in Figures \[multislice\](a-b), respectively, alongside a plot of interface width *versus* crystal thickness in Figure \[multislice\](c). A comparison of the maps in Figures \[multislice\](a-b) shows that the 50 nm-thick sample exhibits a higher background than the 10 nm-thick sample, as indicated by the diffuse bands of contrast between lattice sites. Figure \[multislice\](a) shows line profiles taken from both the $A$- and $B$-site columns of the 10 nm-thick sample; here we find that the ionization signal of all four edges is strongly localized to their respective atomic columns, with a low background between peaks. This case corresponds to the ultrathin limit previously described by Lu *et al.* [@Lu2013], where the electron wave is, in principle, directly convolved with the local EDS ionization potential and is less affected by beam channeling. Using the same logistic fitting procedure we measure the following interface widths: $\delta_\textrm{Sr} = 0.20$, $\delta_\textrm{Ti} = 0.18$, $\delta_\textrm{La} = 0.20$, and $\delta_\textrm{Cr} = 0.18$ nm (all $\pm \, 0.01$ nm). While our simulations cannot account for all factors, such as thickness or strain fluctuations across the interface, we estimate that this result represents the lower limit of intermixing measurements in an extremely thin and abrupt interface; even in the case of such an ideal sample, there is still some artificial intermixing.
![(a–b) Simulated STEM-EDS ionization maps and corresponding $A$- and $B$-site line profiles for abrupt 10 and 50 nm-thick STO / LSCO interfaces, respectively. The line profiles have been normalized to the signal maxima and are inset with a $10 \times$ magnified signal to emphasize channeling effects. (c) Interface width as a function of model thickness for each edge.\[multislice\]](multislice.png){width="\textwidth"}
Line profiles from the 50 nm-thick model, shown in Figure \[multislice\](b), reveal very different behavior from the 10 nm-thick case. The background between lattice sites has significantly increased, resulting from the redistribution of probe intensity within the sample caused by channeling. Owing to the intrinsic delocalization of lower energy edges, the background for the Sr $L_\alpha$ and La $L_\alpha$ edges is nearly 75% of the signal maximum, while the background for the Ti $K_\alpha$ and Cr $K_\alpha$ edges is less than 50% of the maximum. The choice of ionization edge may therefore affect the spatial distribution of EDS ionization potential, as has been observed in STEM-EELS mapping [@Wang2008], as well as the accuracy of composition quantification based on peak area fitting [@Lu2014]. Furthermore, the $A$-site line profiles show that a substantial fraction ($10-15\%$) of the signal can be delocalized across even a perfectly abrupt interface. We find that the La $L_\alpha$ signal is delocalized across the interface to Sr positions, as marked by the arrows in Figure \[multislice\](b); a similar, albeit much reduced, effect is seen for the $B$-sites. This delocalization and subsequent channeling to heavier sites can arise from both thermally and elastically scattered electrons, which are then able to go on to ionize other atoms [@Forbes2012]. As expected, delocalization results in a sizable increase in the apparent interface width, yielding: $\delta_\textrm{Sr} = 0.54 \pm 0.02$, $\delta_\textrm{Ti} = 0.26 \pm 0.01$, $\delta_\textrm{La} = 0.60 \pm 0.03$, and $\delta_\textrm{Cr} = 0.24 \pm 0.01$ nm. We again find that delocalization is much more pronounced for the Sr $L_\alpha$ and La $L_\alpha$ signals, tripling $\delta$ from $0.20$ to $0.54-0.60$ nm. This width corresponds to more than a unit cell of artificial intermixing (measurement error), even in the case of a perfectly abrupt interface structure. On the other hand, the Ti $K_\alpha$ and Cr $K_\alpha$ signals show only a negligible increase in artificial interface width compared to the 10 nm-thick case.
A comparison to our experiments shows that, even though the measured profiles are wider than the simulations, there is a consistent trend toward an artificially broadened interface in thicker regions. In agreement with our simulations, we find that the La $L_\alpha$ and Sr $L_\alpha$ signals exhibit the largest increase in broadening, but the other species exhibit similar trends. Most importantly, we observe that the absolute value of the interface width can vary as much as 2–5 unit cells, making it difficult to extract a meaningful picture of the interface in thicker mapping regions. While such regions may only be 40–50 nm thick, that is more than sufficient to introduce noticeable measurement errors. In summary, our results highlight two main difficulties faced in the quantification of interfaces: electron probe channeling can greatly alter the on- and off-column signal, leading to sizable artificial intermixing over several unit cells; further, the measurement error also depends on the intrinsic delocalization associated with a chosen ionization edge.
Conclusions {#conclusions .unnumbered}
===========
In light of these results, great care must be taken to conduct meaningful studies of interfaces using the aberration-corrected STEM-EDS technique. We find that delocalization effects are minimized in the thinnest regions and that the best results are obtained below 25–30 nm; however, this imposes a severe limit on sample preparation and signal collection times for accurate quantification. We also emphasize that multislice simulations should be conducted for each system and ionization edge of interest to aid the interpretation of experimental data. Future GPU-accelerated computing will allow us to produce more rigorous models to better simulate real-world interfaces and help us account for complex delocalization processes. For now it is imperative that novice users understand the limitations of chemical mapping and that steps are taken to prepare sufficiently thin samples, depending on the level of accuracy desired.
Acknowledgments {#acknowledgments .unnumbered}
===============
S.R.S. thanks Drs. Scott Findlay, Yuanyuan Zhu, Despoina Kepaptsoglou, and Matthew Olszta for insightful discussions. This work was supported by the U.S. Department of Energy (DOE), Office of Science, Division of Materials Sciences and Engineering under award \#10122. This work was performed in the Environmental Molecular Sciences Laboratory (EMSL), a national science user facility sponsored by the DOE Office of Biological and Environmental Research (BER) and located at Pacific Northwest National Laboratory (PNNL), operated for DOE by Battelle. Multislice calculations were performed using PNNL Institutional Computing.
[^1]: Phone:+1-509-371-6517; Fax:+1-509-371-6066
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a neural embedding algorithm called Network Vector, which learns distributed representations of nodes and the entire networks simultaneously. By embedding networks in a low-dimensional space, the algorithm allows us to compare networks in terms of structural similarity and to solve outstanding predictive problems. Unlike alternative approaches that focus on node level features, we learn a continuous global vector that captures each node’s global context by maximizing the predictive likelihood of random walk paths in the network. Our algorithm is scalable to real world graphs with many nodes. We evaluate our algorithm on datasets from diverse domains, and compare it with state-of-the-art techniques in node classification, role discovery and concept analogy tasks. The empirical results show the effectiveness and the efficiency of our algorithm.'
author:
- |
Hao Wu\
USC ISI\
[email protected] Kristina Lerman\
USC ISI\
[email protected]
bibliography:
- 'hao.bib'
title: 'Network Vector: Distributed Representations of Networks with Global Context'
---
Introduction
============
Applications in network analysis, including network pattern recognition, classification, role discovery, and anomaly detection, among others, critically depend on the ability to measure similarity between networks or between individual nodes within networks. For example, given an email communications network within an enterprise, one may want to classify individuals according to their functional roles, or given one individual, find another one playing a similar role. Computing network similarity requires going beyond comparing networks at a node level to measuring their structural similarity. To characterize network structure, traditional approaches extract features such as node degrees, clustering coefficients, eigenvalues, the lengths of shortest paths and so on [@berlingerio2012netsimile; @henderson2012rolx; @gilpin2013guided]. However, these hand-crafted features are usually heterogeneous and it is often not clear how to integrate them within a learning framework. In addition, some graph features, such as eigenvalues, are computationally expensive and do not scale well in tasks involving large networks. Recent advances in distributed representation of nodes [@perozzi2014deepwalk; @tang2015line; @grover2016node2vec] in networks created an alternate framework for unsupervised feature learning of nodes in networks. These methods are based on the idea of preserving local neighborhoods of nodes with neural network embeddings. An objective is defined on the proximity between nodes in exploring the network neighborhood with various strategies, mainly Depth-First Search (DFS) and Breadth-First Search (BFS). The objective is optimized using single layer neural network for efficient training. However, the embeddings used for feature representation limit scope to the local context of nodes, without directly exploiting the global context of the network. To represent the whole network, these approaches require us to integrate the representations of all nodes, for example, by averaging their representations. However, not all nodes contribute equally to the global representation of the network, and in order to account for their varying importance, aggregation schemes need to weigh nodes, which adds an extra layer of complexity to the learning task.
To address above-mentioned challenge we describe a neural network algorithm called Network Vector, which learns distributed representations of networks that account for their global context. The algorithm is scalable to real world networks with large numbers of nodes and can be applied to generic networks such as social networks, knowledge graphs, and citation networks. Networks are compressed into real-valued vectors that preserve the network structure, so that the learned distributed representations can be used to effectively measure network similarity. Specifically, given two networks, even those with different size and topology, the distance between the learned vector representations can be used to measure their structural similarity. In addition, this approach allows us to compare individual nodes by looking at the similarity of their ego-networks, i.e., networks that contain the focal node and all their neighbors and connections between them. Our approach is inspired by Paragraph Vector [@le2014distributed] that learns distributed representations of texts of variable length such as sentences and documents [@le2014distributed]. By exchanging the notions of ordered “word" sequences in sentences and “nodes" in paths along edges on networks. We learn network representations in a similar way of learning representations of sentences and documents. Specifically, we sample sequences of nodes from a network using random walks, same as in [@perozzi2014deepwalk; @grover2016node2vec]. In contrast to existing approaches, the likelihood of next node in a random walk sequence predicted by our algorithm depends not only on the previous nodes, but also on the global context of the network. The global context vector representation of the network is learned to maximize the average predicted likelihood of nodes in random walk sequences sampled from the network. The learned representations can be used as the signatures of the networks for comparison, or as features for classification and other predictive tasks.
We evaluate the algorithm on several real world datasets from a diversity of domains, including citation network of knowledge concepts in Wikipedia, email interaction network, legal citations network, social network of bloggers, protein-protein interaction network and language network. We focus on predictive tasks including role discovery in networks that aims to identify individual nodes serving similar roles, inference of analogous relations between concept pairs in Wikipedia and multi-label node classification. We compare Network Vector with state-of-the-art feature learning algorithm node2vec [@grover2016node2vec], LINE [@tang2015line], DeepWalk [@perozzi2014deepwalk] and feature-based baselines such as node degrees, clustering coefficients and eigenvalues. Experiments demonstrate the superior performance of Network Vector, due to its capacity of learning the global context of the network.
In summary, our contributions are summarized as follows:
1. We propose Network Vector algorithm, a distributed feature learning algorithm for representing an entire network and its nodes simultaneously. We define an objective function that preserves the local neighborhood of nodes and the global context of the entire network.
2. We evaluate Network Vector on role discovery, concept analogy and node multi-label classification tasks. Experiments on several benchmark datasets from diverse domains show its effectiveness and efficiency.
Related Work
============
Our algorithm builds its foundation on learning distributed representations of concepts [@hinton1986learning] . Distributed representations encode structural relationships between concepts and are typically learned using back-propagation through neural networks. Recent advances in natural language processing have successfully adopted distributed representation learning and introduced a family of neural language models [@bengio2003neural; @mnih2007three; @mikolov2010recurrent; @mikolov2013efficient; @mikolov2013distributed] to model word sequences in sentences and documents. These approaches embed words such that words in similar contexts tend to have similar representations in latent space. By exchanging the notions of nodes in a network and words in a document, recent research [@perozzi2014deepwalk; @tang2015line; @cao2015grarep; @grover2016node2vec] attempt to learn node representations in a network in a similar way of learning word embeddings in neural language models. Our work follows this line of approaches in which nodes in a neighborhood will have similar embeddings in vector space. Different node sampling strategies are explored for characterizing the neighborhood structure. For example, DeepWalk [@perozzi2014deepwalk] samples node sequences from a network using a stream of short first-order random walks, and model them just like word sequences in documents using neural embeddings. LINE [@tang2015line] samples nodes in pairwise manner and model the first-order and second-order proximity between them. GrapRep [@cao2015grarep] extends LINE to exploit structural information beyond second-order proximity. To offer a flexible node sampling scheme, node2vec [@grover2016node2vec] utilizes second-order random walks, and combines Depth-First Search (DFS) and Breadth-First Search (BFS) strategies to explore the local neighborhood structure.
However, existing approaches only consider the local network structures (i.e., the neighborhoods of nodes) in learning node embeddings, but exploit little information of the global structure of the network. Although recent approach GrapRep [@cao2015grarep] attempts to capture long distance relationship between two different nodes, it limits scope to a fixed number of hops. More importantly, existing approaches focus on node representations, and it requires additional effort to compute the representation of the entire network. The simple scheme of averaging the representations of all nodes to represent the network is by no means a good choice as it ignores the statistics of node frequency and their roles in the network. In contrast, we introduce a notion called the *global network vector*, which aims to represent the structural properties of an entire network. The global vector representation of the network acts as a memory which is asked to contribute to the prediction of a node accompanying with the node’s neighbors, and updated to maximize the predictive likelihood. As a result, our algorithm can simultaneously learn the global representation of a network and the representations of nodes in the network. This is inspired by Paragraph Vector [@le2014distributed], which learns a continuous vector to represent a piece of text with variable-length, such as sentences, paragraphs and documents.
Network Vector
==============
We consider the problem of embedding nodes of a network and the entire network into a low-dimensional vector space. Let $G = \{V, E\}$ denote a graph, and $V$ is the set of vertices and $E = V \times V$ is the set of edges with weights $W$. The goal of our approach is to map the entire graph to a low-dimensional vector, represented by ${\bf{v}}_G \in \mathbb{R}^d$, and map each node $i$ to a unique vector ${\bf{v}}_i \in \mathbb{R}^d$ in the same vector space. Although the dimensionality of network representation ${\bf{v}}_G$ can be different from that of node representations ${\bf{v}}_i$ in theory, we adopt the same dimensionality $d$ for the ease of computation in real world applications. Suppose that there are $M$ graphs given (e.g., ego-networks of $M$ persons of interest in a social network) and $N$ distinct nodes in the corpus, then there are $(M + N) \times d$ parameters to be learned.
A Neural Architecture
---------------------
Our approach of modeling networks is motivated by learning distributed representations of variable-length texts, e.g., sentences and documents [@le2014distributed]. The concept of “words" in a document [@le2014distributed] is replaced by “nodes" in a network in our modeling. The goal is to predict a node given other nodes in its local context as well as the global context of the network. Text has a linear property that the local context of a word can be naturally defined by surrounding words in ordered sequences. However, networks are not linear. In order to characterize the local context of a node, without loss of generality, we sample node sequences from the given network with second-order random walks in [@grover2016node2vec], which offer a flexible notion of a node’s local neighborhood by combining Depth-First Search (DFS) and Breadth-First Search (BFS) strategies. Our learning framework can easily adopt higher-order random walks, but with higher computation cost. Each random walk starts from an arbitrary root node and generates an ordered sequence of nodes with second-order Markov chains. Specifically, consider node $v_a$ that has been visited in the previous step, and the random walk currently reaches node $v_b$. Consecutively, the next node $v_c$ will be sampled in random walks, with probability: $$\label{eq:transition_prob}
P(v_c | v_a, v_b) = \frac{1}{Z} M_{bc}^a W_{bc}$$ where $M_{bc}^a$ is the unweighted transition probability of moving from node $v_b$ to $v_c$ given $v_a$, $W_{bc}$ is the weight of edge $(v_b, v_c)$, and $Z$ is the normalization term. We define $M_{bc}^a$ as: $$M_{bc}^a = \begin{cases}
\frac{1}{p} &\text{if}\ d(v_a, v_c) = 0\\
1 &\text{if}\ d(v_a, v_c)=1\\
\frac{1}{q} &\text{if}\ d(v_a, v_c)=2
\end{cases}$$ where $d(v_a, v_c)$ is the shortest path distance between $v_a$ and $v_c$. The parameters $p$ and $q$ control how the random walk biases toward visited nodes in previous step and nodes that are further away. The random walk terminates when $l$ vertices are sampled, and the procedure repeats $r$ times for each root node.
Figure \[fig:networkvector\] illustrates a neural network architecture for learning the global network vector and node vectors simultaneously. A sliding window $(v_1, \cdots, v_n)$ with fixed-length $n$ is repeatedly sampled over node sequences. The algorithm predicts the target node $v_n$ given preceding nodes $v_{1:n-1}$ as local context and the entire network $G$ as the global context, with probability $P(v_n | v_{1:n-1}, G)$. Formally, the probability distribution of a target node is defined as: $$\label{eq:objective}
P( v_n | v_{1:n-1}; G) = \frac {1} {Z_c} \exp [ -E( G, v_{1:n-1}; v_n) ]$$ where ${Z_c} = \sum_{v_m \in V} \exp [ -E( G, v_{1:n-1}; v_m) ] $ is the normalization term. We extend the scalable version of Log-Bilinear model [@mnih2007three], called vector Log-Bilinear model (vLBL) [@mnih2013learning]. In our model, the energy function $E(G, v_{1:n-1}; v_n)$ is specified as: $$E( G, v_{1:n-1}; v_n) = -\hat{\mathbf{v}}^\top\mathbf{v}_{n}$$ where $\hat{\mathbf{v}}$ is the predicted representation of the target node: $$\hat{\mathbf{v}} = {\bf{v}}_G + \left(\sum_{i=1}^{n-1} \mathbf{c}_i \odot \mathbf{v}_i \right)$$ Here $\odot$ denotes the Hadamard (element-wise) product, and $\mathbf{c}_i$ is the weight vector for the context node in position $i$. $\mathbf{c}_i$ parameterizes the context nodes at different hops away from the target node in random walks. The global network vector ${\bf{v}}_G$ is shared across all sliding windows of node sequences. After being trained, the global network vector preserves the structural information of the network, and can be used as feature input for the network. In our model, in order to impose symmetry in feature space of nodes, and activate more interactions between the feature vector ${\bf{v}}_G$ and the node vectors, we use the same set of feature vectors for both the target nodes and the context nodes. This is different from [@mnih2013learning], where two separated sets of representations are used for the target node and the context nodes respectively. In practice, we find our approach improves the performance of Network Vector.
Learning with Negative Sampling
-------------------------------
The global network vector $\mathbf{v}_G$, the node vectors $\mathbf{v}_i$ and the position-dependent context parameters $\mathbf{c}_i$ are initialized with random values, and optimized by maximizing the objective in Eq. (\[eq:objective\]). Stochastic gradient ascent is performed to update the set of parameters $\boldsymbol \theta = \{\mathbf{v}_G, \mathbf{v}_i, \mathbf{c}_i\}$: $$\label{eq:gradient}
\begin{split}
\Delta \boldsymbol{\theta} &= \epsilon {\nabla_{\boldsymbol{\theta}}} \log P(v_n|v_{1:n-1}, G) \\
& = \epsilon \frac{\partial}{\partial \boldsymbol \theta} \left[\frac{\exp(\hat{\mathbf{v}}^\top\mathbf{v}_{n})} {\sum_{v_m=1}^N \exp(\hat{\mathbf{v}}^\top\mathbf{v}_{m})}\right]
\end{split}$$ where $\epsilon$ is the learning rate. The computation involves the normalization term and is proportional to the number of distinct nodes $N$. The complexity of computation is expensive and impractical in real applications. In our approach, we adopt negative sampling [@mikolov2013distributed] for optimization. Negative sampling represents a simplified version of noise contrastive estimation [@mnih2012fast], and trains a logistic regression to distinguish between data samples of ${v}_n$ from “noise" distribution. Our objective is to maximize $$\label{eq:ng}
\log \sigma (\hat{\mathbf{v}}^\top\mathbf{v}_{n}) + \sum_{m=1}^k \mathbb{E}_{v_m} \sim P_n (v) \left[\log \sigma (-\hat{\mathbf{v}}^\top\mathbf{v}_{m}) \right]$$ where $\sigma (x) = 1 / (1 + \exp(-x)) $ is the sigmoid function. $P_n (v) $ is the global unigram distribution of the training data acting as the noise distribution where we draw $k$ negative samples of nodes. Negative sampling allows us to train our model efficiently that no longer requires explicitly normalized in Eq. (\[eq:gradient\]), and hence are more scalable.
An Inverse Architecture
-----------------------
The architecture in Figure \[fig:networkvector\] utilizes the linear combination of the global network vector and the context node vectors to predict the target node in a sliding window. Another way of training the global network vector is to model the likelihood of observing a sampled node $v_t$ from the sliding window conditioned on the feature vector ${\bf{v}}_G$, given by $$\label{eq:inverse_probability}
P( v_t | G) = \frac {1} {Z_G} \exp [ -E( G; v_t) ]$$ where $Z_G$ is the normalization term specific to the feature representation of $G$. The energy function $E( G; v_t)$ is: $$E( G; v_t) = -{\mathbf{v}}_G^\top\mathbf{v}_{t}$$ This architecture is a counterpart of the Distributed Bag-of-Words version of Paragraph Vector [@le2014distributed]. However, this architecture ignores the order of the nodes in the sliding window and perform poorly in practice when it is used alone. We extend the framework by simultaneously training network and node vectors using a Skip-gram [@mikolov2013efficient; @mikolov2013distributed] like model. The model additionally maximizes the likelihood of observing the local context $v_{{t-n}:{t+n}}$ (excluding $v_t$) for the target node $v_t$, conditioned on the feature representation of $v_t$. Unfortunately, modeling the joint distribution of a set of context nodes is not tractable. This problem can be relaxed by assuming the node in different context positions are conditionally independent given the target word: $$P(v_{{t-n}:{t+n}} | v_t) = \prod_{i=t-n}^{t+n} P(v_i | v_t)$$ where $P(v_i | v_t) = \frac {1} {Z_t} \exp [ -E(v_t; v_i) ]$. The energy function is: $$E(v_t; v_i) = \mathbf{v}_t^\top (\mathbf{c}_i \odot \mathbf{v}_i)$$ The objective is to maximize the log-likelihood of the product of the probabilities, $P( v_t | G) $ and $P(v_{{t-n}:{t+n}} | v_t)$
Complexity Analysis
-------------------
The computation of Network Vector consists of two key parts: sampling of node sequences with random walks and optimization of vectors. For each node sequence of fixed length $l$, we start from a randomly chosen root node. At each step, the walk visits a new node based on the transition probabilities $P(v_c | v_a, v_b) $ in Eq. (\[eq:transition\_prob\]). The transition probabilities $P(v_c | v_a, v_b) $ can be precomputed and stored in memory using $O(|E|^2/|V|)$ space. Sampling a new node in the walk can be efficiently done in $O(1)$ time using alias sampling [@walker1977efficient]. The overall time complexity is $O(r|V|l)$ for repeating $r$ times of random walks of fixed length $l$ by taking each node as root.
The time complexity of optimization with negative sampling in Eq. (\[eq:ng\]) is proportional to the dimensionality of vectors $d$, the length of context window $n$ and the number of negative samples $k$. It takes $O(dnk)$ time for nodes within the sliding window $(v_1, \cdots, v_n)$. The introduced global vector ${\bf{v}}_G$ requires $O(dk)$ time to optimize, same as any other node vectors in the sliding window. Given $r$ random walks of fixed length $l$ starting from every node, the overall time complexity is $O(dnkr|V|l)$. To store the node vectors and the global network vector, it requires $O(d|V|+d)$ space.
The Property of Network Vector
------------------------------
The property of the global network vector ${\bf{v}}_G$ in the architecture (as shown in Figure \[fig:networkvector\]) can be explained by looking at the objective in Eq. (\[eq:objective\]). ${\bf{v}}_G$ is part of the input to the neural network, and can be viewed as a term that helps to represent the distribution of the target node $v_n$. The relevant part ${\bf{v}}_G^\top{\bf{v}}_n$ is related logarithmically to the probability $P( v_n | v_{1:n-1}; G)$. Therefore, the more frequently a particular $v_n$ is observed in the data, the larger the value ${\bf{v}}_G^\top{\bf{v}}_n$ will have, and hence ${\bf{v}}_G$ will be closer to ${\bf{v}}_n$ in vector space. The training objective is to maximize the logarithm of the product of all probabilities $P( v_n | v_{1:n-1}; G)$, and the value is related to ${\bf{v}}_G^\top{\bar{\bf{v}}}_n$, where ${\bar{\bf{v}}}_n$ is the expected vector that can be obtained by averaging all observed ${\bf{v}}_n$ in the data. It is also true for Eq. (\[eq:inverse\_probability\]) in the inverse architecture where the global network vector ${\bf{v}}_G$ is the only input to the neural network, in order to predict every node $v_t$.
### Karate Network
As an illustrative example, we apply Network Vector to the classic Karate network [@zachary1977information]. The nodes in the network represent members in a karate club, and the edges are social links between the members outside the club. There are 34 nodes and 78 undirected edges in total. We use the inverse architecture to train the vectors. Figure \[fig:karateexample\] shows the output of our method in two dimensional space. We use green circles to denote nodes and orange circle to denote the entire graph. The size of a node is proportional to its degree in the graph. We can see that the learned global network vector is close to these high-degree nodes, such as node 1 and 34, which serve as the hubs of two splits of the club. The resulting global vector mostly represent the backbone nodes (e.g., hubs) in the network and compensates the lack of global information in local neighborhoods.
\[!ht\]
[|c|c|c|c|c]{} **ID** &**Title (with url)** &**Page** &**Year** &\
93272 &[ [Chicago & Grand Trunk Ry. Co. v. Wellman](https://supreme.justia.com/cases/federal/us/143/339/) ]{}&143 U.S. 339 &1892 &\
99622 &[ [F. S. Royster Guano Co. v. Virginia](https://supreme.justia.com/cases/federal/us/253/412/) ]{} &253 U.S. 412 &1920 &\
103222 &[ [Coleman v. Miller](https://supreme.justia.com/cases/federal/us/307/433/)]{} &307 U.S. 433 &1939 &\
109380 &[ [Buckley v. Valeo](https://supreme.justia.com/cases/federal/us/424/1/)]{} &424 U.S. 1 &1976 &\
118093 &[ [Arizonans for Official English v. Arizona](https://supreme.justia.com/cases/federal/us/520/43/)]{} &520 U.S. 43 &1997 &\
110578 &[ [Ridgway v. Ridgway](https://supreme.justia.com/cases/federal/us/454/46/)]{} &454 U.S. 46 &1981 &\
91704 &[ [Yick Wo v. Hopkins](https://supreme.justia.com/cases/federal/us/118/356/)]{} &118 U.S. 356 &1886 &\
98094 &[ [Weeks v. United States](https://supreme.justia.com/cases/federal/us/232/383/)]{} &232 U.S. 383 &1914 &\
101741 &[ [Stromberg v. California](https://supreme.justia.com/cases/federal/us/283/359/)]{} & 283 U.S. 359 &1931 &\
101957 &[ [Powell v. Alabama](https://supreme.justia.com/cases/federal/us/287/45/)]{} &287 U.S. 45 &1932 &\
103050 &[ [Johnson v. Zerbst](https://supreme.justia.com/cases/federal/us/304/458/)]{} &304 U.S. 458 &1938 &\
105547 &[ [Roth v. United States](https://supreme.justia.com/cases/federal/us/354/476/)]{} &354 U.S. 476 &1957 &\
### Legal Citation Networks
Given a citation network of documents, for example, scientific papers or legal opinions, we want to identify similar documents. These could be groundbreaking works that serve to open new fields of discourse in science and law, or foundational works that span disciplines but have less of an impact on discourse, such as “methods” papers in science.
For the purpose of case study, we collected a large digitized record of federal court opinions from the CourtListener project[^1]. The most cited legal decisions from the United States Supreme Court are selected and ego-networks of citations are constructed for these legal cases. Two distinct graph patterns are observed. One is “Citations have a few giant hubs" and “Citations are well connected". We list a few examples in Table \[table:legal\], where the titles of the cases with different citation patterns are colored as red and blue, respectively. The ego-networks of the first six cases listed in Table \[table:legal\] have just a few giant hubs which are linked by many other cases. For example, the case “Buckley v. Valeo, 424 U.S. 1 (1976)" is a landmark decision in American campaign finance law. The case “Coleman v. Miller, 307 U.S. 433 (1939)" is a landmark decision centered on the Child Labor Amendment, which was proposed for ratification by Congress in 1924. These cases are generally centered on a specific topic, and their citations may have a narrowed topic. There are only a few hubs cited frequently by others and the citations generally do not cite each other. On the other side, the ego-networks of the last six cases listed in Table \[table:legal\] have citations that are well connected. For example, “Yick Wo v. Hopkins, 118 U.S. 356 (1886) " was the first case where the United States Supreme Court ruled that a law that is race-neutral on its face, but is administered in a prejudicial manner; The case “Stromberg v. California, 283 U.S. 359 (1931)" is a landmark in the history of First Amendment constitutional law to include a protection of the substance of the First Amendment. These cases are influential in the history and cited by many diverse subsequent legal decisions, which usually cite each other.
Our Network Vector algorithm is used to learn two-dimensional embeddings from the ego-networks of the legal cases, and their projections are shown as open dots in the right figure of Table \[table:legal\]. The structures of the ego-networks for four sampled Supreme Court legal cases (Case IDs: 110578, 93272,101957,105547) are also illustrated in the figure. Note that the ego network includes the case itself (does not show in the figure), and all the cases it cites (smaller circle on the right) as well as all the cases that cite it (larger circle on the left). Lines represent citations among these cases. The two groups of ego-networks contrast each other. Compared to the ego-networks in red boxes, which is cited by unrelated legal cases, there is clearly more coherence in discourse related to the cases in blue boxes, as indicated by citations among other Supreme Court cases that cite this one. Although the differences between these two ego-networks could be captured in a standard way, by features related to the degree distribution of the ego-networks, or their clustering coefficients, the distinctions between other ego-networks may be more subtle necessitating a new approach for evaluating their similarity. In this representation, the position of the case in the learned space captures the similarity of the structure of their ego-networks. Cases that are more similar to the ego-networks in red boxes fall in the top half of the 2-D plane (red open dots); while cases similar to those in blue boxes fall in the bottom half (blue open dots). Thus, distances between the learned representations of the ego-networks of legal cases can be used to quantitatively capture their similarity.
Experiments
===========
Network Vector learns representations of network nodes and the entire network simultaneously. We evaluate both representations on predictive tasks. First, we apply Network Vector to a setting where only local information about nodes, such as their immediate neighbors, is available. We learn representations for ego-networks of a few nodes using Network Vector and evaluate on role discovery in social networks and concept analogy in encyclopedia. Second, when the information of node connectivities in the entire network is available, we may learn node representations using Network Vector, where the additional global vector for the network is used to help in learning high-quality node representations. The resulting node representations are evaluated on multi-label classification.
--------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
Method p@1 p@5 p@10 p@1 p@5 p@10 p@1 p@5 p@10
Degrees+Clustering+Eigens 0.160 0.149 0.146 0.090 0.102 0.083 0.210 0.200 0.196
node2vec 0.231 0.224 0.218 **0.290** 0.280 0.268 0.500 **0.498** 0.474
Network Vector **0.607** **0.560** **0.522** **0.290** **0.298** **0.281** **0.520** **0.498** **0.483**
--------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
Role Discovery
--------------
Roles reflect individuals’ functions within social networks. For example, email communication network within an enterprise reflects employees’ responsibilities and organizational hierarchies. An engineer’s interactions with her team are different from those of a senior manager’s. In the Wikipedia network, each article cites other concepts that explain the meaning of the article’s concept. Some concepts may “bridge" the network by connecting different concept categories. For example, the concept *Bat* belongs to the category *Mammals*, however since a bat resembles a bird, it refers to many similar articles about the category *Birds*.
### Datasets
We use the following datasets in the evaluation:
- [Enron Email Network:]{} It contains email interaction data from about 150 users, mostly senior management of Enron. There are about half million emails communicated by 85,601 distinct email addresses[^2]. We have 362,467 links left after removing duplicates and self links. Each of the email addresses belonging to Enron employees has one of 9 different positions: *CEO*, *President*, *Vice President*, *Director*, *Managing Director*, *Manager*, *Employee*, *In House Lawyer* and *Trader*. We use the positions as roles. This categorization is fine-grained. In order to understand how the feature representations can reflect the properties of different stratum in the corporation, we also use coarse-grained labels *Leader* (aggregates *CEO*, *President*, *Vice President*), *Manager* (aggregates *Director*, *Managing Director*, *Manager*) and *Employee* (includes *Employee*, *In House Lawyer* and *Trader*) to divide the users into 3 roles.
- [Wikipedia for Schools Network:]{} We use a subset of articles available at Wikipedia for Schools[^3]. This datasetcontains 4,604 articles and 119,882 links between them. The articles are categorized by subjects. For example, the article about *Cat* is categorized as *subject.Science.Biology.Mammals*. We use one of 15 second-level category names (e.g., *Science* in the case of *Cat*) as the role label.
### Methods for Comparison
For real-world networks, such as email, information about all connectivities of nodes may not be fully available, e.g., for privacy reasons. For this reason, we explore prediction task with local information (i.e., immediate neighbors). For each node, we first generate its ego-network, which represents the induced subgraph of its immediate neighbors, and learn global vector representations for the set of ego-networks through Network Vector. We use the architecture as in Eq. (\[eq:objective\]). In our experiments, we repeat $\gamma=10$ times for root node initialization in random walks and the length of each random walks is fixed as $l=80$. For comparison, we evaluate the performance of Network Vector against the following network feature-based algorithms [@berlingerio2012netsimile]:
- [Degrees]{}: number of nodes and edges, average node degree, maximum “in" and “out" node degrees. The degree features are aggregated to form the representations of the ego-networks.
- [Clustering Coefficients]{}: measure the degree to which nodes tend to cluster. We compute global clustering coefficient and average clustering coefficient of nodes for representing each ego-network.
- [Eigens]{}: For each ego-network, we compute 10 largest eigenvalues of its adjacency matrix.
- [node2vec]{} [@grover2016node2vec]: This approach learns low-dimensional feature representations of nodes in a network by interpolating between BFS and DFS for sampling node sequences. A parameter $p$ and $q$ is introduced to control the likelihood of revisiting a node in walks, and to dis/encourage outward exploration, resulting in BFS/DFS like sampling strategy. It’s interesting to note when $p=1$ and $q=1$, node2vec boils down to DeepWalk [@perozzi2014deepwalk], which utilizes uniform random walks. We adapt node2vec, and use the mean of learned node vectors to represent each ego-network.
### Results
Given a node’s ego-network, we rank other nodes’ ego-networks by their distance to it in vector space of feature representations. Table \[tb:rolediscovery\] shows the average precision of retrieved nodes with the same roles (class labels) at cut-off $k=1, 5, 10$. For simplicity, Cosine similarity is used to compute the distance between two nodes. From the result, we can see how the global context allows Network Vector outperform node2vec in role discovery. However, the performance gain is dependent on different datasets. We observe Network Vector performs slightly better than node2vec on Enron email interaction network, while the improvement of performance is over 150% on Wikipedia network. Compared to the combination of Degrees, Clustering Coefficients and Eigenvalues, the improvement of the two learning algorithms Network Vector and node2vec are outstanding, with over 100% performance gain in all cases.
Concept Analogy
---------------
We also evaluate the feature representations of ego-networks on the analogy task. For Wikipedia network, we follow the word analogy task defined in [@mikolov2013efficient]. Given a pair of Wikipedia articles describing two concepts $(a,b)$, and an article describing another concept $c$. The task aims to find a concept $d$ such that $a$ is to $b$ as $c$ is to $d$. For example, *Europe* is to *euro* as *USA* is to *dollar*. This analogy task can be solved by finding the concept that is closest to ${\bf{v}}_b - {\bf{v}}_a + {\bf{v}}_c$ in vector space, where the distance is computed using Cosine similarity.
There are 1,632 semantic tuples in Wikipedia network matched for the semantic pairs in [@mikolov2013efficient]. We use them as evaluation benchmark. Table \[tb:analogy\] shows the accuracy of hitting the answer $d$ within cut-off $k=1, 5, 10$ positions in the ranking list. From the results, we can see Network Vector performs much better than the baseline, which use degree, clustering coefficients and eigenvalues of the adjacency matrix. The combination of heterogeneous features (degrees, clustering coefficients and eigenvalues) in different scale causes the difficulty to utilize an efficient distance metric. However, Network Vector does not suffer from this problem by automating the feature learning using an objective function. In this task, we empirically fix the dimensionality of vectors as 100 and context window as 10.
**Method** **Hit@1** **Hit@5** **Hit@10**
--------------------------- ------------ ------------ ------------
Degrees 0.0147 0.0423 0.0717
Clustering Coefficients 0.0006 0.0043 0.0086
Eigens 0.0025 0.0074 0.0116
Degrees+Clustering+Eigens 0.0153 0.0453 0.0803
node2vec 0.2450 0.5098 0.6150
Network Vector **0.2849** **0.5619** **0.6930**
: Performance of concept analogy task. \[tb:analogy\]
Multi-label Classification
--------------------------
Multi-label classification is a challenge task, where each node may have one or multiple labels. A classifier is trained to predict multiple possible labels for each test node. In our Network Vector algorithm, the global representation of entire network serves as additional context along with local neighborhood in learning node representations.
### Datasets
To understand whether the global representation helps learning better node representation, we perform multi-label classification with the same benchmarks and experimental procedure as [@grover2016node2vec] using the same datasets:
- [BlogCatalog [@zafarani2009social; @tang2009relational]]{}: This is a network of social relationships provided by bloggers on the BlogCatalog website. The labels represent the interests of bloggers on a list of topic categories. There are 10,312 nodes, 333,983 edges in the network and 39 distinct labels for nodes.
- [Protein-Protein Interactions (PPI) [@breitkreutz2008biogrid; @grover2016node2vec]]{}: This is a subgraph of the entire PPI network for Homo Sapiens. The node labels are obtained from hallmark gene sets [@liberzon2011molecular] and represent biological states. There are 3,890 nodes, 76,584 edges in the network and 50 distinct labels for nodes.
- [Wikipedia Cooccurrences [@mahoney2009large; @grover2016node2vec]]{}: This is a network of words appearing in the first million bytes of the Wikipedia dump. The edge weight is defined by the cooccurrence of two words within a 2-length slide window. The Part-of-Speech (POS) tags [@marcus1993building] inferred using the Stanford POS-Tagger [@toutanova2003feature] are used as labels. There are 4,777 nodes, 184,812 edges in the network and 40 distinct labels for nodes.
### Methods for Comparison
We compare the node representations learned by Network Vector against the following feature learning methods for node representations:
- [Spectral clustering]{} [@tang2011leveraging]: This method learns the $d$-smallest eigenvectors of the normalized graph Laplacian matrix, and utilize them as the $d$-dimensional feature representations for nodes.
- [DeepWalk]{} [@perozzi2014deepwalk]: This method learns $d$-dimensional feature representations using Skip-gram [@mikolov2013efficient; @mikolov2013distributed] from node sequences, that are generated by uniform random walks from the source nodes on graph.
- [LINE]{} [@tang2015line]: This method learns $d$-dimensional feature representations by sampling nodes at 1-hop and 2-hop distance from the source nodes in BFS-like manner.
- [node2vec]{} [@grover2016node2vec]: We use the original node2vec algorithm with optimal parameter settings of $(p, q)$ reported in [@grover2016node2vec].
Network Vector utilizes only first-order or second-order proximity between nodes in two-layer neural embedding framework. The first layer computes the context feature vector, and the second layer computes the probability distribution of target nodes. It is similar to other neural embedding based feature learning methods DeepWalk, LINE and node2vec. For fair comparison, we exclude recent approaches GraRep [@cao2015grarep], HNE [@chang2015heterogeneous] and SDNE [@wang2016structural]. It is because GraRep utilizes information from network neighborhoods beyond second-order proximity, and both HNE and SDNE employ deep neural networks that have multiple layers (more than two). GraRep, HNE and SDNE are less computational efficient and cannot scale well, as compared to DeepWalk, LINE, node2vec and our algorithm Network Vector.
--------------------------- ------------- ------------ ------------
**Algorithm**
BlogCatalog PPI Wikipedia
Spectral Clustering 0.0405 0.0681 0.0395
LINE 0.0784 0.1447 0.1164
DeepWalk 0.2110 0.1768 0.1274
node2vec (p\*, q\*) 0.2581 0.1791 0.1552
Network Vector (p=q=1) 0.2473 0.1938 0.1388
Network Vector (p\*, q\*) **0.2607** **0.1985** **0.1765**
settings (p\*, q\*) 0.25, 0.25 4, 1 4, 0.5
Gain over DeepWalk 12.4% 9.6% 8.9%
Gain over ndoe2vec 1.0% 9.7% 13.7%
--------------------------- ------------- ------------ ------------
: Macro-F1 scores for multi-label classification with a balanced $50-50$ split between training and testing data. Results of Spectral clustering, DeepWalk, LINE and node2vec are reported in node2vec paper. \[tb:multi-label-classification\]
For fair comparison, we use the inverse architecture of Network Vector, which is Skip-gram [@mikolov2013efficient] like and similar to that of node2vec. The parameter settings for Network Vector are in favor of node2vec, and exactly the same as in [@grover2016node2vec]. Specifically, we set $d = 128$, $r = 10$, $l = 80$, and a context size $n = 10$, and are aligned with typical values used for DeepWalk and LINE. A single pass of the data (one epoch) is used for optimization. In order to perform multi-label classification, the learned node representations from each approach are used as feature input to a one-vs-rest logistic regression with L2 regularization. Our experiments are repeated for 10 random equal splits of train and test data, and average results are reported.
### Results
Macro-F1 scores are used as evaluation metrics, and Table \[tb:multi-label-classification\] shows the results. We run Network Vector with node sequences generated by biased random walks from node2vec. The default parameter setting $(p=1, q=1)$ used in DeepWalk and the optimal parameter setting of node2vec reported in [@grover2016node2vec] are used.
From the results, we can see Network Vector outperforms node2vec using the same biased random walks, and DeepWalk using the same uniform random walks. It is evident that the global representation of the entire network allows Network Vector to exploit the global structure of the networks to learn better node representations. Network Vector achieves a slight performance gain, $1.0\%$ over node2vec, and a significant $12.4\%$ gain over DeepWalk on BlogCatalog. As we can see on PPI, The gain of Network Vector over node2vec and DeepWalk are significant and similar, $9.6\%$ and $9.7\%$ respectively. In the case of Wikipedia word cooccurrence network, Network Vector outperforms node2vec with a decent margin, achieving $13.7\%$ performance gain, while with a less gain, $8.9\%$ over DeepWalk. Overall, sampling strategies even with optimal parameter settings $(p, q)$ in node2vec are limited in exploration of local neighborhood of the source nodes, but cannot exploit the global network structure well. Network Vector overcomes the limitation of locality. By utilizing an additional global vector to memorize the collective information from all the local neighborhoods of nodes even within 2-hops, Network Vector learns improved node representations.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_blog_macro "fig:"){width="25.00000%"} ![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_blog_micro "fig:"){width="25.00000%"}
![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_PPI_macro "fig:"){width="25.00000%"} ![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_PPI_micro "fig:"){width="25.00000%"}
![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_wiki_macro "fig:"){width="25.00000%"} ![Performance of Network Vector and node2vec on varying the parameter $q$ when fixing $p=\infty$ to encourage reaching unvisited nodes in random walks. Macro-F1 and Micro-F1 scores for multi-label classification with a balanced $50\%$ train-test split are reported. \[fig:q\]](q_wiki_micro "fig:"){width="25.00000%"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
### Parameter Sensitivity
In order to understand how Network Vector improves in learning node representations with biased random walks in fine-grained settings, we evaluate performance while varying the parameter settings of $(p, q)$. We fix $p=\infty$ to discourage revisiting sampled nodes at the previous step in random walks, and varying the value $q$ in the range from $2^{-4}$ to $2^4$ to perform DFS-like sampling in various degrees.
Figure \[fig:q\] shows the comparison results for Network Vector and ndoe2vec in both Macro-F1 and Micro-F1 scores. As we can see, Network Vector consistently outperforms node2vec in different parameter settings of $q$ in all the three datasets. However, we observe on BlogCatalog, Network Vector achieves relatively larger gains over node2vec when $q$ is large that the random walks is biased towards BFS-like sampling, as compared to that when $q$ is small that the sampling is more DFS-like. It is mainly because when the random walks is biased towards nodes close to the source nodes, the global information of network structure that are exploited by Network Vector can compensate more for locality information using BFS-like sampling. However, when $q$ is small, the random walks is biased towards sampling nodes far away from the source nodes, and explore information close to the global network structure. Hence, Network Vector is not quite helpful in this case. We can see similar patterns of performance margin between Network Vector and node2vec when $q$ tends to be large in word cooccurrence network of Wikipedia. However, in the case of PPI, the performance gains achieved by Network Vector over node2vec are stable even various values of $q$ are used. The reason is probably because the biological states of proteins in a protein-protein interaction network exhibit a high degree of homophily, since proteins in local neighborhood usually organize together to perform similar functions. Hence, the global network structure is not quite informative to predict the biological states of proteins as we set a large value of $q$.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Performance of Network Vector and node2vec on varying the faction of labeled data for training. \[fig:labeleddata\] ](labeleddata_macro "fig:"){width="25.00000%"} ![Performance of Network Vector and node2vec on varying the faction of labeled data for training. \[fig:labeleddata\] ](labeleddata_micro "fig:"){width="25.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------
### Effect of Training Data
To see the effect of training data, we compare performance while varying the fraction of labeled data from $10\%$ to $90\%$. Figure \[fig:labeleddata\] shows the results on PPI. The parameters $(p, q)$ is fixed using optimal values $(4, 1)$. As we can see, when using more labeled data, the performance of node2vec and Network Vector generally increases. Network Vector achieves the largest gain over node2vec of $9.0\%$ in Macro-F1 score and $10.3\%$ at $40\%$ labeled data. When only $10\%$ labeled data is used, Network Vector only yields $1.5\%$ gain in Macro-F1 score, and $7.1\%$ in Micro-F1 score. We have similar observations on BlogCatalog and Wikipedia datasets, and the results are not shown.
Conclusion
==========
We have presented Network Vector, an algorithm for learning distributed representations of nodes and networks simultaneously. By embedding the network in a lower-dimensional vector space, our algorithm allows for quantitative comparison of networks. It also allows for the comparison of individual network nodes, since each node can be represented by its ego-network—a network containing the node itself, its network neighbors, and all connections between them.
In contrast to existing network embedding methods, which only learn representations of component nodes, Network Vector directly learns the representation of an entire network. Learning a representation of a network allows us to evaluate the similarity between two networks or two individual nodes, which enables us to answer questions that were difficult to address with existing methods. For instance, given a node in a network, for example, a manager within an organization, we can identify other people serving a similar role within that organization. Also, given a connection, denoting some relationship between two people within a social network, we could find another pair in an analogous relationship. Beyond social networks, we can also answer new questions about knowledge networks that connect concepts or documents to each others, for example, Wikipedia and citations networks. This can be useful especially in cases where the contents of documents is not available for privacy or other reasons, but the network of interactions exists.
For the networks in which content is available for the nodes, the learning method could be extended to account for it. For example, for knowledge networks, the approach could be combined with text to learn representations of networks that will give a more fine-grained view of their similarity. Additionally, other non-textual attributes could also be included in the learning algorithm. The flexibility of such learning algorithms make them ideal candidates for applications requiring similarity comparison of different types of objects.
[^1]: <https://www.courtlistener.com/>
[^2]: <http://www.cs.cmu.edu/~enron/>
[^3]: <http://schools-wikipedia.org/>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a phase-connected timing solution for the nearby isolated neutron star [RX J1308.6+2127]{} ([RBS 1223]{}). From dedicated [*Chandra*]{} observations as well as archival [*Chandra*]{} and [*XMM-Newton*]{} data spanning a period of five years, we demonstrate that the 10.31-sec pulsations are slowing down steadily at a rate of $\dot P={\ensuremath{1.120(3) \times 10^{-13}}}\mbox{ s s}^{-1}$. Under the assumption that this is due to magnetic dipole torques, we infer a characteristic age of 1.5 Myr and a magnetic field strength of ${\ensuremath{3.4 \times 10^{13}}}$ G. As with [RX J0720.4$-$3125]{}, the only other radio-quiet thermally emitting isolated neutron star for which a timing solution has been derived, the field strength is roughly consistent with what was inferred earlier from the presence of a strong absorption feature in its X-ray spectrum. Furthermore, for both sources the characteristic age is in excess of the cooling age inferred from standard cooling models. The sources differ, however, in their timing noise: while [RX J0720.4$-$3125]{} showed considerable timing noise, [RX J1308.6+2127]{} appears relatively stable.'
author:
- 'D. L. Kaplan and M. H. van Kerkwijk'
title: 'A Coherent Timing Solution for the Nearby Isolated Neutron Star RX J1308.6+2127/RBS 1223'
---
Introduction
============
The sample of nearby, radio-quiet, isolated neutron stars discovered by [*ROSAT*]{} (for a review, e.g., @haberl04) is of particular interest because of the unambiguous presence of strong, broad absorption features in the X-ray spectra of most sources. Since the spectra appear thermal, these features almost certainly arise in the neutron-star atmospheres. The features have usually been interpreted under the assumption of a pure hydrogen composition (not unlikely, given the short settling times for neutron stars) and a strong magnetic field (as suggested by the long, 3–10 s spin periods; @hh98b), with the absorption reflecting either the proton cyclotron line or transitions between bound states of neutral hydrogen (e.g., @hsh+03 [@vkkd+04]).
In order to help determine the nature of the absorption features, as well as to elucidate what sets the isolated neutron stars apart from young rotation-powered pulsars, we have started a program to obtain phase-connected timing solutions, and use these to estimate ages and magnetic field strengths. Earlier, we presented our first results, for [RX J0720.4$-$3125]{} [@kvk05]; here, we consider a second source, [RX J1308.6+2127]{}.
[RX J1308.6+2127]{} (also known as [RBS 1223]{} and [1RXS J130848.6+212708]{}) was identified as a possible nearby isolated neutron star by @shs+99. The identification was confirmed with the detection of a 5.16-s X-ray periodicity [@hhss02] and a very faint ($V\approx 28$ mag) probable optical counterpart with no radio emission [@kkvk02]. By comparing the periods measured from [*Chandra*]{} and archival [*ROSAT*]{}data, @hhss02 inferred a spin-down rate of $\dot
P={\ensuremath{(0.7-2.0) \times 10^{-11}}}\mbox{ s s}^{-1}$, implying a very strong magnetic field of ${\gtrsim}10^{14}$ G. @haberl04, however, showed that the 5.16-s periodicity was in fact the first harmonic, implying a true period of 10.31 s (see also @hsh+03), and that the spin-down rate inferred earlier was likely erroneous. This was confirmed by @shhm05, who attempted to determine a phase-connected timing solution, but failed due to cycle-count ambiguities.
Here, we show that with additional [*Chandra*]{} observations specifically obtained for timing purposes, we can obtain an unambiguous timing solution. We describe our analysis of the [*Chandra*]{} data, as well as of archival [*ROSAT*]{}, [*Chandra*]{}, and [*XMM-Newton*]{} data, in § \[sec:obs\], and use these to obtain a timing solution in § \[sec:timing\]. Since our analysis closely follows the one used for [RX J0720.4$-$3125]{} [@kvk05], we focus primarily on those aspects of the analysis that differ. We discuss the implications of our result in § \[sec:discuss\].
Observations {#sec:obs}
============
Our primary data are eight observations taken with the Advanced CCD Imaging Spectrometer [ACIS; @gbf+03] aboard the *Chandra X-ray Observatory* (*CXO*). These were designed for timing accuracy, consisting of two sets of four exposures in the Continuous-Clocking (CC) mode geometrically spaced over a period of about two weeks and separated by about half a year. We combined these with data from other [*Chandra*]{} observations, as well as from observations with *XMM-Newton* and [*ROSAT*]{}. A log of all observations[^1] is given in Table \[tab:obs\].
[lrcrrl]{} HRI & 704082 & 1998 Oct 01 & 4.8 & 498& 50824.2143496(30)\
ACIS $1/8$ & 731 & 2000 Jun 24 & 9.5 & 7395& 51719.5182790(5)\
PN/sw & 377-U2 & 2001 Dec 31 & 18.0 & 10633& 52274.2594926(10)\
PN/ff & 561-S5 & 2003 Jan 01 & 27.0 & 66219& 52640.4325929(5)\
MOS1 & 561-S3 & 2003 Jan 01 & 29.0 & 14469& 52640.4265055(9)\
MOS2 & 561-S4 & 2003 Jan 01 & 29.0 & 14925& 52640.4265042(7)\
PN/ff & 743-S3 & 2003 Dec 30 & 30.0 & 74587& 53003.4668264(6)\
MOS1 & 743-S1 & 2003 Dec 30 & 32.0 & 16086& 53003.4607372(7)\
MOS2 & 743-S2 & 2003 Dec 30 & 32.0 & 16523& 53003.4607388(9)\
HRC & 4595 & 2004 Mar 30 & 90.1 & 33823& 53095.3847872(16)\
ACIS CC & 5522 & 2005 Feb 14 & 16.0 & 8953& 53415.6878535(10)\
& 5523 & 2005 Feb 15 & 5.7 & 3254& 53416.5968822(17)\
& 5524 & 2005 Feb 19 & 5.2 & 2922& 53420.1717683(14)\
& 5525 & 2005 Mar 10 & 5.6 & 2923& 53439.0475044(11)\
ACIS CC & 5526 & 2005 Jul 09 & 15.1 & 7684& 53560.2591584(8)\
& 5527 & 2005 Jul 10 & 5.1 & 2876& 53561.2549594(17)\
& 5528 & 2005 Jul 14 & 5.2 & 2937& 53565.7697939(18)\
& 5529 & 2005 Jul 29 & 5.2 & 3048& 53580.7839447(14)\
For the [*Chandra*]{} data, we processed the level-1 event lists to the level-2 stage following standard procedures and the latest calibration set (`CALDB` version 3.1.0). For the ACIS data, we extracted events within $1\arcsec$ of the source with energies between 0.2 and 2.0 keV, and then applied a clock correction of $284.7\,\,\mu{\rm
s}$ [@dhm03]. For the HRC-S/LETG data, we extracted zeroth-order events from a circle with radius of $1\farcs3$, and first-order events using the standard LETG spectral extraction windows, but limited to $10\,\mbox{\AA}\leq\lambda\leq65\,\mbox{\AA}$. Finally, we used the `axbary` program to barycenter all of the events (using the X-ray position: $\alpha_{\rm
J2000}=13{\ensuremath{^{\rm h}}}08{\ensuremath{^{\rm m}}}48\farcs27$, $\delta_{\rm
J2000}=+21\degr27\arcmin06\farcs8$; @kkvk02)
For the [*XMM-Newton*]{} data, we used the standard procedures `emchain` and `epchain` (`XMMSAS` version 6.5.0) to reprocess the observations. Next, we extracted events within 375 of the source position (using standard quality and pattern selections) with energies from 0.12 keV to 1.2 keV, and used `barycen` to convert the arrival times to the solar-system barycenter. And finally, for the [*ROSAT*]{} HRI data, we extracted the events within $12\farcs5$ of the source. We barycentered these using the `FTOOLS` programs `ABC` and `BCT`, and converted the event times from Coordinated Universal Time (UTC) to Barycentric Dynamical Time (TDB) using the corrections supplied in @allen [p. 14].
Timing Analysis {#sec:timing}
===============
Our goal is to use times-of-arrival (TOAs) to infer a phase-coherent timing solution in which each cycle of the source is accounted for. To measure TOAs, we need an initial reference period. We determined this from the ACIS CC data-sets using a $Z_2^2$ test [@bbb+83], which combines power from the 10.31-s fundamental with that from the 5.16-s harmonic (we could not detect significant power in any higher harmonics). We calculated $Z_2^2$ both for the individual sets of four observations (observations 5522–5525 and 5526–5529) as well as for the eight observations combined. The overall $Z_2^2$ spectrum has a single, well-defined peak, and all possible aliases are at considerably lower significances. The measured period is $P=10.31252293(9)$ s (here and below, numbers in parentheses indicate the formal 1-$\sigma$ uncertainties in the last digit unless otherwise indicated). Using the period derived above, we constructed binned light curves (with 16 phase bins) for all of the observations. Unlike [RX J0720.4$-$3125]{}, [RX J1308.6+2127]{}has non-sinusoidal pulsations and hence determining TOAs by fitting a single sinusoid would be inappropriate. Instead we fit for both the first harmonic and fundamental, with a possible phase shift between them: $$N_{i} = A\left[ \sin\left(2\pi(f_0 t_i + \phi_0)\right)
+ r_2\sin\left(4\pi(f_0 t_i + \phi_0+\Delta \phi_2)\right)\right]+C,
\label{eqn:fitfh}$$ where $N_{i}$ is the number of counts in bin $i$, $f_0=0.096969481$ Hz is the frequency of the fundamental from above, $t_i$ the time of bin $i$ on the interval $[0,P)$, and the parameters are amplitude $A$, phase $\phi_0$, amplitude ratio $r_2$, phase offset $\Delta
\phi_2$, and constant offset $C$. As can be seen in Figure \[fig:pn\], this model provides a good fit to even the highest-quality lightcurves, which are those derived from the long EPIC-PN observations. The inferred phase offset and relative amplitude are similar for both observations, with $\Delta
\phi_2\approx 0.03$, and $r_2\approx 2.8$.
Assuming that the pulse profile is not varying, we measured arrival times for all of the data using the above model, but with $\Delta\phi_2$ and $r_2$ fixed at 0.03 and 2.8, respectively, which generally provides a good fit to the light curves (see Fig. \[fig:pn\] for the EPIC-PN data). Our resulting arrival times are listed in Table \[tab:obs\]. Here, we assigned these times to the time of the maximum of the model lightcurve that follows the lowest minimum and is nearest to the middle of the observation (see Fig. \[fig:pn\]; this is the spectrally harder maximum from @shhm05). For completeness, we note that the TOAs derived using the above model are consistent with those one would find using a simpler model consisting of only the first harmonic (i.e., a sinusoid at $P=5.16$ s): the mean absolute difference is $0.2\sigma$.
We determined a timing solution from the TOAs using an iterative procedure, as in @kvk05. We started with the [*Chandra*]{} ACIS CC TOAs and found that we were able to fit these without the need for a frequency derivative and with no cycle ambiguity, consistent with the power spectrum analysis. However, between the first ACIS CC TOA and the preceding TOA ([*Chandra*]{} HRC) there is a gap of 320 days and we found that a constant-frequency model led to a poor fit ($\chi^2=202.7$ for 7 degrees of freedom). Including a frequency derivative ${\ensuremath{\dot \nu}}$, the cycle count over the gap became ambiguous by $\pm 1$ cycle. The three possibilities—2683550, 2683551, and 2683552 cycles—lead to three possible solutions for ${\ensuremath{\dot \nu}}$: ${\ensuremath{2.1 \times 10^{-15}}}\mbox{ Hz s}^{-1}$, ${\ensuremath{1.0 \times 10^{-15}}}\mbox{ Hz
s}^{-1}$, and ${\ensuremath{-1.1 \times 10^{-15}}}\mbox{ Hz s}^{-1}$ . Fortunately, the addition of the remaining data eliminates the first two of those solutions: with the [*XMM-Newton*]{} TOAs (and the remaining ACIS point), they have $\chi^2=439.1$ and 1107.6 with TOA rms of 0.041 s and 0.099 s, respectively, while the third has $\chi^2=18.7$ with a TOA rms of 0.010 s (all for 14 degrees of freedom).
The timing solution derived from the [*Chandra*]{} and [*XMM-Newton*]{} TOAs is presented in Table \[tab:ephem\]. As indicated by the value of $\chi^2$, the data are well reproduced by simple spin-down; the addition of a cubic ([$\ddot \nu$]{}) term leads to only an insignificant reduction in $\chi^2$, from 18.7 to 18.6, and the inferred value of ${\ensuremath{\ddot \nu}}={\ensuremath{-7(15) \times 10^{-26}}}\mbox{ Hz s}^{-2}$ is consistent with zero.
[p[1.5in]{} c]{}\[t\] Dates (MJD) & 51720–53581\
$t_{0}$ (MJD)& 53415.687853(2)\
$\nu$ (Hz) & 0.0969694896(2)\
$\dot \nu$ ([$\mbox{Hz s}^{-1}$]{}) & ${\ensuremath{-1.053(3) \times 10^{-15}}}$\
TOA rms (s) & 0.010\
$\chi^2$/DOF & 18.7/14=1.34\
$P$ (s)& 10.31252206(2)\
$\dot P$ ([$\mbox{s s}^{-1}$]{})& ${\ensuremath{1.120(3) \times 10^{-13}}}$\
$\dot E$ ($\mbox{erg s}^{-1}$)& ${\ensuremath{4.0 \times 10^{30}}}$\
$B_{\rm dip}$ (G) & ${\ensuremath{3.4 \times 10^{13}}}$\
$\tau_{\rm char}$ (yr)& ${\ensuremath{1.5 \times 10^{6}}}$\
Unlike the [*Chandra*]{} and [*XMM-Newton*]{} data, however, the [*ROSAT*]{} HRI point is slightly discrepant from the fit in Table \[tab:ephem\], exceeding it by 0.15(3) cycles. If we include it in the fit, the overall solution does not change drastically (${\ensuremath{\dot \nu}}$ becomes ${\ensuremath{-1.050(2) \times 10^{-15}}}\mbox{ Hz s}^{-1}$) and the deviation does decrease to $0.11(3)$ cycles, but the overall $\chi^2$ increases to $43.5$ for 15 degrees of freedom (with a TOA rms of 0.0271 s). The deviation could be intrinsic (a glitch, timing noise, precession, changes in pulse profile), but might also be instrumental, perhaps related to differences in energy responses between the different instruments. Unfortunately, the sampling is too sparse to determine a unique solution. As an example, we show in Figure \[fig:resid\] a cubic fit that matches the HRI point reasonably well, with $\chi^2=22.5$ for 14 degrees of freedom, TOA rms of 0.011 s, and ${\ensuremath{\ddot \nu}}={\ensuremath{-2.5(5) \times 10^{-25}}}\mbox{ Hz s}^{-2}$. We stress, however, that, in essence, this solution approximates the [*Chandra*]{} and [*XMM-Newton*]{} data with a ${\ensuremath{\ddot \nu}}=0$ model and then adjusts ${\ensuremath{\ddot \nu}}$ to pass through the HRI point. Hence, it does not gives much additional insight.
Discussion & Conclusions {#sec:discuss}
========================
The spin period and spin-down rate we derive for [RX J1308.6+2127]{} are quite similar to those of [RX J0720.4$-$3125]{}, the only other isolated neutron star with a timing solution [@kvk05], placing both of them above the pulsar “death-line” in a $P$-$\dot P$ diagram despite their lack of radio emission [@kkvk02; @kvkm+03]. Hence, assuming both spin down by magnetic dipole radiation, [RX J1308.6+2127]{} and [RX J0720.4$-$3125]{} have similar inferred magnetic field strengths ($B_{\rm dip}=3.4$ and ${\ensuremath{2.4 \times 10^{13}}}$ G, respectively), characteristic ages ($\tau_{\rm char}=1.5$ and 1.9 Myr), and spin-down luminosities ($\dot E=4.0$ and ${\ensuremath{4.7 \times 10^{30}}}\mbox{ ergs s}^{-1}$). Both sources also have thermal X-ray spectra with temperatures of $\sim\!1\times10^6~$K, but superposed on these are rather different absorption features: in [RX J1308.6+2127]{}, the absorption feature is centered near or below 300 eV and is very wide and strong, with an equivalent width of $>\!150$ eV [@hsh+03]; in [RX J0720.4$-$3125]{}, on the other hand, no absorption was apparent until 2004, when an absorption feature with only a slightly higher energy but much smaller equivalent width ($\sim\!40$ eV) appeared [@hztb04; @dvvmv04]. If the magnetic fields, line energies, and blackbody temperatures are all similar, what would account for the significant difference in line strength?
The interpretation of the absorption lines is still a matter of debate. For [RX J1308.6+2127]{}, @hsh+03 suggested proton cyclotron absorption in a field of (2–6)$\times 10^{13}$ G, which could match both the energy and strength of the observed feature (for a gravitational redshift of $z=0.3$, the observed line energy would be at 100–300 eV). As the feature in [RX J0720.4$-$3125]{} was so much weaker but at a similar energy, @vkkd+04 argued that for that source the proton cyclotron line was below the observed band and that the absorption was due to the $0\rightarrow2$ transition between tightly-bound states of neutral hydrogen in a $\sim\!{\ensuremath{2 \times 10^{13}}}$ G field.
Qualitatively, the field strengths inferred from our timing measurements agree with the above expectations: the field of [RX J1308.6+2127]{} is stronger than that of [RX J0720.4$-$3125]{}, while a weaker field would be expected if the absorption in both sources were due to the same mechanism. Quantitatively, a larger difference between the two sources was expected. However, this discrepancy may simply reflect orientation and/or substructure in the magnetic field—higher-order multipoles near the surface would affect the emission properties but not the spin-down rate. Such effects may also be responsible for the difference in pulse profile: double-peaked for [RX J1308.6+2127]{}, and sinusoidal for [RX J0720.4$-$3125]{}. Phase-resolved X-ray spectroscopy coupled with observations of more sources are probably the best ways to disentangle these effects, and efforts are underway [e.g., @hztb04; @shhm05].
Another open issue is that the characteristic ages of [RX J0720.4$-$3125]{} and now [RX J1308.6+2127]{} are three to four times larger than the values of $\sim\!0.5$ Myr one would expect for simple cooling models [@hh98b; @kkvkm02; @zhc+02] and, in the case of [RX J0720.4$-$3125]{}, tracing the object back along its trajectory to a likely birth location (@mzh03, @kaplan04). Indeed, this age discrepancy may extend to the other isolated neutron stars discovered by [*ROSAT*]{}: all have similar temperatures and thus likely similar cooling ages, and most also have similar periods (3–10 s) and, based on the similar energies at which they show X-ray absorption features (0.3–0.7 keV), similar magnetic field strengths, implying similar characteristic ages.
We first consider whether the discrepancy could result from the characteristic age being an overestimate. In general, for a spin-down torque $\propto\nu^n$, the pulsar’s spin-down age is given by $t_{\rm sd} = \left[P/(n-1)\dot P\right]\left[1-\left(P_0/P\right)^{n-1}\right]$, where $P_0$ is the initial spin period and $n=\nu\ddot\nu/\dot\nu^2$ is the “braking index,” equal to 3 under the assumption of magnetic dipole radiation [e.g., @mt77 p. 111]. For $P_0\ll P$ and $n=3$, one recovers the characteristic age $\tau_{\rm char}\equiv P/2\dot P$, but if $P_0$ is not much smaller than the current period $P$, the characteristic age is an overestimate. While we cannot exclude the required birth periods of 7–8 s, there is as yet no concrete indication that neutron stars are born with periods longer than 100 ms [e.g., @kh02; @ghs05].
For [RX J0720.4$-$3125]{}, this led us to discuss the possibility that the neutron star formed in a binary system and accreted matter before being ejected in a second supernova; at ejection, the neutron star would still be hot from the accretion, but spinning slowly [@kvk05]. A variation on this model would involve accretion from a residual debris disk such as that recently discovered around the anomalous X-ray pulsar 4U 0142+61 [@wck05], but this assumes that the accretion disk persists for a sufficient time and can affect the spin-down of the neutron stars, both of which are far from clear. In either case, if accretion played a role, it might explain why the spectral properties of the isolated neutron stars are rather different from those of the radio pulsar population.
Unfortunately, the above solution for the age discrepancy is not unique. First, it is possible that the spin-down was not due to a constant dipole but that the magnetic field decayed (effectively, this implies $n>3$). Second, the cooling age could be incorrect due to non-standard cooling. For instance, the energy released by magnetic field decay might keep a neutron star hotter (@hk98; note, however, that these authors do not expect significant effects for the field strengths we infer) and longer cooling times are also expected for a light neutron star [@ygk+04]. A prolonged cooling timescale seems somewhat unlikely in light of the kinematic age estimates, though it should be kept in mind that these may also not be unique: for [RX J0720.4$-$3125]{}, several different birth places are possible [@mzh03; @kaplan04].
Finally, when fitted with just a simple spin-down model, [RX J0720.4$-$3125]{} showed significant timing residuals of 0.31 s (root-mean-square), much larger than the uncertainties on the TOAs. In contrast, for [RX J1308.6+2127]{} the root-mean-square residuals are only 0.010 s (excluding the HRI point), consistent with measurement errors. Since the sources were observed with the same instruments, and have similar spectral shapes, the differences in timing behaviors are likely not instrumental. Instead, they may well reflect the fact that both the spectrum and pulse profile of [RX J0720.4$-$3125]{} are varying with time [@dvvmv04; @vdvmv04], while [RX J1308.6+2127]{} appears to be stable.
We acknowledge support through Chandra grant GO5-6050A.
[31]{} natexlab\#1[\#1]{}
, R. [et al.]{} 1983, , 128, 245
, A. N. 2000, Allen’s [A]{}strophysical [Q]{}uantities, 4th edn. (New York: AIP)
Davis, W., Holmes, J., & Myers, R. 2003, in Proc. 2003 Chandra Calibration Workshop (Cambridge: CXC), http://cxc.harvard.edu/ccw/proceedings/03\_proc
, C. P., [Vink]{}, J., [M[' e]{}ndez]{}, M., & [Verbunt]{}, F. 2004, , 415, L31
, G. P., [Bautz]{}, M. W., [Ford]{}, P. G., [Nousek]{}, J. A., & [Ricker]{}, G. R. 2003, , 4851, 28
, E. V., [Halpern]{}, J. P., & [Seward]{}, F. D. 2005, , 627, 390
, F. 2004, Advances in Space Research, 33, 638, (astro-ph/0302540)
, F., [Schwope]{}, A. D., [Hambaryan]{}, V., [Hasinger]{}, G., & [Motch]{}, C. 2003, , 403, L19
, F., [Zavlin]{}, V. E., [Tr[" u]{}mper]{}, J., & [Burwitz]{}, V. 2004, , 419, 1077
, V., [Hasinger]{}, G., [Schwope]{}, A. D., & [Schulz]{}, N. S. 2002, , 381, 98
, J. S. & [Hernquist]{}, L. 1998, , 297, L69
, J. S. & [Kulkarni]{}, S. R. 1998, , 506, L61
, D. L. 2004, Ph.D. Thesis, California Institute of Technology
, D. L., [Kulkarni]{}, S. R., & [van Kerkwijk]{}, M. H. 2002, , 579, L29
, D. L., [Kulkarni]{}, S. R., [van Kerkwijk]{}, M. H., & [Marshall]{}, H. L. 2002, , 570, L79
, D. L. & [van Kerkwijk]{}, M. H. 2005, , 628, L45
, D. L., [van Kerkwijk]{}, M. H., [Marshall]{}, H. L., [Jacoby]{}, B. A., [Kulkarni]{}, S. R., & [Frail]{}, D. A. 2003, , 590, 1008
, V. M. & [Helfand]{}, D. J. 2002, in ASP Conf. Ser. 271: Neutron Stars in Supernova Remnants, ed. P. O. Slane & B. M. Gaensler (San Fransisco: ASP), 3, (astro-ph/0201183)
, R. P. [et al.]{} 1997, , 3114, 53
, R. N. & [Taylor]{}, J. H. 1977, [Pulsars]{} (San Francisco : W. H. Freeman, c1977.)
, C., [Zavlin]{}, V. E., & [Haberl]{}, F. 2003, , 408, 323
, A. D., [Hambaryan]{}, V., [Haberl]{}, F., & [Motch]{}, C. 2005, , 441, 597
, A. D., [Hasinger]{}, G., [Schwarz]{}, R., [Haberl]{}, F., & [Schmidt]{}, M. 1999, , 341, L51
, L. [et al.]{} 2001, , 365, L18
, M. J. L. [et al.]{} 2001, , 365, L27
, M. H., [Kaplan]{}, D. L., [Durant]{}, M., [Kulkarni]{}, S. R., & [Paerels]{}, F. 2004, , 608, 432
, J., [de Vries]{}, C. P., [M[' e]{}ndez]{}, M., & [Verbunt]{}, F. 2004, , 609, L75
, Z., [Chakrabarty]{}, D., & [Kaplan]{}, D. L. 2005, , in press
, D. G., [Gnedin]{}, O. Y., [Kaminker]{}, A. D., [Levenfish]{}, K. P., & [Potekhin]{}, A. Y. 2004, Advances in Space Research, 33, 523, (astro-ph/0306143)
, S., [Haberl]{}, F., [Cropper]{}, M., [Zavlin]{}, V. E., [Lumb]{}, D., [Sembay]{}, S., & [Motch]{}, C. 2002, , 334, 345
, M. V., [David]{}, L. P., [Harnden]{}, F. R., & [Kearns]{}, K. 1995, , 2518, 304
[^1]: We do not use data from [*ROSAT*]{} observation 703848, since we, like @hhss02 and @shhm05, found there were insufficient counts to extract a period or an arrival time.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Xia Li, Zhen Hu, Robert C. Qiu, [^1]'
bibliography:
- 'dsoref/compressed.bib'
title: MIMO UWB Radar System with Compressive Sensing
---
Acknowledgment {#acknowledgment .unnumbered}
==============
This work is funded by National Science Foundation through two grants (ECCS-0901420 and ECCS-0821658), and Office of Naval Research through two grants (N00010-10-1- 0810 and N00014-11-1-0006). This work is partly funded by an AFOSR subcontract though the prime contractor (RNET Technologies, Inc.) on the CIRE (Center for Innovative Radar Engineering) contract FA8650- 10-D-1750 with AFRL/RY.
[^1]: X. Li, Z. Hu, and R. C. Qiu are with the Department of Electrical and Computer Engineering, Center for Manufacturing Research, Tennessee Technological University, Cookeville, TN, 38505, e-mail: [email protected]; [email protected]; [email protected].
|
{
"pile_set_name": "ArXiv"
}
|
---
address: |
$^1$Novosibirsk State University, 630090, Novosibirsk-90, Russia\
\
$^2$Landau Institute for Theoretical Physics, 142432, Chernogolovka, Russia\
\
$^3$Department of Mathematics, University of Arizona, Tucson, AZ, 857201, USA\
\
$^4$Physical Institute of RAS, Leninskiy prospekt, 53, Moscow, 119991, Russia
author:
- 'A.I.Dyachenko$^{1,2,}$[^1], D.I.Kachulin$^{1}$, V.E.Zakharov$^{1,2,3,4}$'
title: 'Evolution of One-Dimensional Wind-Driven Sea Spectra'
---
Introduction
============
We perform numerical simulation of evolution of surface waves spectra that has been excited by wind. One of the motivation of writing this article is purely practical. Measure of nonlinearity of wave at the surface of deep water is their average steepness $\mu = <\nabla\eta^2>$, where $\eta(\vec r,t)$ is the shape of the surface. Characteristic value of $\mu$ in real sea is moderate, $\mu \simeq 0.06-0.07$. However, even at small steepness “white capping” (or wave breaking) happens occasionally, due to what waves loose energy. This phenomenon is not studied yet, either experimentally nor theoretically. Nevertheless in the operational models of wind waves prediction heuristics formulas for rate of wave decay (due to this phenomenon) are widely used. They were introduced about thirty years before [@Komen84; @WAMDI88] and little has changed since then. In our opinion they have no serious justification. The goal of this article is to check these heuristic formulas by numerical experiments not assuming statistical description.
To study “white capping” model with one horizontal dimension is enough. If steepness is moderate, $\mu \leq 0.07$, one can use dynamical “Zakharov equation” [@Z68], which is greatly simplified in 1-D case. It reduces to the simple Hamiltonian system which is very convenient for numerical simulation [@DZ2011-1; @DZ2012-2]. Canonical transformation resulting to this model is described in detail in [@DKZ2016]. In the framework of this model we perform numerical simulations for very long time (hundreds of thousands of characteristic wave period) and make sure that heuristic formulas [@Komen84; @WAMDI88] give to large rates of energy decay. It makes to treat used below wave prediction operational models highly critical.
Another motivation for this work is the desire to describe (possible more in detail) phenomenon of “white capping” for the waves with so moderate steepness. This work is not finished yet, but we established a most important fact - wave breaking is preceded by “freak wave” which actually breaks. “Freak waves” appear naturally as a result of modulational instability [@Z68], but even stable spectra of moderate amplitude are able to generate them. Although freak waves are now rare events separated by time interval it tens of thousands of wave periods.
Finally, in this article we come back to the old question about integrability of the free surface hydrodynamics of the deep water. Hypothesis of integrability was formulated in the paper [@DZ94] the result which was a key when deriving compact equation [@DZ2011-1; @DZ2012-2]. Since then it was argued both against integrability [@DLZ95; @DKZ2013] and in favor integrability [@ZD00; @DZ08]. In our experiments we observed behavior which is typical for integrable systems. Dynamics of the wave field was quasi-periodical. Spectra averaged over great time ((of the order of hundred thousands wave periods) have changed little, loosing 15% of their energy due to arising rare “freak waves”.
As an initial condition we used experimental and often cited in the oceanographic literature JONSWAP spectrum [@Hass73] with the wind speed 12m/sec.
JONSWAP Spectrum
================
Hasselmann et al, in [@Hass73], have analyzed data collected during the Joint North Sea Wave Observation Project - JONSWAP, and found that the wave spectrum is never fully developed. It continues to develop through non-linear, wave-wave interactions even for very long times and distances. They therefore proposed a spectrum in the form: $$\begin{aligned}
\label{JS}
S_J(\omega)d\omega &=& \alpha\frac{g^2}{\omega^5}\mbox{exp}^{\left [ -\frac{5}{4}(\frac{\omega_0}{\omega})^4\right ]} \gamma^{r}d\omega,\cr
r &=& \mbox{exp}^{\left [ -\frac{(\omega - \omega_p)^2}{2\sigma^2\omega_p^2})\right ]}\end{aligned}$$ $$\begin{aligned}
\nonumber
\alpha &=& 0.076(\frac{U_{10}^2}{F g})^{0.22},\cr
\omega_p &=& 22\left ( \frac{g^2}{U_{10} F}\right )^{\frac{1}{3}}, \cr
\gamma &=& 3.3, \cr
\sigma &=& \begin{cases}
0.07 \mbox{ if } \omega \leq \omega_p
\cr 0.09 \mbox{ if } \omega \leq \omega_p
\end{cases}\end{aligned}$$ Here $U_{10}$ - wind speed at the altitude 10m, $F$ - is fetch, e.i. distance from the shore. Spectra with different wind seeps are shown in Figure \[FIG\_JONSWAP\].
In the present article we study the relaxation of this developed sea state.
Compact equation for water waves
================================
We start with well known Hamiltonian for water waves $$\begin{gathered}
\label{Ham4}
H = \frac{1}{2}\int {g\eta^2 + \psi \hat k\psi} dx
-\frac{1}{2}\int \{(\hat k\psi)^2 -(\psi_x)^2\}\eta dx +\\
+\frac{1}{2}\int \{\psi_{xx} \eta^2 \hat k\psi + \psi \hat k(\eta \hat k(\eta \hat k\psi))\} dx + \dots\end{gathered}$$ which is expanded up to the fourth order as a function of Hamiltonian variables $\eta$ and $\psi$ (see [@Z68]): After introducing complex canonical variables $a_k$ $$\eta_k\! =\! \sqrt{\frac{\omega_k}{2g}}(a_k+a^*_{-k}) \hspace{.2cm}
\psi_k \!=\! -i\sqrt{\frac{g}{2\omega_k}}(a_k-a^*_{-k}) \hspace{1em}
\omega_k = \sqrt{gk}$$ in the articles [@DZ2011-1; @DZ2012-2] we applied canonical transformation to the Hamiltonian variable $a_k$ to introduce normal canonical variable $b(x,t)$:
1. $\eta_k, \psi_k \Rightarrow \mbox{normal canonical variable } a_k$
2. $a_k \Rightarrow b_k$
This transformation explicitly uses vanishing of four-wave interaction and possibility to consider surface waves moving in the same direction, see [@DZ2011-1; @DZ2012-2]. For this variable $b(x,t)$ Hamiltonian (\[Ham4\]) acquires nice and elegant form: $$\begin{gathered}
\label{SPACE_NICE}
{\cal H} = \int\!b^*\hat\omega_k bdx + \\
+ \frac{1}{2}\int\!\left |\frac{\partial b}{\partial x}\right |^2
\left [\frac{i}{2}\left ( b \frac{\partial b^*}{\partial x} - b^*\frac{\partial b}{\partial x}\right ) -\hat K|b|^2 \right ] dx.\end{gathered}$$
Corresponding equation of motion is the following: $$\begin{gathered}
\label{MotionSPACE}
i\frac{\partial b}{\partial t} = \hat\omega_k b
+\frac{i}{4}\hat P^+\left [ b^* \frac{\partial}{\partial x} ({b'}^2) -
\frac{\partial}{\partial x}( {b^*}' \frac{\partial}{\partial x}b^2) \right ]- \\
-\frac{1}{2} \hat P^+ \left [ b \cdot \hat k(|b'|^2) - \frac{\partial}{\partial x}(b'\hat k (|b|^2))\right ],\end{gathered}$$ Eigenvalue of the projection operator $\hat P^+$ in the Fourier-space is step-function: $$\nonumber
P^+_k=\theta(k) =\left\{\begin{array}{ll}
1, & k>0; \\
0, & k\leq 0. \\
\end{array}\right.$$ Transformation from $b(x,t)$ to physical variables $\eta(x,t)$ and $\psi(x,t)$ can be recovered from canonical transformation. It has been derived in [@DKZ2016]. Here we write this transformation up to the second order: $$\begin{gathered}
\label{EPX_up_2_2order}
\eta(x) = \frac{1}{\sqrt{2}g^{\frac{1}{4}}}(\hat k^{\frac{1}{4}}b(x)+\hat k^{\frac{1}{4}}b(x)^*)+\frac{\hat k}{4\sqrt{g}}[\hat k^{\frac{1}{4}}b(x) - \hat k^{\frac{1}{4}}b^*(x)]^2,\\
\psi(x) = -i\frac{g^{\frac{1}{4}}}{\sqrt{2}}(\hat k^{-\frac{1}{4}}b(x)-\hat k^{-\frac{1}{4}}b(x)^*)+ \\
+\frac{i}{2}[\hat k^{\frac{1}{4}}b^*(x)\hat k^{\frac{3}{4}}b^*(x) -
\hat k^{\frac{1}{4}}b(x)\hat k^{\frac{3}{4}}b(x)] + \cr
+\frac{1}{2}\hat H[\hat k^{\frac{1}{4}}b(x)\hat k^{\frac{3}{4}}b^*(x)+
\hat k^{\frac{1}{4}}b^*(x)\hat k^{\frac{3}{4}}b(x)].\end{gathered}$$ Here $\hat H$ - is Hilbert transformation with eigenvalue $i\textbf{sign}(k)$.
Kinetic equation
================
Along with simulation in the framework of equation (\[MotionSPACE\]) we have been solving the same initial problem with simple quasi-linear model $$\label{FP}
\frac{\partial |b_k|^2}{\partial t} = -\gamma_{diss}|b_k|^2 .$$ performing averaging by time and wave numbers, so that $$\nonumber
|b_k|^2 \rightarrow n_k = <|b_k|^2>,$$ [*here expressions for $\gamma_{diss}$ are taken from the article [@WAMDI88] (these are formulas 2.10 and 2.16), namely*]{}: $$\begin{aligned}
\label{W34}
\gamma_1^{WAM3} &=& 3.33\cdot 10^{-5}\bar\omega(\frac{\omega}{\bar\omega})^2(\frac{\bar\alpha}{\bar\alpha_{PM}})^2,\cr
\gamma_2^{WAM3} &=& 2.33\cdot 10^{-5}\hat\omega(\frac{\omega}{\hat\omega})^2(\frac{\hat\alpha}{\hat\alpha_{PM}})^2.\end{aligned}$$ Here $\bar\omega$ and $\hat\omega$ mean averaging over spectrum, $\alpha$ is an integral wave steepness, and $$\bar\alpha_{PM} = E\bar\omega^4 g^{-2},$$ $$\bar\alpha_{PM} = 4.57\cdot 10^{-3}$$ is the theoretical value of $\bar\alpha$ for a Pierson-Moskowitz spectrum [@PM64], and $$\hat\alpha_{PM} = 0.66\cdot\bar\alpha_{PM}.$$ $E$ - is the total energy (surface elevation variance). [*Again, all these definitions are taken from operational models from [@Komen84; @WAMDI88]*]{}. More recent models have just slight corrections to them. In this case model (\[FP\]) is equivalent to the well-known Hasselmann kinetic equation [@HASS1962] (see also [@DLZ95]) because wind pumping is absent and collision term $S_{nl}$ (due to the result of [@DZ94]) identically equal to zero. In both models we add artificial damping $$\label{GD}
\Gamma_d(k) = \begin{cases}
\alpha k^4 & \mbox{ if highest harmonics of } b_k \mbox{ are $10^4$ times}
\\ & \mbox{ greater then roundoff errors}
\\ 0 & \mbox{ in the other case}
\end{cases}$$ with $\alpha = 0.9 / \tau k_{max}^4$. It provides dissipation of extreme waves due to wave breaking. We calculated effective damping due to wave breaking, $<\gamma_{diss}>$, plugging results of calculations in the framework of (\[MotionSPACE\]) into the equation (\[FP\]). Another words we define $<\gamma_{diss}>$ as following: $$\label{FP1}
\gamma_{diss} = -\frac{1}{|b_k|^2}\frac{\partial |b_k|^2}{\partial t}.$$
Evolution of the JONSWAP spectrum
=================================
We study relaxation of developed sea with different wind speeds - $U_{10} = 9m/sec$, $U_{10} = 12m/sec$ and $U_{10} = 20m/sec$. However in this article we show results of simulation for $U_{10} = 12m/sec$ only. The others are very similar. Periodic domain of the length $L=10000$ meters was used for numerical simulations. Initial conditions for $b_k$ where chosen according to JONSWAP spectrum: $$\label{bk0}
|b_k|^2=\sqrt{\frac{2g}{\omega_k}} |\eta_k|^2 = S_J(k)\frac{2\pi}{L}\frac{g}{\omega_k}.$$ Phases of $b_k$ were chosen randomly in the interval $[0:2\pi]$. Fetch $F$ was equal to 157000 meters.
We observed much smaller dissipation than predicts WAM3 model. For the wind velocity $U_{10}=12 m/sec$ energy density both in our numeric experiment and calculated according to [@Komen84; @WAMDI88] are shown in Figures \[FIG\_02\].
Energy density is measured in oceanographic units $$\frac{\mbox{energy density}}{g}=\mbox{meter}^2.$$ The average steepness $\mu$ is calculated as following: $$\mu = \sqrt{\int_{-\infty}^\infty k^2 |\eta_k|^2 dk}.$$ In the picture one can see initial fast relaxation of energy in numerical experiment. It is due to dissipation of long tail $\simeq \omega^{-5}$ of JONSWAP spectrum in k-space (see (\[JS\])). After initial relaxation there are rare events of energy dissipation in our experiment. Average steepness is also shown in the Figure \[FIG\_02\].
One of this rare events, wave breaking, taking place at $\mbox{time}\simeq 93340$, is shown in detail in the Figure \[FIG\_03\]. One can see oscillation of the amplitude of the extreme wave.
Spectrum $S(k)$ along with zoomed profile of the surface at $\mbox{time}\simeq 93340$ is shown in the Figure \[FIG\_04\]. One can see the amplitude of the extreme (freak) wave more that 3 times large then for nearby waves.
Also in the Figure \[FIG\_04\] energy spectrum $S(k)$ after wave breaking is shown. It does not have tail in large wavenumbers.
Great difference between numerical results and prediction of WAM3 model is seen in Figure \[FIG\_05\]. Both of them had the same initial condition. However at the final time the spectra are very different. WAM3 predicts much more energy dissipation. It is also seen in Figure \[FIG\_02\].
One can see that relaxation of energy is sufficiently long process. During hundreds of thousands seconds it decreases by $\simeq 20\%$. During this time we calculated average $<\gamma_{diss}>$ according to (\[FP1\]). To make it smooth enough time of averaging was few hours (10000 sec). In the Figure \[FIG\_06\] there are dissipations according to (\[W34\]) [*(or equations (2.10) and (2.16) in [@WAMDI88])*]{} plotted by dotted and double dotted lines. $<\gamma_{diss}>$ calculated with the use of dynamical equation with (\[GD\]) shown solid line.
One can see that numerical experiment gives much less value of dissipation. Moreover, dissipation is absent in the core of spectral density where $k_0\simeq 0.06-0.07 m^{-1}$.
Conclusion
==========
The main result of our work is the fact that heuristic formulas for damping rate of free wind sea (\[W34\]) due to “white capping” dramatically exaggerates the role of this effect. Especially convincing is Figure \[FIG\_06\] showing that in the region of spectral maximum dissipation of energy is practically absent. Increase of $<\gamma_{diss}>$ with increasing of wave number indicates that damping is concentrated in the region of large wave numbers. it means that “white capping” leads primarily to vanishing of the spectra “tails” and smoothing of the wave field. We stress that our simulations describe sea evolution during few days after “switch off” wind. During this time sea lost no more then 20% of the energy. Similar picture of slow energy dissipation was observed in [@ZKP09]. Because “dissipation function” $\gamma_{diss}$ plays a key role in the massively used operational models the inevitable conclusion is that these models need to be fundamentally reviewed.
Our simulations are another argument in favor of integrability of deep water hydrodynamics. Others arguments in this favor are given in [@ZD00] and are very serious. There it is shown that exact system of Euler equations describing potential flow of deep water with a free surface can have any number of commuted integrals of motion. Weak point of this argument is the question about completeness of tis system of integrals. In the article [@DKZ2013] it is shown that model (\[MotionSPACE\]) is not integrable. But nonintegrability arises in the fifth order of the perturbation theory where equation (\[MotionSPACE\]) strictly speaking is not applicable. The most serious arguments contained in [@DLZ95], where indicated the non-existence of higher integrals.
Acknowledgments
===============
This work was supported by the Grant “Wave turbulence: theory, numerical simulation, experiment” \#14-22-00174 of Russian Science Foundation.
Numerical simulation was performed on the Informational Computational Center of the Novosibirsk State University.
[99]{}
G.J. Komen, S. Hasselmann and K. Hasselmann, Journal of Physical Oceanography, [**14**]{}, 1271-1285 (1984).
The WAMDI group, Journal of Physical Oceanography, [**18**]{}, 1775-1810 (1988).
V.E. Zakharov, Journal of Applied Mechanics and Technical Physics, [**9(2)**]{}, 190 (1968)
A.I. Dyachenko and V.E. Zakharov, JETP Letters, [**93(12)**]{}, 701 (2011)
A.I. Dyachenko and V.E. Zakharov, European Journal of Mechanics - B/Fluids, [**32**]{}, 17 (2012)
A.I. Dyachenko, D.I. Kachulin and V.E. Zakharov, in the book <<Extreme Ocean Waves>>, 2nd ed., Springer, Eds. E.Pelinovsky and C. Harif (Eds), (2016)
A.I. Dyachenko and V.E.Zakharov, Phys. Lett. A [**190**]{}, 144-148 (1994)
A.I. Dyachenko, Y.V. Lvov and V.E. Zakharov, Physica D, [**87**]{}, 233-261 (1995)
A.I. Dyachenko, D.I. Kachulin and V.E. Zakharov, JETP Letters, [**98(1)**]{}, 43-47 (2013)
V.E. Zakharov and A.I. Dyachenko, ArXiv:1206.2046
A.I. Dyachenko and V.E. Zakharov, JETP Letters, [**88(5)**]{}, 307 (2008) K. Hasselmann et al, Ergaenzungsheft zur Deutschen Hydrographischen Zeitschrift, Reihe A(8), [**12**]{}, 95 pp. (1973) A.I. Dyachenko, D.I. Kachulin and V.E. Zakharov, J. Phys. Conf. Ser, [**510**]{}, 012050 (2014)
K. Hasselmann, J. fluid Mech, [**12**]{}, 481-500, (1962)
W. Pierson and L. Moskowitz, Journal of Geophysical Reserch, [**69(24)**]{}, 5181-5190 (1964)
V.E. Zakharov, A.O. Korotkevich and A.O. Prokofiev, AIP Conf. Proc. [**1168**]{}, 1229-1231 (2009)
[^1]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider the effects that accretion from the interstellar medium onto the particles of an N-body system has on the rate of two-body relaxation. To this end, we derive an accretion-modified relaxation time by adapting Spitzer’s two-component model to include the damping effects of accretion. We consider several different mass-dependencies and efficiency factors for the accretion rate, as well as different mass ratios for the two components of the model.
The net effect of accretion is to accelerate mass segregation by increasing the average mass $\bar{m}$, since the relaxation time is inversely proportional to $\bar{m}$. Under the assumption that the accretion rate increases with the accretor mass, there are two additional effects that accelerate mass segregation. First, accretion acts to increase the range of any initial mass spectrum, quickly driving the heaviest members to even higher masses. Second, accretion acts to reduce the velocities of the accretors due to conservation of momentum, and it is the heaviest members that are affected the most. Using our two-component model, we quantify these effects as a function of the accretion rate, the total cluster mass, and the component masses. We conclude by discussing the implications of our results for the dynamical evolution of primordial globular clusters, primarily in the context of black holes formed from the most massive stellar progenitors.
author:
- |
Nathan Leigh$^{1}$, Alison Sills$^{2}$, Torsten Böker$^{1}$ [^1]\
$^{1}$European Space Agency, Space Science Department, Keplerlaan 1, 2200 AG Noordwijk, The Netherlands\
$^{2}$McMaster University, Department of Physics and Astronomy, 1280 Main St. W., Hamilton, Ontario, Canada, L8S 4M1
title: 'Modifying two-body relaxation in N-body systems by gas accretion'
---
\[firstpage\]
globular clusters: general – stellar dynamics – stars: formation – black hole physics.
Introduction {#intro}
============
For most of the life of a massive star cluster, two-body relaxation is the dominant physical mechanism driving its evolution [e.g. @henon60; @henon73; @spitzer87; @heggie03; @gieles11]. That is, the cumulative effects of long-range gravitational interactions between stars act to alter their orbits within the cluster. These interactions push the cluster toward a state of energy equipartition in which all objects have roughly the same kinetic energy. Consequently, the velocities of the most massive objects decrease, and they accumulate in the central regions of the cluster. Similarly, the velocities of the lowest mass objects increase, and they are subsequently dispersed to wider orbits. This mechanism, called mass segregation, also contributes to the escape of stars from their host cluster across the tidal boundary, with the probability of ejection increasing with decreasing stellar mass. Therefore, two-body relaxation acts to slowly modify the radial distribution of stellar masses within clusters, and can cause very dynamically evolved clusters to be severely depleted of their low-mass stars [e.g. @vonhippel98; @demarchi10; @leigh12].
Energy equipartition is an idealized state that should arise after the cumulative effects of many long-range interactions. In a real star cluster with a full spectrum of stellar masses, however, equipartition may not actually be achievable [e.g. @binney87; @heggie03]. As mentioned, the tendency towards energy equipartition reduces the velocities of the heaviest stars, causing them to sink in to the central cluster regions. Here, they are re-accelerated by the central cluster potential and gain kinetic energy. As this process proceeds, it leads to a contraction of the core and subsequently a shorter central relaxation time [e.g. @spitzer87; @heggie03]. A shorter relaxation time leads to a faster rate of energy transfer from heavier to lighter stars. Eventually, this makes the heaviest stars evolve away from equipartition.
This was first demonstrated by @spitzer69 using analytic techniques and a number of simplifying assumptions. @spitzer69 adopted a two-component system with masses $m_{\rm 1}$ and $m_{\rm 2}$ (where $m_{\rm 1} > m_{\rm 2}$), forming sub-systems with total masses $M_{\rm 1}$ and $M_{\rm 2}$. Provided that $M_{\rm 1} \ll M_{\rm 2}$, Spitzer derived the conditional requirement for a cluster to achieve energy equipartition in equilibrium. Based on this, Spitzer argued that energy equipartition could not be achieved in a cluster with a realistic mass spectrum, since there should always be enough mass in the heavier species for it to form a sub-system in the central cluster regions that decouples dynamically from the lighter species. This is commonly called the Spitzer instability [e.g. @spitzer87; @heggie03; @portegieszwart04].
A particularly compelling example of the Spitzer instability involves stellar-mass black holes (BHs) in globular clusters (GCs). @phinney91 first argued that BHs formed from the most massive stars should rapidly segregate into the core where they decouple dynamically from the rest of the cluster to form a distinct sub-system. Three-body scattering events then lead to the formation of BH-BH binaries, which in turn encounter other BHs and BH-BH binaries. These 3- and 4-body interactions are sufficiently energetic to eject the BHs from the cluster. In the end, most BHs are expected to be ejected, leaving only a handful behind.
This picture has recently been challenged in the literature. In particular, several authors have argued that the Spitzer instability should break down before most BHs are ejected [e.g. @moody09], and that the time-scale for all BHs to be ejected could exceed a Hubble time in some clusters [e.g. @downing10]. This view is supported by recent claims in the literature that stellar-mass BHs may be present in GCs in surprising numbers. For instance, @strader12 recently reported two flat-spectrum radio sources in M22, which appear to be accreting stellar-mass BHs. This suggests that this cluster could contain on the order of $\sim 5 - 100$ stellar-mass BHs. If BHs were indeed efficiently dynamically ejected, this, in turn, would suggest that a more substantial population of BHs once existed in M22, and likely other GCs as well.
The emerging picture for the formation of massive GCs involves multiple episodes of star formation [e.g. @piotto07; @gratton12; @conroy11; @conroy12]. In this context, @leigh13 recently considered the implications of the mass growth of BHs formed from massive progenitors belonging to the first generation due to accretion from the interstellar medium. The authors argued that, in principle, BHs could deplete a significant fraction of the available gas reservoir within $\lesssim 10^8$ years. If BHs were indeed to accrete efficiently from the ISM, they should not only grow in mass, but their velocities should also decrease due to conservation of momentum. This should preferentially accelerate the process of mass segregation for the BHs, causing them to rapidly accumulate in the central regions of the cluster if they did not form there in the first place. This could accelerate the dynamical decoupling of the BH sub-population from the rest of the system, and hence the phase of dynamical BH ejections due to the Spitzer instability.
In this paper, we consider how accretion from the interstellar medium affects the rate of mass segregation in a star cluster. We are especially interested in the implications for BHs in primordial GCs. Thus, we re-visit Spitzer’s two-component model to derive an accretion-modified relaxation time. We argue that the rate of mass segregation should be affected by accretion in the following way. First, assuming the accretion rate increases with the accretor mass, accretion acts to increase the range of any initial mass spectrum, driving the heaviest members to higher masses the fastest. Second, accretion acts to reduce the velocities of the accretors due to conservation of momentum, and it is the heaviest members whose velocities are reduced the fastest. Both of these effects exacerbate the Spitzer instability, and should accelerate the rate of mass segregation in a primordial star cluster.
In order to better quantify this qualitative picture, we present our adapted version of Spitzer’s two-component model in Section \[method\]. Specifically, we derive an accretion-modified relaxation time, as well as the critical accretion rate at which the rates of mass segregation due to both two-body relaxation and accretion are equal. We present our results in Section \[results\] for several different assumptions regarding the total cluster mass and accretion rate. In Section \[discussion\], we discuss the implications of our results for both star formation and stellar remnants in primordial globular clusters. We summarize our results in Section \[summary\].
Method
======
In this section, we present our analytic derivation of an accretion-modified relaxation time for a two-component model star cluster, and derive the critical accretion rate required for the mass segregation timescales due to two-body relaxation and accretion to be equal. We begin by summarizing briefly the relevant background related to both two-body relaxation and accretion.
Two-body relaxation {#relax}
-------------------
Consider a two-component model for a star cluster with component masses $m_1$ and $m_2$, such that $m_1 > m_2$. The populations for these two species have total masses $M_1$ and $M_2$ with $M_1 \ll M_2$. We let $v^2$ denote the initial mean square speed of *both* species, since at birth the cluster is not in a state of energy equipartition. The e-folding time for the tendency to equipartition bears a striking resemblance to the relaxation time [@heggie03]. Thus, to order of magnitude, the relaxation time can be approximated by calculating the time required for the mean square speed of the heavier species to fall from $v^2$ to a value $\sim m_2v^2/m_1$.
If the potential well of the lighter species is modelled using a parabolic profile, then equipartition will lead to the heavier species being confined to a region of size $\sim r_h\sqrt{m_2/m_1}$ [@heggie03]. The total mass of the heavier species within this region is $M_1$, whereas that for the lighter species is $M_2(m_2/m_1)^{3/2}$. At this point, however, it is not clear whether or not the lighter species remains the dominant mass component in this region. If not, the heavier species becomes increasingly affected by its own self-gravity, and can decouple dynamically from the remainder of the system. Consequently, it may only be possible to achieve equipartition provided [@spitzer69; @heggie03]: $$\label{eqn:spitzer-criterion}
M_1 \le M_2\Big( \frac{m_2}{m_1} \Big)^{3/2}.$$ This is known as Spitzer’s criterion.
In general, the half-mass relaxation time approximates the rate of two-body relaxation throughout the entire cluster. In GCs, it ranges from roughly a few hundred million years to the age of the Universe or longer, and is approximated by [@spitzer87]: $$\label{eqn:t-rh}
\tau_{\rm rh} [yr] = 1.7 \times 10^5[r_{\rm h} [pc]]^{3/2}N^{1/2}[\bar{m} [M_{\odot}]]^{-1/2},$$ where $r_{\rm h}$ is once again the half-mass radius, $N$ is the total number of stars within $r_{\rm h}$, and $\bar{m}$ is the average stellar mass. We assume that the value of $r_{\rm h}$ remains constant in time. This is reasonable since simulations have shown that $r_{\rm h}$ changes by not more than a factor of a few over the course of a typical cluster’s lifetime [@henon73; @spitzer87; @heggie03; @webb12]. The timescale for mass segregation due to two-body relaxation for an object of mass $m$ is then approximately [@vishniac78]: $$\label{eqn:tau-seg-time}
\tau_{\rm seg,2body}(m) = \frac{\bar{m}}{m}\tau_{\rm rh}.$$
Accretion {#accrete}
---------
A strict theoretical upper limit for the accretion rate is given by the Bondi-Hoyle limit [@bondi44]. In this approximation, the accretion is spherically symmetric, and the forces due to gas pressure are insignificant compared to gravitational forces. The background gas is treated as uniform and either stationary or moving with constant velocity relative to the accretor. This assumption gives reasonable accretion rates in the low-density, low-angular momentum regime. That is, provided the properties of the gas are such that the density, velocity, and total angular momentum are low, at least in the vicinity of the accretor, the Bondi-Hoyle limit approximately describes the true accretion rate [e.g. @fryxell88; @ruffert94; @ruffert97; @foglizzo99].
For large accretor masses and high gas densities, the Bondi-Hoyle rate can become extremely high. Here, pressure forces could play an important role in reducing the accretion rate. Indeed, this occurs if the outward continuum radiation force balances the inward gravitational force. This limit, called Eddington-limited accretion, gives considerably more modest accretion rates in the high gas density regime [@eddington26; @eddington30]. The Eddington rate should also be much closer to the true accretion rate if the gas contains significant angular momentum, and accretion proceeds mainly via angular momentum re-distribution within a disk [e.g. @rybicki79].
In general, theoretical studies have shown that the exact accretion rate can deviate substantially from the idealized cases described by Bondi-Hoyle and Eddington-limited accretion. For example, when an accretor is radiating at above the Eddington luminosity, significant amounts of gas can be expelled at high velocities due to the intense winds that are initiated [e.g. @king03]. This contributes to a reduction in the overall accretion rate. On the other hand, at very high accretion rates, photon-trapping can occur. This makes accretion disks radiatively inefficient and provides a means of circumventing the Eddington limit [@paczynsky80].
Numerical studies have also revealed the sensitivity of the accretion rate to small-scale gas dynamics. Recent work by @krumholz06 considered gas accretion onto point masses in a supersonically turbulent medium characterized by background density and velocity distributions that vary in both time and space. The authors show that in this regime, the accretion rate can either be described by the classical Bondi-Hoyle approximation, or a vorticity-dominated flow. In the latter case, the accretion rate can be significantly reduced relative to what is predicted by the Bondi-Hoyle limit [@krumholz04; @krumholz05]. Even more recently, @park13 studied the growth and luminosity of BHs in motion with respect to their surrounding medium using two-dimensional axis-symmetric numerical simulations. The authors show that the accretion rate can actually increase with increasing BH velocity, contrary to the naive predictions of simple analytic theory [@hoyle39].
In summary, theoretical work has shown that a wide range of accretion rates are possible. We will use the Bondi-Hoyle and Eddington limits for the accretion rate (combined with an accretion efficiency parameter) since these provide two different dependences on the accretor mass. However, the derivation presented in the subsequent section can be used to model other mass-dependences for the accretion rate as well.
Deriving the Relaxation Time {#derivation}
----------------------------
We are interested in the mass segregation timescale due to two-body relaxation for the heavier species in the two-component model described in Section \[relax\]. To first order, this is approximated by: $$\label{eqn:tau-seg-time}
\tau_{\rm seg,2body}(m_{\rm 1},t) = \frac{\bar{m}(t)}{m_{\rm 1}(t)}\tau_{\rm rh}(t),$$ where $\tau_{\rm rh}(t)$ denotes the half-mass relaxation time obtained by using the total number of objects $N = N_{\rm 1} + N_{\rm 2}$ and average object mass $\bar{m}(t) = (m_{\rm 1}(t)N_{\rm 1} + m_{\rm 2}(t)N_{\rm 2})/(N_{\rm 1} + N_{\rm 2})$ in Equation \[eqn:t-rh\].
We calculate the time-dependence for the mass of an object belonging to species $i$ as follows. First, assuming an object of initial mass $m_{\rm i}(0)$ accretes at a rate $\dot{m}_{\rm i} = dm_{\rm i}/dt$ for a total time $t$, we have: $$\label{eqn:mass-time}
m_{\rm i}(t) = m_{\rm i}(0) + \int_{0}^{t} \dot{m}_{\rm i}dt.$$ For the accretion rate, we assume a mass-dependence of the form: $$\label{eqn:acc-rate}
\dot{m}_{\rm i} = {\lambda}{\delta}m_{\rm i}^{\rm \epsilon},$$ where $\lambda$, $\delta$, and $\epsilon$ are all free parameters. The power-law exponent $\epsilon$ decides the mass-dependence for the accretion rate. The accretion coefficent $\delta$ is derived according to the physical assumptions that decide the rate of accretion. For example, adopting the Bondi-Hoyle approximation implies $\epsilon = 2$, and [e.g. @bondi44; @maccarone12; @leigh13]: $$\label{eqn:mdot-BH}
\delta = 7 \times 10^{-8} {M_{\odot}^{-1}}{\rm yr}^{-1} \Big( \frac{n}{\rm 10^6 cm
^{-3}} \Big) \Big( \frac{\sqrt{c_{\rm s}^2 + v^2}}{\rm 10^6 cm s^{-1}} \Big)^{-3},$$ where $n$ is the particle number density, $c_{\rm s}$ is the sound speed, and $v^2$ is the root-mean-square speed. Similarly, assuming Eddington-limited accretion implies $\epsilon = 1$, and [@eddington26; @eddington30]: $$\begin{gathered}
\label{eqn:mdot-Edd}
\delta = \frac{4{\pi}G}{{\eta}{\kappa}c} \\
= 2.2 \times 10^{-8} {\rm yr}^{-1},
\end{gathered}$$ where $c$ is the speed of light, ${\eta}c^2$ is the accretion yield from unit mass, and ${\kappa}$ is the electron scattering opacity. We take $\kappa = 0.34$ cm$^{2}$ g$^{-1}$, which assumes a hydrogen mass fraction of $X = 0.7$. We further adopt $\eta = 0.1$, which is reasonable if the accretors are BHs since this parameter quantifies the amount of energy radiated away during accretion. The accretion efficiency parameter $\lambda$, on the other hand, is left as a free parameter in our model.
In order to solve for the accretor mass as a function of time, Equation \[eqn:mass-time\] can be re-written such that the integral is with respect to mass. Provided $\epsilon > 1$, this gives for the mass of an object belonging to species $i$ at time $t$: $$\label{eqn:mass-time2}
m_{\rm i}(t) = \Big( m_{\rm i}(0)^{1-\epsilon} + {\lambda}{\delta}(1-\epsilon)t \Big)^{1/(1-\epsilon)}.$$ Similarly, for $\epsilon = 1$, we have: $$\label{eqn:mass-time3}
m_{\rm i}(t) = m_{\rm i}(0)e^{\rm {\lambda}{\delta}t}.$$ We arrive at the mass segregation timescale due to two-body relaxation for the heavier species by substituting either Equation \[eqn:mass-time2\] or Equation \[eqn:mass-time3\] for both species 1 and 2 into Equation \[eqn:tau-seg-time\].
In addition to two-body relaxation, accretion should also affect the stellar velocities via conservation of momentum. Provided the accretion rate increases with increasing accretor mass, this will reduce the kinetic energy of the heavier species faster than the lighter species, in rough analogy with the effects of two-body relaxation. Thus, accretion-induced changes in the stellar velocities should *accelerate* the rate of mass segregation. We calculate the time needed for accretion to change the velocities of the heavier species from a root-mean-square speed of $v^2$ to a value of $m_{\rm 2}/m_{\rm 1}v^2$. This corresponds to the time for accretion to affect the velocities of the heavier species by roughly the same amount as is done by two-body relaxation over a single relaxation time.
For a given accretion rate and initial masses, we calculate the time needed for the mean square speed of the heavier species to reach a value $m_{\rm 2}(t)v^2/m_{\rm 1}(t)$ using conservation of momentum, from an initial value $v$. The time needed for the heavier species to reach equipartition via accretion can then be written: $$\label{eqn:tau-acc}
\tau_{\rm seg,acc}(m_1,t) = \int_{\rm m_1(t)}^{\rm m_{1,f}(t)}\frac{dm_1}{\dot{m}_1},$$ where the final mass is $m_{\rm 1,f} = \sqrt{m_{\rm 1}^3/m_{\rm 2}}$ by conservation of momentum.
Substituting $m_{\rm 1,f}(t) = \sqrt{m_{\rm 1}(t)^3/m_{\rm 2}(t)}$ into Equation \[eqn:tau-acc\], we arrive at the timescale needed for accretion to push the heavier species into approximate equipartition: $$\label{eqn:tau-acc2}
\tau_{\rm seg,acc}(m_{\rm 1},t) = \frac{m_{\rm 1}(t)^{1-\epsilon}\Big( (m_{\rm 1}(t)/m_{\rm 2}(t))^{(1-\epsilon)/2} - 1 \Big)}{{\lambda}{\delta}(1 - \epsilon)},$$ for $\epsilon > 1$. Similarly, for $\epsilon = 1$ or Eddington-limited accretion, we obtain using Equation \[eqn:mass-time3\]: $$\label{eqn:tau-acc3}
\tau_{\rm seg,acc}(m_{\rm 1},t) = \frac{\ln(m_{\rm 1}(t)/m_{\rm 2}(t))}{2{\lambda}{\delta}}.$$ We consider accretion rates with a mass-dependence such that $\epsilon \ge 1$, since this includes both Bondi-Hoyle and Eddington-limited accretion, as well as intermediate and even steeper mass accretion rates.
The total rate (taken to be the inverse of the total mass segregation timescale $\tau_{\rm seg,tot}$) at which the heavier species achieves mass segregation can be written as the sum of the rate of two-body relaxation and the rate at which accretion pushes the heavier species to equipartition. Re-arranging this equation, we arrive at the total accretion-modified mass segregation timescale for the heavier species: $$\label{eqn:t-eq}
\tau_{\rm seg,tot}(m_{\rm 1},t) = \frac{\tau_{\rm seg,acc}(m_{\rm 1},t)\tau_{\rm seg,2body}(m_{\rm 1},t)}{\tau_{\rm seg,acc}(m_{\rm 1},t) + \tau_{\rm seg,2body}(m_{\rm 1},t)}.$$
Deriving the critical accretion rate {#critical}
------------------------------------
To derive the critical accretion rate $\delta_{\rm crit}$ at which two-body relaxation and accretion drive the mass segregation process at the same rate, we set $\tau_{\rm seg,acc} = \tau_{\rm seg,2body}$ and solve for $\delta$ as a function of $\epsilon$, $\lambda$, $m_{\rm 1}$, $m_{\rm 2}$, and $t_{\rm rh}$. This gives for $\epsilon > 1$: $$\label{eqn:delta2}
\delta_{\rm crit} = \frac{m_{\rm 1}(t)^{2-\epsilon}\Big( (m_{\rm 1}(t)/m_{\rm 2}(t))^{(1-\epsilon)/2} - 1 \Big)}{{\lambda}(1 - \epsilon)\bar{m}(t)t_{\rm rh}(t)}.$$
The procedure is similar for $\epsilon = 1$, except we use Equation \[eqn:tau-acc3\] instead of Equation \[eqn:tau-acc2\]. This gives for Eddington-limited accretion: $$\label{eqn:delta3}
\delta_{\rm crit} = \frac{m_{\rm 1}(t)\ln(m_{\rm 1}(t)/m_{\rm 2}(t))}{2{\lambda}\bar{m}(t)t_{\rm rh}(t)}.$$
In Section \[results\], we will use Equation \[eqn:delta2\] and Equation \[eqn:delta3\] in order to study the interplay between our assumptions regarding the gas properties, which affect the accretion rate, and our assumption for the total cluster mass, which determines the rate of two-body relaxation.
Accretion efficiency {#lambda-time}
--------------------
Given our limited understanding of the precise physics of accretion onto a BH, it is not possible to reliably define a functional form for the accretion efficiency parameter $\lambda$. In principle, any realistic accretion model should include a *time-dependence* for $\lambda(t)$.[^2] For example, fluctuations in the local gas density due to turbulence, a gradual or even sudden depletion of the available gas reservoir, or dynamical interactions between accreting objects may cause the accretion efficiency to vary over time.
Our analysis is easily modified to treat time-dependent accretion rates by substituting an appropriate choice for $\lambda(t)$ into either Equation \[eqn:tau-acc2\] or \[eqn:tau-acc3\], and then solving for $\tau_{\rm seg,acc}$. Plausible choices for $\lambda(t)$ may either oscillate or decline (steadily or abruptly) in time. The first case, i.e. an oscillating accretion efficiency, is more easily understood, because under these circumstances, our analysis can simply be interpreted as discussing the *time-averaged* accretion efficiency parameter. Thus, in the subsequent sections, we assume an oscillating (or constant) accretion efficiency parameter, and discuss only the time-averaged value.
For example, the function $\lambda(t) = (1+{\rm sin}({\pi}t/t_{\rm 0}))$ oscillates between 0 and 2 with a frequency of $2/t_{\rm 0}$. In this case, the time-averaged value for $\lambda(t)$ is equal to 1, so that the time-averaged value for $\tau_{\rm seg,acc}$ remains the same as for a constant $\lambda = 1$.
Accretion efficiency parameters that oscillate in time should be suitable to cases where the accretors have alternating “on” and “off” phases. This may well be the case with accreting BHs, since the radiation emitted due to accretion can heat the surrounding gas, which in turn decreases the accretion rate [e.g. @blaes95]. In this case, the source of energetic photons responsible for heating the gas is turned off, allowing the gas to cool and accretion to re-start in a “feedback regulated” loop [e.g. @king03; @yuan09].
Accretion efficiency parameters that decline in time should be appropriate to cases where the available gas reservoir is depleted over time. This could arise gradually if the gas is used to form stars, or if significant quantities of gas are accreted by BHs. Alternatively, the gas reservoir could be depleted suddenly, e.g. due to, energy injected from supernovae, stellar winds, or winds from accreting compact objects. In either case, Equations \[eqn:tau-acc2\] and \[eqn:tau-acc3\] should include the explicit time-dependence for the accretion efficiency parameter. This will contribute to an increase in $\tau_{\rm seg,acc}$ with time, since the decreasing gas mass should translate into a decreasing gas density, and hence accretion rate. In Section \[results\], we will assume that the amount of gas lost from the system is negligible over the calculated mass segregation timescales, and our interpretation of $\lambda$ as a time-independent quantity remains valid. This is reasonable provided the mass segregation timescales due to accretion are much less than the timescale for gas depletion. As we will show in Section \[results\], the current picture for the formation of globular clusters and their multiple populations is consistent with this scenario [e.g. @krause12; @krause13; @leigh13].
Results
=======
In this section, we present the results of our analytic two-component model for an accretion-modified two-body relaxation time. Our aim is to quantify the relative rates at which two-body relaxation and gas accretion drive a star cluster towards mass segregation, as a function of our assumptions for the gas properties, component masses, and total system mass. To this end, we present the time evolution of all three mass segregation timescales, namely $\tau_{\rm seg,acc}$, $\tau_{\rm seg,2body}$, and $\tau_{\rm seg,tot}$, and discuss the critical accretion rate required for the mass segregation timescales due to two-body relaxation and accretion to be equal as a function of the mass-dependence for the accretion rate.
Time evolution of the mass segregation timescales {#time-evolution}
-------------------------------------------------
We begin by quantifying the relative rates of mass segregation due to two-body relaxation and gas accretion for different model assumptions. Specifically, we consider several different mass ratios and total system masses for our two-component model, as well as different mass-dependences for the rate of accretion. This is meant to quantify the sensitivity of the two different mass segregation mechanisms to the cluster and gas properties that decide their rates.
First, we describe our assumptions for the two-component model star cluster, which are needed in order to calculate $\tau_{\rm seg,2body}$. We adopt $m_{\rm 2} = 1$ M$_{\odot}$ for the lighter species, but consider two different masses for the heavier species, namely $m_{\rm 1} = 10$ M$_{\odot}$ and $m_{\rm 1} = 50$ M$_{\odot}$. We assume a population size of $N_{\rm 1} = 10^2$ for the heavier species, but vary the population size of the lighter species by considering the values $N_{\rm 2} = 10^5, 10^6, 10^7$. The component masses and population sizes are chosen to represent reasonable mass ratios between the average stellar and BH masses, and to ensure that the Spitzer criterion (i.e. Equation \[eqn:spitzer-criterion\]) is initially satisfied. We adopt a half-mass radius for our model cluster of $r_{\rm h} = 10$ pc, and note that assuming a lower value for the half-mass radius would only shorten the calculated mass segregation timescales.
Next, we describe our assumptions for the properties of the accreted gas, which are needed to calculate $\tau_{\rm seg,acc}$, and therefore $\tau_{\rm seg,tot}$. We assume a uniform time-independent gas density throughout the cluster, so that the accretion rate changes only with the stellar mass. We further assume that the gas is always at rest relative to the accretor when calculating the final accretor velocity using conservation of momentum. For the accretors, we adopt a root mean-square-speed of $v = 10$ km s$^{-1}$, which is guided by the relation $v = \sqrt{2GM/5r_h}$ [@binney87] for a total cluster mass $M \sim 10^5-10^6$ M$_{\odot}$. For the gas, we assume a sound speed of $c_{\rm s} = 10$ km s$^{-1}$, and a particle number density of $n = 10^6$ cm$^{-3}$. These assumptions should be reasonable for what is expected in a massive primordial GC for the first $\sim 10^8$ years [e.g. @dercole08; @maccarone12; @conroy12; @krause12; @krause13; @leigh13].
We show our results for two different mass-dependencies for the accretion rate. The left panels in Figure \[fig:tau-seg-BH\] show our results assuming $\epsilon = 2$ in Equation \[eqn:acc-rate\], which corresponds to Bondi-Hoyle accretion. We use Equation \[eqn:mdot-BH\] for $\delta$, and $\lambda = 0.1$ for the accretion efficiency parameter. The panels to the right in Figure \[fig:tau-seg-BH\] show our results assuming Eddington-limited accretion, which means that $\epsilon = 1$ in Equation \[eqn:acc-rate\] and we use Equation \[eqn:mdot-Edd\] for $\delta$.
The main conclusion to be drawn from Figure \[fig:tau-seg-BH\] is that, for all but the least massive clusters and the lowest accretion rates considered here, the rate of mass segregation due to accretion can actually exceed the rate due to two-body relaxation. The timescale at which this occurs is on the order of $\sim 10^8$ years. Interestingly, this timescale is similar to the total time thought to be required for multiple episodes of star formation to occur in primordial GCs [e.g. @conroy11; @conroy12]. Thus, our results suggest that accretion from the ISM could significantly affect both the spatial and velocity distributions of the heaviest objects in a primordial GC *before* the gas reservoir is depleted. For a typical primordial GC, this should be the case provided the average accretion rate is greater than $\sim 5-10$% of the Eddington-limited rate, assuming the mass-dependence for the accretion rate is linear. Similarly, if the accretion rate scales with the square of the accretor mass, then accretion from the ISM is non-negligible as long as the average accretion rate is greater than 1-10% of the Bondi-Hoyle rate.
The critical accretion rate {#critical2}
---------------------------
In this section, we calculate the critical accretion rate required for the rates of mass segregation due to two-body relaxation and gas accretion to be equal. Our aim is to quantify the relative importance of the different parameters for the cluster and gas properties in establishing a balance between the competing effects of two-body relaxation and accretion.
In Figure \[fig:delta-crit\], we show the critical accretion rate $\delta_{\rm crit}$ as a function of the mass of the heavier species $m_{\rm 1}$. These results are calculated using Equations \[eqn:delta2\] and \[eqn:delta3\], which correspond to Bondi-Hoyle (blue) and Eddington-limited (red) accretion, respectively. In both cases, we assume a constant mass for the lighter species of $m_{\rm 2} = 1$ M$_{\odot}$, a constant population size for the heavier species of $N_{\rm 1} = 10^2$, and a constant accretion efficiency parameter $\lambda = 1.0$. In order to vary the rate of two-body relaxation without affecting the rate of mass segregation due to accretion, we consider three different population sizes for the lighter species, namely $N_{\rm 2} = 10^5, 10^6, 10^7$.
![The critical accretion rate $\delta_{\rm crit}$ at which the mass segregation timescales due to two-body relaxation and accretion are equal, shown as a function of the mass of the heavier species $m_{\rm 1}$. The blue curves show the results assuming Bondi-Hoyle accretion, whereas the red curves are for Eddington-limited accretion. The dashed, solid, and dash-dotted lines correspond to $N_{\rm 1} = 10^5, 10^6, 10^7$, respectively. For both sets of curves, we assume a constant mass for the lighter species of $m_{\rm 2} = 1$ M$_{\odot}$, and we set $\lambda = 1.0$ for the accretion efficiency parameter. \[fig:delta-crit\]](fig3.eps){width="\columnwidth"}
Figure \[fig:delta-crit\] shows that for the case of Eddington-limited accretion, the critical accretion rate increases with increasing accretor mass. This is because, as the accretor mass increases, the mass segregation timescale due to two-body relaxation decreases faster than the mass segregation timescale due to accretion. In the case of Bondi-Hoyle accretion, however, the critical accretion rate depends only very weakly on the accretor mass, which is due to the fact that the accretion rate scales as the square of the accretor mass. We emphasize that with the exception of the Eddington-limited rate at large accretor masses, the critical accretion rates are comparable to, or even smaller than, those observed in nearby star-forming regions [e.g. @mckee07].
Discussion
==========
In this section, we discuss the implications of our results for mass segregation in primordial globular clusters, in particular with regards to black holes.
Enhanced mass segregation {#primordial}
-------------------------
One of the key conclusions arising from our analysis is that accretion should accelerate the rate at which a star cluster becomes mass segregated compared to two-body relaxation alone. In fact, accretion can dominate over two-body relaxation in massive clusters for accretion rates that are below the Bondi-Hoyle or Eddington-limited rates by one or even two orders of magnitude. This is because the relaxation time increases with the cluster mass, whereas the mass segregation timescale due to accretion is independent of the cluster mass (assuming that the gas properties are independent of the cluster mass). Our results suggest that two-body relaxation should dominate the mass segregation process in low-mass primordial clusters with global relaxation times $\lesssim 10^7-10^8$ years and hence total cluster masses $\lesssim 10^4-10^5$ M$_{\odot}$, provided that our models assumptions are valid. In this regime, accretion should only have a small effect on the rate of mass segregation, and long-range gravitational interactions should alter the accretors’ velocities faster than they are reduced by the accretion process. In more massive clusters, however, the damping effects of accretion could play a significant role in accelerating the rate of mass segregation.
The effects of a realistic mass spectrum {#mass-spec}
----------------------------------------
Our assumption of a two-component model serves to demonstrate the effects of accretion from the ISM on a cluster’s dynamical evolution. The qualitative nature of our results should hold if a realistic mass spectrum is adopted instead. Accretion can modify the distribution of velocities on relatively short timescales in gas-embedded clusters. How exactly the velocities become modified depends on several parameters, in particular the mass spectrum, the total cluster mass, the properties of the gas, and the accretion rate.
In general, we expect accretion to amplify or exacerbate the Spitzer instability. This is due to the mass-dependence of the accretion rate, and the fact that typically $\epsilon > 0$ in Equation \[eqn:acc-rate\] (i.e. the accretion rate), which causes the more massive component to grow in mass the fastest. Thus, according to Equation \[eqn:spitzer-criterion\], Spitzer’s criterion should typically break down sooner as a result of accretion. However, accretion also acts to reduce the velocities of the accretors due to conservation of momentum, and this should most strongly impact the most massive objects due once again to the mass-dependence of the accretion rate. This can actually serve to combat the effects of the Spitzer instability by inhibiting the most massive objects from decoupling dynamically from the rest of the system once they have segregated to the central regions of the cluster. Clearly, a more sophisticated treatment will be needed in future studies in order to properly quantify these effects and their implications for the Spitzer instability, and a cluster’s ability to achieve energy equipartition.
Gas properties and the accretion rate {#gas-prop}
-------------------------------------
We stress that our results depend sensitively on our assumption for the accretion rate, which is poorly constrained, both theoretically and observationally. Indeed, the accretion efficiency parameter $\lambda$ adopted in Equation \[eqn:acc-rate\] is needed to account for the many sources of uncertainty in the gas properties, and hence the accretion rate. For example, our assumption of a uniform, time-independent gas density is an over-simplification. For one, stellar winds and supernovae could create over- and under-densities in the form of sheets and/or filaments, and the efficiency of these processes should fluctuate in time given the presence of a realistic mass function combined with stellar evolution and the cluster dynamics [e.g. @krause12; @krause13]. These effects could contribute to a reduction in the accretion rate by increasing the relative velocity between the gas and the accretors, or by reducing the gas density along the trajectories of the accretors. Realistic hydrodynamical simulations of star cluster formation will be needed in order to properly quantify these effects and their implications for the accretion rate.
Our results can be used to guide the parameter space relevant to these future studies. In particular, we have placed a lower limit on the minimum accretion rate required for accretion to significantly affect the distribution of stellar velocities on timescales shorter than the relaxation time, as a function of the cluster and gas properties. Specifically, the results of our simple model suggest that, for a typical primordial GC, the average accretion rate cannot be much less than $\sim 5-10$% of the Eddington-limited rate, assuming the mass-dependence for the accretion rate is linear. Similarly, for our model assumptions, the average accretion rate cannot be much less than $1-10$% of the Bondi-Hoyle rate if the accretion rate scales with the square of the accretor mass.
We have adopted the same root-mean-square speed for all models, independent of the total cluster mass. This is a reasonable assumption since the root-mean-square speed scales as $v \propto (M/r_{\rm h})^{0.5}$, and $r_{\rm h}$ itself depends weakly on the total cluster mass. Thus, in total, the root-mean-square speed depends only very weakly on the total cluster mass. Nevertheless, if the accretion rate scales inversely with the velocity of the accretor, as is the case with the Bondi-Hoyle approximation, then the dependence of the root-mean-square speed on the total cluster mass should contribute to a decrease in the accretion rate with increasing cluster mass. A proper treatment of this effect is beyond the scope of this paper, however, it should certainly be considered in future studies.
Black hole dynamics {#BH-dynamics}
-------------------
The results presented in this paper are especially relevant for black holes in primordial globular clusters, since they should be the most massive objects in the cluster within a few Myrs of its formation. Recent evidence suggests that there should be a substantial gas reservoir in GCs for the first $\sim 10^8$ years [e.g. @conroy11], albeit perhaps intermittently, and that nearly all BHs should form from the most massive cluster members within the first few Myrs [e.g. @maeder09]. It follows that the BHs could have on the order of $10^8$ years to accrete gas from the ISM. Additionally, since any BHs formed from progenitors more massive than $\sim 50$ M$_{\odot}$ are only slightly less massive than the progenitors themselves and do not experience natal kicks [@fryer12], these BHs should have both the shortest mass segregation timescales due to two-body relaxation and the highest accretion rates (ignoring BH winds and/or Compton heating; see below).
The key point is that accretion should act to reduce the mass segregation times of BHs in primordial GCs, and that this effect could be dramatic. Beyond this, more detailed modeling will be needed to determine the fates of the BHs. In particular, should the increased rate of mass segregation contribute to accelerating the onset of the hypothesized phase of dynamical BH ejections? Or could the damping effects of accretion be so dramatic that the BHs are driven to merge [e.g. @davies11]? If so, the formation of an intermediate-mass BH (IMBH) could be the inevitable result. Alternatively, it could be that black hole winds are sufficiently powerful to eject the bulk of the gas from the cluster. Another possibility is that the gas in the immediate vicinity of the BHs becomes very hot due to, for example, Compton heating [@blaes95; @yuan09], such that the accretion rate becomes drastically reduced and BH growth is severely limited?
A better understanding of how the presence of significant quantities of gas modifies the black hole dynamics in a primordial GC could help to constrain the initial cluster conditions. For example, if massive BHs should inevitably merge in the presence of gas but no IMBHs are observed in present-day GCs, does this necessarily imply that the BHs never formed in the first place? If so, this would suggest that stars with masses $\gtrsim 50$ M$_{\odot}$ must have been rare. This could be the case, for instance, if massive primordial GCs were assembled from the mergers of many low-mass sub-clumps, as opposed to a single monolothic collapse. This is because the mass of the most massive cluster member correlates with the total cluster mass [e.g. @kirk11; @kirk12], and hence the massive stellar progenitors of the most massive BHs are unlikely to form in low-mass clusters.
Summary
=======
In this paper, we have considered the effects of accretion from the interstellar medium on the rate of two-body relaxation in a star cluster. To do this, we derived an accretion-modified relaxation time by adapting Spitzer’s two-component model to include the effects of accretion. We considered several different mass-dependencies and efficiency factors for the accretion rate, as well as different mass ratios for the two components of the model.
We have shown that accretion acts to increase the rate of mass segregation. This is because the relaxation time is inversely proportional to the average mass, which increases due to accretion. There are two additional effects that accelerate the mass segregation process, assuming that the accretion rate increases with the accretor mass. First, accretion acts to increase the range of any initial mass spectrum, quickly driving the heaviest members to even higher masses. Second, accretion acts to reduce the velocities of the accretors due to conservation of momentum, and it is the heaviest members that are affected the most. Using our two-component model, these effects have been quantified as a function of the accretion rate, the total cluster mass, and the component masses. We have discussed our results in the context of the dynamical evolution of primordial globular clusters and their black hole sub-populations.
Acknowledgments {#acknowledgments .unnumbered}
===============
We kindly thank an anonymous referee for helpful suggestions that improved our manuscript, as well as Cole Miller for useful discussions. AS is supported by NSERC.
[99]{}
Baumgardt H., Makino J. 2003, MNRAS, 340, 227 Baumgardt H., De Marchi G., Kroupa P. 2008, ApJ, 685, 247 Binney J., Tremaine S., 1987, Galactic Dynamics (Princeton: Princeton University Press) Blaes O., Warren O., Madau P. 1995, ApJ, 454, 370 Bondi H., Hoyle F. 1944, MNRAS, 104, 273 Conroy C., Spergel D. N. 2011, ApJ, 726, 36 Conroy C. 2012, ApJ, 758, 21 Davies M. B., Miller M. C., Bellovary J. M. 2011, ApJ, 740, 42 De Angeli F., Piotto G., Cassisi S., Busso G., Recio-Blanco A., Salaris M., Aparicio A., Rosenberg A. 2005, AJ, 130, 116 De Marchi G., Paresce F., Pulone L. 2007, ApJ, 656, L65 De Marchi G., Paresce F., Portegies Zwart S. 2010, ApJ, 718, 105 D’Ercole A., Vesperini E., D’Antona F., McMillan S. L. W., Recchi S. 2008, MNRAS, 391, 825 Downing J. M. B., Benacquista M. J., Giersz M., Spurzem R. 2010, MNRAS, 407, 1946 Dopita M. A., Smith G. H. 1986, ApJ, 304, 283 Eddington A. S. 1926, The Internal Constitution of the Stars (Cambridge: Cambridge University Press) Eddington A. S. 1930, MNRAS, 90, 279 Fall S. M., Zhang Q. 2001, ApJ, 561, 751 Foglizzo T., Ruffert M. 1999, A&A, 347, 901 Fryer C. L., Kalogera V. 2001, ApJ, 554, 548 Fryer C. L., Belczynski K., Wiktorowicz G., Dominik M., Kalogera V., Holz D. E. 2012, ApJ, 749, 91 Fryxell B. A., Taam R. E. 1988, ApJ, 335, 862 Gieles M., Heggie D., Zhao H. 2011, MNRAS, accepted Gratton R., Carretta E., Bragaglia A. 2012, Astronomy & Astrophysics Review, in press (arXiv:1201.6526) Harris, W. E. 1996, AJ, 112, 1487 (2010 update) Heggie D. C., Hut P. 2003, The Gravitational Million-Body Problem: A Multidisciplinary Approach to Star Cluster Dynamics (Cambridge: Cambridge University Press) Heggie D. C., Giersz M. 2008, MNRAS, 389, 1858 Heggie D. C., Giersz M. 2009, MNRAS, 397, 46 Henon M. 1960, Annales d’Astrophysique, 23, 668 Henon M. 1973, Dynamical Structure and Evolution of Dense Stellar Systems, ed. L. Martinet & M. Mayor (Geneva Obs.) Pringle J. E. 2006, MNRAS, 373, L90 Hoyle F., Lyttleton R. A. 1939, in Proceedings of the Cambridge Philosophical Society, 35 King A. R., Pounds K. A. 2003, MNRAS, 345, 657 Kirk H., Myers P. C. 2011, ApJ, 727, 64 Kirk H., Myers P. C. 2012, ApJ, 745, 131 Krause M., Charbonnel C., Decressin T., Meynet G., Prantzos N., Diehl R. 2012, A&A, 546, L5 Krause M., Charbonnel C., Decressin T., Meynet G., Prantzos N.. 2013, A&A, accepted (arXiv:1302.2494) Krumholz M. R., McKee C. F., Klein R. I. 2004, ApJ, 611, 399 Krumholz M. R., McKee C. F., Klein R. I. 2005, ApJ, 618, 757 Krumholz M. R., McKee C. F., Klein R. I. 2006, ApJ, 638, 369 Leigh N. W., Umbreit S., Sills A., Knigge C., Glebbeek E., Sarajedini A. 2012, MNRAS, 422, 1592 Leigh N. W., Böker T., Maccarone T. J., Perets H. B. 2013, MNRAS, 429, 2997 Maccarone T. J., Zurek D. R. 2012, MNRAS, 423, 2 Maeder A. 2009, Physics, Formation and Evolution of Rotating Stars. Berlin: Springer-Verlag Marks M., Kroupa P., Baumgardt H. 2008, MNRAS, 386, 2047 Marks M., Kroupa P. 2010, MNRAS, 406, 2000 McKee C. F., Ostriker E. C. 2007, ARA&A, 45, 565 Moody K., Sigurdsson S. 2009, ApJ, 690, 1370 Paczynsky B., Wiita P. J. 1980, A&A, 88, 23 Park K., Ricotti M. 2013, ApJ, submitted (arXiv:1211.0542) Phinney S. E., Sigurdsson S. 1991, Nature, 349, 220 Piotto G., Bedin L. R., Anderson J., King I. R., Cassisi S., Milone A. P., Villanova S., Pietrin- ferni A., Renzini A. 2007, ApJ, 661, L53 Portegies Zwart S. F., Baumgardt H., Hut P., Makino J., McMillan S. L. W. 2004, Nature, 428, 724 Ruffert M. 1994, ApJ, 427, 342 Ruffert M. 1997, A&A, 317, 793 Rybicki G. B., Lightman A. P. 1979, Radiative Processes in Astrophysics (New York: Wiley-Interscience) Spitzer L. Jr. 1969, ApJ, 158, 139 Spitzer L. Jr. 1987, Dynamical Evolution of Globular Clusters (Princeton, NJ: Princeton Univ. Press) Strader J., Chomiuk L., Maccarone T. J., Miller-Jones J. C. A., Seth A. C. 2012, Nature, 490, 71 Tremaine S. D., Ostriker J. P., Spitzer L. Jr. 1975, ApJ, 196, 407 Tutukov A. V. 1978, A&A, 70, 57 Vesperini E., Heggie D. C. 1997, MNRAS, 289, 898 Vishniac E. T. 1978, ApJ, 223, 986 von Hippel T., Sarajedini A. 1998, AJ, 116, 1789 Webb J. J., Harris W. E., Sills A. 2012, ApJ, 759, 39 Yuan F., Xie F., Ostriker J. P. 2009, ApJ, 691, 98 Zonoozi A. H., Kupper A. H. W., Baumgardt H., Haghi H., Kroupa P., Hilker M. 2011, MNRAS, 411, 1989 Zhang Q., Fall S. M. 1999, ApJ, 527, 81
\[lastpage\]
[^1]: E-mail: [email protected] (NL), [email protected] (AS), [email protected] (TB)
[^2]: Alternatively, the time-dependence can be absorbed directly into the parameter $\delta$ in Equations \[eqn:tau-acc2\] and \[eqn:tau-acc3\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use the backstepping method to study the stabilization of a 1-D linear transport equation on the interval $(0,L)$, by controlling the scalar amplitude of a piecewise regular function of the space variable in the source term. We prove that if the system is controllable in a periodic Sobolev space of order greater than $1$, then the system can be stabilized exponentially in that space and, for any given decay rate, we give an explicit feedback law that achieves that decay rate.'
author:
- '[C]{}hristophe [Z]{}hang'
title: 'Internal rapid stabilization of a 1-D linear transport equation with a scalar feedback'
---
**Keywords.** Backstepping, transport equation, Fredholm transformations, stabilization, rapid stabilization, internal control.
Introduction
============
We study the linear 1-D hyperbolic equation $$\label{GeneralSystem}
\left\{\begin{aligned}
y_t + y_x + a(x) y& = u(t) \tilde{\varphi} (x), \ x\in [0,L], \\
y(t,0) & = y(t,L), \ \forall t \geq 0,
\end{aligned}\right.$$ where $a$ is continuous, real-valued, $\tilde{\varphi}$ is a given real-valued function that will have to satisfy certain conditions, and at time $t$, $y(t, \cdot)$ is the state and $u(t)$ is the control. As the system can be transformed into $$\label{System}
\left\{\begin{aligned}
\alpha_t + \alpha_x + \mu \alpha & = u(t) \varphi (x), \ x\in [0,L], \\
\alpha(t,0) & = \alpha(t,L), \ \forall t \geq 0,
\end{aligned}\right.$$ through the state transformation $$\alpha(t,x) := e^{\int_0^x a (s)ds - \mu x} y(x,t),$$ where $\mu=\int_0^L a(s)ds$, and with $$\varphi(x):= e^{\int_0^x a (s)ds- \mu x} \tilde{\varphi} (x),$$ we will focus on systems of the form in this article.
Notations and definitions
-------------------------
We note $\ell^2$ the space of summable square series $\ell^2 ({\mathbb{Z}})$. To simplify the notations, we will note $L^2$ the space $L^2(0,L)$ of complex-valued $L^2$ functions, with its hermitian product $$\label{ScalarProduct}
\langle f, g \rangle = \int_0^L f(x)\overline{g(x)}dx, \quad \forall f, g\in L^2,$$ and the associated norm $\|\cdot\|$. We also use the following notation $$e_n(x)=\frac1{\sqrt{L}} e^{\frac{2i\pi}{L} n x}, \quad \forall n \in {\mathbb{Z}},$$ the usual Hilbert basis for $L^2$. For a function $f\in L^2$, we will note $(f_n) \in \ell^2$ its coefficients in this basis: $$f=\sum_{n \in {\mathbb{Z}}} f_n e_n.$$
Note that with this notation, we have $$\bar{f}=\sum_{n \in {\mathbb{Z}}} \overline{f_{-n}} e_n,$$ so that, in particular, if $f$ is real-valued: $$f_{-n} = \overline{f_n}, \quad \forall n \in {\mathbb{Z}}.$$
Functions of $L^2$ can also be seen as $L$-periodic functions on ${\mathbb{R}}$, by the usual $L$-periodic continuation: in this article, for any $f\in L^2$ we will also note $f$ its $L$-periodic continuation on ${\mathbb{R}}$.
We will use the following definition of the convolution product on $L$-periodic functions: $$\label{Conv}
f \star g = \sum_{n\in {\mathbb{Z}}} f_n g_n e_n= \int_{0}^L f(s)g(\cdot -s) ds \in L^2,\quad \forall f, g\in L^2,$$ where $g(x-s)$ should be understood as the value taken in $x-s$ by the $L$-periodic continuation of $g$.
Let us now note $\mathcal{E}$ the space of finite linear combinations of the $(e_n)_{n \in {\mathbb{Z}}}$. Then, any sequence $(f_n)_{n \in {\mathbb{Z}}}$ defines an element $f$ of $\mathcal{E}^\prime$: $$\langle e_n, f \rangle =\overline{f_n}.$$ On this space of linear forms, derivation can be defined by duality: $$f^\prime= \left(\frac{2i\pi n}L f_n \right), \quad \forall f \in {\mathcal{E}^\prime}.$$ We also define the following spaces:
Let $m \in {\mathbb{N}}$. We note $H^m$ the usual Sobolev spaces on the interval $(0,L)$, equipped with the Hermitian product $$\langle f, g\rangle_m = \int_0^L \partial^m f \overline{\partial^m g}, \quad \forall f,g \in H^m,$$ and the associated norm $\|\cdot\|_m$.
For $m\geq1$ we also define $H^m_{(pw)}$ the space of piecewise $H^m$ functions, that is, $f\in H^m_{(pw)}$ if there exists a finite number $d$ of points $(\sigma_j)_{1 \leq j \leq d} \in [0, L]$ such that, noting $\sigma_0:=0$ and $\sigma_{d+1}:=L$, $f$ is $H^m$ on every $[\sigma_j, \sigma_{j+1}]$ for $0\leq j \leq d$. This space can be equipped with the norm $$\|f\|_{m, pw}:=\sum_{j=0}^d \|f_{|[\sigma_j, \sigma_{j+1}]} \|_{H^m(\sigma_j, \sigma_{j+1})}.$$
For $s>0$, we also define the periodic Sobolev space $H^s_{per}$ as the subspace of $L^2$ functions $f=\sum_{n \in {\mathbb{Z}}} f_n e_n$ such that $$\sum_{n \in {\mathbb{Z}}} \left(1+\left|\frac{2i\pi n}{L}\right|^{2s}\right)|f_n|^2 < \infty.$$ $H^s$ is a Hilbert space, equipped with the Hermitian product $$\langle f, g\rangle_s = \sum_{n \in {\mathbb{Z}}} \left(1+\left|\frac{2i\pi n}{L}\right|^{2s}\right) f_n \overline{g_n}, \quad \forall f,g \in H^s,$$ and the associated norm $\|\cdot\|_s$, as well as the Hilbert basis $$(e_n^s):=\left(\frac{e_n}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\right).$$
Note that for $m \in {\mathbb{N}}$, $H^m_{per}$ is a closed subspace of $H^m$, with the same scalar product and norm, thanks to the Parseval identity. Moreover, $$H^m_{per}=\left\{f\in H^m, \quad f^{(i)}(0)=f^{(i)}(L), \forall i \in \{0, \cdots, m\}\right\}.$$
Main result
-----------
To stabilize , we will be considering linear feedbacks of the form $$\langle \alpha(t), F \rangle = \sum_{n\in {\mathbb{Z}}} \overline{F_n} \alpha_n(t)=\int_{0}^L \bar{F}(s) \alpha(s)ds$$ where $F\in {\mathcal{E}^\prime}$ and $(F_n) \in {\mathbb{C}}^{\mathbb{Z}}$ are its Fourier coefficients, and $F$ is real-valued, that is, $$F_{-n} = \overline{F_n}, \quad \forall n \in {\mathbb{Z}}.$$ In fact, the integral notation will appear as purely formal, as the $(F_n)$ will have a prescribed growth, so that $F \notin L^2$. The associated closed-loop system now writes $$\label{ClosedLoop}
\left\{\begin{aligned}
\alpha_t + \alpha_x +\mu \alpha & = \langle \alpha (t), F \rangle \varphi (x), \ x\in [0,1], \\
\alpha(t,0) & = \alpha(t,L), \ \forall t \geq 0.
\end{aligned}\right.$$ This is a linear transport equation, which we seek to stabilize with an internal, scalar feedback, given by a real-valued feedback law. This article aims at proving the following class of stabilization results:
\[MainResult\] Let $m \geq 1$. Let $\varphi \in H^m_{(pw)} \cap H^{m-1}_{per}$ such that $$\label{phiGrowth}
\frac{c}{\sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2m}}} \leq |\varphi_n| \leq \frac{C}{\sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2m}}}, \quad \forall n \in {\mathbb{Z}},$$ where $c,C >0$ are the optimal constants for these inequalities. Then, for every $\lambda \geq 0$ there exists a stationary feedback law $F$ such that for all $\alpha_0 \in H^{m}_{per}$ the closed-loop system has a solution $\alpha(t)$ which satisfies $$\| \alpha(t) \|_{m} \leq \left(\frac{C}{c}\right)^2e^{(\mu+\lambda) L} e^{-\lambda t} \|\alpha_0\|_{m},\quad \forall t \geq 0.$$
The growth restriction on the Fourier coefficients of $\varphi$ can be written, more intuitively, and for some other constants $c^\prime, C^\prime >0$, $$\frac{c^\prime}{1+\left|\frac{2i \pi n}{L}\right|^{m}} \leq |\varphi_n| \leq \frac{C^\prime}{1+\left|\frac{2i \pi n}{L}\right|^{m}}, \quad \forall n \in {\mathbb{Z}},$$ and corresponds to the necessary and sufficient condition for the controllability of system in $H^{m}_{per}$. Indeed, this system satisfies an observability inequality in $H^{m}_{per}$ if and only if $\varphi$ satisfies . The controllability of system , in turn, will allow us to “shift its poles”, using the so-called backstepping method.
On the other hand, the additional regularity $\varphi \in H^m_{(pw)}$ gives us the following equality, by iterated integration by parts on each interval $[\sigma_j, \sigma_{j+1}]$: $$\label{DAphi}
\varphi_n=\tau^\varphi_n \frac{(-1)^{m-1}}{\left(\frac{2i\pi}{L}n\right)^m} + \frac{(-1)^m}{\left(\frac{2i\pi}{L}n\right)^m} \sum_{j=0}^{d}\left\langle \chi_{[\sigma_j, \sigma_{j+1}]}\partial^m \varphi, e_n\right\rangle, \quad \forall n\in {\mathbb{Z}}^\ast,$$ where $$\tau^\varphi_n:= \frac1{\sqrt{L}}\left( \partial^{m-1}\varphi(L)-\partial^{m-1}\varphi(0) + \sum_{j=1}^d e^{-\frac{2i\pi}L n\sigma_j} (\partial^{m-1} \varphi (\sigma_j^-) - \partial^{m-1}\varphi(\sigma_j^+))\right), \quad \forall n \in {\mathbb{Z}}^\ast,$$ and we can set $$\tau^\varphi_0:=1.$$ Note that, thanks to condition , there exists $C_1, C_2 >0$ such that $$C_1 \leq |\tau^\varphi_n| \leq C_2, \quad n\in {\mathbb{Z}},$$ so that these numbers are the eigenvalues of a diagonal isomorphism of any Sobolev space into itself, which we note $\tau^\varphi$. Also, note that $\tau^\varphi_n \neq 0$, and thus, $\varphi \notin H^m_{per}$. Finally, note that $$\label{ResteDA}
\left(\sum_{j=0}^{d}\left\langle \chi_{[\sigma_j, \sigma_{j+1}]}\partial^m \varphi, e_n\right\rangle\right) \in \ell^2.$$
The backstepping method: a finite-dimensional example {#FiniteD}
-----------------------------------------------------
Consider the finite-dimensional control system $$\label{FiniteDControlSys}
\dot{x}=Ax+ Bu(t), \quad x\in {\mathbb{C}}^n, A \in \mathcal{M}_n ({\mathbb{C}}), B\in\mathcal{M}_{n,1} ({\mathbb{C}}).$$ Suppose that is controllable. Then, it is well known (see for example [@CoronBook]) that for every polynomial $P \in {\mathbb{C}}[X]$ there exists a feedback $K\in\mathcal{M}_{1,n} ({\mathbb{R}})$ such that $P$ is the characteristic polynomial of $A+BK$.
This pole-shifting property for controllable systems can be formulated in another way, by trying to invertibly transform system into another system with shifted poles, namely $$\label{FiniteDTarget}
\dot{x}=(A-\lambda I)x+ Bv(t),$$ which is asymptotically stable for a large enough $\lambda$.
Suppose that $x(t)$ is a solution of system with $u(t)=Kx(t) + v(t)$ for some control function $v$. Such a transformation $T$ would map into $$\dot{(Tx)} = T\dot{x}=T(A+BK)x + TB v(t).$$ In order for $Tx$ to be a solution of , we need $$T(A+BK)x + TBv(t)= (A-\lambda I)Tx + Bv(t),$$ hence the conditions $$\label{FiniteDOpEq}
\begin{aligned}
T(A+BK)&=AT-\lambda T, \\
TB&=B,
\end{aligned}$$ for which one has the following theorem (see for example [@Coron15], or [@Schrodinger] for a different proof, more adaptable to the context of PDEs):
There exists a unique pair $(T,K)$ satisfying conditions .
The controllability of is crucial here, as it allows to build a basis for the space state, in which $T$ can then be constructed. The Hautus test gives the invertibility of $T$, and the uniqueness is given by the $TB=B$ condition.
This other approach to pole-shifting, which links controllability to stabilization, can be used in infinite dimensions. In our case, the controllability of will have the same importance: it will also allow us to build some sort of basis for the state space, and find a general form for the backstepping transformation, depending on $F$.
Related results
---------------
To investigate the stabilization of infinite-dimensional systems, there are three main types of approaches.
The first type of approach relies on abstract methods, such as the Gramian approach and the Riccati equations (see for example [@Vest; @Urquiza; @Komornik]). Although quite powerful, it seems that these methods fail to obtain the stabilization of nonlinear systems from the stabilization of their linearized systems.
The second approach relies on Lyapunov functions. Many results on the boundary stabilization of first-order hyperbolic systems, linear and nonlinear, have been obtained using this approach: see for example the book [@BastinCoronBook], and the recent results in [@Hayat1; @Hayat2]. However, this approach can be limited, as it is sometimes impossible to obtain an arbitrary decay rate using Lyapunov functions (see [@CoronBook Remark 12.9, page 318] for a finite dimensional example).
The third approach, which we will be using in this article, is the backstepping method. This name originally refers to a way of designing, in a recursive way, more effective feedback laws, for systems for which one already has a Lyapunov function and a feedback law which globally asymptotically stabilizes the system, see [@CoronBook; @Sontag] for an overview of the finite-dimensional case, and [@CoronAN] or [@KrsticLiu2] for applications to partial differential equations. Another way of applying this approach to partial differential equations was then developed in [@KBHeat] and [@KBParab]: when applied to the discretization of the heat equation, the backstepping approach yielded a change of coordinates which was equivalent to a Volterra transform of the second kind. Backstepping then took yet another successful form, consisting in mapping the system to stable target system, using a Volterra transformation of the second kind (see [@KrsticBook] for a comprehensive introduction to the method): $$f(t,x) \mapsto f(t,x)-\int_0^x k(x,y) f(t,y) dy.$$ This was used to prove a host of results on the boundary stabilization of partial differential equations: let us cite for example [@KrsticWave] and [@KrsticWave2] for the wave equation, [@XiangKdVNull; @XiangKdVFinite] for the Korteweg-de Vries, [@BastinCoronBook chapter 7] for an application to first-order hyperbolic systems, and also [@BCKV], which combines the backstepping method with Lyapunov functions to prove finite-time stabilization in $H^2$ for a quasilinear $2\times 2$ hyperbolic system.
In some cases, the method was used to obtained stabilization with an internal feedback. This was done in [@KrsticHaraTsub] and [@Woittennek] for parabolic systems, and [@YuXuJiangGanesan] for first-order hyperbolic systems. The strategy in these works is to first apply a Volterra transformation as usual, which still leaves an unstable source term in the target, and then apply a second invertible transformation to reach a stable target system. Let us note that in the latter reference, the authors study a linear transport equation and get finite-time stabilization. However, their controller takes a different form than ours, and several hypotheses are made on the space component of the controller so that a Volterra transform can be successfully applied to the system. This is in contrast with the method in this article, where the assumption we make on the controller corresponds to the exact null-controllability of the system.
In this paper, we use another application of the backstepping method, which uses another type of linear transformations, namely, Fredholm transformations: $$f(t,x) \mapsto \int_0^L k(x,y) f(t,y) dy.$$ These are more general than Volterra transformations, but they require more work: indeed, Volterra transformations are always invertible, but the invertibility of a Fredholm transform is harder to check. Even though it is sometimes more involved and technical, the use of a Fredholm transformation proves more effective for certain types of control: for example, in [@CoronLu1] for the Korteweg-de Vries equation and [@CoronLu2] for a Kuramoto-Sivashinsky, the position of the control makes it more appropriate to use a Fredholm transformation. Other boundary stabilization results using a Fredholm transformation can be found in [@CHO1] for integro-differential hyperbolic systems, and in [@CHO2] for general hyperbolic balance laws. Fredholm transformations have also been used in [@Schrodinger], where the authors prove the rapid stabilization of the Schrödinger equation with an internal feedback.
The backstepping method has the advantage of providing explicit feedback laws, which makes it a powerful tool to prove other related results, such as null-controllability or small-time stabilization (stabilization in an arbitrarily small time). This is done in [@CoronNguyen], where the authors give an explicit control to bring a heat equation to $0$, then a time-varying, periodic feedback to stabilize the equation in small time. In [@XiangKdVFinite], the author obtains the same kind of results for the Korteweg-de Vries equation.
Structure of the article
------------------------
The structure of this article is as follows: in Section 2, after some formal calculations, and using a formal $TB=B$ condition, we build candidates for the backstepping transformation. Using the properties of Riesz bases, we prove that such candidates are indeed invertible, under some conditions on the feedback coefficients $(F_n)$. For consistency, we then determine the feedback law $(F_n)$ such that the corresponding transformation indeed satisfies a weak form of the $TB=B$ condition. Then, in Section 3, we check that the corresponding transformation indeed satisfies an operator equality analogous to , making it a valid backstepping transformation. We check the well-posedness of the closed-loop system for the feedback law obtained in Section 2, which allows us to prove the stability result. Finally, Section 4 gives a few remarks on the result, as well as further questions on this stabilization problem.
Definition and properties of the transformation
===============================================
Let $\lambda^\prime >0$, and $m \geq 1$. Let $\varphi \in H^m \cap H^{m-1}_{per}$ be a real-valued function satisfying . We consider the following target system: $$\label{Target1}
\left\{\begin{aligned}
z_t+z_x +\lambda^\prime z& =0, \quad x\in (0,L), \\
z(t,0)& =z(t,L), \quad t\geq 0.\end{aligned}\right.$$ Then it is well-known that, taking $\alpha_0 \in L^2$, the solution to with initial condition $\alpha_0$ writes $$z(t,x)= e^{-\lambda^\prime t} \alpha_0 (x -t), \quad \forall (t,x) \in {\mathbb{R}}^{+} \times (0,L).$$ Hence,
For all $s \geq 0$, the system is exponentially stable for $\|\cdot \|_s$, for initial conditions in $H^s_{per}$.
Kernel equations {#FindKernel}
----------------
As mentioned in the introduction, we want to build backstepping transformations $T$ as a kernel operator of the Fredholm type: $$f(t,x) \mapsto \int_0^L k(x,y) f(t,y) dy.$$ To have an idea of what this kernel looks like, we can do the following formal computation for some Fredholm operator $T$: first the boundary conditions $$\left(\int_{0}^L k(0,y)\alpha(y) dy\right)=\left(\int_{0}^L k(L,y)\alpha(y) dy\right),$$ then the equation of the target system, for $x\in [0,L]$: [$$\begin{array}{rcl}
0&=& \left(\int_{0}^L k(x,y)\alpha(y) dy\right)_t + \left(\int_{0}^L k(x,y)\alpha(y) dy\right)_x +\lambda^\prime \left(\int_{0}^L k(x,y)\alpha(y) dy\right) \\
&=& \left(\int_{0}^L k(x,y)\alpha_t(y) dy\right) + \left(\int_{0}^L k_x(x,y)\alpha(y) dy\right)+\lambda^\prime \left(\int_{0}^L k(x,y)\alpha(y) dy\right) \\
&=& \left(\int_{0}^L k(x,y)(-\alpha_x(y) - \mu \alpha(y) + \langle \alpha, F \rangle \varphi(y)) dy\right)+ \left(\int_{0}^L (k_x(x,y)+\lambda^\prime k(x,y))\alpha(y) dy\right)
\end{array}$$]{} [$$\begin{array}{rcl}
&=& \left(\int_{0}^L k_y(x,y)\alpha(y) dy\right) - (k(x,L) \alpha (L)-k(x,0)\alpha(0))+ \left(\int_{0}^L k(x,y)\langle \alpha, F \rangle \varphi(y)) dy\right)+\\
& & \left(\int_{0}^L (k_x(x,y)+(\lambda^\prime-\mu) k(x,y))\alpha(y) dy\right)\\
&=& \left(\int_{0}^L k(x,y)\left(\int_0^L \bar{F}(s)\alpha(s) ds\right) \varphi(y)) dy\right)- (k(x,L) \alpha (L)-k(x,0)\alpha(0))\\
& & + \left(\int_{0}^L (k_y(x,y)+k_x(x,y)+(\lambda^\prime-\mu) k(x,y))\alpha(y) dy\right)\\
&=& \left(\int_{0}^L \bar{F}(s)\left(\int_0^L k(x,y) \varphi(y) dy\right) \alpha(s) ds\right)- (k(x,L) \alpha (L)-k(x,0)\alpha(0))\\
& & + \left(\int_{0}^L (k_y(x,y)+k_x(x,y)+(\lambda^\prime-\mu) k(x,y))\alpha(y) dy\right).
\end{array}$$]{} Now, suppose we have the formal $TB=B$ condition $$\int_0^L k(x,y) \varphi(y) dy = \varphi(x), \quad \forall x \in [0,L].$$ Then, we get, noting $\lambda:=\lambda^\prime-\mu$, $$\left(\int_{0}^L \left(k_y(x,y)+k_x(x,y)+\lambda k(x,y) + \varphi(x)\bar{F}(y)\right)\alpha(y) dy\right)- (k(x,L) \alpha (L)-k(x,0)\alpha(0)) = 0.$$ Hence the kernel equation: $$\label{Kernel1}
\left\{\begin{aligned}
k_x+k_y+\lambda k&=- \varphi(x)\bar{F}(y), \\
k(0,y) & =k(L,y), \\
k(x,0)& =k(x,L),\end{aligned}\right.$$ together with the $TB=B$ condition $$\label{TB}
\left\langle k(x, \cdot), \varphi(\cdot) \right\rangle = \varphi(x), \quad \forall x \in [0,L].$$
Construction of Riesz bases for Sobolev spaces
----------------------------------------------
To study the solution to the kernel equation, we project it along the variable $y$. Let us write heuristically $$k(x,y)=\sum_{n \in {\mathbb{Z}}} k_{n}(x)e_n(y),$$ so that we get the projected kernel equations $$\label{kn1}
k_{n}^\prime + \lambda_n k_{n} = -\overline{F_{-n}} \varphi,$$ where $$\label{lambda_n}
\lambda_n=\lambda + \frac{2i\pi}{L} n.$$ Note that $$\label{ElementaryCoeff}
\frac{2i\pi p}{L} \frac1{\lambda_{n+p}} + \lambda_n \frac1{\lambda_{n+p}} = 1,\quad \forall n, p \in {\mathbb{Z}}.$$ Now consider the $L^2$ function given by $$\label{Elementary}
\Lambda^{\lambda}_n(x)=\frac{\sqrt{L}}{1-e^{-\lambda L}}e^{-\lambda_n x}, \quad \forall n \in {\mathbb{Z}}, \quad\forall x \in [0, L).$$ Then, for all $m\geq 0$, $\Lambda^{\lambda}_n \in H^m$, and we have $$\langle \Lambda^{\lambda}_n , e_p \rangle = \frac1{\sqrt{L}}\int_0^L\frac{\sqrt{L}}{1-e^{-\lambda L}}e^{-\lambda_n x} e^{-\frac{2i\pi p}{L}x} dx = \frac{1}{1-e^{-\lambda L}}\int_0^L e^{-\lambda_{n+p} x}= \frac1{\lambda_{n+p}}, \quad \forall n, p \in {\mathbb{Z}},$$ so that, using , $$(\Lambda^{\lambda}_n)^\prime + \lambda_n \Lambda^{\lambda}_n = \sum_{p \in {\mathbb{Z}}} e_p \ \textrm{in} \ {\mathcal{E}^\prime}.$$
In ${\mathcal{E}^\prime}$, $\sum_{p \in {\mathbb{Z}}} e_p$ is the equivalent of the Dirac comb, or the “Dirac distribution” on the space of functions on $[0,L]$. So, in a sense, $\Lambda^{\lambda}_n$ is the elementary solution of .
Let us now define, in analogy with the elementary solution method, $$\label{kndef}
k_{n, \lambda}=-\overline{F_{-n}} \Lambda^{\lambda}_n \star \varphi \in H^m_{per}, \quad \forall n \in {\mathbb{Z}}.$$ The regularity comes from the definition of the convolution product, and , and one can check, using , that $k_{n, \lambda}$ is a solution of .
The next step to build an invertible transformation is to find conditions under which $(k_{n, \lambda})$ is some sort of basis. More precisely we use the notion of Riesz basis (see [@Christensen Chapter 4])
A Riesz basis in a Hilbert space $H$ is the image of an orthonormal basis of $H$ by an isomorphism.
\[Riesz\] Let $H$ be a Hilbert space. A family of vectors $(f_k)_{k \in {\mathbb{N}}} \in H$ is a Riesz basis if and only if it is complete (i.e., $ \overline{\textup{Span}(f_k)}=H $) and there exists constants $C_1, C_2 >0$ such that, for any scalar sequence $(a_k)$ with finite support, $$\label{RieszBounds}
C_1 \sum |a_k|^2 \leq \left\|\sum a_k f_k \right\|_H^2 \leq C_2 \sum |a_k|^2.$$
Let us now introduce the following growth condition:
Let $s \geq 0$, $(u_n)\in {\mathbb{C}}^{\mathbb{Z}}$ (or $u\in {\mathcal{E}^\prime}$) . We say that $(u_n)$ (or $u$) has $s$-growth if $$\label{FGrowth}
c\sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2s}}\leq |u_n| \leq C\sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2s}}, \quad \forall n \in {\mathbb{Z}},$$ for some $c,C>0$. The optimal constants for these inequalities are called growth constants.
The inequalities can also be written, more intuitively, and for some other positive constants, $$c\left(1+|n|^{s}\right)\leq |u_n| \leq C\left(1+|n|^{s}\right), \quad \forall n \in {\mathbb{Z}}.$$
We can now establish the following Riesz basis properties for the $(k_{n, \lambda})$:
\[RieszFam\] Let $s\geq 0$. If $(F_n)$ has $s$-growth, then the family of functions $$(k_{n, \lambda}^s):= \left(\frac{k_{n, \lambda}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\right)$$ is a Riesz basis for $H^{m}_{per}$.
We use the characterization of Riesz bases given in Proposition \[Riesz\]. First, let us prove the completeness of $(k_{n, \lambda}^s)$. Let $f\in H^m_{per}$ be such that $$\langle f, k_{n, \lambda}^s \rangle_{m} =0, \quad \forall n \in {\mathbb{Z}}.$$ Then for all $n\in {\mathbb{Z}}$ we get $$0=\langle \Lambda_{n}^{\lambda} \star \varphi, f \rangle_{m}=\sum_{p\in {\mathbb{Z}}} \frac{\overline{f_p} \varphi_p}{\lambda_{n+p}} \left(1+\left|\frac{2i \pi p}{L}\right|^{2m}\right)=\left\langle \Lambda_{n}^{\lambda}, \sum_{p\in Z} \left(1+\left|\frac{2i \pi p}{L}\right|^{2m}\right)f_p\overline{\varphi_p} e_p \right\rangle,$$ as, thanks to , and using the fact that $f\in H^m_{per}$, $$\sum_{p\in Z} \left(1+\left|\frac{2i \pi p}{L}\right|^{2m}\right)f_p\overline{\varphi_p} e_p \in L^2.$$ Now, $(\Lambda_{n}^{\lambda})$ is a complete family of $L^2$, as it is a Riesz basis, so that $$f_p \varphi_p = 0, \quad \forall p \in {\mathbb{Z}}.$$ Recalling condition , this yields $$f_p = 0, \quad \forall p \in {\mathbb{Z}},$$ which proves the completeness of $(k_{n, \lambda}^s)$.
Now let $I \subset {\mathbb{Z}}$ be a finite set, and $(a_n) \in {\mathbb{C}}^I$. Then, [$$\begin{array}{rcl}
\left\|\sum_{n\in I} a_n k_{n, \lambda}^s\right\|_m^2 & =& \left\| \sum_{n\in I} -a_n \frac{\overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \Lambda^{\lambda}_n \star \varphi \right\|_{m}^2 \\
&=& \left\| \sum_{n\in I} a_n \frac{\overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \sum_{p \in {\mathbb{Z}}} \frac{\varphi_p}{\lambda_{n+p}} e_p\right\|_{m}^2 \\
&=& \left\| \sum_{p \in {\mathbb{Z}}} \varphi_p\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\lambda_{n+p}\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} e_p\right\|_{m}^2 \\
&=& \sum_{p \in {\mathbb{Z}}} \left(1+\left|\frac{2\pi p}{L} \right|^{2m}\right) |\varphi_p|^2\left|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\lambda_{n+p}\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \right |^2.
\end{array}$$]{} Now, using condition , we have $$c^2 \sum_{p \in {\mathbb{Z}}}\left|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\lambda_{n+p}\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \right|^2
\leq \left\|\sum_{n\in I} a_n k_{n, \lambda}^s\right\|_{m}^2
\leq C^2 \sum_{p \in {\mathbb{Z}}}\left|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\lambda_{n+p}\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \right|^2,$$ where $c, C>0$ are the decay constants in condition .
This last inequality can be rewritten $$c^2 \left\|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\Lambda^{\lambda}_n \right\|^2
\leq \left\|\sum_{n\in I} a_n k_{n, \lambda}^s\right\|_{m}^2
\leq C^2 \left\|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}} \Lambda^{\lambda}_n \right\|^2 ,$$ as $$\Lambda^{\lambda}_n=\sum_{p\in {\mathbb{Z}}} \frac1{\lambda_{n+p} }e_p.$$ We now use the fact that $(\Lambda^{\lambda}_n)$ is a Riesz basis of $L^2$: indeed, it is the image of the Hilbert basis $(e_n)$ by the isomorphism $$\Lambda^{\lambda} : f\in L^2 \mapsto\frac{\sqrt{L}}{1-e^{-\lambda L}} e^{-\lambda \cdot} f.$$ The norms of $\Lambda^{\lambda}$ and its inverse are rather straightforward to compute using piecewise constant functions, we have $${{\vert\kern-0.25ex\vert\kern-0.25ex\vert \Lambda^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} =\frac{\sqrt{L}}{1-e^{-\lambda L}},$$ $${{\vert\kern-0.25ex\vert\kern-0.25ex\vert (\Lambda^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} = \frac{1-e^{-\lambda L}}{\sqrt{L}} e^{\lambda L},$$ so that $$\frac{1}{{{\vert\kern-0.25ex\vert\kern-0.25ex\vert (\Lambda^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} ^2}\sum_{n\in I} \left|\frac{ a_n \overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\right|^2 \leq \left\|\sum_{n\in I} \frac{ a_n \overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\Lambda^{\lambda}_n \right\|^2 \leq {{\vert\kern-0.25ex\vert\kern-0.25ex\vert \Lambda^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} ^2\sum_{n\in I} \left|\frac{ a_n \overline{F_{-n}}}{\sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2s}}}\right|^2,$$ and we finally get, using the fact that $(F_n)$ has $s$-growth, $$c^2C_1^2 \frac{1}{{{\vert\kern-0.25ex\vert\kern-0.25ex\vert (\Lambda^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} ^2}\sum_{n\in I} |a_n|^2 \leq \left\|\sum_{n\in I} a_n k_{n, \lambda}^s\right\|_{m}^2 \leq C^2C_2^2 {{\vert\kern-0.25ex\vert\kern-0.25ex\vert \Lambda^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} ^2\sum_{n\in I} |a_n|^2.$$ where $C_1, C_2 >0$ are the growth constants of $(F_n)$, so that the constants in the inequalities above are optimal. Hence, using again point 2. of Proposition \[Riesz\], $(k_{n, \lambda}^s)$ is a Riesz basis of $H^{m}_{per}$.
We now have candidates for the backstepping transformation, under some conditions on $F$:
\[TInvert\] Let $m\in {\mathbb{N}}^\ast$, and $F$ such that $(F_n)$ has $m$-growth, with growth constants $C_1, C_2>0$. Define $$\label{T}
\quad T^{\lambda} \alpha := \sum_{n\in {\mathbb{Z}}} \sqrt{1+\left|\frac{2i\pi n}{L}\right|^{2m}} \alpha_n k_{-n, \lambda}^{m} \in H^m_{per}, \quad \forall \alpha \in H^{m}_{per},$$ where $\alpha=\sum_{n\in {\mathbb{Z}}} \alpha_n e_n$. Then, $T^{\lambda}:H^{m}_{per} \rightarrow H^m_{per}$ is an isomorphism. Moreover, $$\label{Norms}
\begin{aligned}
{{\vert\kern-0.25ex\vert\kern-0.25ex\vert T^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} &= \frac{CC_2\sqrt{L}}{1-e^{-\lambda L}}, \\
{{\vert\kern-0.25ex\vert\kern-0.25ex\vert (T^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} &= \frac{1-e^{-\lambda L}}{cC_1 \sqrt{L}} e^{\lambda L}.
\end{aligned}$$
The invertibility of $T^{\lambda}$ is clear thanks to the Riesz basis property of $( k_{-n, \lambda}^{m})$, and comes from the fact that, as mentioned at the end of the proof of Proposition \[RieszFam\], all the constants in the inequalities are optimal.
Definition of the feedback law
------------------------------
In order to further determine the feedback law, and define our final candidate for the backstepping transformation, the idea is now to return to the $TB=B$ condition , as we have used it in the formal computations of section \[FindKernel\], in the equation . However, in this case, $\varphi \notin H^m_{per}$, and so it is not clear whether $T^\lambda \varphi$ is well-defined.
We can nonetheless obtain a $TB=B$ condition in some weak sense: indeed, let us set $$\varphi^{(N)}:= \sum_{n=-N}^N \varphi_n e_n \in H^m_{per}, \quad \forall N \in {\mathbb{N}}.$$ Then, $$\varphi^{(N)} \xrightarrow[N \to \infty]{H^{m-1}} \varphi$$ and $$\begin{array}{rcl}
T^{\lambda} \varphi^{(N)} &= &\sum_{n=-N}^N -\varphi_n \overline{F_{n}} \Lambda^{\lambda}_{-n} \star \varphi \\
&=& \sum_{n=-N}^N \sum_{p \in {\mathbb{Z}}} \frac{-\varphi_n \overline{F_{n}} \varphi_p}{\lambda_{-n+p}} e_p \\
&=& \sum_{p \in {\mathbb{Z}}} \varphi_p \left(\sum_{n=-N}^N \frac{-\varphi_n \overline{F_{n}} }{\lambda_{-n+p}}\right) e_p.
\end{array}$$ Now, notice that one can apply the Dirichlet convergence theorem for Fourier series (see for example [@Kahane]) to $\Lambda^{\lambda}_p, p \in {\mathbb{Z}}$ at $0$: $$\sum_{n=-N}^N \frac{1 }{\lambda_{-n+p}} = \sum_{n=-N}^N \frac{1 }{\lambda_{n+p}} \xrightarrow[N \to \infty]{} \frac{\Lambda^{\lambda}_p (0) + \Lambda^{\lambda}_p (L)}2 = \frac{\sqrt{L}}{1-e^{-\lambda L}} \frac{1+e^{-\lambda L}}{2}.$$ Let us note $$K(\lambda):=\frac{2}{\sqrt{L}} \frac{1-e^{-\lambda L}}{1+e^{-\lambda L}},$$ and set $$\label{Feedback}
F_n:=-\frac{K(\lambda)}{\overline{\varphi_n}}, \quad \forall n \in {\mathbb{Z}}.$$ This defines a feedback law $F\in {\mathcal{E}^\prime}$ which is real-valued, as $\varphi$ is real-valued, and which has $m$-growth thanks to condition , so that $T^{\lambda}$ is a valid backstepping transformation. Moreover, $$\label{WeakTB}
\langle T^{\lambda} \varphi^{(N)}, e_p \rangle = \varphi_p K(\lambda) \sum_{n=-N}^N \frac{1 }{\lambda_{-n+p}}\xrightarrow[N \to \infty]{} \varphi_p, \quad \forall p \in {\mathbb{Z}},$$ which corresponds to the $TB=B$ condition in some weak sense.
With this feedback law, the backstepping transformation now writes $$\label{TFeedback}
T^{\lambda} \alpha = \sum_{n\in {\mathbb{Z}}} \alpha_n k_{-n, \lambda}, \quad \forall \alpha \in H^{m}_{per},$$ and $$\label{NormsFeedback}
\begin{aligned}
{{\vert\kern-0.25ex\vert\kern-0.25ex\vert T^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} &= \frac{CK(\lambda)\sqrt{L}}{c(1-e^{-\lambda L})}, \\
{{\vert\kern-0.25ex\vert\kern-0.25ex\vert (T^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} &= \frac{C(1-e^{-\lambda L})}{cK(\lambda) \sqrt{L}} e^{\lambda L}.
\end{aligned}$$
Regularity of the feedback law
------------------------------
Finally, in order to study the well-posedness of the closed-loop system corresponding to , we need some information on the regularity of $F$.
Let us first begin by a general lemma for linear forms with coefficients that have $m$-growth:
\[Fcont\] Let $m \geq 0$, and $G\in {\mathcal{E}^\prime}$ with $m$-growth.
Then, for all $s>1/2$, $G$ is defined on $H^{m+s}_{per}$, is continuous for $\|\cdot\|_{m+s}$, but not for ${\|\cdot\|}_{m+\sigma}$, for $-m \leq \sigma < 1/2$.
In particular, the feedback law $F\in {\mathcal{E}^\prime}$ defined by defines a linear form on $H^{m+1}_{per}$ which is continuous for ${\|\cdot\|}_{m+1}$ but not for ${\|\cdot\|}_{m}$.
Let $s > 1/2$, and let $\alpha\in H^{m+s}_{per}$. Using the growth conditions , we can do the following computations for $ \alpha \in H^{m+s}_{per}$: $$\begin{array}{rcl}
\sum_{n\in {\mathbb{Z}}} |G_n||\alpha_n|
&\leq & C\sum_{n\in {\mathbb{Z}}} \sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2m}}|\alpha_n| \\
&\leq & C^\prime \sum_{n\in {\mathbb{Z}}} \frac{1}{1+|n|^s} \sqrt{1+\left|\frac{2i \pi n}{L}\right|^{2m+2s}} |\alpha_n|\\
&\leq & C^\prime \left(\sqrt{\sum_{n\in {\mathbb{Z}}} \frac{1}{(1+|n|^{s})^2}}\right) \|\alpha\|_{m +s}
\end{array}$$ where $C, C^\prime>0$ are constants that do not depend on $\alpha$, and where the last inequality is obtained using the Cauchy-Schwarz inequality. Thus $G$ is defined on $H^{m+s}_{per}$ by $$\langle \alpha, G \rangle := \sum_{n\in {\mathbb{Z}}} G_n\alpha_n, \quad \forall \alpha \in H^{m+s}_{per},$$ and $G$ is continuous on $H^{m+s}_{per}$.
On the other hand, let $-m \leq \sigma < 1/2$, and consider, for $N \geq 1$, $$\gamma^{(N)}:= \sum_{|n|\geq N}\frac{1}{\overline{G_n}\left(1+|n|^{1+s}\right) } e_n \in H^{m +s}_{per}.$$ We have $$\|\gamma^{(N)} \|_{m + \sigma}^2 = \sum_{|n|\geq N} \frac{\left(1+\left|\frac{2i \pi n}{L}\right|^{2m+2\sigma}\right)}{|G_n|^2}\frac{1}{\left(1+|n|^{1 + s}\right)^2 } \leq C\sum_{|n|\geq N} \frac{1}{1+|n|^{2+2s-2\sigma} }$$ for some constant $C>0$. Then, $$\begin{array}{rcl}
|\langle \gamma^{(N)}, G \rangle | & = & \sum_{|n|\geq N} \frac1{1+|n|^{1 + s}} \\
&\geq& c\sum_{|n| \geq N} |n|^{1+s-2\sigma} \frac1{1+|n|^{2+2s-2\sigma}} \\
&\geq & c N^{1+s-2\sigma} \sum_{|n| \geq N} \frac1{1+|n|^{2+2s-2\sigma}} \\
&\geq & c^\prime N^{1+s-2\sigma} \sqrt{\sum_{|n| \geq N} \frac1{1+|n|^{2+2s-2\sigma}} }\|\gamma^{(N)}\|_{m+\sigma}
\end{array}$$ for some constants $c, c^\prime >0$. Now, we know that there exists constants $c^{\prime\prime}, C^{\prime\prime} >0$ such that $$\frac{c^{\prime\prime}}{N^{1+2s-2\sigma}}\leq \sum_{|n| \geq N} \frac1{1+|n|^{2+2s-2\sigma}} \leq \frac{C^{\prime\prime}}{N^{1+2s-2\sigma}},$$ So that $$N^{1+s-2\sigma}\sqrt{\sum_{|n| \geq N} \frac1{1+|n|^{2+2s-2\sigma}} } \geq c^{\prime\prime} N^{\frac12-\sigma} \xrightarrow[N \to \infty]{} \infty.$$ This proves that $G$ is not continuous for${\|\cdot\|}^{m + \sigma}$.
Let us now give a more precise description of the domain of definition and regularity of $F$. Recalling the identity , we can derive the following identity for $F_n$: $$\label{DAF} F_n = \frac{K(\lambda)}{\tau^\varphi_n} \left(\frac{2i \pi n}{L}\right)^m -\frac{K(\lambda)}{\tau^\varphi_n} \left(\frac{2i \pi n}{L}\right)^m \frac{\overline{\displaystyle{\sum_{j=0}^{d}\left\langle \chi_{[\sigma_j, \sigma_{j+1}]}\partial^m \varphi, e_n\right\rangle}}}{\tau^\varphi_n-\overline{\displaystyle{\sum_{j=0}^{d}\left\langle \chi_{[\sigma_j, \sigma_{j+1}]}\partial^m \varphi, e_n\right\rangle}}}, \quad \forall n \in {\mathbb{Z}}^\ast,$$ so that $$\label{RegularPart}\left(\frac1{\left(\frac{2i\pi n}{L}\right)^m}\left(F_n - \frac{K(\lambda)}{\tau^\varphi_n} \left(\frac{2i \pi n}{L}\right)^m \right)\right)_{n\in {\mathbb{Z}}^\ast} \in \ell^2.$$
Let us then note $$h_n:=\frac{K(\lambda)}{\tau^\varphi_n} \left(\frac{2i \pi n}{L}\right)^m,\quad \forall n \in {\mathbb{Z}},$$ and $h$ the associated linear form in ${\mathcal{E}^\prime}$.
\[FeedbackReg\] The linear form $h$ defines the following linear form on $\tau^\varphi(H^{m+1}_{(pw)})$, continuous for $\|\cdot\|_{m+1,pw}$: $$\label{SingularPart} \langle \alpha, h \rangle=(-1)^m \sqrt{L}\frac{K(\lambda)}2 \partial^m\left((\tau^\varphi)^{-1}\alpha\right)(0)+\partial^m\left((\tau^\varphi)^{-1}\alpha\right)(L), \quad \forall \alpha \in H^{m+1}.$$ Moreover, $\tilde{F}:=F-h$ is continuous for $\|\cdot\|_m$, so that $F$ is defined on $\tau^\varphi(H^{m+1}_{(pw)})\cap H^m_{per}$, and is continuous for $\|\cdot\|_{m+1, pw}$, but not $\|\cdot\|_m$.
It is clear, by definition of $H^m_{per}$, and using , that for $\alpha \in H^m_{per}$, the expression $$\langle \alpha, F-h \rangle = \sum_{n\in {\mathbb{Z}}} \alpha_n (\overline{F_n}-\overline{h_n}) = \sum_{n \neq 0} \left(\frac{2i\pi n}{L}\right)^m\alpha_n \frac1{\left(\frac{2i\pi n}{L}\right)^m}(\overline{F_n}+\overline{h_n}) + \frac{K(\lambda)\alpha_0}{\varphi_0}$$ defines a continuous linear form on $H^m_{per}$.
On the other hand, let $\alpha \in \tau^\varphi (H^{m+2}_{(pw)})$, then $$\sum_{n=-N}^N \alpha_n \overline{h_n} = (-1)^m \sqrt{L}K(\lambda)\sum_{n=-N}^N \left(\frac{2i \pi n}{L}\right)^m\frac{\alpha_n}{\tau^\varphi_n}\frac1{\sqrt{L}}$$ we can use the Dirichlet convergence theorem (see [@Kahane]) on $\partial^m \left((\tau^\varphi)^{-1}\alpha\right) \in H^2_{(pw)}$ at $0$, so that $$\begin{aligned}\sum_{n=-N}^N \alpha_n \overline{h_n} = (-1)^m \sqrt{L}K(\lambda) &\sum_{n=-N}^N \left(\frac{2i \pi n}{L}\right)^m\frac{\alpha_n}{\tau^\varphi_n}\frac1{\sqrt{L}} \\
& \xrightarrow[N \to \infty]{} (-1)^m \sqrt{L}\frac{K(\lambda)}2 \left( \partial^m\left((\tau^\varphi)^{-1}\alpha\right)(0)+\partial^m\left((\tau^\varphi)^{-1}\alpha\right)(L)\right).\end{aligned}$$ Now, we know that $H^{m+2}_{(pw)}$ is dense in $H^{m+1}_{(pw)}$ for the $H^{m+1}_{(pw)}$ norm. As $\tau^\varphi$ is a sum of translations, it is continuous for $\|\cdot\|_{m+1, pw}$, so that $\tau^\varphi(H^{m+2}_{(pw)})$ is dense in $\tau^\varphi(H^{m+1}_{(pw)})$ for $\|\cdot\|_{m+1, pw}$.
Moreover, using the Sobolev inequality for $H^1$ and $L^\infty$ (see for example [@Brezis Chapter 8, Theorem 8.8]), we get the continuity of $h$ for $\|\cdot\|_{m+1, pw}$, so that we can extend it from $\tau^\varphi(H^{m+2}_{(pw)})$ to $\tau^\varphi(H^{m+1}_{(pw)})$ by density. We also get that $h$ is not continuous for $\|\cdot\|_{m}$, as $\alpha \in H^m \mapsto \partial^m\alpha(0)$ and $\alpha \in H^m \mapsto \partial^m\alpha(L)$ are not continuous for $\|\cdot\|_{m}$.
Thus, $F=\tilde{F}+h$ is defined on $\tau^\varphi(H^{m+1}_{(pw)})\cap H^m_{per}$, is continuous for $\|\cdot\|_{m+1}$ but not for $\|\cdot\|_m$.
Well-posedness and stability of the closed-loop system
======================================================
Let $m \geq 1$, $\varphi \in H^m_{(pw)} \cap H^{m-1}_{per}$ satisfying growth condition . Let the feedback law $F$ be defined by .
Operator equality
-----------------
Now that we have completely defined the feedback $F$ and the transformation $T^\lambda$, let us check that we have indeed built a backstepping tranformation. As in the finite dimensional example of subsection \[FiniteD\], this corresponds to the formal operator equality $$T(A+BK)=(A-\lambda I) T.$$
Let us define the following domain: $$\label{Domains}
\quad D_{m} := \left\{ \alpha \in \tau^\varphi(H^{m+1}_{(pw)})\cap H^m_{per}, \quad -\alpha_x - \mu \alpha +\langle \alpha, F \rangle \varphi \in H^m_{per}\right\}.$$
Notice that, as $\varphi \in H^m_{(pw)}$, the condition $ \alpha \in H^{m+1}_{(pw)} \supset \tau^\varphi(H^{m+1}_{(pw)}) $ is necessary for $-\alpha_x - \mu \alpha +\langle \alpha, F \rangle \varphi$ to be in $H^m_{per}$ . Let us first check the following property:
\[Dense\] For $m \geq 1$, $D_{m}$ is dense in $H^{m}_{per}$ for ${\|\cdot\|}_{m}$.
It is clear that $H^{m+1}_{per} \subset \tau^\varphi\left(H^{m+1}_{(pw)}\right)$, so that $$\mathcal{K}_m:=\left\{\alpha \in H^{m+1}_{per}, \ \langle \alpha, F \rangle=0 \right\} \subset D_m.$$ Now, by Lemma \[Fcont\], as $F$ has $m$-growth, $\mathcal{K}_m$ is dense in $H^{m +1}_{per}$ for ${\|\cdot\|}_{m}$, as the kernel of the linear form $F$ which is not continuous for ${\|\cdot\|}_{m}$. As $H^{m +1}_{per}$ is dense in $H^{m}_{per}$, then $D_{m}$ is dense in $H^{m}_{per}$ for ${\|\cdot\|}_{m}$.
Now, on this dense domain, let us establish the operator equality:
\[OpEqProp\] $$\label{OpEq}
T^{\lambda} (-\partial_x - \mu I + \langle \cdot, F \rangle \varphi) \alpha = (-\partial_x - \lambda^\prime I)T^{\lambda} \alpha \quad \textrm{in} \ H^{m}_{per}, \quad \forall \alpha \in D_{m}.$$
First let us rewrite in terms of $\lambda$: $$T^{\lambda} (-\partial_x + \langle \cdot, F \rangle \varphi) \alpha = (-\partial_x - \lambda I)T^{\lambda} \alpha \quad \textrm{in} \ H^{m}_{per}, \quad \forall \alpha \in D_{m}(F).$$
Let $\alpha \in D_{m}$. By definition of the domain $D_m$, the left-hand side of is a function of $H^m_{per} \subset {\mathcal{E}^\prime}$, and by construction of $T^{\lambda}$, the right-hand side of is a function of $H^{m-1}_{per} \subset {\mathcal{E}^\prime}$. To prove that these functions are equal, it is thus sufficient to prove their equality in ${\mathcal{E}^\prime}$. Let us then write each term of the equality against $e_n$ for $n\in {\mathbb{Z}}$. One has
[$$\begin{array}{rcl}
\left\langle (-\partial_x - \lambda I)T^{\lambda} \alpha, e_n \right\rangle &=& \left\langle T^{\lambda}\alpha, \frac{2i \pi n}L e_n \right\rangle -\lambda \left\langle T^{\lambda} \alpha , e_n \right\rangle\\
&=& -\lambda_n \langle T^{\lambda}\alpha, e_n \rangle.\\
\end{array}$$]{}
Let us now prove that $$\label{weakOpEq}
\langle T^{\lambda} (-\alpha_x + \langle \alpha, F \rangle \varphi), e_n\rangle = -\lambda_n \langle T^{\lambda}\alpha, e_n \rangle, \quad \forall n\in {\mathbb{Z}}.$$
Now, as we only have $\alpha_x \in H^{m-1}_{per}$, $T^{\lambda}\alpha_x$ is not defined *a priori*. In order to allow for more computations, let us define $$\begin{aligned}
\alpha^{(N)}:= \sum_{n=-N}^N \alpha_n e_n,& \quad \forall N \in {\mathbb{N}}, \\
\varphi^{(N)}:= \sum_{n=-N}^N \varphi_n e_n,& \end{aligned}$$ so that we have, by property of the partial Fourier sum of a $H^m_{per}$ function, $$-\alpha^{(N)}_x + \langle \alpha, F \rangle \varphi^{(N)} \xrightarrow[N \to \infty]{H^m} -\alpha_x + \langle \alpha, F \rangle \varphi,$$ so that in particular, $$\label{weakOpEq1}\langle T^{\lambda} (-\alpha^{(N)}_x + \langle \alpha, F \rangle \varphi^{(N)}), e_n\rangle \xrightarrow[N \to \infty]{} \langle T^{\lambda} (-\alpha_x + \langle \alpha, F \rangle \varphi), e_n\rangle$$
Let $N \in {\mathbb{N}}$. We can now write [$$\begin{array}{rcl}
\langle T^{\lambda} (-\alpha^{(N)}_x + \langle \alpha, F \rangle \varphi^{(N)}), e_n\rangle & = & -\langle T^{\lambda} \alpha^{(N)}_x , e_n\rangle + \langle \alpha, F \rangle \langle T^{\lambda}\varphi^{(N)}, e_n\rangle \\
&=& -\left\langle \sum_{p=-N}^N \frac{2i\pi p}L \alpha_p k_{-p, \lambda}, e_n \right\rangle + \langle \alpha, F \rangle \langle T^{\lambda}\varphi^{(N)}, e_n\rangle.
\end{array}$$]{}
Now, using , we get $$\frac{2i\pi p}L k_{-p, \lambda} = (k_{-p, \lambda})_x + \lambda k_{-p, \lambda} + \overline{F_p}\varphi,$$ so that $$-T^{\lambda} \alpha^{(N)}_x = \sum_{p=-N}^N \alpha_p \left((k_{-p, \lambda})_x +\lambda k_{-p, \lambda} +\overline{F_p}\varphi\right).$$ Hence $$-\langle T^{\lambda} \alpha^{(N)}_x , e_n\rangle = - \left\langle \left(T^{\lambda} \alpha^{(N)}\right)_x, e_n \right\rangle - \lambda \left\langle T^\lambda \alpha^{(N)}, e_n \right\rangle - \langle \alpha^{(N)}, F \rangle \varphi_n,$$ and finally, $$\label{weakOpEq2}
\begin{aligned}\langle T^{\lambda} (-\alpha^{(N)}_x + \langle \alpha, F \rangle \varphi^{(N)}), e_n\rangle =
-\lambda_n \left\langle T^{\lambda} \alpha^{(N)}, e_n \right\rangle &+ \left( \langle \alpha - \alpha^{(N)} , F \rangle \right)\varphi_n \\
&+ \langle \alpha, F \rangle \left(\left\langle T^{\lambda}\varphi^{(N)}- \varphi, e_n \right\rangle \right).
\end{aligned}$$ To deal with the third term of the right-hand side of this equality, recall that we have chosen a feedback law so that the $TB=B$ condition holds. Thus, $$\label{OpEqTerm1}\left\langle T^{\lambda}\varphi^{(N)}- \varphi, e_n \right\rangle \xrightarrow[N \to \infty]{} 0.$$ To deal with the second term, recall that $F$ is the sum of a regular part $\tilde{F}$ and a singular part $h$: $$\langle \alpha - \alpha^{(N)} , F \rangle = \left\langle \alpha - \alpha^{(N)}, \tilde{F} \right\rangle + \langle\alpha - \alpha^{(N)}, h \rangle.$$ Now, by definition of $\alpha^{(N)}$ and continuity of $\tilde{F}$ for ${\|\cdot\|}_{m}$, $$\label{OpEqTerm21}\left\langle \alpha - \alpha^{(N)}, \tilde{F} \right\rangle \xrightarrow[N\to \infty]{} 0.$$ On the other hand, for all $N \in {\mathbb{N}}$, [$$\label{hOtherForm}\begin{array}{rcl}
\langle \alpha^{(N)}, h \rangle &=& K(\lambda) \sum_{n=-N}^{N} \frac{\alpha_n}{\tau^\varphi_n} \left(\frac{2i \pi n}{L}\right)^m \\
&=& \frac{K(\lambda)}2 \sum_{n=-N}^{N} \left(\frac{\alpha_n}{\tau^\varphi_n} + (-1)^{m} \frac{\alpha_{-n}}{\overline{\tau^\varphi_n}}\right) \left(\frac{2i \pi n}{L}\right)^m. \\
&=&\frac{K(\lambda)}2 \partial^{m-1} \tilde{\tau}^\varphi \alpha^{(N)}_x (0) ,
\end{array}$$]{} where $$\tilde{\tau}^\varphi f = \sum_{n\in {\mathbb{Z}}} \left(\frac{f_n}{\tau^\varphi_n} + (-1)^{m-1} \frac{f_{-n}}{\overline{\tau^\varphi_n}}\right) e_n, \quad \forall f \in L^2.$$ Now, notice that, by definition of $\tau^\varphi$ and $D_m$, $$\label{RegA+BK}
\tilde{\tau}^\varphi \left(-\alpha_x - \mu \alpha +\langle \alpha, F \rangle \varphi\right) \in H^m_{per}.$$
Moreover, using , we have for $n \in {\mathbb{Z}}^\ast$: [$$\begin{array}{rcl}
\frac{\varphi_n}{\tau^\varphi_n} + (-1)^{m-1} \frac{\varphi_{-n}}{\overline{\tau^\varphi_n}} &=& \frac{\varphi_n}{\tau^\varphi_n} + (-1)^{m-1} \overline{\frac{\varphi_{n}}{\tau^\varphi_n}} \\
&=& \frac{(-1)^{m-1}+(-1)^{m-1}(-1)^{m-1}(-1)^m }{\left(\frac{2i\pi}{L}n\right)^m} + \frac{r_n}{\left(\frac{2i\pi}{L}n\right)^m} \\
&=& \frac{r_n}{\left(\frac{2i\pi}{L}n\right)^m},
\end{array}$$]{} where $r_n \in \ell^2$. Hence, $\tilde{\tau}^\varphi \varphi \in H^m_{per}$. This, together with , yields $$\tilde{\tau}^\varphi \alpha_x \in H^m_{per}.$$ This implies that $$\tilde{\tau}^\varphi \alpha^{(N)}_x \xrightarrow[N\to \infty]{H^{m}} \tilde{\tau}^\varphi\alpha_x,$$ as $\tilde{\tau}^\varphi \alpha^{(N)}_x$ is the partial sum of $\tilde{\tau}^\varphi \alpha_x$.
Hence, by continuity of $\alpha \mapsto \partial^{m-1} \alpha(0)$ for ${\|\cdot\|}_{m}$, implies that $$\label{OpEqTerm22}\left\langle \alpha - \alpha^{(N)}, h \right\rangle \xrightarrow[N\to \infty]{} 0.$$
Finally, , , , , and the continuity of $T^\lambda$ yield $$\langle T^{\lambda} (-\alpha^{(N)}_x + \langle \alpha, F \rangle \varphi^{(N)}), e_n\rangle \xrightarrow[N\to \infty]{} -\lambda_n \left\langle T^\lambda \alpha, e_n \right\rangle.$$ This, put together with , gives by unicity of the limit, which in turn proves .
When $\varphi \in H^m$, $\tau^\varphi$ is simply $(-1)^{m-1}(\partial^{m-1} \varphi(L)-\partial^{m-1} \varphi(0)) Id$, $F$ is defined on $H^{m+1}\cap H^m_{per}$, and $\tilde{\tau}^\varphi \alpha$ is simply the symmetrisation $\alpha + (-1)^m \alpha(L-\cdot)$, which is $H^m_{per}$ if $\alpha\in H^m$.
Well-posedness of the closed-loop system
----------------------------------------
The operator equality we have established in the previous section means that $T^{\lambda}$ transforms, if they exist, solutions of the closed-loop system with a well-chosen feedback into solutions of the target system. Let us now check that the closed-loop system in question is indeed well-posed.
The operator $A+BK:=-\partial_x -\mu \alpha + \langle F, \cdot \rangle \varphi$ defined on $D_{m}$ generates a $C^0$-semigroup on $H^{m}_{per}$.
We know from Lemma \[Dense\] that $A+BK$ is densely defined. We use the Lumer-Phillips theorem (see for example [@Pazy Section 1.4]): if $A+BK$ is densely defined, closed, and if $A+BK$ and its adjoint are both dissipative on their respective domains, then $A+BK$ generates a $C^0$-semigroup.
Let us prove that $A+BK$ is closed. Let $$\alpha^{(N)} \in D_{m} \xrightarrow[N \to \infty]{H^{m}} \alpha \in H^{m}_{per}$$ such that $$(A+BK)\alpha^{(N)} \xrightarrow[N \to \infty]{H^{m}} \beta \in H^{m}_{per}$$ i.e. $$\label{ConvergencetoBeta}
-\alpha^{(N)}_x + \mu \alpha^{(N)} + \langle \alpha^{(N)}, F \rangle \varphi \xrightarrow[N \to \infty]{H^{m}} \beta.$$
Now, as $$-\alpha^{(N)}_x \xrightarrow[N\to \infty]{H^{m-1}} -\alpha_x,$$ the first two terms of the left-hand side of converge, which implies that $$\langle \alpha^{(N)}, F \rangle \varphi \xrightarrow[N\to \infty]{H^{m-1}} \gamma \varphi$$ and even $$\langle \alpha^{(N)}, F \rangle \varphi \xrightarrow[N\to \infty]{H^{m}_{(pw)}} \gamma \varphi$$ for some limit $\gamma \in {\mathbb{C}}$, so that $$\label{alphaLimit}-\alpha^{(N)}_x \xrightarrow[N \to \infty]{H^{m}_{(pw)}} \beta -\mu \alpha - \gamma \varphi.$$ Now, notice that implies that $$\varphi=\tau^\varphi \left(\sum_{n\in {\mathbb{Z}}^\ast} \frac{(-1)^{m-1}}{\left(\frac{2i\pi}{L}n\right)^m} e_n + r\right),$$ where $r \in H^m_{per}$. Moreover, it is well-known that $$\partial^m\sum_{n\in {\mathbb{Z}}^\ast} \frac{(-1)^{m-1}}{\left(\frac{2i\pi}{L}n\right)^m} e_n = \sum_{n\in {\mathbb{Z}}^\ast} \frac{(-1)^{m-1}}{\frac{2i\pi}{L}n} e_n$$ is a piecewise constant function, so in particular $$\sum_{n\in {\mathbb{Z}}^\ast} \frac{(-1)^{m-1}}{\left(\frac{2i\pi}{L}n\right)^m} e_n \in H^m_{(pw)}$$ which means that $\varphi\in \tau^\varphi \left(H^m_{(pw)}\right) \cap H^{m-1}_{per}$, and thus, by , $$-\alpha_x + \mu \alpha + \gamma \varphi = \beta,$$ $$\alpha \in \tau^\varphi \left(H^{m+1}_{(pw)}\right)\cap H^m_{per},$$ $$\alpha^{(N)} \xrightarrow[N \to \infty]{H^{m +1}_{(pw)}} \alpha.$$ Then, using Proposition \[FeedbackReg\], we get, by continuity of $F$ for the $H^{m+1}_{(pw)}$ norm: $$\langle \alpha, F \rangle=\gamma.$$ Hence, $\alpha \in D_{m}$ and $(A+BK)\alpha = \beta$, and $A+BK$ is closed.
Finally, to study the behavior of the adjoint, let us consider the norm $\|\cdot\|_{T^{\lambda}} := \|T^{\lambda} \cdot \|_{m}$, which is equivalent and its associated scalar product $\langle\cdot, \cdot \rangle_{T^{\lambda}}:=\langle T^{\lambda} \cdot, T^{\lambda} \cdot \rangle_{m}$. Then, for $\alpha \in D_{m}, \beta\in H^{m}_{per}$, we get, using the operator equality of Proposition \[OpEqProp\]: $$\begin{array}{rcl}
\langle (A+BK)\alpha, \beta \rangle_{T^{\lambda}} &=& \langle T^{\lambda} (-\alpha_x+\mu \alpha + \langle \alpha, F \rangle \varphi), T^{\lambda} \beta \rangle_{m}\\
&=&\langle -\partial_x T^{\lambda} \alpha - \lambda^\prime T^{\lambda} \alpha , T^{\lambda} \beta \rangle_m \\
&=& \langle T^{\lambda} \alpha, \partial_x T^{\lambda}\beta \rangle_{m} - \lambda^\prime \langle \alpha, \beta \rangle_{T^{\lambda}},
\end{array}$$ so that the adjoint of $(A+BK)$ for this scalar product is defined by $$\begin{array}{c}
D_{m}^\ast = \{\beta \in H^{m}_{per}, \quad T^{\lambda} \beta \in H^{m +1}\}, \\
(A+BK)^\ast=(T^{\lambda})^{-1}\partial_x \ T^{\lambda} - \lambda^\prime I \ \textrm{on} \ D_{m}^\ast.
\end{array}$$ Thus, for all $\alpha \in D_{m}, \gamma \in D_{m}^\ast$: $$\left\langle (A+BK)\alpha, \alpha\right\rangle_{T^{\lambda}} = \langle \partial_x T^{\lambda}\alpha, T^{\lambda}\alpha\rangle_{m} - \lambda^\prime \|\alpha\|_{T^{\lambda}},$$ and $$\left\langle (A+BK)^\ast\gamma, \gamma\right\rangle_{T^{\lambda}} = \langle \partial_x T^{\lambda}\gamma, T^{\lambda}\gamma\rangle_{m} - \lambda^\prime \|\gamma\|_{T^{\lambda}}.$$ Now, as for any $f \in H^{m+1}_ {per}$, we have, by integration by parts, $$\mathfrak{Re}\left(\left\langle \partial_x f, f\right\rangle_{m}\right)=0,$$ both $A+BK$ and $(A+BK)^\ast$ are dissipative.
Stability of the closed-loop system
-----------------------------------
We can now prove Theorem \[MainResult\].
Let $m \geq 1$, $m \geq 0$. Let $(F_n)\in {\mathbb{C}}^{\mathbb{Z}}$ with $m$-growth and growth constants $C_1, C_2>0$. Let us first consider trajectories with initial data in $D_{m}$.
System with initial data $\alpha_0 \in D_{m}$ has a unique solution $\alpha(t) \in C^0({\mathbb{R}}^{+}, D_{m}) \cap C^1({\mathbb{R}}^{+}, H^{m}_{per})$, and we have [$$\begin{array}{rcl}
\frac{d}{dt} T^{\lambda} \alpha &=& T^{\lambda} \frac{d}{dt} \alpha \\
&=& T^{\lambda}(-\alpha_x + \mu\alpha +\langle \alpha, F \rangle\varphi) \\
&=& (-\partial_x - \lambda^\prime I)T^{\lambda}\alpha.
\end{array}$$]{} Thus, $T^{\lambda}\alpha (t) \in H^{m}_{per}$ is a solution of the target system and satisfies for $t \geq 0$, $$\|T^{\lambda} \alpha (t) \|_{m} \leq e^{-\lambda^\prime t} \|T^{\lambda}\alpha_0\|_{m}.$$ Using , we then get, for $t \geq 0$, [$$\begin{array}{rcl}
\| \alpha(t) \|_{m} &\leq& {{\vert\kern-0.25ex\vert\kern-0.25ex\vert (T^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}}\|T^{\lambda} \alpha (t) \|_{m} \\
&\leq & {{\vert\kern-0.25ex\vert\kern-0.25ex\vert (T^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}}\ e^{-\lambda^\prime t} \|T^{\lambda}\alpha_0\|_{m}\\
&\leq& {{\vert\kern-0.25ex\vert\kern-0.25ex\vert (T^{\lambda})^{-1}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} {{\vert\kern-0.25ex\vert\kern-0.25ex\vert T^{\lambda}
\vert\kern-0.25ex\vert\kern-0.25ex\vert}} e^{-\lambda^\prime t} \|\alpha_0\|_{m} \\
&\leq & \left(\frac{C}{c}\right)^2 e^{\lambda L }e^{-\lambda^\prime t} \|\alpha_0\|_{m}
\end{array}$$]{} This proves the exponential stability of solutions to system with initial data in $D_{m}$. As the constant in this last inequality does not depend on the initial conditions, by density of $D_{m}$ in $H^{m}_{per}$, any solution to system with initial data in $H^{m}_{per}$ satisfies the last inequality with the same constant.
Application
-----------
Let $m=1, \lambda >0$, and let us suppose, to simplify the computations, that $a\equiv 0$. Define $$\label{phiEx}
\varphi(x)=L-x, \quad \forall x \in (0, L),$$ so that $\varphi\in H^1$ but is not periodic, and satisfies , with $$\begin{aligned}
\varphi_n=-\frac{i L^{\frac32}}{2\pi n},& \quad \forall n \in {\mathbb{Z}}^\ast, \\
\varphi_0= \frac{L^{\frac32}}2.&
\end{aligned}$$ Then, $$\label{Fex}
\langle \alpha, F \rangle = -\frac{2K(\lambda)}{L^{\frac32}} \alpha_0 - K(\lambda) \alpha_x(0), \quad \forall \alpha \in H^{2} \cap H^1_{per},$$ and $$D_1=\left\{\alpha \in H^2 \cap H^1_{per}, \quad \frac{2K(\lambda)}{L^{\frac32}} \alpha_0 + \left(\frac1L - K(\lambda)\right)\alpha_x(0)-\frac1L \alpha_x (L)=0 \right\},$$ so that $$\label{Systemex}
\left\{\begin{aligned}
\alpha_t + \alpha_x &= \left(-\frac{2K(\lambda)}{L^{\frac32}} \alpha_0 - K(\lambda) \alpha_x(0) \right)\varphi (x), \ x\in [0,L], \\
\alpha(t,0) & = \alpha(t,L), \ \forall t \geq 0,
\end{aligned}\right.$$
has a unique solution for initial conditions in $D_1$.
The backstepping transformation can be written as: $$\label{Tex}
\alpha = \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} \alpha_x - \frac{2K(\lambda)}{L^2} \alpha_0\right)\right) \star \varphi , \quad \forall \alpha \in H^1_{per}T^\lambda.$$ Let $\alpha(t) \in D_1$ be the solution of the closed loop system with initial condition $\alpha^0 \in D_1$, and let us note $z(t):=T^\lambda \alpha(t)$, then [$$\begin{array}{rcl}
z_t &=& \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} \alpha_{xt} - \frac{2K(\lambda)}{L^2} \alpha_0^\prime\right)\right) \star \varphi. \\
&=& \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} (-\alpha_{xx}+ \langle \alpha, F \rangle \varphi_x) - \frac{2K(\lambda)}{L^2} \alpha_0^\prime\right)\right) \star \varphi. \\
&=& \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} (-\alpha_{xx}- \langle \alpha, F \rangle) - \frac{2K(\lambda)}{L^2} \alpha_0^\prime\right)\right) \star \varphi. \\
z_x&=& \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(-e^{-\lambda x}\frac{K(\lambda)}{\sqrt{L}} \alpha_{xx} \right) \star \varphi - \lambda z \\
z_t+z_x + \lambda z &=& \frac{\sqrt{L}}{1-e^{-\lambda L}}\left(e^{-\lambda x}\left(\frac{K(\lambda)}{\sqrt{L}} \langle \alpha, F \rangle - \frac{2K(\lambda)}{L^2} \alpha_0^\prime\right)\right) \star \varphi.
\end{array}$$]{} By projecting the closed loop system on $e_0$, we get $$\alpha_0^\prime = \langle \alpha, F\rangle \varphi_0 = \langle \alpha, F \rangle \frac{L^{\frac32}}2$$ so that $$z_t+z_x + \lambda z =0.$$ In particular, $$\label{energyDissip}
\frac{d}{dt}\|z\|^2_1=-2\lambda\|z\|^2_1.$$ Let us now set $$V(\alpha):= \|z\|^2_1, \quad \forall \alpha \in H^1_{per}.$$
Now, notice that [$$\begin{array}{rcl}
\|T^\lambda \alpha \|^2_1&=& \frac{L}{(1-e^{-\lambda L})^2} \sum_{n\in {\mathbb{Z}}} \left(1+\left|\frac{2i\pi n}{L}\right|^2\right) |\varphi_n|^2 \left|\left\langle e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} \alpha_x - \frac{2K(\lambda)}{L^2} \alpha_0\right), e_n \right\rangle\right|^2 \\
&\geq& C \left\| e^{-\lambda x}\left(-\frac{K(\lambda)}{\sqrt{L}} \alpha_x - \frac{2K(\lambda)}{L^2} \alpha_0\right) \right\|^2 \\
&\geq& C e^{2\lambda L} \left\| -\frac{K(\lambda)}{\sqrt{L}} \alpha_x - \frac{2K(\lambda)}{L^2} \alpha_0 \right\|^2 \\
&\geq& C^\prime K(\lambda)^2 e^{2\lambda L} \|\alpha\|_1^2.
\end{array}$$]{} Together with , this shows that $V$ is a Lyapunov function, and is exponentially stable.
Further remarks and questions
=============================
Controllability and the $TB=B$ condition
----------------------------------------
In the introduction we have mentioned that the growth constraint on the Fourier coefficients of $\varphi$ actually corresponds to the exact null controllability condition in some Sobolev space for the control system . As we have mentioned in the finite dimensional example, the controllability condition is essential to solve the operator equation: in our case, formal computations lead to a family of functions that turns out to be a Riesz basis precisely thanks to that rate of growth. Moreover, that rate of growth is essential for the compatibility of the $TB=B$ condition and the invertibility of the backstepping transformation. Indeed, as the transformation is constructed formally using a formal $TB=B$ condition, that same $TB=B$ condition fixes the value of the coefficients of $F_n$, giving them the right rate of growth for $T^\lambda$ to be an isomorphism.
It should be noted that, while in [@Schrodinger] the $TB=B$ condition is well-defined, in our case, it only holds in a rather weak sense. This is probably because of a lack of regularization, indeed in [@Schrodinger] the backstepping transformation has nice properties, as it can be decomposed in Fredholm form, i.e. as the sum of a isomorphism and a compact operator. Accordingly, the Riesz base in that case is quadratically close to the orthonormal base given by the eigenvectors of the Laplacian operator. That is not the case for our backstepping transformation, as it is closely linked to the operator $\Lambda^\lambda$, which does not have any nice spectral properties.
Nonetheless, it appears that thanks to some information on the regularity of $F$, a weak sense is sufficient and allows us to prove the operator equality by convergence.
In that spirit, it would be interesting to investigate if a backstepping approach is still valid if the conditions on $\varphi$ are weakened. For example, if we suppose approximate controllability instead of exact controllability, i.e. $$\varphi_n \neq 0, \quad \forall n \in {\mathbb{Z}},$$ then $F$ can still be defined using a weak $TB=B$ condition. However, it seems delicate to prove, in the same direct way as we have done, that $T^\lambda$ is an isomorphism, as we only get the completeness of the corresponding $\left(k_{n, \lambda}\right)$, but not the Riesz basis property.
Null-controllability and finite-time stabilization
--------------------------------------------------
As we have mentioned in the introduction, one of the advantages of the backstepping method is that it can provide an explicit expression for feedbacks, thus allowing the construction of explicit controls for null controllability, as well as time-varying feedbacks that stabilize the system in finite time $T>0$.
The general strategy (as is done in [@CoronNguyen], [@XiangKdVNull]) is to divide the interval $[0,T]$ in smaller intervals $[t_n, t_{n+1}]$, the length of which tends to $0$, and on which one applies feedbacks to get exponential stabilization with decay rates $\lambda_n$, with $\lambda_n \rightarrow \infty$. Then, for well-chosen $t_n, \lambda_n$, the trajectory thus obtained reaches $0$ in time $T$. Though this provides an explicit control to steer the system to $0$, the norm of the operators applied successively to obtain the control tends to infinity. As such, it does not provide a reasonably regular feedback. However, the previous construction of the control can be used, with some adequate modifications (see [@CoronNguyen] and [@XiangKdVFinite]) to design a time-varying, periodic feedback, with some regularity in the state variable, which stabilizes the system in finite time.
Let us first note that, due to the hyperbolic nature of the system, there is a minimal control time, and thus small-time stabilization cannot be expected. Moreover, even for $T>L$, the estimates we have established on the backstepping transforms prevent us from applying the strategy we have described above: indeed, for any sequences $(t_n) \rightarrow T$, $\lambda_n \rightarrow \infty$, we have $$\|\alpha(t)\|_{m} \leq \prod_{k=0}^n \left(\frac{C}{c}\right)^{2n} e^{n\mu L} exp\left(\sum_{k=0}^n -\lambda_k(t_{k+1}-t_k -L)\right) \|\alpha_0\|_{m}, \quad \forall t \in [t_n, t_{n+1}],$$ where $c, C$ are the decay constants in . Moreover, as $t_{k+1}-t_k \rightarrow 0$, we have $$exp\left(\sum_{k=0}^n -\lambda_k(t_{k+1}-t_k -L)\right) \xrightarrow[n \to \infty]{} \infty.$$
Another approach could be to draw from [@CHO2] and apply a second transformation to design a more efficient feedback law. It would also be interesting to adapt the strategy in [@YuXuJiangGanesan], inspired from [@KrsticHaraTsub], to our setting.
Nonlinear systems
-----------------
Finally, another prospect, having obtained explicit feedbacks that stabilize the linear system, is to investigate the stabilization of nonlinear transport equations. This has been done in [@CoronLu2], where the authors show that the feedback law obtained for the linear Korteweg-de Vries equation also stabilize the nonlinear equation. However, as in [@Schrodinger], the feedback law we have obtained is not continuous in the norm for which the system is stabilize. This would require some nonlinear modifications to the feedback law in order to stabilize the nonlinear system.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is partially supported by ANR project Finite4SoS (ANR-15-CE23-0007), and the French Corps des Mines.
The author would like to thank Jean-Michel Coron, for bringing the problem to his attention, his constant support, and his valuable remarks, as well as Amaury Hayat and Shengquan Xiang for discussions on this problem.
[9]{}
Dejan M. Bosković, Andras Balogh, and Miroslav Krstić. Backstepping in infinite dimension for a class of parabolic distributed parameter systems. , 16(1):44–75, 2003.
Georges Bastin and Jean-Michel Coron. , volume 88 of [*Progress in Nonlinear Differential Equations and their Applications*]{}. Birkhäuser/Springer, \[Cham\], 2016. Subseries in Control.
Andras Balogh and Miroslav Krstić. . , 8(2):165–175, 2002.
Haim Brezis. . Universitext. Springer, New York, 2011.
Ole Christensen. . Applied and Numerical Harmonic Analysis. Birkhäuser/Springer, \[Cham\], second edition, 2016.
Jean-Michel Coron and Brigitte d’Andréa Novel. Stabilization of a rotating body beam without damping. , 43(5):608–618, 1998.
Jean-Michel Coron, Ludovick Gagnon, and Morgan Morancey. Rapid stabilization of a linearized bilinear 1-[D]{} [S]{}chrödinger equation. , 115:24–73, 2018.
Jean-Michel Coron, Long Hu, and Guillaume Olive. Stabilization and controllability of first-order integro-differential hyperbolic equations. , 271(12):3554–3587, 2016.
Jean-Michel Coron, Long Hu, and Guillaume Olive. Finite-time boundary stabilization of general linear hyperbolic balance laws via [F]{}redholm backstepping transformation. , 84:95–100, 2017.
Jean-Michel Coron and Qi Lü. Local rapid stabilization for a [K]{}orteweg-de [V]{}ries equation with a [N]{}eumann boundary control on the right. , 102(6):1080–1120, 2014.
Jean-Michel Coron and Qi Lü. Fredholm transform and local rapid stabilization for a [K]{}uramoto-[S]{}ivashinsky equation. , 259(8):3683–3729, 2015.
Jean-Michel Coron and Hoai-Minh Nguyen. Null controllability and finite time stabilization for the heat equations with variable coefficients in space in one dimension via backstepping approach. , 225(3):993–1023, 2017.
Jean-Michel Coron. , volume 136 of [*Mathematical Surveys and Monographs*]{}. American Mathematical Society, Providence, RI, 2007.
Jean-Michel Coron. Stabilization of control systems and nonlinearities. In [*Proceedings of the 8th [I]{}nternational [C]{}ongress on [I]{}ndustrial and [A]{}pplied [M]{}athematics*]{}, pages 17–40. Higher Ed. Press, Beijing, 2015.
Jean-Michel Coron, Rafael Vazquez, Miroslav Krstić, and Georges Bastin. Local exponential [$H^2$]{} stabilization of a [$2\times 2$]{} quasilinear hyperbolic system using backstepping. , 51(3):2005–2035, 2013.
Amaury Hayat. . preprint, October 2017. <https://hal.archives-ouvertes.fr/hal-01613139>
Amaury Hayat. preprint, July 2018. <https://hal.archives-ouvertes.fr/hal-01790104/>
Jean-Pierre Kahane and Pierre-Gilles Lemarié-Rieusset. . Cassini, 1998.
Miroslav Krstić, Bao-Zhu Guo, Andras Balogh, and Andrey Smyshlyaev. Output-feedback stabilization of an unstable wave equation. , 44(1):63–74, 2008.
Vilmos Komornik. Rapid boundary stabilization of linear distributed systems. , 35(5):1591–1613, 1997.
Miroslav Krstić and Andrey Smyshlyaev. , volume 16 of [*Advances in Design and Control*]{}. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2008. A course on backstepping designs.
Wei Jiu Liu and Miroslav Krstić. Backstepping boundary control of [B]{}urgers’ equation with actuator dynamics. , 41(4):291–303, 2000.
Wei Jiu Liu and Miroslav Krstić. Boundary feedback stabilization of homogeneous equilibria in unstable fluid mixtures. , 80(6):982–989, 2007.
Amnon Pazy. , volume 44 of [*Applied Mathematical Sciences*]{}. Springer-Verlag, New York, 1983.
Andrey Smyshlyaev, Eduardo Cerpa, and Miroslav Krstić. Boundary stabilization of a 1-[D]{} wave equation with in-domain antidamping. , 48(6):4014–4031, 2010.
Eduardo D. Sontag. , volume 6 of [*Texts in Applied Mathematics*]{}. Springer-Verlag, New York, second edition, 1998. Deterministic finite-dimensional systems.
Daisuke Tsubakino, Miroslav Krstić, and Shinji Hara. Backstepping control for parabolic pdes with in-domain actuation. , pages 2226–2231, 2012.
Jose Manuel Urquiza. Rapid exponential feedback stabilization with unbounded control operators. , 43(6):2233–2244, 2005.
Ambroise Vest. Rapid stabilization in a semigroup framework. , 51(5):4169–4188, 2013.
Frank Woittennek, Siqian Wang, and Torsten Kn[ü]{}ppel. . , 47(3):5175–5180, 2014. 19th IFAC World Congress.
Shengquan Xiang. . preprint, February 2017. <https://hal.archives-ouvertes.fr/hal-01468750>
Shengquan Xiang. Small-time local stabilization for a [K]{}orteweg–de [V]{}ries equation. , 111:64–69, 2018.
Xin Yu, Chao Xu, Huacheng Jiang, Arthi Ganesan, and Guojie Zheng. Backstepping synthesis for feedback control of first-order hyperbolic [PDE]{}s with spatial-temporal actuation. , pages Art. ID 643640, 13, 2014.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the influence of spontaneous symmetry breaking on the decoherence of a many-particle quantum system. This decoherence process is analyzed in an exactly solvable model system that is known to be representative of symmetry broken macroscopic systems in equilibrium. It is shown that spontaneous symmetry breaking imposes a fundamental limit to the time that a system can stay quantum coherent. This universal timescale is $t_{spon} \simeq 2\pi N \hbar / (k_B T)$, given in terms of the number of microscopic degrees of freedom $N$, temperature $T$, and the constants of Planck ($\hbar$) and Boltzmann ($k_B$).'
author:
- 'Jasper van Wezel, Jeroen van den Brink and Jan Zaanen'
title: An intrinsic limit to quantum coherence due to spontaneous symmetry breaking
---
[*Introduction.*]{} The relation between quantum physics at microscopic scales and the classical behavior of macroscopic bodies has been a puzzle in physics since the days of Einstein and Bohr. This subject has revived in recent years both due to experimental progress, making it possible to study this problem empirically, and because of its possible implications for the use of quantum physics as a computational resource [@Bennett00]. This ‘micro-macro’ connection actually has two sides. Under equilibrium conditions it is well understood in terms of the mechanism of spontaneous symmetry breaking. But in the dynamical realms its precise nature is still far from clear. The question is, can spontaneous symmetry breaking play a role in a [*dynamical*]{} reduction of quantum physics to classical behavior? This is a highly non trivial question as spontaneous symmetry breaking is intrinsically associated with the difficult problem of many particle quantum physics. Here, we will analyze a tractable model system which is known to be representative of macroscopic systems in equilibrium, to find the surprising outcome that [*spontaneous symmetry breaking imposes a fundamental limit to the time that a system can stay quantum coherent*]{} [@Caldeira81; @Chuang95]. This universal timescale turns out to be $t_{spon} \simeq 2\pi N \hbar / (k_B T)$.
This result is surprising in the following sense. Consider a macroscopic body at room temperature; $\hbar / (k_B T) \simeq 10^{-14}$ seconds which is quite a short time. However, multiplying it with Avogadro’s number $N \simeq 10^{24}$, $t_{spon}$ becomes $ \simeq 10^{10}$ seconds, a couple of centuries. Given all other sources of decoherence for such a large macroscopic body, this is surely not a relevant timescale. However, quantum systems of contemporary interest are typically much smaller. Let us for instance consider a flux state qubit with a squid the size of one cubic micron and a temperature of the order of one Kelvin [@Chiorescu03]. The time $t_{spon}$ then turns out to be of order of seconds, a coherence timescale which might well be reached in the near future. The counterintuitive feature of this intrinsic decoherence mechanism linked to equilibrium classicality is that it starts to matter when systems become small.
[*Spontaneous symmetry breaking.*]{} In main-stream quantum measurement theory, the nature of the classical machine executing the measurement is treated rather casually – it is just asserted to exist, according to daily observations. However, eventually this machine is also subjected to the laws of quantum physics. After all, it is made of microscopic stuff similar to the small quantum system on which the machine acts. The description of this machine typically involves $10^{24}$ strongly interacting quantum particles and this is not an easy problem. In fact, the very existence of the machine seems to violate the basic laws of quantum physics. The most fundamental difference between quantum- and classical physics lies in the role of symmetry. Dealing with an exact quantum mechanical eigenstate, all configurations equivalent by symmetry should have the same status in principle, while in a classical state one of them is singled out. For example, given that space is translationally invariant, the measurement machine should be in an eigenstate of total momentum, being spread out with equal probability over all of space. In the classical limit however it takes a definite locus. The explanation of this ‘spontaneous symmetry breaking’ in terms of the singular nature of the thermodynamic limit is one of the central achievements of quantum condensed matter physics [@Anderson72]. One imagines a symmetry breaking ‘order parameter field’ $h$ (e.g., a potential singling out a specific position in space). Upon sending $h$ to zero before taking the thermodynamic limit ($N \rightarrow \infty$) one finds the exact quantum groundstate respecting the symmetry. However, taking the opposite order of limits one finds that the classical state becomes fact.
What does all of this have to do with the dynamical phenomenon of decoherence? Decoherence refers to the fact that the quantum information encoded in some microscopic state entangles in the course of its time evolution with environmental degrees of freedom. Since this information cannot be recovered ‘for any practical purpose’, one should trace out the environment from the density matrix with the effect that the reduced density matrix will reveal a mixed state. The crucial point is that spontaneous symmetry breaking is intrinsically linked to the presence of a spectrum of ‘environmental states’. In a rigorous fashion, the quantum information carried by these states cannot be retrieved when the body is macroscopic. This so-called ‘thin spectrum’ is so sparse that it even ceases to influence the partition function [@Kaplan90]. The question we wish to address in the remainder is, to what extent this [*thin spectrum can be a source of decoherence, intrinsically associated with the fact that quantum measurements need classical measurement machines*]{}.
Given that spontaneous symmetry breaking involves the a-priori untractable problem of a near-infinity of interacting quantum degrees of freedom, this question cannot be answered in full generality. However, some time ago it was discovered that the mechanism of spontaneous symmetry breaking reveals itself in representative form in a simple, integrable model. This model is the Lieb-Mattis long-ranged quantum Heisenberg antiferromagnet [@Lieb62], given by the Hamiltonian $$\begin{aligned}
H_{LM} = \frac{2|J|}{N} {\bf S}_A \cdot {\bf S}_B - h (S_A^z - S_B^z),\end{aligned}$$ It is defined for a bipartite lattice with $A$ and $B$ sublattices, where ${\bf S}_{A/B}$ is the total spin on the $A/B$ sublattice with $z$-projection $S_{A/B}^z$, and $h$ is the symmetry breaking field, in this case a staggered magnetic field acting on the staggered magnetization $M^z = S_A^z - S_B^z$. The particularity of the Lieb-Mattis Hamiltonian is that every spin on sublattice $A$ is interacting with [*all*]{} spins on sublattice $B$ and vice versa, with interaction strength $2|J|/N$ (which depends on the total number of sites $N$ so that the system is extensive). This very simple Hamiltonian accurately describes the thin spectrum encountered in more complicated Hamiltonians, like the nearest neighbor Heisenberg antiferromagnet, the BCS superconductor, or the harmonic crystal [@Bernu92; @Anderson58]. Moreover, in this Hamiltonian the singular nature of the thermodynamic limit can be explicitly demonstrated [@Kaplan90; @Kaiser89]. We will therefore use the Lieb-Mattis magnet as a model for a measurement machine.
[*Measurement scheme.*]{} Our scheme for quantum measurement using this Lieb-Mattis magnet explicitly keeps track of the particular role of the thin spectrum. We start out preparing the Lieb-Mattis machine built from $N$ spins at time $t < t_0$ in the symmetry broken Néel ground state ($\langle M^z \rangle \neq 0$) of $H_{LM}$. The microscopic quantum system to be measured is isolated at $t<t_0$ and consists of two qubits (qubits $a$ and $b$, each with two $S=1/2$ states) in a maximally entangled singlet state, $|qubit \rangle = \frac{1}{\sqrt{2}}\left[ | \uparrow_a \downarrow_b \rangle - | \downarrow_a \uparrow_b \rangle \right]$. At time $t=t_0$ we instantaneously include qubit $a$ ($b$) in the Lieb-Mattis (infinite range) interactions of the spins on the $A$ $(B)$ sublattice of the Lieb-Mattis machine. We then follow the exact time evolution of the coupled $N+2$ particle system at $t > t_0$: $$\begin{aligned}
H_{t<t_0} &=& H_{LM}+ {\bf S}_a \cdot {\bf S}_b \nonumber \\
H_{t>t_0} &=& \frac{2|J|}{N+2} {\bf S}_{A+a} \cdot {\bf S}_{B+b} - h (S_{A+a}^z - S_{B+b}^z),\end{aligned}$$ where ${\bf S}_{A+a}$ is ${\bf S}_A + {\bf S}_a$, and ${\bf S}_{B+b}$ is ${\bf S}_B + {\bf S}_b$.
To obtain further insight in how this quantum measurement works, let us first see what would happen if the measurement machine would be semi-classical, i.e. described in terms of a spin wave expansion. This starts with assuming a maximally polarized staggered magnetization $\langle M^z \rangle$ for the Lieb-Mattis measurement machine. By linearizing the equations of motion one then obtains the spin waves that are characterized by a ’plasmon’ gap due to the long range nature of the interactions. Stronger, because of the infinite range of the interactions their spectrum is dispersionless and it is easily demonstrated that in fact the spin waves do not give rise to perturbative quantum corrections to the staggered magnetization –from this perspective, the classical Néel state appears to be an exact eigenstate. It is now immediately clear what happens at times $t > t_0$. At $t<t_0$ the system was prepared in a product state of the spin singlet qubit and the $N$-spin Néel ground state of the Lieb-Mattis antiferromagnet: $|\psi_{t<t_0} \rangle = |0\rangle_{N} \times |qubit\rangle$. When at $t=t_0$ the interaction between the micro and macro system is switched on, the $N+2$ spin system can either be in its Néel ground state $|0 \rangle_{N+2}$ or in an excited state where both spins $a$ and $b$ are misaligned relative to the magnetization on the respective sublattices with which they interact (Fig. 1). This state corresponds to a two magnon excited state and since the magnons do not propagate, this excited state $|2 \rangle_{N+2}$ also appears to be an exact eigen state. Hence, the semi-classical wavefunction is simply $|\psi^{\rm sc}_{t=t_0} \rangle = \left[ |0 \rangle_{N+2} - |2\rangle_{N+2} \right] / \sqrt{2}$ and the time evolution at $t > t_0$ is characterized by a coherent oscillation between the two states. Since the state $|2\rangle$ is distinguishable from the ground state $|0\rangle$, it is in principle measurable by slowly switching on interactions with other environmental degrees of freedom, and eventually the wavefunction will collapse. The outcome of this experiment would be the usual Rabi oscillations, with a frequency that is proportional to $E_2-E_0$, the energy difference between the two states. Thus in a semi-classical description there is no loss of quantum coherence.
One recognizes in the above the typical way that canonical measurement machines are conceptualized in quantum measurement theory. The surprise is now that even for this (in a sense, extremely ‘classical’) Lieb-Mattis measurement machine the preceding semi-classical analysis is only exact when [*the machine is infinitely large*]{}! The construction turns out to be subtly flawed when $N$ is finite and $T>0$. The culprit is the thin spectrum which is completely disregarded in the semi-classical analysis. To reveal the decohering effect of the thin spectrum, the Lieb-Mattis model should be solved exactly. This can in fact easily be done by first introducing the operator of [*total*]{} spin ${\bf S}={\bf S}_A+{\bf S}_B$. Taking $h=0$, the Hamiltonian can then be written as $(J/N) ( S^2 - S^2_A -S^2_B)$ and accordingly the eigenstates are $|S_A, S_B, S, M \rangle$ where $S,M$ denote total spin and its z-axis projection, while $S_A$ and $S_B$ refer to the total sublattice spin quantum numbers. $S_A$ and $S_B$ are maximally polarized in the ground state. Lowering $S_A$ or $S_B$ corresponds to exciting a magnon carrying an energy $J$. One sees immediately that the [*true*]{} ground state of the system is an overall $S=0$ [*singlet*]{}, i.e. a state characterized by $\langle M^z \rangle = 0$. One also infers the presence of a tower of [*total*]{} $S$ states characterized by an energy scale $E_{thin}=J/N$, and this is the thin spectrum. For a finite staggered magnetic field $h$, the situation changes drastically; $h$ couples the states in the thin spectrum and it is easy to show that the ground state becomes a wave packet of thin spectrum states and in this case $E_{thin}=\sqrt{Jh}$. This groundstate does carry a finite staggered magnetization: it is the antiferromagnetic Néel state. One can now straightforwardly demonstrate the singular nature of the thermodynamic limit [@Kaplan90; @Kaiser89]. By sending first $h \rightarrow 0$ and then $N \rightarrow \infty$ one obtains the exact total singlet groundstate, respecting the spin rotational symmetry. Upon taking the opposite order one finds the fully polarized Néel antiferromagnet of the semi-classical expansion.
[*Exact time evolution.*]{} Let us now reconsider our quantum measurement, taking full account of the thin spectrum states (Fig. 2). For $t<t_0$ the Lieb-Mattis machine is described by the following thermal density matrix, assuming that ${\rm k_B T} \ll J$ so that magnon excitations can be neglected, $$\begin{aligned}
\rho_{t < t_0} = \frac{1}{Z} \sum_{n=0}^{N-1} e^{-\frac{E^n_0}{\rm k_B T } } |0,n \rangle \times |qubit \rangle \langle 0,n| \times \langle qubit|, \end{aligned}$$ where $Z$ is the partition function, the thin spectrum states are labeled by $n$ and have an energy $E_0^n$. Switching on the Lieb-Mattis interaction between the qubits and the machine’s sublattices at $t = 0$ we find that the density matrix at $t > t_0$ becomes, $$\begin{aligned}
\rho_{t>t_0} &=& U \rho_{t=t_0} U^{\dagger} \nonumber \\
&=& \frac{1}{2Z} \sum_{n=0}^{N-1} \left. e^{-\frac{E^n_0}{\rm k_B T} }\right. \left[ |0,n\rangle \langle 0,n| + |2,n\rangle\langle 2,n| \right. \nonumber \\
&+& e^{-i(E^n_2-E^n_0)(t-t_0)/\hbar } \left. \left( |0,n\rangle\langle 2,n| +h.c.\right) \right], \end{aligned}$$ where $U$ is the exact time evolution operator and the states now describe the $(N+2)$-particle Lieb-Mattis model. Given their unobservable nature [@Kaplan90], we trace over the thin spectrum states in this density matrix. The off-diagonal matrix elements of this reduced density matrix are now $$\begin{aligned}
\rho^{OD}_{t>t_0}=\frac{e^{-2iJ(t-t_0)/\hbar}}{2Z} \sum_{n=0}^{N-1} e^{-\frac{E^n_0}{\rm k_B T} }
e^{-i(E^n_2-E^n_0-2J)(t-t_0)/\hbar }, \end{aligned}$$ where the phase factor associated with the two-magnon state is taken out of the summation. The absolute value $|\rho^{OD}_t|$ is the measure for the time dependent entanglement between states $|0\rangle$ and $|2\rangle$. It can be evaluated exactly for any given $N$ and the result is shown in Fig. 3. The vanishing of this matrix element in the course of the time evolution signals decoherence and we find that this is associated with a characteristic timescale of a remarkably universal nature: under the physical conditions that $E_{thin} \ll {\rm k_B T} \ll J$ and $J/N < hN$ we find that the decoherence time due to spontaneous symmetry breaking becomes completely independent of the energy scales characterizing the system: $t_{spon} = { 2 \pi N \hbar}/{ \rm k_B T}$, the result we announced in the beginning.
The fact that the reduced density matrix at $t>t_0$ describes a mixed state, while at $t<t_0$ the system was in a pure state, could lead to the conclusion that the present mechanism for decoherence is irreversible. But irreversibility is at odds with unitary time evolution [@Adler02]. We actually do find that after a certain time $t_{rec}$ the system returns to a pure state again, with exactly the same reduced density matrix it started with at $t=t_0$. Thus the decoherence is in fact reversible, see fig 3. This recurrence time depends on the energy scales of the Lieb-Mattis measurement machine in a quite remarkable way: $t_{rec}/ t_{spon}= {\rm k_B T} / E_{thin}$. Under the physical condition that the typical level splittings in the thin spectrum are very small compared to temperature, the recurrence takes infinitely long so that for all practical purposes the thin spectrum acts as a truly dissipative bath turning quantum information into an increase of classical entropy.
[*Origin of decoherence.*]{} Given that decoherence via the thin spectrum requires temperature to be finite, it is tempting to associate $t_{spon}$ with the thermal fluctuations of the order parameter in the finite system, as described by spin wave theory. However, this is [*not*]{} the case because these thermal fluctuations invoke the thermal excitation of the magnon states. These are exponentially suppressed by Boltzmann factors $e^{-J/(k_B T)}$, which depend on the energy scale $J$ of the individual interactions. The origin of $t_{spon}$ is more subtle: it is due to the hidden thin spectrum that reflects the zero point fluctuations of the order parameter as a whole. This thin spectrum does not carry any thermodynamic weight, and turns into a heat bath destroying quantum information if temperature is finite.
It is remarkable that the coherence time is such a universal timescale, independent of the detailed form of the thin spectrum –which, after all, is determined by the parameters $J$ and $h$ in the Lieb-Mattis Hamiltonian. Physically one can think of this universal timescale as arising from two separate ingredients. First, the energy of a thin spectrum state $|n \rangle$ changes when magnons appear. The change is of the order of $n E_{thin}/N$, where $E_{thin}$ is the characteristic level spacing of the thin spectrum that we happen to be considering. The fact that each thin state shifts its energy somewhat at $t>t_0$ leads to a phase shift of each thin state and in general these phases interfere destructively, leading to dephasing and decoherence. The larger $n E_{thin}/N$, the faster this dynamics. But in order for this dephasing to occur, it is necessary for a finite number of thin states to actually participate in the dynamics of decoherence. Since temperature is finite (but always small compared to the magnon energy) a finite part of the thin spectrum is available for the dynamics. Thin spectrum states with an excitation energy higher than $k_B T$ are suppressed exponentially due to their Boltzmann weights. The maximum number of thin states that do contribute is roughly determined by the condition that $n^{max} \sim k_B T/E_{thin}$. Putting the ingredients together, we find that the highest energy scale that is available to the system to decohere is approximately $ \frac{k_B T}{E_{thin}} \frac{E_{thin}}{N}$. All together, the thin spectrum drops out of the equations. The fastest time scale at which the dynamics take place is given by the inverse of this energy scale, converted into time: one finds the decoherence time $t_{spon} \sim \frac{2 \pi \hbar N}{k_B T}$.
[*Conclusions.*]{} To what extent is the Lieb-Mattis machine representative of a general classical measurement machine displaying a broken continuous symmetry? In fact the Lieb Mattis machine is the best case scenario for the kind of measurement machine envisaged in main stream quantum measurement theory, as its behavior is extremely close to semi-classical due to the presence of the infinite range interactions. Machines characterized by short range interactions carry massless Goldstone modes and these will surely act as an additional heat bath limiting the coherence time. It is of course not an accident that the most ’silent’ systems are qubits based on superconducting circuitry, which have a massive Goldstone spectrum in common with the Lieb-Mattis system. We have demonstrated here that even under these most favorable circumstances quantum coherence eventually has to come to an end, because of the unavoidable condition that even the most classical measurement machines are subtly influenced by their quantum origin. These effects become noticeable in the mesoscopic realms and we present it as a challenge to the experimental community to measure $t_{spon}$.
[99]{} C.H. Bennett and D.P. DiVincenzo, Nature [**404**]{}, 247 (2000). A.O. Caldeira and A.J. Leggett, Phys. Rev. Lett., [**46**]{}, 211 (1981). I.L. Chuang, R. Laflamme, P.W. Shor and W.H. Zurek, Science [**270**]{}, 1633 (1995). I. Chiorescu, Y. Nakamura, C.J.P.M. Hormans and J.E. Mooij, Science, [**299**]{}, 1869 (2003). P.W. Anderson, Science, [**177**]{}, 393 (1972). T.A. Kaplan, W. Von der Linden, and P. Horsch, Phys. Rev. B, [**42**]{}, 4663 (1990). E. Lieb, and D. Mattis, J. Math. Phys., [**3**]{}, 749 (1962). B. Bernu, C. Lhuillier and L. Pierre, Phys. Rev. Lett., [**69**]{}, 2590 (1992), B. Bernu, P. Lecheminant, C. Lhuillier, and L. Pierre, Phys. Rev. B, [**50**]{}, 10048 (1994). P.W. Anderson, Phys. Rev., [**112**]{}, 1900 (1958). C. Kaiser, and I. Peschel, J. Phys. A, [**22**]{}, 4257 (1989). A.E. Allahverdyan, R. Balian, Th.M. Nieuwenhuizen, Phys. Rev. Lett., [**92**]{}, 120402 (2004). S.L. Adler, Stud. Hist. Philos. M. P., [**34**]{}, B 135 (2003). A.J. Leggett, J. Phys. Cond. Mat., [**14**]{}, R415 (2002). J. Zaanen, Nature, [**430**]{}, 512 (2004), and references therein.
{width="0.8\columnwidth"}
{width="0.4\columnwidth"}
{width="0.5\columnwidth"} {width="0.5\columnwidth"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This article studies a direct numerical approach for fractional advection-diffusion equations (ADEs). Using a set of cubic trigonometric B-splines as test functions, a differential quadrature (DQ) method is firstly proposed for the 1D and 2D time-fractional ADEs of order $(0,1]$. The weighted coefficients are determined, and with them, the original equation is transformed into a group of general ordinary differential equations (ODEs), which are discretized by an effective difference scheme or Runge-Kutta method. The stability is investigated under a mild theoretical condition. Secondly, based on a set of cubic B-splines, we develop a new Crank-Nicolson type DQ method for the 2D space-fractional ADEs without advection. The DQ approximations to fractional derivatives are introduced and the values of the fractional derivatives of B-splines are computed by deriving explicit formulas. The presented DQ methods are evaluated on five benchmark problems and the concrete simulations of the unsteady propagation of solitons and Gaussian pulse. In comparison with the existing algorithms in the open literature, numerical results finally illustrate the validity and accuracy.'
author:
- 'X. G. Zhu'
- 'Y. F. Nie'
- 'W. W. Zhang'
date: 'Received: date / Accepted: date'
title: 'An efficient differential quadrature method for fractional advection-diffusion equation'
---
Introduction {#intro}
============
The differential equations with a fractional derivative serve as superior models in subjects as diverse as astrophysics, chaotic dynamics, fractal network, signal processing, continuum mechanics, turbulent flow and wave propagation [@ref02; @ref01; @19; @20]. This type of equations admit the non-local memory effects in mathematical mechanism, thereby filling in a big gap that the classical models can not work well for some of the natural phenomena like anomalous transport. In general, the exact solutions can seldom be represented as closed-form expressions by using elementary functions that presents a tough challenge to derive a sufficiently valid method concerned with analytic approximations, so a keen interest has been attracted to design robust algorithms to investigate them in numerical perspectives.
In this article, we aim to construct an efficient method to numerically solve the general problems:
- 1D time-fractional ADEs $$\begin{aligned}
\frac{\partial^\alpha u(x,t)}{\partial t^\alpha}+\kappa\frac{\partial u(x,t)}{\partial x}
&-\varepsilon\frac{\partial^2 u(x,t)}{\partial x^2} =f(x,t), \label{eq01}\end{aligned}$$ with $0<\alpha\leq1$, $\kappa, \varepsilon \geq 0$, $a\leq x\leq b$, $t>0$, and the initial and boundary conditions $$\begin{aligned}
& u(x,0)=\psi(x),\quad a\leq x\leq b, \label{eq02}\\
& u(a,t)=g_1(t), \quad u(b,t)=g_2(t), \quad t>0; \label{eq03}\end{aligned}$$
- 2D time-fractional ADEs $$\begin{gathered}
\begin{aligned}\label{eq04}
&\frac{\partial^\alpha u(x,y,t)}{\partial t^\alpha}+\kappa_x\frac{\partial u(x,y,t)}{\partial x}+\kappa_y\frac{\partial u(x,y,t)}{\partial y}\\
&-\varepsilon_x\frac{\partial^2 u(x,y,t)}{\partial x^2}-\varepsilon_y\frac{\partial^2 u(x,y,t)}{\partial y^2} =f(x,y,t),
\end{aligned}\end{gathered}$$ with $0<\alpha\leq1$, $\kappa_x, \kappa_y, \varepsilon_x,\varepsilon_y \geq 0$, $(x,y)\in\Omega$, $t>0$, and the initial and boundary conditions $$\begin{aligned}
& u(x,y,0)=\psi(x,y),\quad (x,y)\in\Omega, \label{eq05}\\
& u(x,y,t)=g(x,y,t), \ \ (x,y)\in\partial\Omega,\quad t>0, \label{eq07}\end{aligned}$$ where $\Omega=\{(x,y):a\leq x\leq b, c\leq y\leq d\}$ and $\partial\Omega$ denotes its boundary;
- 2D space-fractional ADEs without advection $$\begin{gathered}
\begin{aligned}\label{xz01}
&\frac{\partial u(x,y,t)}{\partial t}-\varepsilon_x\frac{\partial^{\beta_1} u(x,y,t)}{\partial x^{\beta_1}}\\
&\qquad\qquad-\varepsilon_y\frac{\partial^{\beta_2} u(x,y,t)}{\partial y^{\beta_2}} =f(x,y,t),
\end{aligned}\end{gathered}$$ with $1<\beta_1, \beta_2\leq2$, $\varepsilon_x,\varepsilon_y \geq 0$, $(x,y)\in\Omega$, $t>0$, and the initial and boundary conditions $$\begin{aligned}
& u(x,y,0)=\psi(x,y),\quad (x,y)\in\Omega, \label{xz02}\\
& u(x,y,t)=0, \ \ (x,y)\in\partial\Omega,\quad t>0, \label{xz03}\end{aligned}$$ where $\Omega$ and $\partial\Omega$ are given as above.
In Eqs. (\[eq01\]), (\[eq04\]), the time-fractional derivatives are defined in Caputo sense, i.e., $$\begin{aligned}
\frac{\partial^\alpha u(x,t)}{\partial t^\alpha}=\frac{1}{\Gamma(1-\alpha)}
\int^t_0\frac{\partial u(x,\xi)}{\partial \xi}\frac{d\xi}{(t-\xi)^\alpha},\end{aligned}$$ while in Eq. (\[xz01\]), the space-fractional derivatives are defined in Riemann-Liouville sense, i.e., $$\begin{aligned}
\frac{\partial^{\beta_1} u(x,y,t)}{\partial x^{\beta_1}}&=\frac{1}{\Gamma(2-\beta_1)}\frac{\partial^2 }{\partial x^2}
\int^x_a\frac{u(\xi,y,t)d\xi}{(x-\xi)^{\beta_1-1}},\\
\frac{\partial^{\beta_2} u(x,y,t)}{\partial y^{\beta_2}}&=\frac{1}{\Gamma(2-\beta_2)}\frac{\partial^2 }{\partial y^2}
\int^y_c\frac{u(x,\xi,t)d\xi}{(y-\xi)^{\beta_2-1}},\end{aligned}$$ and $\frac{\partial^\alpha u(x,y,t)}{\partial t^\alpha}$ is an analog of $\frac{\partial^\alpha u(x,t)}{\partial t^\alpha}$, where $\Gamma(\cdot)$ is the Gamma function. It is noted that Eqs. (\[eq01\])-(\[eq03\]), (\[eq04\])-(\[eq07\]), and (\[xz01\])-(\[xz03\]) reduce into the classical 1D or 2D advection-diffusion equations if $\alpha=1$, $\beta_1=\beta_2=2$ are fixed.
In recent decades, fractional ADEs have been notable subjects of intense research. Except for a few analytic solutions, various numerical methods have been done for Eqs. (\[eq01\])-(\[eq03\]) without advection, covering implicit difference method [@ref09], high-order finite element method (FEM) [@ref12], Legendre wavelets and spectral Galerkin methods [@08; @ref11], direct discontinuous Galerkin method [@ref13], quadratic spline collocation method [@ref22], cubic B-spline collocation method (CBCM) [@ref23], orthogonal spline collocation method [@13], pseudo-spectral method [@11], high-order compact difference method [@ref10], implicit radial basis function (RBF) meshless method [@12], nonpolynomial and polynomial spline methods [@14]. In [@ref21; @ref15; @ref17; @06], the algorithms based on shifted fractional Jacobi polynomials, Sinc functions and shifted Legendre polynomials, Haar wavelets and the third kind Chebyshev wavelets functions were well developed via the integral operational matrix or collocation strategy for Eqs. (\[eq01\])-(\[eq03\]) with variable coefficients. The Gegenbauer polynomial spectral collocation method was proposed in [@ref18] for the same type of equations and a Sinc-Haar collocation method can be found in [@ref24]. Uddin and Haq considered a radial basis interpolation approach [@ref16]. Cui established a high-order compact exponential difference scheme [@02]. Razminia et al. proposed a DQ method for time-fractional diffusion equations by using Lagrangian interpolation polynomials as test functions [@15]. Shirzadi et al. solved the 2D time-fractional ADEs with a reaction term via a local Petrov-Galerkin meshless method [@07]. Gao and Sun derived two different three-point combined compact alternating direction implicit (ADI) schemes for Eqs. (\[eq04\])-(\[eq07\]) [@16], both of which own high accuracy. High-dimensional space-fractional ADEs are challenging topics in whether analytic or numerical aspects due to the complexity and huge computing burden. The application of an numerical method to Eqs. (\[xz01\])-(\[xz03\]) did not have large diffusion; for the conventional algorithms, we refer the readers to [@wx02; @wx01; @wx03; @wx04; @wx05; @wx06] and references therein.
The trigonometric B-splines are a class of piecewise-defined functions constructed from algebraic trigonometric spaces, which have got recognition since 1964. They are preferred to the familiar polynomial B-splines since they often yield less errors when served as basis functions in interpolation theory. Nevertheless, using these basis splines to set up numerical algorithms is in its infancy and the related works are limited [@18; @17]. In this study, a DQ method for the general ADEs is developed with its weighted coefficients calculated based on cubic trigonometric B-spline (CTB) functions. The basis splines are slightly modified for brevity and a few advantages. Difference schemes and Runge-Kutta Gill’s method are introduced to discretize the resulting ODEs. The condition ensuring the stability of the time-stepping DQ method is discussed and found to be mild. Also, we propose a new cubic B-splines based DQ method for the 2D space-fractional diffusion equations by introducing the DQ approximations to fractional derivatives. The weights are determined by deriving explicit formulas for the fractional derivatives of B-splines via a recursive technique of partial integration. The approaches in presence are straight forward to apply and simple in implementation; numerical results highlight the superiority over some previous algorithms.
The remainder is organized as follows. In Section \[s1\], we outline some basic definitions and the cubic spline functions that are very useful hereinafter. In Section \[s2\], how to determine the weighted coefficients based on these CTB functions is studied and a time-stepping DQ method is constructed for Eqs. (\[eq01\])-(\[eq03\]) and Eqs. (\[eq04\])-(\[eq07\]). Section \[s3\] elaborates on its stable analysis. In Section \[s4\], we suggest a spline-based DQ method for Eqs. (\[xz01\])-(\[xz03\]) based on a set of modified cubic B-splines by explicitly computing the values of their fractional derivatives at sampling points. Some numerical examples are included in Section \[s5\], which manifest the effectiveness of our methods. The last section devotes to a conclusion.
Preliminaries {#s1}
=============
Let $M, N\in \mathbb{Z}^+$ and a time-space lattice be $$\begin{aligned}
\Omega_{\tau}&=\{t_n:t_n=n\tau,\ 0\leq n\leq N \},\\
\Omega_h&=\{x_i:x_i=a+ih,\ 0\leq i\leq M\},\end{aligned}$$ with $\tau=T/N $, $h=(b-a)/M$ on $(0,T]\times[a,b]$. Then, some auxiliary results are introduced for preliminaries.
Fractional derivatives and their discretizations
------------------------------------------------
Given a good enough $f(x,t)$, the formulas $$\begin{aligned}
&{^C_0}D^\alpha_tf(x,t)=\frac{1}{\Gamma(m-\alpha)}
\int^t_0\frac{\partial^m f(x,\xi)}{\partial \xi^m}\frac{d\xi}{(t-\xi)^{1+\alpha-m}}, \\% \label{eq09}\\
&{^{RL}_0}D^\alpha_tf(x,t)=\frac{1}{\Gamma(m-\alpha)}\frac{\partial^m}{\partial t^m}
\int^t_0\frac{f(x,\xi)d\xi}{(t-\xi)^{1+\alpha-m}}, %\label{eq10}\end{aligned}$$ define the $\alpha$-th Caputo and Riemann-Liouville derivatives, respectively, where $m-1<\alpha<m$, $m\in\mathbb{Z}^+$, and particularly, in the case of $\alpha=m$, both of them degenerate into the $m$-th integer-order derivative.
The two frequently-used fractional derivatives are equivalent with exactness to an additive factor, i.e., $$\label{eq11}
{^C_0}D^\alpha_tf(x,t)={^{RL}_0}D^\alpha_tf(x,t)-\sum^{m-1}_{l=0}\frac{f^{(l)}(x,0)t^{l-\alpha}}{\Gamma(l+1-\alpha)};$$ see [@23; @ref01] for references. Utilizing ${^{RL}_0}D^\alpha_t t^l=\frac{\Gamma(l+1)t^{l-\alpha}}{\Gamma(l+1-\alpha)}$ and a proper scheme to discretize the Riemann-Liouville derivatives on the right side of Eq. (\[eq11\]), a difference scheme for Caputo derivative can be $$\begin{aligned}
\label{eq12}
\begin{aligned}
{^C_0}D^\alpha_tf(x,t_n)&\cong \frac{1}{\tau^\alpha}\sum_{k=0}^{n}\omega^\alpha_kf(x,t_{n-k}) \\
&\ -\frac{1}{\tau^\alpha}\sum^{m-1}_{l=0}\sum^{n}_{k=0}\frac{\omega^\alpha_k f^{(l)}(x,0)t_{n-k}^{l}}{l!}.
\end{aligned}\end{aligned}$$ with several valid alternatives of the discrete coefficients $\{\omega^\alpha_k\}_{k=0}^{n}$ [@66]. Typically, we have $$\label{ez06}
\omega^\alpha_k=(-1)^k\binom\alpha k=\frac{\Gamma{(k-\alpha)}}{\Gamma{(-\alpha)}\Gamma{(k+1)}}, \ \ k\geq0,$$ whose truncated error is $\mathscr{R}_\tau=\mathscr{O}(\tau)$, and $$\label{eq08}
\omega^\alpha_k=\bigg(\frac{11}{6}\bigg)^\alpha\sum^k_{p=0}\sum^p_{q=0}\mu^q\overline{\mu}^{p-q}l^\alpha_ql^\alpha_{p-q}l^\alpha_{k-p},\ \ k\geq0,$$ with $\mu=\frac{4}{7+\sqrt{39}\textrm{i}}$, $\overline{\mu}=\frac{4}{7-\sqrt{39}\textrm{i}}$, $\textrm{i}=\sqrt{-1}$, and $$\label{ez08}
l^\alpha_0=1,\ \ l^\alpha_k=\bigg(1-\frac{\alpha+1}{k}\bigg)l^\alpha_{k-1},\ \ k\geq1,$$ in which case, the truncated error fulfills $\mathscr{R}_\tau=\mathscr{O}(\tau^3)$. Actually, Eq. (\[ez08\]) is the recursive relation of Eq. (\[ez06\]). Moreover, the coefficients $\{\omega^\alpha_k\}_{k=0}^{n}$ in Eq. (\[ez06\]) satisfy
- $\omega^\alpha_0=1, \quad \omega^\alpha_k< 0$, $\forall k\geq 1$,
- $\sum_{k=0}^{\infty}\omega^\alpha_k=0, \quad \sum_{k=0}^{n-1}\omega^\alpha_k>0$.
These properties are easily obtained from [@ref01].
Reset $0<\alpha<1$, (\[eq12\]) thus turns into $$\begin{aligned}
\label{eq15}
\begin{aligned}
{^C_0}D^\alpha_tf(x,t_n)&= \frac{1}{\tau^\alpha}\sum_{k=0}^{n-1}\omega^\alpha_kf(x,t_{n-k})\\
&\ -\frac{1}{\tau^\alpha}\sum_{k=0}^{n-1}\omega^\alpha_kf(x,0)+\mathscr{R}_\tau.
\end{aligned}\end{aligned}$$ It is noteworthy that Eq. (\[eq15\]) gives a smooth transition to the classic schemes when $\alpha=1$, for instance, Eq. (\[eq15\]) would be the four-point backward difference scheme if $\alpha=1$ and $\{\omega^\alpha_k\}_{k=0}^{n}$ are chosen to be the ones in Eq. (\[eq08\]), because these coefficients also fulfill $\sum_{k=0}^{\infty}\omega^\alpha_k=0$ and vanish apart from $\omega^\alpha_0$, $\omega^\alpha_1$, $\omega^\alpha_2$ and $\omega^\alpha_3$.
Cubic spline functions
----------------------
Let $x_{-i}=a-ih$, $x_{M+i}=b+ih$, $i=1,2,3$ be the six ghost knots outside $[a,b]$. Then the desirable CTB basis functions $\{CTB_m(x)\}_{m=-1}^{M+1}$ are defined as [@17; @21] $$\begin{aligned}
CT{B_m}(x) = \frac{1}{\chi }\left\{ \begin{array}{l}
\phi_1(x),\qquad x \in [{x_{m - 2}},{x_{m - 1}})\\
\phi_2(x),\qquad x \in [{x_{m - 1}},{x_m})\\
\phi_3(x), \qquad x \in [{x_m},{x_{m + 1}})\\
\phi_4(x),\qquad x \in [{x_{m + 1}},{x_{m + 2}})\\
0, \qquad\quad\quad {\rm{otherwise}}
\end{array} \right.\end{aligned}$$ where $$\begin{aligned}
&\phi_1(x)={p^3}({x_{m - 2}}), \\
&\phi_2(x)=q({x_{m + 2}}){p ^2}({x_{m-1}})+p^2({x_{m - 2}})q ({x_m})\\
&\qquad\quad\quad+p({x_{m-2}})p ({x_{m - 1}})q({x_{m + 1}}),\\
&\phi_3(x)=p({x_{m - 2}}){q^2}({x_{m + 1}})+q^2({x_{m + 2}})p ({x_m})\\
&\qquad\quad\quad+ p({x_{m - 1}})q({x_{m + 1}})q({x_{m + 2}}),\\
&\phi_4(x)={q ^3}({x_{m + 2}}),\end{aligned}$$ with the notations $$\begin{aligned}
&p({x_m})=\sin\bigg({\frac{{x-{x_m}}}{2}}\bigg),\\
&q({x_m})=\sin\bigg({\frac{{{x_m}-x}}{2}}\bigg),\\
&\chi =\sin \bigg({\frac{h}{2}}\bigg)\sin(h)\sin\bigg({\frac{{3h}}{2}}\bigg).\end{aligned}$$ The values of $CTB_m(x)$ at each knot are given by
$$\begin{aligned}
\label{eq16}
CT{B_m}(x_i)=\left\{
\begin{aligned}
&\sin^2\bigg(\frac{h}{2}\bigg)\csc(h)\csc\bigg(\frac{3h}{2}\bigg),\ i=m \pm 1 \\
&\frac{2}{1+2\cos(h)},\ \ i = m \\
&0,\ \ \rm{otherwise}
\end{aligned}
\right.\end{aligned}$$
and the values of $CTB'_m(x)$ at each knot are given by $$\begin{aligned}
\label{eq17}
CT{B'_m}(x_i)=\left\{
\begin{aligned}
&\frac{3}{4}\csc \bigg( {\frac{{3h}}{2}} \bigg),\ \ i=m - 1\\
&-\frac{3}{4}\csc \bigg( {\frac{{3h}}{2}} \bigg),\ \ i = m+1\\
&0.\ \ \rm{otherwise}
\end{aligned}
\right.\end{aligned}$$
Using the same grid information, the cubic B-spline basis functions $\{B_m(x)\}_{m=-1}^{M+1}$ are defined by $$\begin{aligned}
{B_m}(x) = \frac{1}{h^3}\left\{ \begin{array}{l}
\varphi_1(x),\qquad x \in [{x_{m - 2}},{x_{m - 1}})\\
\varphi_2(x),\qquad x \in [{x_{m - 1}},{x_m})\\
\varphi_3(x), \qquad x \in [{x_m},{x_{m + 1}})\\
\varphi_4(x),\qquad x \in [{x_{m + 1}},{x_{m + 2}})\\
0, \qquad\quad\quad {\rm{otherwise}}
\end{array} \right.\end{aligned}$$ with the piecewise functions $$\begin{aligned}
&\varphi_1(x)=(x-x_{m - 2})^3, \\
&\varphi_2(x)=(x-x_{m - 2})^3-4(x-x_{m - 1})^3,\\
&\varphi_3(x)=(x_{m + 2}-x)^3-4(x_{m + 1}-x)^3,\\
&\varphi_4(x)=(x_{m + 2}-x)^3.\end{aligned}$$
Both $\{CTB_m(x)\}_{m=-1}^{M+1}$ and $\{B_m(x)\}_{m=-1}^{M+1}$ are locally compact and twice continuously differentiable on $[a,b]$. Since the knots $x_{-1}$, $x_{M+1}$ lie beyond $[a,b]$ and the weights in relation to the B-splines at both ends do not participate in practical computation, hereunder, as in [@MJ] for cubic B-splines, we modify the CTBs by $$\begin{aligned}
\label{ref22}
\left\{\begin{aligned}
&MT{B_0}(x) = CT{B_0}(x) + 2CT{B_{ - 1}}(x),\\
&MT{B_1}(x) = CT{B_1}(x) - CT{B_{ - 1}}(x),\\
&MT{B_m}(x) = CT{B_m}(x),\;\;m = 2,3, \ldots ,M - 2,\\
&MT{B_{M - 1}}(x) = CT{B_{M - 1}}(x) - CT{B_{M + 1}}(x),\\
&MT{B_M}(x) = CT{B_M}(x) + 2CT{B_{M + 1}}(x),
\end{aligned} \right.\end{aligned}$$ for simplicity, which will result in a strictly tri-diagonal algebraic system after discretization on the uniform grid $\Omega_h$. $\{MTB_m(x)\}_{m=0}^{M}$ are also linearly independent and constitute a family of basis elements of a spline space.
Description of CTB based DQ method {#s2}
==================================
On a 2D domain $[a,b]\times[c,d]$, letting $M_x, M_y\in \mathbb{Z}^+$, we add a spatial lattice with equally spaced grid points with spacing of $h_x=(b-a)/M_x$ in $x$-axis and $h_y=(d-c)/M_y$ in $y$-axis, i.e., $$\begin{aligned}
\Omega_x&=\{x_i:x_i=a+ih_x,\ 0\leq i\leq M_x\},\\
\Omega_y&=\{y_j:y_j=c+jh_y,\ 0\leq j\leq M_y\}.\end{aligned}$$ DQ method is understood as a numerical technique for finding the approximate solutions of differential equations that reduces the original problem to those of solving a system of algebraic or ordinary differential equations via replacing the spatial partial derivatives by the representative weighted combinations of the functional values at certain grid points on the whole domain [@26]. The key procedure of such method lies in the determination of its weights and the selection of the test functions whose derivative values are explicit at the prescribed discrete grid points. As requested, we let $$\begin{aligned}
\label{eq18}
\frac{\partial ^su(x_i,t)}{\partial x^s}
\cong\sum\limits_{j=0}^M {a_{ij}^{(s)}u(x_j,t)},\ \ 0\leq i\leq M,\end{aligned}$$ while for 2D problems, we let $$\begin{aligned}
&\frac{\partial ^su(x_i,y_j,t)}{\partial x^s}
\cong\sum\limits_{m=0}^{M_x} {a_{im}^{(s)}u(x_m,y_j,t)}, \label{eq19}\\
&\frac{\partial ^su(x_i,y_j,t)}{\partial y^s}
\cong\sum\limits_{m=0}^{M_y} {b_{jm}^{(s)}u(x_i,y_m,t)},\label{eq20}\end{aligned}$$ where $s\in\mathbb{Z}^+$, $0\leq i\leq M_x$, $0\leq j\leq M_y$ and $a_{ij}^{(s)}$, $a_{im}^{(s)}$, $b_{jm}^{(s)}$ are the weighted coefficients allowing us to approximate the $s$-th derivatives or partial derivatives at the given grid points in the DQ methods.
The calculation of weighted coefficients
----------------------------------------
In the sequel, we apply $\{MTB_m(x)\}_{m=0}^{M}$ to calculate the 1D, 2D unknown weights. Putting $s=1$ and substituting these basis splines into Eq. (\[eq18\]), we get $$\begin{aligned}
\frac{\partial MT{B_m}(x_i)}{\partial x}=\sum^M_{j=0}a_{ij}^{(1)}MT{B_m}(x_j),\ \ 0\leq i,m\leq M,\end{aligned}$$ with the weighted coefficients of the first-order derivative $a_{ij}^{(1)}$, $0\leq i,j\leq M$, yet to be determined. In view of (\[ref22\]) and the properties (\[eq16\])-(\[eq17\]), some manipulations on the above equations yield the matrix-vector forms $$\begin{aligned}
\label{eq21}
\left\{\begin{aligned}
&\textbf{A}\textbf{a}^{(1)}_0=\textbf{Z}_0,\\
&\textbf{A}\textbf{a}^{(1)}_1=\textbf{Z}_1,\\
&\qquad\ \vdots \\
&\textbf{A}\textbf{a}^{(1)}_M=\textbf{Z}_M,\\
\end{aligned} \right.\end{aligned}$$ where $\textbf{A}$ is the $(M+1)\times(M+1)$ coefficient matrix $$\begin{aligned}
\textbf{A}=\left( \begin{array}{cccccc}
A_0+2A_1& A_1 & & & &\\
0& A_0 & A_1 & & & \\
& A_1 & A_0 & A_1 & & \\
& &\ddots&\ddots&\ddots& \\
& & & A_1 & A_0 & 0\\
& & & & A_1 & A_0+2A_1
\end{array} \right),\end{aligned}$$ $$\begin{aligned}
A_0=\frac{2}{1+2\cos(h)},\ \
A_1=\sin^2\bigg(\frac{h}{2}\bigg)\csc(h)\csc\bigg(\frac{3h}{2}\bigg),\end{aligned}$$ $\textbf{a}^{(1)}_k$, $0\leq k\leq M$, are the weighted coefficient vectors at $x_k$, i.e., $\textbf{a}^{(1)}_k=[a_{k0}^{(1)},a_{k1}^{(1)},\ldots,a_{kM}^{(1)}]^{\rm{T}}$, and the right-side vectors $\textbf{Z}_k$ at $x_k$, $0\leq k\leq M$, are as follows $$\begin{aligned}
&\textbf{Z}_0 = \left(\begin{array}{c}
-2z\\ 2z\\ 0\\ 0\\ \vdots \\ 0\\ 0
\end{array} \right),\ \
\textbf{Z}_1=\left(\begin{array}{c}
-z\\ 0\\ z\\ 0\\ \vdots \\ 0\\ 0
\end{array}\right), \cdots, \\
&\textbf{Z}_{M-1}=\left( \begin{array}{c}
0\\ 0\\ \vdots \\ 0\\ -z\\ 0\\ z
\end{array}\right),\ \
\textbf{Z}_M=\left( \begin{array}{c}
0\\ 0\\ \vdots \\ 0\\ 0\\ -2z\\ 2z
\end{array}\right),\end{aligned}$$ with $z=\frac{3}{4}\csc\left(\frac{3h}{2}\right)$, respectively. Thus, $a_{ij}^{(1)}$ are obtained by solving Eqs. (\[eq21\]) for each point $x_i$. There are two different way to derive the weighted coefficients $a_{ij}^{(2)}$ of the second-order derivative: (i) do a similar fashion as above by putting $s=2$ in Eq. (\[eq18\]) and solve an algebraic system for each grid point; (ii) find the weighted coefficients $a_{ij}^{(s)}$, $s\geq2$, corresponding to the high-order derivatives in a recursive style [@27], i.e., $$\begin{aligned}
&a_{ij}^{(s)}=s\Bigg( {a_{ii}^{(s-1)}a_{ij}^{(1)}-\frac{a_{ij}^{(s-1)}}{x_i-x_j}}\Bigg),\ \ i\ne j,\ 0\leq i\leq M,\\
&a_{ii}^{(s)}=-\sum\limits_{j=0,j\ne i}^M {a_{ij}^{(s)}} ,\ \ i = j,\end{aligned}$$ which includes $s=2$ as a special case. The former would be less efficient since the associated equations have to be solved as priority, so the latter one will be selected during our entire computing process. Proceeding as before via replacing $\Omega_h$ by $\Omega_x$, $\Omega_y$ leads to a 2D generalization to get $a^{(1)}_{im}$, $b^{(1)}_{jm}$ of the first-order partial derivatives with regard to variables $x$, $y$ in Eqs. (\[eq19\])-(\[eq20\]) and by them, the following relationships can further be applied, i.e., $$\begin{aligned}
&a_{im}^{(s)}=s\Bigg( {a_{ii}^{(s-1)}a_{im}^{(1)}-\frac{a_{im}^{(s-1)}}{x_i-x_m}}\Bigg),\ \ i\ne m,\ 0\leq i\leq M,\\
&a_{ii}^{(s)}=-\sum\limits_{m=0,m\ne i}^{M_x}{a_{im}^{(s)}} ,\ \ i = m,\\
&b_{jm}^{(s)}=s\Bigg( {b_{jj}^{(s-1)}b_{jm}^{(1)}-\frac{b_{jm}^{(s-1)}}{y_j-y_m}}\Bigg),\ \ j\ne m,\ 0\leq j\leq M,\\
&b_{jj}^{(s)}=-\sum\limits_{m=0,m\ne j}^{M_y}{b_{jm}^{(s)}} ,\ \ j = m,\end{aligned}$$ to calculate $a^{(s)}_{im}$, $b^{(s)}_{jm}$ with $s\geq2$.
A point worth noticing is that $A_0,\ A_1>0$, when $0<h<1$, $0<h_x,\ h_y<1$. Since $A_0,\ A_1$ can be deemed to be the functions of $h$, we obtain their derivatives $$\begin{aligned}
A'_0&=\frac{4\sin(h)}{(1+2\cos(h))^2},\\
A'_1&=\frac{\sec(\frac{h}{2})\tan(\frac{h}{2})(5+6\cos(h))}{4(1+2\cos(h))^2}.\end{aligned}$$ On letting $0<h<1$, both are proved to be larger than zero, i.e., $A_0,\ A_1$ are the increasing functions with respect to $h$. On the other hand, there exist $A_0(0)=0.6667$, $A_1(1)=0.2738$. Then, it suffices to show $$\frac{2}{1+2\cos(h)}>2\sin^2\bigg(\frac{h}{2}\bigg)\csc(h)\csc\bigg(\frac{3h}{2}\bigg),$$ which implies $A_0>2A_1$, and thus $\textbf{A}$ is a strictly diagonally dominant tri-diagonal matrix. Hence, Thomas algorithm can be applied to tackle the algebraic equations as Eqs. (\[eq21\]), which simply requires the arithmetic operation cost $\mathscr{O}(M+1)$ and would greatly economize on the memory and computing time in practice.
Construction of CTB based DQ method
-----------------------------------
In this subpart, a DQ method based on $\{MTB_m(x)\}_{m=0}^{M}$ (MCTB-DQM) is constructed for Eqs. (\[eq01\])-(\[eq03\]) and Eqs. (\[eq04\])-(\[eq07\]). Let $s=1$, $2$. The substitution of the weighted sums (\[eq18\]), (\[eq19\])-(\[eq20\]) into the main equations gives
$$\begin{aligned}
\frac{\partial^\alpha u(x_i,t)}{\partial t^\alpha}+\kappa\sum\limits_{j=0}^M {a_{ij}^{(1)}u(x_j,t)}
-\varepsilon\sum\limits_{j=0}^M {a_{ij}^{(2)}u(x_j,t)}=f(x_i,t),\end{aligned}$$
with $i=0,1,\cdots,M$, and
$$\begin{aligned}
&\frac{\partial^\alpha u(x_i,y_j,t)}{\partial t^\alpha}\!+\!\kappa_x\sum\limits_{m=0}^{M_x}{a_{im}^{(1)}u(x_m,y_j,t)}
\!+\!\kappa_y\sum\limits_{m=0}^{M_y}{b_{jm}^{(1)}u(x_i,y_m,t)}\\
&\!-\!\varepsilon_x\sum\limits_{m=0}^{M_x}{a_{im}^{(2)}u(x_m,y_j,t)}
\!-\!\varepsilon_y\sum\limits_{m=0}^{M_y}{b_{jm}^{(2)}u(x_i,y_m,t)}\!=\!f(x_i,y_j,t),\end{aligned}$$
with $i=0,1,\cdots,M_x$, $j=0,1,\cdots,M_y$, which are indeed a group of $\alpha$-th ODEs associated with the boundary constraints (\[eq03\]), (\[eq07\]), and involve $\alpha\in(0,1)$ and $\alpha=1$ as two separate cases. In what follows, we employ $$\begin{aligned}
&u_i^n=u(x_i,t_n),\quad u_{ij}^n=u(x_i,y_j,t_n),\\
&f^n_{i}=f(x_i,t_n),\quad f^n_{ij}=f(x_i,y_j,t_n),\\
&g_1^n=g_1(t_n),\quad g_2^n=g_2(t_n),\quad g_{ij}^n=g(x_i,y_j,t_n),\end{aligned}$$ for the ease of exposition, where $n=0,1,\cdots,N$.
### The case of fractional order
Discretizing the ODEs above by the difference scheme (\[eq15\]) and imposing boundary constraints, we have $$\begin{aligned}
\left\{\begin{aligned}\label{eq22}
&\omega^\alpha_0U_i^n+\kappa\tau^\alpha\sum\limits_{j=1}^{M-1}a_{ij}^{(1)}U_j^n
-\varepsilon\tau^\alpha\sum_{j=1}^{M-1}a_{ij}^{(2)}U_j^n\\
&=-\sum_{k=1}^{n-1}\omega^\alpha_kU_i^{n-k}\!+\!\sum_{k=0}^{n-1}\omega^\alpha_kU_i^0+ \tau^\alpha G^n_i,
\end{aligned}\right.\end{aligned}$$ with $i=1,2,\cdots,M-1$ and $$G^n_i=f^n_i-\kappa\big(a_{i0}^{(1)}g^n_1+a_{iM}^{(1)}g^n_2\big)+\varepsilon\big(a_{i0}^{(2)}g^n_1+a_{iM}^{(2)}g^n_2\big),$$ for Eqs. (\[eq01\])-(\[eq03\]), and the following scheme
$$\begin{aligned}
\left\{\begin{aligned}\label{eq23}
&\omega^\alpha_0U_{ij}^n+\kappa_x\tau^\alpha\sum_{m=1}^{M_x-1}a_{im}^{(1)}U_{mj}^n+\kappa_y\tau^\alpha\sum_{m=1}^{M_y-1}b_{jm}^{(1)}U_{im}^n\\
&\ -\varepsilon_x\tau^\alpha\sum_{m=1}^{M_x-1}a_{im}^{(2)}U_{mj}^n-\varepsilon_y\tau^\alpha\sum_{m=1}^{M_y-1}b_{jm}^{(2)}U_{im}^n\\
&\ =-\sum_{k=1}^{n-1}\omega^\alpha_kU_{ij}^{n-k}+\sum_{k=0}^{n-1}\omega^\alpha_kU_{ij}^0+ \tau^\alpha G^n_{ij},
\end{aligned}\right.\end{aligned}$$
with $i=1,2,\cdots,M_x-1$, $j=1,2,\cdots,M_y-1$, and
$$\begin{aligned}
G^n_{ij}\!=&f^n_{ij}\!-\!\kappa_x\big(a_{i0}^{(1)}g^n_{0j}+a_{iM_x}^{(1)}g^n_{M_xj}\big)
\!-\!\kappa_y\big(b_{j0}^{(1)}g^n_{i0}+b_{jM_y}^{(1)}g^n_{iM_y}\big)\\
&+\varepsilon_x\big(a_{i0}^{(2)}g^n_{0j}+a_{iM_x}^{(2)}g^n_{M_xj}\big)
+\varepsilon_y\big(b_{j0}^{(2)}g^n_{i0}+b_{jM_y}^{(2)}g^n_{iM_y}\big),\end{aligned}$$
for Eqs. (\[eq04\])-(\[eq07\]). The Eqs. (\[eq22\])-(\[eq23\]) can further be rewritten in matrix-vector forms, for instance, letting $$\begin{aligned}
\textbf{U}^n&=[U^n_{11},\ldots,U^n_{M_{x}-1,1},U^n_{12},\ldots,U^n_{M_{x}-1,M_{y}-1}]^T,\\
\textbf{G}^n&=[G^n_{11},\ldots,G^n_{M_{x}-1,1},G^n_{12},\ldots,G^n_{M_{x}-1,M_{y}-1}]^T,\end{aligned}$$ for Eqs. (\[eq23\]), we have $$\label{ez81}
\omega^\alpha_0\textbf{U}^n+\tau^\alpha\textbf{K}\textbf{U}^n=-\sum_{k=1}^{n-1}\omega^\alpha_k\textbf{U}^{n-k}
+\sum_{k=0}^{n-1}\omega^\alpha_k\textbf{U}^0+\tau^\alpha\textbf{G}^n,$$ where $$\textbf{K}=\kappa_x\textbf{I}_y\otimes \textbf{W}^{1}_x+\kappa_y \textbf{W}^{1}_y\otimes\textbf{I}_x
-\varepsilon_x\textbf{I}_y\otimes \textbf{W}^{2}_x-\varepsilon_y \textbf{W}^{2}_y\otimes\textbf{I}_x,$$ with $\textbf{I}_x$, $\textbf{I}_y$ being the identity matrices in $x$- and $y$-axis, “$\otimes$” being Kronecker product, and $$\begin{aligned}
\textbf{W}^{c}_{z}
=\left( \begin{array}{llll}
w^{(c)}_{11}&w^{(c)}_{12} &\cdots &w^{(c)}_{1,M_z-1}\\
w^{(c)}_{21}&w^{(c)}_{22} &\cdots &w^{(c)}_{2,M_z-1} \\
\vdots &\vdots &\ddots&\vdots \\
w^{(c)}_{M_z-1,1}&w^{(c)}_{M_z-1,2}&\cdots &w^{(c)}_{M_z-1,M_z-1}
\end{array} \right), \ c=1,2,\end{aligned}$$ in which, $z=x$ if $w=a$ while $z=y$ if $w=b$. The initial states are got from Eqs. (\[eq02\]), (\[eq05\]). As a result, the approximate solutions are obtained via performing the iteration in Eqs. (\[eq22\])-(\[eq23\]) until the last time level by rewriting them in matrix-vector forms first.
### The case of integer order
When $\alpha=1$, despite $\omega^1_0=1.8333$, $\omega^1_1=-3$, $\omega^1_2=1.5$, $\omega^1_3=-0.3333$, being the coefficients of the four-point backward difference scheme, the initial values with the errors of the same convergent rate are generally necessary to start Eqs. (\[eq22\])-(\[eq23\]). However, this situation would not happen if $\{\omega_k^\alpha\}_{k=0}^n$ in Eq. (\[ez06\]) are applied. In such a case, to make the algorithm to be more cost-effective, we use Runge-Kutta Gill’s method to handle those ODEs instead, which is explicit and fourth-order convergent. Rearrange the ODEs in a unified form $$\label{ez09}
\frac{\partial\textbf{u}}{\partial t}=\textbf{F}(\textbf{u}),$$ then the DQ method is constructed as follow
$$\begin{aligned}
\label{eq26}
&\textbf{U}^{n}=\textbf{U}^{n-1}+\frac{1}{6}\Big[K_1+(2-\sqrt{2})K_2+(2+\sqrt{2})K_3+K_4\Big],\nonumber\\
&K_1=\tau \textbf{F}\big(t_{n-1},\textbf{U}^{n-1}\big),\nonumber\\
&K_2=\tau \textbf{F}\bigg(t_{n-1}+\frac{\tau}{2},\textbf{U}^{n-1}+\frac{K_1}{2}\bigg),\\
&K_3=\tau \textbf{F}\bigg(t_{n-1}+\frac{\tau}{2},\textbf{U}^{n-1}+\frac{\sqrt{2}-1}{2}K_1+\frac{2-\sqrt{2}}{2}K_2\bigg),\nonumber\\
&K_4=\tau \textbf{F}\bigg(t_{n-1}+\tau,\textbf{U}^{n-1}-\frac{\sqrt{2}}{2}K_2+\frac{2+\sqrt{2}}{2}K_3\bigg),\nonumber\end{aligned}$$
where $\textbf{u}$, $\textbf{U}^{n}$, $n=1,2,\ldots,N$, are the unknown vectors and $\textbf{F}(\cdot)$ stands for the matrix-vector system corresponding to the weighted sums in ODEs and contains $a^{(s)}_{ij}$ or $a^{(s)}_{im}$, $b^{(s)}_{jm}$, $s=1$, $2$, as its elements. Meanwhile, the boundary constraints (\[eq03\]), (\[eq07\]) must be imposed on $\textbf{F}(\cdot)$ in the way as they are done for the fractional cases before we can fully run the procedures for Eqs. (\[eq26\]).
![The eigenvalues of the weighted matrices generated by DQ method when $a=c=0$, $b=d=2$: (a) $\textbf{W}^1_z$; (b) $\textbf{W}^2_z$.[]{data-label="zfig0"}](0_a.eps "fig:"){width="3.3in"}\
![The eigenvalues of the weighted matrices generated by DQ method when $a=c=0$, $b=d=2$: (a) $\textbf{W}^1_z$; (b) $\textbf{W}^2_z$.[]{data-label="zfig0"}](0_b.eps "fig:"){width="3.3in"}
Stability analysis {#s3}
==================
This part makes a attempt to study the matrix stability of Eqs. (\[ez09\]) and the numerical stability of Eqs. (\[eq22\])-(\[eq23\]). When $\alpha=1$, we rewrite Eqs. (\[ez09\]) by $$\label{ez12}
\frac{\partial\textbf{u}}{\partial t}=-\textbf{K}\textbf{u} + \textbf{Q},$$ where $\textbf{Q}$ is a vector containing the non-homogenous part and the boundary conditions, and $\textbf{K}$ is the weighted matrix mentioned before. We discuss the homogeneous case. The numerical stability of an algorithm for the ODEs generated by a DQ method relies on the stability of the ODEs themselves. Only when their solutions are stable can a well-known method such as Runge-Kutta Gill’s method yield convergent solutions. It is enough to show their stability that the real parts of the eigenvalues of $-\textbf{K}$ are all negative. Denote the row vector consisting of the eigenvalues of $\textbf{W}^{c}_z$ by $\boldsymbol{\lambda}_z^c$, with $z=x,y$ and $c=1,2$. In view of the properties of Kronecker product, the eigenvalues of $\textbf{W}^{c}_y\otimes\textbf{I}_x$, $\textbf{I}_y\otimes \textbf{W}^{c}_x$ are $\boldsymbol{\lambda}^{c}_y\otimes\textbf{e}_x$ and $\textbf{e}_y\otimes\boldsymbol{\lambda}^{c}_x$ (see [@xz99]), respectively, therefore, we have the eigenvalues of $-\textbf{K}$ in Eq. (\[ez12\]), i.e., $$\boldsymbol{\lambda}=-\kappa_x\textbf{e}_y\otimes\boldsymbol{\lambda}^{1}_x-\kappa_y \boldsymbol{\lambda}^{1}_y\otimes\textbf{e}_x
+\varepsilon_x\textbf{e}_y\otimes\boldsymbol{\lambda}^{2}_x+\varepsilon_y \boldsymbol{\lambda}^{2}_y\otimes\textbf{e}_x,$$ where $\textbf{e}_x$, $\textbf{e}_y$ are the row vectors of sizes $M_x+1$ and $M_y+1$, respectively, with all of their components being 1. The exact solution of ODEs is related to $\textrm{Re}\{\boldsymbol{\lambda}\}$ and the condition $\textrm{Re}\{\boldsymbol{\lambda}\}\leq 0$ is easy to meet because $\boldsymbol{\lambda}_z^2$ are always verified to be real and negative while $\boldsymbol{\lambda}_z^1$ be complex with their real parts being very close to zero; see Fig. \[zfig0\] for example. More than that, we notice that the foregoing analysis is also valid for the 1D cases and the phenomena appearing in Fig. \[zfig0\] would be enhanced as the grid numbers increase. Hence, we come to a conclusion that the ODEs are stable in most cases.
The discussion about the numerical stability of a fully discrete DQ method is difficulty and still sparse [@TS0; @TS1]. In the sequel, we show the conditionally stable nature of Eqs. (\[eq22\])-(\[eq23\]) in the context of $L_2$-norm $||\cdot||$ and the analysis is not just applicable to the fractional case. Without loss of generality, consider the 2D cases and the discrete coefficients $\{\omega^\alpha_k\}_{k=0}^{n}$ in Eq. (\[ez06\]). Let $\tilde{\textbf{U}}^0$ be the approximation of initial values $\textbf{U}^0$. Then
$$\label{ez80}
\tilde{\textbf{U}}^n+\tau^\alpha\textbf{K}\tilde{\textbf{U}}^n=-\sum_{k=1}^{n-1}\omega^\alpha_k\tilde{\textbf{U}}^{n-k}
+\sum_{k=0}^{n-1}\omega^\alpha_k\tilde{\textbf{U}}^0+\tau^\alpha\textbf{G}^n. %\omega^\alpha_0$$
On subtracting Eq. (\[ez80\]) from Eq. (\[ez81\]) and letting $\textbf{e}^n=\textbf{U}^n-\tilde{\textbf{U}}^n$, we have the perturbation equation
$$\label{ez82}
\textbf{e}^n=-\sum_{k=1}^{n-1}\omega^\alpha_k(\textbf{I}+\tau^\alpha\textbf{K})^{-1}\textbf{e}^{n-k}
+\sum_{k=0}^{n-1}\omega^\alpha_k(\textbf{I}+\tau^\alpha\textbf{K})^{-1}\textbf{e}^0,$$
where $\textbf{I}$ is the identity matrix in the same size of **K**. To prove $||\textbf{e}^n||\leq||\textbf{e}^0||$, we make the assumption $$\label{ez83}
||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||\leq 1.$$ When $n=1$, by taking $||\cdot||$ on both sides of Eq. (\[ez82\]), $||\textbf{e}^1||\leq ||\textbf{e}^0||$ is trivial due to $\omega^\alpha_0=1$. Let $$||\textbf{e}^m||\leq ||\textbf{e}^0||, \quad m=1,2,\ldots,n-1.$$ Using mathematical induction, it thus follows from the properties of $\{\omega^\alpha_k\}_{k=0}^{n}$ stated in Section \[s1\] that
$$\begin{aligned}
||\textbf{e}^n||&\!=\!\Bigg|\Bigg|\!-\!\sum_{k=1}^{n-1}\omega^\alpha_k(\textbf{I}\!+\!\tau^\alpha\textbf{K})^{-1}\textbf{e}^{n-k}
+\sum_{k=0}^{n-1}\omega^\alpha_k(\textbf{I}\!+\!\tau^\alpha\textbf{K})^{-1}\textbf{e}^0\Bigg|\Bigg| \\
&\!\leq\!\Bigg(1\!-\!\sum_{k=0}^{n-1}\omega^\alpha_k\!+\!\sum_{k=0}^{n-1}\omega^\alpha_k\Bigg)
||(\textbf{I}\!+\!\tau^\alpha\textbf{K})^{-1}||\max\limits_{0\leq m\leq n-1}||\textbf{e}^m|| \\
&=||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||\max\limits_{0\leq m\leq n-1}||\textbf{e}^m|| \leq ||\textbf{e}^0||.\end{aligned}$$
Hereinafter, we proceed with a full numerical investigation on the assumption (\[ez83\]) to explore the potential factors which may lead to $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||> 1$. At first, if $\tau^\alpha$ continuously varies from 1 to 0, there holds $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||\rightarrow 1$. However, this process can affect the maximal ratio of the coefficients of advection and diffusivity to keep (\[ez83\]); we leave this case to the end of the discussion. To be more representative, we take $\tau=1.0\times10^{-3}$, $\alpha=0.5$, and $M_x=M_y=5$, unless otherwise stated. The main procedures are divided into three steps: (i) fixing $\varepsilon_x,\varepsilon_y$, and $\Omega$, let $\kappa_x,\kappa_y$ vary and the values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ as the function of $\kappa_x,\kappa_y$ are plotted in (a), (b) of Fig. \[zfig1\]; (ii) fixing $\kappa_x,\kappa_y$, and $\Omega$, let $M_x$, $M_y$ vary and the results are plotted in (c) of Fig. \[zfig1\], where $\kappa_x=\kappa_y=500$; (iii) fixing $\kappa_x,\kappa_y,\varepsilon_x,\varepsilon_y$, let $a=c=0$ and $b$, $d$ vary, and the corresponding results are presented in (d) of Fig. \[zfig1\]. It is worthy to note that $\Omega$ is the unit square except the case of (iii), and the parameters of the same types in $x$- and $y$- axis are used as the same, for example, $\varepsilon_x=\varepsilon_y$. Now, we consider the influence brought by $\tau$. Resetting $\tau=1.0\times10^{-10}$, let $\varepsilon_x=\varepsilon_y=1$ and $\kappa_x, \kappa_y$ vary. The behavior of objective quantity is plotted in subfigure (e), from which, we see that the critical ratio between $\kappa_x, \kappa_y$ and $\varepsilon_x, \varepsilon_y$ to maintain (\[ez83\]) is about 40, far less than the case of (i), and can further be improved by increasing $M_x, M_y$.
![The values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ versus the variation of various factors: $\kappa_x,\kappa_y$, $\varepsilon_x,\varepsilon_y$, $M_x, M_y$, and $b, d$.[]{data-label="zfig1"}](1_a.eps "fig:"){width="3.3in"}\
![The values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ versus the variation of various factors: $\kappa_x,\kappa_y$, $\varepsilon_x,\varepsilon_y$, $M_x, M_y$, and $b, d$.[]{data-label="zfig1"}](1_b.eps "fig:"){width="3.3in"}\
![The values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ versus the variation of various factors: $\kappa_x,\kappa_y$, $\varepsilon_x,\varepsilon_y$, $M_x, M_y$, and $b, d$.[]{data-label="zfig1"}](1_c.eps "fig:"){width="3.3in"}\
![The values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ versus the variation of various factors: $\kappa_x,\kappa_y$, $\varepsilon_x,\varepsilon_y$, $M_x, M_y$, and $b, d$.[]{data-label="zfig1"}](1_d.eps "fig:"){width="3.3in"}\
![The values of $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ versus the variation of various factors: $\kappa_x,\kappa_y$, $\varepsilon_x,\varepsilon_y$, $M_x, M_y$, and $b, d$.[]{data-label="zfig1"}](1_e.eps "fig:"){width="3.3in"}
From the foregoing discussion and figures, we summarize the conclusions as follows: (i) if $\varepsilon_x$, $\varepsilon_y$ are not small, the tolerant ranges of $\kappa_x$, $\kappa_y$ to guarantee (\[ez83\]) is quite loose and when $\varepsilon_x$, $\varepsilon_y\rightarrow\infty$, $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ can be very close to zero; (ii) if $\kappa_x$, $\kappa_y$ are larger than $\varepsilon_x$, $\varepsilon_y$ and $\varepsilon_x$, $\varepsilon_y$ themselves are small, $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ can be larger than 1, however, such issue can be remedied by increasing the grid numbers; (iii) in general, the larger $M_x$, $M_y$, the smaller $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$; (iv) when the computational domain expands, $||(\textbf{I}+\tau^\alpha\textbf{K})^{-1}||$ grows at a speed, which may result in the invalidation of (\[ez83\]) if $\varepsilon_x$, $\varepsilon_y$ and $M_x, M_y$ remain unchanged; (v) when $\tau\rightarrow0$, the critical ratio between $\kappa_x, \kappa_y$ and $\varepsilon_x, \varepsilon_y$ to maintain this assumption appears to decrease, but it would be enhanced as the spatial grid is refined.
Consequently, the assumption is meaningful and essentially a mild theoretical restriction in practise.
Description of cubic B-spline DQ method {#s4}
=======================================
In this section, a robust DQ method (MCB-DQM) based on the modified cubic B-splines $\{MB_m(x)\}_{m=0}^{M}$ is established for Eqs. (\[xz01\])-(\[xz03\]) by introducing the DQ approximations to fractional derivatives. In the light of the essence of traditional DQ methods, we consider $$\begin{aligned}
&\frac{\partial ^{\beta_1}u(x_i,y_j,t)}{\partial x^{\beta_1}}\cong\sum\limits_{m=0}^{M_x} {a_{im}^{(\beta_1)}u(x_m,y_j,t)}, \label{xz07}\\
&\frac{\partial ^{\beta_2}u(x_i,y_j,t)}{\partial y^{\beta_2}}\cong\sum\limits_{m=0}^{M_y} {b_{jm}^{(\beta_2)}u(x_i,y_m,t)},\label{xz08}\end{aligned}$$ for fractional derivatives in constructing our DQ algorithm, where $0\leq i\leq M_x$, $0\leq j\leq M_y$ and the weighted coefficients $a_{im}^{(\beta_1)}$, $b_{jm}^{(\beta_2)}$ satisfy
$$\begin{aligned}
\frac{\partial^{\beta_1} M{B_k}(x_i)}{\partial x^{\beta_1}}=\sum^{M_x}_{m=0}a_{im}^{(\beta_1)}M{B_k}(x_m),\ \ 0\leq i,k\leq M_x,\label{xz17}\\
\frac{\partial^{\beta_2} M{B_k}(y_j)}{\partial y^{\beta_2}}=\sum^{M_y}_{m=0}b_{jm}^{(\beta_2)}M{B_k}(y_m),\ \ 0\leq j,k\leq M_y.\label{xz18}\end{aligned}$$
The validation of Eqs. (\[xz07\])-(\[xz08\]) is ensured by the linear properties of fractional derivatives. $a_{im}^{(\beta_1)}$, $b_{jm}^{(\beta_2)}$ are then determined by tackling the resulting algebraic problems from the above equations for each axis if the values of $\{MB_m(x)\}_{m=0}^{M}$ at all sampling points are known.
The explicit formulas of fractional derivatives
-----------------------------------------------
It is the weakly singular integral structure that makes it difficult to calculate the values of the fractional derivatives for a function as B-spline at a sampling point. In the text that follows, we concentrate on the explicit expressions of the $\beta$-th ($1<\beta\leq2$) Riemann-Liouville derivative of $\{B_m(x)\}_{m=-1}^{M+1}$ with a recursive technique. Since these basis splines are piecewise and locally compact on four consecutive subintervals, we have $$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_m(x)\!=\!\left\{ \begin{array}{l}
0,\qquad\qquad\qquad\ \, x \in [x_0,x_{m-2})\\
{^{RL}_{x_{m-2}}}D^\beta_x\varphi_1(x), \qquad x \in [{x_{m - 2}},{x_{m - 1}})\\
{^{RL}_{x_{m-2}}}D^\beta_{x_{m-1}}\varphi_1(x)\\
\quad +{^{RL}_{x_{m-1}}}D^\beta_x\varphi_2(x), \ \, x \in [{x_{m - 1}},{x_m})\\
{^{RL}_{x_{m-2}}}D^\beta_{x_{m-1}}\varphi_1(x)\\
\quad +{^{RL}_{x_{m-1}}}D^\beta_{x_m}\varphi_2(x)\\
\quad +{^{RL}_{x_m}}D^\beta_{x}\varphi_3(x), \quad \ x \in [{x_m},{x_{m + 1}})\\
{^{RL}_{x_{m-2}}}D^\beta_{x_{m-1}}\varphi_1(x)\\
\quad +{^{RL}_{x_{m-1}}}D^\beta_{x_m}\varphi_2(x)\\
\quad +{^{RL}_{x_m}}D^\beta_{x_{m+1}}\varphi_3(x)\\
\quad +{^{RL}_{x_{m+1}}}D^\beta_{x}\varphi_4(x), \ \ x \in [{x_{m + 1}},{x_{m + 2}})\\
{^{RL}_{x_{m-2}}}D^\beta_{x_{m-1}}\varphi_1(x)\\
\quad +{^{RL}_{x_{m-1}}}D^\beta_{x_m}\varphi_2(x)\\
\quad +{^{RL}_{x_m}}D^\beta_{x_{m+1}}\varphi_3(x)\\
\quad +{^{RL}_{x_{m+1}}}D^\beta_{x_{m+2}}\varphi_4(x), \ \, x \in [{x_{m + 2}},{x_M}]
\end{array} \right.\end{aligned}$$ with $2\leq m\leq M-2$. The compact supports of $B_{M-1}(x)$, $B_{M}(x)$, and $B_{M+1}(x)$ partially locate on the outside of $[x_0,x_M]$, so do $B_{-1}(x)$, $B_{0}(x)$, and $B_{1}(x)$; nevertheless, $B_{M-1}(x)$, $B_{M}(x)$, and $B_{M+1}(x)$ can be thought of as the special cases of the aforementioned argument, so are omitted here. Further, we have $$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_{-1}(x)\!=\!\left\{ \begin{array}{l}
{^{RL}_{x_0}}D^\beta_x\varphi_4(x), \quad \ x \in [x_0,x_1)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_4(x), \quad \ x \in [x_1,{x_M}]
\end{array} \right.\end{aligned}$$ $$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_0(x)\!=\!\left\{ \begin{array}{l}
{^{RL}_{x_0}}D^\beta_x\varphi_3(x), \quad x \in [x_0,x_1)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_3(x) \\
\quad +{^{RL}_{x_1}}D^\beta_x\varphi_4(x), \ \ \ x \in [x_1,x_2)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_3(x) \\
\quad +{^{RL}_{x_1}}D^\beta_{x_2}\varphi_4(x), \ \ x \in [x_2,{x_M}]
\end{array} \right.\end{aligned}$$ $$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_1(x)\!=\!\left\{ \begin{array}{l}
{^{RL}_{x_0}}D^\beta_x\varphi_2(x), \quad x \in [x_0,x_1)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_2(x) \\
\quad +{^{RL}_{x_1}}D^\beta_x\varphi_3(x), \ \ \ x \in [x_1,x_2)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_2(x) \\
\quad +{^{RL}_{x_1}}D^\beta_{x_2}\varphi_3(x)\\
\quad +{^{RL}_{x_2}}D^\beta_{x}\varphi_4(x), \ \ x \in [x_2,x_3)\\
{^{RL}_{x_0}}D^\beta_{x_1}\varphi_2(x) \\
\quad +{^{RL}_{x_1}}D^\beta_{x_2}\varphi_3(x)\\
\quad +{^{RL}_{x_2}}D^\beta_{x_3}\varphi_4(x). \ \ x \in [x_3,{x_M}]
\end{array} \right.\end{aligned}$$ On the other hand, as the integrands of the integration in fractional derivatives, $\varphi_i(x)$, $i=1,2,3,4$, are cubic polynomials, for which, the order shrinks by one each time integration by parts is applied. Being aware of this, we can eliminate the weakly singular integrations by repeating integration by parts four times for each $\varphi_i(x)$ to derive the fully explicit formulas. The derivation processes are lengthy and tedious, we therefore outline the specific expressions of ${^{RL}_{x_0}}D^\beta_xB_m(x)$ in **Appendix**.
Construction of cubic B-spline DQ method
----------------------------------------
Use the early notations for brevity. On using DQ approximations (\[xz07\])-(\[xz08\]) to handle fractional derivatives, Eq. (\[xz01\]) is transformed into a set of first-order ODEs $$\begin{aligned}
\begin{aligned}\label{xz09}
&\frac{\partial u(x_i,y_j,t)}{\partial t}-\varepsilon_x\sum\limits_{m=0}^{M_x}{a_{im}^{(\beta_1)}u(x_m,y_j,t)}\\
&\qquad \ -\varepsilon_y\sum\limits_{m=0}^{M_y}{b_{jm}^{(\beta_2)}u(x_i,y_m,t)}=f(x_i,y_j,t),
\end{aligned}\end{aligned}$$ with $i=0,1,\cdots,M_x$, $j=0,1,\cdots,M_y$. Imposing the boundary constraint (\[xz03\]) and applying the Crank-Nicolson scheme, we thus obtain the following DQ scheme $$\begin{aligned}
\left\{
\begin{aligned}\label{xz12}
&U_{ij}^n-\frac{\tau\varepsilon_x}{2}\sum_{m=1}^{M_x-1}a_{im}^{(\beta_1)}U_{mj}^n-\frac{\tau\varepsilon_y}{2}\sum_{m=1}^{M_y-1}b_{jm}^{(\beta_2)}U_{im}^n\\
& =U_{ij}^{n-1}+\frac{\tau\varepsilon_x}{2}\sum_{m=1}^{M_x-1}a_{im}^{(\beta_1)}U_{mj}^{n-1} \\
&\qquad\quad+\frac{\tau\varepsilon_y}{2}\sum_{m=1}^{M_y-1}b_{jm}^{(\beta_2)}U_{im}^{n-1}+\tau f^{n-1/2}_{ij},
\end{aligned}\right.\end{aligned}$$ where $i=1,2,\cdots,M_x-1$, $j=1,2,\cdots,M_y-1$. It is visible that DQ methods are truly meshless and convenient in implementation. Due to the insensitivity to dimensional changes, (\[xz12\]) can easily be generalized to the higher-dimensional space-fractional problems, but do not cause the rapid increase of computing burden.
Illustrative examples {#s5}
=====================
In this section, a couple of numerical examples are carried out to gauge the practical performance of MCTB-DQM and new MCB-DQM. In order to check their accuracy, we compute the errors by using the norms $$\begin{aligned}
&e_\infty(M)\cong\max_{i}\Big|u^n_i-U^n_i\Big|,\\
&e_2(M)\cong\sqrt{\frac{1}{M}\sum^{M-1}_{i=1}\Big|u^n_i-U^n_i\Big|^2},\\
&e_N(M)\cong\sqrt{\sum^{M-1}_{i=1}\Big|u^n_{i}-U^n_{i}\Big|^2\bigg/\sum^{M-1}_{i=1}\Big|U_{i}^0\Big|^2},\\
&e_\infty(M_x,M_y)\cong\max_{i,j}\Big|u^n_{ij}-U^n_{ij}\Big|,\\
&e_2(M_x,M_y)\cong\sqrt{\frac{1}{M_xM_y}\sum^{M_x-1}_{i=1}\sum^{M_y-1}_{j=1}\Big|u^n_{ij}-U^n_{ij}\Big|^2},\end{aligned}$$ where $e_N(M)$ is termed by a normalized $L_2$-norm. As to $\{\omega^\alpha_k\}_{k=0}^{n}$ in Eqs. (\[eq22\])-(\[eq23\]), we use (\[ez06\]) in the first and fifth examples and (\[eq08\]) in the others but not the last two ones. In the computation, our algorithms are implemented on Matlab platform in a Lenovo PC with Intel(R) Pentium(R) G2030 3.00GHz CPU and 4 GB RAM. The obtained results are comparatively discussed with the early works available in the open literature.\
*Example 6.1.* Let $\kappa=1$, $\varepsilon=2$; the Eqs. (\[eq01\])-(\[eq03\]) with $\psi(x)=\exp(x)$, $g_1(t)=E(t^\alpha)$, $g_2(t)=eE(t^\alpha)$ and homogeneous forcing term are considered on $[0,1]$, where $E(t^\alpha)$ is the well-known *Mittag-Leffler function* $$E_\alpha(z)=\sum_{k=0}^{\infty}\frac{z^k}{\Gamma(\alpha k+1)}, \quad 0<\alpha<1.$$ It is verified that its solution is $u(x,t)=\exp(x)E(t^\alpha)$. In order to show the convergence of MCTB-DQM, we fix $\tau=1.0\times10^{-5}$ so that the temporal errors are negligible as compared to spatial errors. The numerical results at $t=0.1$ for various $\alpha$ are displayed in Table \[tab1\]; the convergent rate is shortly written as “Cov. rate”. As one sees, our method is pretty stable and convergent with almost spatial second-order for this problem.\
*Example 6.2.* In this test, we solve a diffusion equation on $[0,1]$ with $\varepsilon=1$, $\psi(x)=4x(1-x)$, zero boundary condition and right side. Its true solution has the form $$u(x,t)=\frac{16}{\pi^3}\sum^\infty_{k=1}\frac{1}{k^3}E_\alpha(-k^2\pi^2t^\alpha)(1-(-1)^k)\sin(k\pi x).$$ For comparison of the numerical results given by FDS-D I, FDS-D II [@28] and the semi-discrete FEM [@29], we choose the same time stepsize $\tau=1.0\times10^{-4}$. Letting $\alpha=0.1$, $0.5$ and $0.95$, the corresponding results of these four methods at $t=1$ are tabulated side by side in Table \[tab2\], from which, we conclude that MCTB-DQM is accurate and produces very small errors as the other three methods as the grid number $M$ increases.\
$\alpha$ $M$ $e_2(M)$ Cov. rate $e_\infty(M)$ Cov. rate
---------- ----- ------------ ----------- --------------- -----------
$0.2$ 8 2.4430e-03 - 3.5200e-03 -
16 6.3696e-04 1.9394 9.2142e-04 1.9337
32 1.6272e-04 1.9688 2.4362e-04 1.9192
64 4.1425e-05 1.9738 6.2649e-05 1.9592
128 1.0765e-05 1.9441 1.5906e-05 1.9777
$0.5$ 8 1.2489e-03 - 1.8283e-03 -
16 3.2655e-04 1.9352 4.8198e-04 1.9235
32 8.3679e-05 1.9644 1.2771e-04 1.9160
64 2.1466e-05 1.9628 3.3008e-05 1.9520
128 5.7295e-06 1.9056 8.4114e-06 1.9724
$0.8$ 8 9.7261e-04 - 1.4452e-03 -
16 2.5487e-04 1.9321 3.8329e-04 1.9147
32 6.5378e-05 1.9629 1.0160e-04 1.9156
64 1.6774e-05 1.9625 2.6330e-05 1.9481
128 4.4723e-06 1.9072 6.7183e-06 1.9705
$\alpha$ $M$ FDS-D I [@28] FDS-D II [@28] FEM [@29] MCTB-DQM
---------- ----- --------------- ---------------- ----------- ------------
$0.1$ 8 9.98e-04 1.00e-03 5.23e-04 2.7445e-04
16 2.44e-04 2.53e-04 1.29e-04 3.6823e-05
32 5.36e-05 6.33e-05 3.21e-05 5.0524e-06
64 5.89e-06 1.55e-05 8.01e-06 9.6533e-07
128 6.08e-06 3.62e-06 2.00e-06 4.6423e-07
$0.5$ 8 7.13e-04 7.13e-04 3.37e-04 1.9258e-04
16 1.79e-04 1.79e-04 8.31e-04 2.6563e-05
32 4.46e-05 4.44e-05 2.07e-05 4.5588e-06
64 1.07e-05 1.06e-05 5.17e-06 1.7866e-06
128 2.23e-06 2.12e-06 1.30e-06 1.4534e-06
$0.95$ 8 1.11e-04 1.11e-04 4.84e-05 2.8976e-05
16 2.83e-05 2.82e-05 1.21e-05 4.0937e-06
32 7.09e-06 7.08e-06 3.05e-06 8.7766e-07
64 1.76e-06 1.75e-06 7.93e-07 4.7682e-07
128 4.29e-07 4.23e-07 2.32e-07 4.2766e-07
*Example 6.3.* Let $\kappa=0$, $\varepsilon=1$; we solve Eqs. (\[eq01\])-(\[eq03\]) with homogeneous initial and boundary values, and $$f(x,t)=\frac{2t^{2-\alpha}\sin(2\pi x)}{\Gamma(3-\alpha)}+4\pi^2t^2\sin(2\pi x),$$ on $[0,1]$. The ture solution is $u(x,t)=t^2\sin(2\pi x)$. The algorithm is first run with $\alpha=0.8$, $\tau=2.0\times10^{-2}$ and $M=50$. In Fig. \[fig1\], we plot the approximate solution and a point to point error distribution at $t=1$, where good accuracy is observed. In Table \[tab3\], we then report a comparison of $e_2(M)$, $e_\infty(M)$ at $t=1$ between MCTB-DQM and CBCM [@ref23], when $\alpha=0.3$. Here, MCTB-DQM uses $\tau=5.0\times10^{-3}$ while CBCM chooses $\tau=1.25\times10^{-3}$. As expected, our approach generates the approximate solutions with a better accuracy than those obtained by CBCM.\
----- ------------ --------------- ------------ ---------------
$e_2(M)$ $e_\infty(M)$ $e_2(M)$ $e_\infty(M)$
8 3.4134e-02 4.8273e-02 9.4300e-03 1.5762e-02
16 8.7334e-03 1.2351e-02 1.1924e-03 2.1670e-03
32 2.1955e-03 3.1048e-03 1.5040e-04 2.8541e-04
64 5.4957e-04 7.7721e-04 1.8925e-05 3.6701e-05
128 1.3739e-04 1.9430e-04 2.3752e-06 4.6559e-06
----- ------------ --------------- ------------ ---------------
*Example 6.4.* We consider a 2D diffusion equation on $[-1,1]\times[-1,1]$ with $\varepsilon_x=\varepsilon_y=1$, which is referred to by Zhai and Feng as a test of a block-centered finite difference method (BCFDM) on nonuniform grids [@31]. The forcing function is specified to enforce $$u(x,y,t)=(1+t^2)\tanh(20x)\tanh(20y).$$ Under $\tau=1.0\times10^{-2}$, $M_x=M_y=60$ and $\alpha=0.5$, we first plot the approximate solution and a point to point error distribution at $t=0.5$ in Fig. \[fig2\]. Then, we compare MCTB-DQM and BCFDM in term of $e_\infty(M_x,M_y)$ at $t=0.5$ in Table \[tab4\]. It is obvious that MCTB-DQM produces significantly smaller errors than BCFDM as the grid number increases despite a smaller time stepsize $\tau=2.5\times10^{-3}$ and the nonuniform girds BCFDM adopts; moreover, MCTB-DQM provides more than quadratic rate of convergence for this problem.\
---- --------------------- ----------- --------------------- -----------
$e_\infty(M_x,M_y)$ Cov. rate $e_\infty(M_x,M_y)$ Cov. rate
12 8.75e-02 - 3.3376e-01 -
24 2.73e-02 1.68 4.6331e-03 6.1707
48 8.26e-03 1.73 3.4566e-04 3.7446
96 2.24e-03 1.88 1.8605e-05 4.2156
---- --------------------- ----------- --------------------- -----------
$\alpha$ Method Real part Imaginary part CPU time
---------- ------------------- ------------ ---------------- ------------
0.2 scheme (\[eq22\]) 7.4588e-04 7.3396e-04 10.355 (s)
0.5 scheme (\[eq22\]) 3.9328e-03 3.7620e-03 10.819 (s)
0.8 scheme (\[eq22\]) 4.5245e-03 4.6714e-03 10.775 (s)
1.0 scheme (\[eq22\]) 2.2229e-03 2.2153e-03 1.7436 (s)
scheme (\[ez09\]) 8.0954e-04 8.1843e-04 0.1521 (s)
*Example 6.5.* In this test, we simulate the solitons propagation and collision governed by the following time-fractional nonlinear Schrödinger equation (NLS): $$\begin{aligned}
\textrm{i}\frac{\partial^\alpha u}{\partial t^\alpha}+\frac{\partial^2 u}{\partial x^2} + \beta|u|^2u=0, \ \ x\in(-\infty,+\infty),\end{aligned}$$ with $\textrm{i}=\sqrt{-1}$ and $\beta$ being a real constant, subjected to the initial values of two Gaussian types:
- mobile soliton $$\label{ez90}
\psi(x)=\textrm{sech}(x)\exp(2\textrm{i}x);$$
- double solitons collision $$\label{ez91}
\psi(x)=\sum_{j=1}^2\textrm{sech}(x-x_j)\exp(\textrm{i}p(x-x_j)).$$
When $\alpha=1$ and $\beta=2$, the NLS with Eq. (\[ez90\]) has the soliton solution $u(x,t)=\textrm{sech}(x-4t)\exp(\textrm{i}(2x-3t))$. As the solutions would generally decay to zero as $|x|\rightarrow \infty$, we truncate the system into a bounded interval $\Omega=[a,b]$ with $a\ll0$ and $b\gg0$, and enforce periodic or homogeneous Dirichlet boundary conditions. Letting $u(x,t)=U(x,t)+\textrm{i}V(x,t)$. Then, the original equation can be recast as a coupled diffusion system $$\begin{aligned}
&\frac{\partial^\alpha U}{\partial t^\alpha}+\frac{\partial^2 V}{\partial x^2}+\beta(U^2+V^2)V=0, \\
&\frac{\partial^\alpha V}{\partial t^\alpha}-\frac{\partial^2 U}{\partial x^2}-\beta(U^2+V^2)U=0.\end{aligned}$$ After applying the scheme (\[eq22\]), nevertheless, a nonlinear system has to be solved at each time step. In such a case, the Newton’s iteration is utilized to treat it and terminated by reaching a solution with tolerant error $1.0\times10^{-12}$ if $\alpha=1$, for which, the Jacobian matrix is $$\begin{aligned}
\textbf{J}=\left( \begin{array}{cc}
2UV&U^2+3V^2\\
-3U^2-V^2&-2UV
\end{array} \right).\end{aligned}$$ When $\alpha\neq1$, because the analytic solutions still remain unknown and the Newton’s procedure relies heavily on its initial values, we instead employ the trust-region-dogleg algorithm built into Matlab to improve the convergence of iteration. At first, taking $\tau=2.0\times10^{-3}$, $M=100$, $\beta=2$, and $\Omega=[-10,10]$, the mean square errors at $t=0.1$ with the initial condition (\[ez90\]) for various $\alpha$ are reported in Table \[tab6\], where the solutions computed by using the coefficients (\[eq08\]) on a very fine time-space lattice, i.e., $\tau=2.5\times10^{-4}$, $M=400$, are adopted as reference solutions ($\alpha\neq1$). As seen from Table \[tab6\], our methods are convergent and applicable to nonlinear coupled problems; besides, the scheme (\[ez09\]) is clearly more efficient than (\[eq22\]) since an extra Newton’s outer loop is avoided. Then, retaking $M=200$ and $\Omega=[-20,20]$, we display the evolution of the amplitude of the mobile soliton created by (\[eq22\]) for $\alpha=0.98$ and $1.0$ in Fig. \[fig8\], respectively. Using the same discrete parameters, we consider the double solitons collision for $\alpha=0.96$ and $1.0$ with $x_1=-6$, $x_2=6$, and $p=\pm2$ in Fig. \[fig9\]. It is easily drawn from these figures that the width and height of the solitons have been significantly changed by the fractional derivative. In particular, when $\alpha=1$, a collision of double solitons without any reflection, transmission, trapping and creation of new solitary waves is exhibited, which says that it is elastic, while in fractional cases, the shapes of the solitons may not be retained after they intersect each other.\
*Example 6.6.* In this test, we simulate an unsteady propagation of a Gaussian pulse governed by a classical 2D advection-dominated diffusion equation on a square domain $[0,2]\times[0,2]$ by using the scheme (\[ez09\]), which has been extensively studied [@ninep02; @32; @33; @cirbf00]. The Gaussian pulse solution is expressed as
$$u(x,y,t)\!=\!\frac{1}{1+4t}\exp\Bigg(-\frac{(x-\!\kappa_xt-\!0.5)^2}{\varepsilon_x(1+4t)}
-\frac{(y-\!\kappa_yt-\!0.5)^2}{\varepsilon_y(1+4t)}\Bigg),$$
and the initial Gaussian pulse and boundary values are taken from the pulse solution. Letting $\kappa_x=\kappa_y=0.8$, $\varepsilon_x=\varepsilon_y=0.01$, we display its true solution at $t=1.25$ with $M_x=M_y=50$ and the used lattice points on problem domain in Fig. \[fig3\], which describe a pulse centred at $(1.5, 1.5)$ with a pulse height of $1/6$. Using the same grid number together with $\tau=5.0\times10^{-3}$, we present the contour plots of the approximate solutions at $t=0$, $0.25$, $0.75$, $1.25$ created by MCTB-DQM in Fig. \[fig5\]. As the graph shows, the pulse is initially centred at $(0.5, 0.5)$ with a pulse height of $1$, then it moves towards a position centred at $(1.5, 1.5)$; during this process, its width and height appear to be continuously varying as the time goes by. Besides, the last contour plot in Fig. \[fig5\] coincides with the true solution plotted in Fig. \[fig3\]. Retaking $\tau=6.25\times10^{-3}$ and $M_x=M_y=80$, we compare our results with those obtained by some previous algorithms as nine-point high-order compact (HOC) schemes [@ninep02; @33], Peaceman-Rachford ADI scheme (PR-ADI) [@pr08], HOC-ADI scheme [@hoc03], exponential HOC-ADI scheme (EHOC-ADI) [@ehoc05], HOC boundary value method (HOC-BVM) [@bvm06], compact integrated RBF ADI method (CIRBF-ADI) [@cirbf00], coupled compact integrated RBF ADI method (CCIRBF-ADI) [@cirbf01], and the Galerkin FEM combined with the method of characteristics (CGFEM) [@cgfem07], at $t=1.25$ in Table \[tab5\]. We implement CGFEM on a quasi-uniform triangular mesh with the meshsize $2.5\times10^{-2}$ by using both Lagrangian P1 and P2 elements. Also, average absolute errors are added as supplements to evaluate and compare their accuracy. As seen from Table \[tab5\], all of these methods are illustrated to be very accurate to capture the Gaussian pulse except the PR-ADI scheme; besides, our method reaches a better accuracy than the others and even shows promise in treating the advection-diffusion equations in the high Péclet number regime.\
\
Method Average Error $e_2(M_x,M_y)$ $e_\infty(M_x,M_y)$
-------------------------- --------------- ---------------- ---------------------
CIRBF-ADI [@cirbf00] 6.742e-06 2.197e-05 1.703e-04
CCIRBF-ADI [@cirbf01] 5.989e-06 1.904e-05 1.427e-04
Noye and Tan [@33] 1.971e-05 1.280e-04 6.509e-04
Kalita et al. [@ninep02] 1.597e-05 1.024e-04 4.477e-04
PR-ADI [@pr08] 3.109e-04 2.025e-03 7.778e-03
HOC-ADI [@hoc03] 9.218e-06 5.931e-05 2.500e-04
EHOC-ADI [@ehoc05] 9.663e-06 6.194e-05 2.664e-04
HOC-BVM [@bvm06] 9.4931e-06 - 2.4766e-04
CGFEM P1 6.3746e-05 1.8849e-04 1.5731e-03
CGFEM P2 1.5667e-05 5.4061e-05 5.8044e-04
MCTB-DQM 9.1512e-07 5.6996e-06 2.2830e-05
*Example 6.7.* In the last test, we consider the 2D space-fractional equations (\[xz01\])-(\[xz03\]) on $[0,1]\times[0,1]$ with $\varepsilon_x=\varepsilon_y=1$, $\psi(x,y)=x^2(1-x)^2y^2(1-y)^2$, and homogeneous boundary values. The source term is manufactured as
$$\begin{aligned}
&f(x,y,t)=-e^{-t}x^2(1-x)^2y^2(1-y)^2\\
&\ -\frac{2e^{-t}x^{2-\beta_1}y^2(1-y)^2}{\Gamma(3-\beta_1)}\Bigg(1
-\frac{6x}{3-\beta_1}+\frac{12x^2}{(3-\beta_1)(4-\beta_1)}\Bigg)\\
&\ -\frac{2e^{-t}x^2(1-x)^2 y^{2-\beta_2}}{\Gamma(3-\beta_2)}\Bigg(1
-\frac{6y}{3-\beta_2}+\frac{12y^2}{(3-\beta_2)(4-\beta_2)}\Bigg)\end{aligned}$$
to enforce the analytic solution $u(x,y,t)=e^{-t}x^2(1-x)^2y^2(1-y)^2$. Letting $\beta_1=1.1$, $\beta_2=1.3$, and $\tau=2.5\times10^{-4}$, we solve the problem via the FEM proposed by [@wx05] and MCB-DQM, and compare their numerical results at $t=0.2$ in Table \[tab8\], where the P1 element and structured meshes are adopted. The data indicate that DQ method converges towards the analytic solution as the grid numbers increase and admits slightly better results than FEM. More importantly, the implemental CPU times of MCB-DQM are less than those of FEM, which confirm its computing efficiency.\
Conclusion {#s6}
==========
The ADEs are the subjects of active interest in mathematical physics and the related areas of research. In this work, we have proposed an effective DQ method for such equations involving the derivatives of fractional orders in time and space. Its weighted coefficients are calculated by making use of modified CTBs and cubic B-splines as test functions. The stability of DQ method for the time-fractional ADEs in the context of $L_2$-norm is performed. The theoretical condition required for the stable analysis is numerically surveyed at length. We test the codes on several benchmark problems and the outcomes have demonstrated that it outperforms some of the previously reported algorithms such as BCFDM and FEM in term of overall accuracy and efficiency.
In a linear space, spanned by a set of proper basis functions as B-splines, any function can be represented by a weighted combination of these basis functions. While all basis functions are defined, the function remains unknown because the coefficients on the front of basis functions are still unknown. However, when all basis functions satisfy Eqs. (\[xz17\])-(\[xz18\]), by virtue of linearity, it can be examined that the objective function satisfies Eqs. (\[xz17\])-(\[xz18\]) as well. This is the essence of DQ methods, which guarantees their convergence.
Despite the error bounds are difficult to determine, the numerical results illustrate that the spline-based DQ method admits the convergent results for the fractional ADEs. The presented approach can be generalized to the higher-dimensional and other complex model problems arising in material science, structural and fluid mechanics, heat conduction, biomedicine, differential dynamics, and so forth. High computing efficiency, low memory requirement, and the ease of programming are its main advantages.
The authors are very grateful to the reviewers for their valuable comments and suggestions. This research was supported by National Natural Science Foundations of China (Nos.11471262 and 11501450).
Method $M_x,\ M_y$ $e_2(M_x,M_y)$ Cov. rate $ e_\infty(M_x,M_y)$ Cov. rate CPU time
------------- ------------- ---------------- ----------- ---------------------- ----------- -----------
FEM [@wx05] 10 6.5781e-05 - 1.5632e-04 - 63.750(s)
15 3.0082e-05 1.9296 7.2433e-05 1.8972 199.37(s)
20 1.7376e-05 1.9079 4.1321e-05 1.9511 494.98(s)
25 1.1404e-05 1.8872 2.6599e-05 1.9741 1062.7(s)
MCB-DQM 10 5.4217e-05 - 1.4763e-04 - 0.3726(s)
15 2.6606e-05 1.7556 6.9553e-05 1.8562 1.1785(s)
20 1.5559e-05 1.8650 4.0088e-05 1.9153 4.0598(s)
25 1.0207e-05 1.8892 2.6163e-05 1.9124 11.267(s)
**Appendix**: **The explicit formulas of the fractional derivatives of cubic B-splines**
The fractional derivatives center at $x_{-1}$, $x_0$, and $x_1$: $$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_{-1}(x)\!=\!\left\{ \begin{array}{l}
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}-\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{6(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_0,x_1)\\
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}-\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{6(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{6(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_1,x_M]
\end{array} \right.\end{aligned}$$
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_0(x)\!=\!\left\{ \begin{array}{l}
\frac{4(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}-\frac{12(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2}
+\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_0,x_1)\\
\frac{4(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}-\frac{12(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2}
+\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_1,x_2)\\
\frac{4(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}-\frac{12(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2}
+\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{6(x-x_2)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_2,x_M]
\end{array} \right.\end{aligned}$$
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_1(x)\!=\!\left\{ \begin{array}{l}
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}+\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_0,x_1)\\
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}+\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_1,x_2)\\
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}+\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}
-\frac{24(x-x_2)^{3-\beta}}{\Gamma(4-\beta)h^3},\quad x \in [x_2,x_3)\\
\frac{(1-\beta)(x-x_0)^{-\beta}}{\Gamma(2-\beta)}+\frac{3(x-x_0)^{1-\beta}}{\Gamma(2-\beta)h}
+\frac{6(x-x_0)^{2-\beta}}{\Gamma(3-\beta)h^2} -\frac{18(x-x_0)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_1)^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_2)^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{6(x-x_3)^{3-\beta}}{\Gamma(4-\beta)h^3}.\quad x \in [x_3,x_M]
\end{array} \right.\end{aligned}$$
The fractional derivatives center at $x_{m}$ with $2\leq m\leq M+1$:
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_m(x)\!=\!\left\{ \begin{array}{l}
0,\qquad\qquad\qquad \, x \in [x_0,x_{m-2})\\
\frac{6(x-x_{m-2})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [{x_{m - 2}},{x_{m - 1}})\\
\frac{6(x-x_{m-2})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m-1})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [{x_{m - 1}},{x_m})\\
\frac{6(x-x_{m-2})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m-1})^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_{m})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [{x_m},{x_{m + 1}})\\
\frac{6(x-x_{m-2})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m-1})^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_{m})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m+1})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [{x_{m + 1}},{x_{m + 2}})\\
\frac{6(x-x_{m-2})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m-1})^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_{m})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{m+1})^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{6(x-x_{m+2})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_{m+2},x_M]
\end{array} \right.\end{aligned}$$
which contain ${^{RL}_{x_0}}D^\beta_xB_{M-1}(x)$, ${^{RL}_{x_0}}D^\beta_xB_M(x)$, and ${^{RL}_{x_0}}D^\beta_xB_{M+1}(x)$ as special cases:
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_{M-1}(x)\!=\!\left\{ \begin{array}{l}
0,\qquad\qquad\qquad \, x \in [x_0,x_{M-3})\\
\frac{6(x-x_{M-3})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_{M - 3},x_{M - 2})\\
\frac{6(x-x_{M-3})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{M-2})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [{x_{M - 2}},x_{M-1})\\
\frac{6(x-x_{M-3})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{M-2})^{3-\beta}}{\Gamma(4-\beta)h^3}
+\frac{36(x-x_{M-1})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_{M-1},x_M]
\end{array} \right.\end{aligned}$$
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_M(x)\!=\!\left\{ \begin{array}{l}
0, \quad x \in [x_0,x_{M-2})\\
\frac{6(x-x_{M-2})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_{M-2},x_{M-1})\\
\frac{6(x-x_{M-2})^{3-\beta}}{\Gamma(4-\beta)h^3}-\frac{24(x-x_{M-1})^{3-\beta}}{\Gamma(4-\beta)h^3}, \quad x \in [x_{M-1},x_{M}]
\end{array} \right.\end{aligned}$$
$$\begin{aligned}
{^{RL}_{x_0}}D^\beta_xB_{M+1}(x)\!=\!\left\{ \begin{array}{l}
0, \quad x \in [x_0,x_{M-1})\\
\frac{6(x-x_{M-1})^{3-\beta}}{\Gamma(4-\beta)h^3}. \quad x \in [x_{M-1},x_M]
\end{array} \right.\end{aligned}$$
[10]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{}
Abbas, M., Majid, A.A., Ismail, A.I.M., Rashid, A.: The application of cubic trigonometric [B]{}-spline to the numerical solution of the hyperbolic problems. Appl. Math. Comput. **239**, 74–88 (2014)
Bellman, R., Casti, J.: Differential quadrature and long-term integration. J. Math. Anal. Appl. **34**(2), 235–238 (1971)
Bhrawy, A., Zaky, M.: A fractional-order [J]{}acobi [T]{}au method for a class of time-fractional [PDE]{}s with variable coefficients. Math. Method Appl. Sci. **39**(7), 1765–1779 (2016)
Chen, M.H., Deng, W.H.: Fourth order difference approximations for space [R]{}iemann-[L]{}iouville derivatives based on weighted and shifted [L]{}ubich difference operators. Commun. Comput. Phys. **16**(2), 516–540 (2014)
Chen, Y.M., Wu, Y.B., Cui, Y.H., Wang, Z.Z., Jin, D.M.: Wavelet method for a class of fractional convection-diffusion equation with variable coefficients. J. Comput. Sci. **1**(3), 146–149 (2010)
Cui, M.R.: A high-order compact exponential scheme for the fractional convection-diffusion equation. J. Comput. Appl. Math. **255**, 404–416 (2014)
Dehghan, M., Mohebbi, A.: High-order compact boundary value method for the solution of unsteady convection–diffusion problems. Math. Comput. Simulat. **79**(3), 683–699 (2008)
, J., Russell, T.F.: Numerical methods for convection-dominated diffusion problems based on combining the method of characteristics with finite element or finite difference procedures. SIAM J. Numer. Anal. **19**(5), 871–885 (1982)
Esmaeili, S., Garrappa, R.: A pseudo-spectral scheme for the approximate solution of a time-fractional diffusion equation. Int. J. Comput. Math. **92**(5), 980–994 (2015)
Gao, G.H., Sun, H.W.: Three-point combined compact alternating direction implicit difference schemes for two-dimensional time-fractional advection-diffusion equations. Commun. Comput. Phys. **17**(2), 487–509 (2015)
Heydari, M.H.: Wavelets [G]{}alerkin method for the fractional subdiffusion equation. J. Comput. Nonlinear Dynam. **11**(6), 061,014 (2016)
Hosseini, S.M., Ghaffari, R.: Polynomial and nonpolynomial spline methods for fractional sub-diffusion equations. Appl. Math. Model. **38**(14), 3554–3566 (2014)
Huang, C.B., Yu, X.J., Wang, C., Li, Z.Z., An, N.: A numerical method based on fully discrete direct discontinuous [G]{}alerkin method for the time fractional diffusion equation. Appl. Math. Comput. **264**, 483–492 (2015)
Izadkhah, M.M., Saberi-Nadjafi, J.: Gegenbauer spectral method for time-fractional convection-diffusion equations with variable coefficients. Math. Method Appl. Sci. **38**(15), 3183–3194 (2015)
Jia, J.H., Wang, H.: A fast finite volume method for conservative space-fractional diffusion equations in convex domains. J. Comput. Phys. **310**, 63–84 (2016)
Ji, C.C., Sun, Z.Z.: A high-order compact finite difference scheme for the fractional sub-diffusion equation. J. Sci. Comput. **64**(3), 959–985 (2015)
Jiang, Y.J., Ma, J.T.: High-order finite element methods for time-fractional partial differential equations. J. Comput. Appl. Math. **235**(11), 3285–3290 (2011)
Jin, B.T., Lazarov, R., Zhou, Z.: Error estimates for a semidiscrete finite element method for fractional order parabolic equations. SIAM J. Numer. Anal. **51**(1), 445–466 (2013)
Kalita, J.C., Dalal, D.C., Dass, A.K.: A class of higher order compact schemes for the unsteady two-dimensional convection–diffusion equation with variable convection coefficients. Int. J. Numer. Meth. Fluids **38**(12), 1111–1131 (2002)
Karaa, S., Zhang, J.: High order [ADI]{} method for solving unsteady convection–diffusion problems. J. Comput. Phys. **198**(1), 1–9 (2004)
Kilbas, A.A., Srivastava, H.M., Trujillo, J.J.: Theory and [A]{}pplications of [F]{}ractional [D]{}ifferential [E]{}quations. Elsevier B. V., Amsterdam (2006)
Laub, A.J.: Matrix Analysis For Scientists & Engineers. Society for Industrial and Applied Mathematics, Philadelphia, PA (2005)
Lin, Y.M., Xu, C.J.: Finite difference/spectral approximations for the time-fractional diffusion equation. J. Comput. Phys. **225**(2), 1533–1552 (2007)
Liu, Q., Gu, Y.T., Zhuang, P., Liu, F., Nie, Y.F.: An implicit [RBF]{} meshless approach for time fractional diffusion equations. Comput. Mech. **48**(1), 1–12 (2011)
Luo, W.H., Huang, T.Z., Wu, G.C., Gu, X.M.: Quadratic spline collocation method for the time fractional subdiffusion equation. Appl. Math. Comput. **276**, 252–265 (2016)
Mittal, R.C., Jain, R.K.: Numerical solutions of nonlinear Burgers’ equation with modified cubic B-splines collocation method. Appl. Math. Comput. **218**(15), 7839-7855 (2012)
Ma, Y.B., Sun, C.P., Haake, D.A., Churchill, B.M., Ho, C.M.: A high-order alternating direction implicit method for the unsteady convection-dominated diffusion problem. Int. J. Numer. Meth. Fluids **70**(6), 703–712 (2012)
Meerschaert, M.M., Scheffler, H.P., Tadjeran, C.: Finite difference methods for two-dimensional fractional dispersion equation. J. Comput. Phys. **211**(1), 249–261 (2006)
Mainardi, F.: Fractals and [F]{}ractional [C]{}alculus [C]{}ontinuum [M]{}echanics. Springer, Verlag (1997)
Nazir, T., Abbas, M., Ismail, A.I.M., Majid, A.A., Rashid, A.: The numerical solution of advection-diffusion problems using new cubic trigonometric [B]{}-splines approach. Appl. Math. Model. **40**(7-8), 4586–4611 (2016)
Noye, B.J., Tan, H.H.: Finite difference methods for solving the two-dimensional advection-diffusion equation. Int. J. Numer. Meth. Fluids **9**(1), 75–98 (1989)
Peaceman, D.W., [Rachford, Jr.]{}, H.H.: The numerical solution of parabolic and elliptic differential equations. J. Soc. Indust. Appl. Math. **3**(1), 28–41 (1955)
Pirkhedri, A., Javadi, H.H.S.: Solving the time-fractional diffusion equation via [S]{}inc-[H]{}aar collocation method. Appl. Math. Comput. **257**, 317–326 (2015)
Podlubny, I.: Fractional [D]{}ifferential [E]{}quations. Academic Press, San Diego, CA (1999)
Razminia, K., Razminia, A., Baleanu, D.: Investigation of the fractional diffusion equation based on generalized integral quadrature technique. Appl. Math. Model. **39**(1), 86–98 (2015)
Roop, J.P.: Computational aspects of FEM approximation of fractional advection dispersion equations onbounded domains in $\mathbb{R}^2$. J. Comput. Appl. Math. **193**(1), 243–268 (2006)
Saadatmandi, A., Dehghan, M., Azizi, M.R.: The [S]{}inc-[L]{}egendre collocation method for a class of fractional convection-diffusion equations with variable coefficients. Commun. Nonlinear Sci. Numer. Simul. **17**(11), 4125–4136 (2012)
Sayevand, K., Yazdani, A., Arjang, F.: Cubic [B]{}-spline collocation method and its application for anomalous fractional diffusion equations in transport dynamic systems. J. Vib. Control **22**(9), 2173–2186 (2016)
Shirzadi, A., Ling, L., Abbasbandy, S.: Meshless simulations of the two-dimensional fractional-time convection-diffusion-reaction equations. Eng. Anal. Bound. Elem. **36**(11), 1522–1527 (2012)
Shlesinger, M.F., West, B.J., Klafter, J.: Lévy dynamics of enhanced diffusion: [A]{}pplication to turbulence. Phys. Rev. Lett. **58**(11), 1100–1103 (1987)
Shu, C., Richards, B.E.: Application of generalized differential quadrature to solve two-dimensional incompressible [N]{}avier-[S]{}tokes equations. Int. J. Numer. Meth. Fluids **15**(7), 791–798 (1992)
Tomasiello, S.: Stability and accuracy of the iterative differential quadrature method. Int. J. Numer. Methods Eng. **58**(9), 1277-1296 (2003)
Tomasiello, S.: Numerical stability of DQ solutions of wave problems. Numer. Algor. **57**(3), 289-312 (2011)
Tian, W.Y., Zhou, H., Deng, W.H.: A class of second order difference approximations for solving space fractional diffusion equations. Math. Comp. **84**(294), 1703-1727 (2015)
Thai-Quang, N., Mai-Duy, N., Tran, C.-D., Tran-Cong, T.: High-order alternating direction implicit method based on compact integrated-[RBF]{} approximations for unsteady/steady convection-diffusion equations. CMES Comput. Model. Eng. Sci. **89**(3), 189–220 (2012)
Tian, Z.F., Ge, Y.B.: A fourth-order compact [ADI]{} method for solving two-dimensional unsteady convection–diffusion problems. J. Comput. Appl. Math. **198**(1), 268–286 (2007)
Tien, C.M.T., Thai-Quang, N., Mai-Duy, N., Tran, C.-D., Tran-Cong, T.: A three-point coupled compact integrated [RBF]{} scheme for second-order differential problems. CMES Comput. Model. Eng. Sci. **104**(6), 425–469 (2015)
Uddin, M., Haq, S.: [RBF]{}s approximation method for time fractional partial differential equations. Commun. Nonlinear Sci. Numer. Simul. **16**(11), 4208–4214 (2011)
Walz, G.: Identities for trigonometric [B]{}-splines with an application to curve design. BIT **37**(1), 189–201 (1997)
Yang, X.H., Zhang, H.X., Xu, D.: Orthogonal spline collocation method for the two-dimensional fractional sub-diffusion equation. J. Comput. Phys. **256**, 824–837 (2014)
Zaslavsky, G.M., Stevens, D., Weitzner, H.: Self-similar transport in incomplete chaos. Phys. Rev. E. **48**(3), 1683–1694 (1993)
Zhu, X.G., Nie, Y.F., Wang, J.G., Yuan, Z.B.: A numerical approach for the Riesz space-fractional Fisher’ equation in two-dimensions. Int. J. Comput. Math. **94**(2), 296-315 (2017)
Zeng, F.H., Li, C.P., Liu, F.W., Turner, I.: The use of finite difference/element approaches for solving the time-fractional subdiffusion equation. SIAM J. Sci. Comput. **35**(6), A2976–A3000 (2013)
Zeng, F.H., Liu, F.W., Li, C.P., Burrage, K., Turner, I., Anh, V.: A Crank-Nicolson ADI spectral method for a two-dimensional Riesz space fractional nonlinear reaction-diffusion equation. SIAM J. Numer. Anal. **52**(6), 2599-2622 (2014)
Zhai, S.Y., Feng, X.L.: A block-centered finite-difference method for the time-fractional diffusion equation on nonuniform grids. Numer. Heat Transf., Part B Fundam. **69**(3), 217–233 (2016)
Zhou, F.Y., Xu, X.Y.: The third kind [C]{}hebyshev wavelets collocation method for solving the time-fractional convection diffusion equations with variable coefficients. Appl. Math. Comput. **280**, 11–29 (2016)
Zhuang, P.H., Liu, F.W.: Implicit difference approximation for the time fractional diffusion equation. J. Appl. Math. & Computing **22**(3), 87–99 (2006)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The population of exoplanetary systems detected by Kepler provides opportunities to refine our understanding of planet formation. Unraveling the conditions needed to produce the observed exoplanet systems will allow us to make informed predictions as to where habitable worlds exist within the galaxy. In this paper, we examine using N-body simulations how the properties of planetary systems are determined during the final stages of assembly, when planets accrete from embryos and planetesimals. While accretion is a chaotic process, trends emerge allowing certain features of an ensemble of planetary systems to provide a memory of the initial distribution of solid mass around a star prior to accretion. We also use EPOS, the Exoplanet Population Observation Simulator, to account for detection biases and show that different accretion scenarios can be distinguished from observations of the Kepler systems. We show that the period of the innermost planet, the ratio of orbital periods of adjacent planets, and masses of the planets are determined by the total mass and radial distribution of embryos and planetesimals at the beginning of accretion. In general, some amount of orbital damping, either via planetesimals or gas, during accretion is needed to match the whole population of exoplanets. Surprisingly, all simulated planetary systems have planets that are similar in size, showing that the “peas in a pod” pattern can be consistent with both a giant impact scenario and a planet migration scenario. The inclusion of material at distances larger than what *Kepler* observes ($>1$ au) has a profound impact on the observed planetary architectures, and thus on the formation and delivery of volatiles to possible habitable worlds.'
author:
- 'Gijs D. Mulders'
- 'David P. O’Brien'
- 'Fred J. Ciesla'
- Dániel Apai
- Ilaria Pascucci
bibliography:
- 'references.bib'
- 'software.bib'
title: 'Earths in Other Solar Systems N-body simulations: the Role of Orbital Damping in Reproducing the Kepler Planetary Systems'
---
Introduction
============
Terrestrial planets form through the piece-wise collisional aggregation of solids spanning many orders of magnitude in size, beginning with micron-sized grains and continuing upward. The final stages of growth involve the accretion of swarms of planetesimals and embryos, bodies measuring hundreds to thousands of kilometers across. It is at these sizes where gravitational interactions lead to changes in the paths of these bodies around their central star, shifting them from the roughly circular orbits they develop within a gaseous protoplanetary disk, to more eccentric, crossing orbits, that result in collisions and thus growth. As protoplanets grow larger in this manner, the gravitational interactions work to space out the larger surviving bodies, isolating them from one another as the initial building blocks are largely cleared by further accretion, scattering out of the system, or collisions with the central star. This leaves behind a fairly stable system of planets, with a small amount of leftover primordial bodies [e.g. @raymond14].
One possible outcome of this evolution is seen in the Solar System today, with planets orbiting at relatively large separations and with populations of small bodies within the asteroid belt, Kuiper Belt, and Oort Cloud. How the Solar System achieved this structure has been the focus of a number of studies that used N-body codes to simulate the final stages of planetary growth, tracking how a distribution of embryos and planetesimals interact and grow over time [e.g. @wetherill94; @chambers98; @Chambers2001Icar; @Raymond2004Icar; @raymond05; @raymond06; @raymond09; @obrien06; @OBrien2014Icar; @Bond2010Icar; @Bond2010ApJ; @CarterBond2012bApJ; @Fischer2014EPSL; @walsh11; @Quintana2007ApJ; @Quintana2014ApJ; @Izidoro2013ApJ; @Izidoro2014ApJ]. The success of these models in understanding the final stages of planet formation has largely been evaluated by their ability to reproduce the properties of the Solar System. That is, key dynamical features (such as the mass of the asteroid belt, the angular momentum of the planets, their mass distribution, the low mass of Mars, and the timing of the last accretion event of the Earth) have been used to constrain the orbital properties of Jupiter and Saturn during terrestrial planet formation [@obrien06; @raymond09], the radial distribution of mass of the planetesimals and embryos [@raymond06], or the extent to which Jupiter migrated while the solar nebula was still present [@walsh11].
These studies have also found that similar initial conditions can yield very different final planetary systems [e.g. @lissauer07; @Fischer2014EPSL]. The chaotic nature of the final assembly phase presents a challenge when comparing our Solar System to the model outcomes. That is, it cannot be said with certainty whether our inability to simultaneously match all of the properties of the Solar System is due to the stochastic nature of the process and infinite number of possible outcomes that could be produced or if the initial conditions used in the models differ from those of reality.
While comparisons to the Solar System is a natural starting point to understand this stage of evolution, the discovery of thousands of exoplanetary systems has shown that our planetary system represents just one possible outcome of the planet formation process [e.g. @batalha14]. The large number of systems discovered now allows us to identify general trends and relationships between key properties that arise during planetary assembly. Thus, these systems provide a new opportunity to test our understanding of the final stages of planet formation by investigating how well our models reproduce general trends in observed exoplanet systems.
Here we report on our study that explores how the key properties of planetary systems relate back to the initial conditions when planetary accretion begins and what conditions were necessary to produce the final architectures of the planetary systems found around sun-like stars. Our goal is not to define or identify the best, single set of conditions that will reproduce the known planetary systems, but instead, to understand the connections and relationships between initial conditions that are responsible for defining the properties of the systems that we see today as a means of understanding formation pathways, and the bulk properties, for those planets [@apaiWP].
Our approach here is to build a statistical framework in which we can model the outcomes of various planet formation scenarios and evaluate them against the known systems of exoplanets. @hansenmurray13 carried out an initial study with these goals, arguing that the known *Kepler* systems could have formed if the final stages of assembly began with a massive disk of planetesimals concentrated within $\sim$1 au of their host stars. In particular, the distributions of orbital period and planet radius, as well as the relative frequencies of observed multi-planet systems produced in the models were consistent with those real systems that had been detected. Subsequent studies have corroborated and expanded on these results, focusing on the dynamical effects of residual gas in these simulations [@dawson15; @dawson16] and explaining the apparent excess of single-transiting systems [@moriarty16].
Currently, the number of Kepler planet candidates has grown to over four thousand [@thompson17] and the survey detection efficiency has been thoroughly characterized [e.g. @burke15; @christiansen16], providing a more robust set of data to describe actual planetary properties to which model systems can be compared to. We explore a wide range of parameter space, and generate a large number of model planetary systems in order to understand the distribution of outcomes possible for a given range of parameters. We then evaluate how well the distribution of planetary properties produced by these initial conditions compares to the Kepler dataset using the Exoplanet Population Observation Simulator [`epos`[^1], @epos1; @epos2]. Again, we do not expect to identify a single set of parameters or scenarios that describes the formation of all known planetary systems; instead, we are evaluating whether particular signposts of the imprints of the initial conditions and process of planetary accretion can be identified within the data sets. Success would then inform us of the formation histories of these planets. In the next section, we describe the various sets of initial conditions and dynamical effects we considered in our simulations. The results of our simulations, which are part of the `Genesis` database of the *Earths in Other Solar Systems* project[^2], are presented in Section 3. We compare the model results to the Kepler system using version 3.0 of the new `epos` tool in Section 4 to evaluate how well our models and assumptions about the final stages of planet formation match the real planetary systems that have been identified to date. We conclude with a discussion and outline of future work.
Dynamical Scenarios
===================
In all of our simulations, we follow previous work by assuming the distribution of solids around a solar mass central star follows a power law given by a surface density $\sigma = \sigma_o \left( \frac{r}{\mathrm{1~au}} \right)^{-p}$, where $\sigma_o$ is the surface density at a semimajor axis of $r=1$ au with index $p$. This prescription defines how much mass is available to build planets and how radially concentrated it is with respect to the star. Some fraction of this mass, $f_{e}$, is assumed to be present in the form of embryos which interact through gravitational interactions with all other bodies in the system. The remaining fraction of solids, $f_{p}=1-f_{e}$, is taken to be present as planetesimals, which are smaller, less massive objects that interact gravitationally with the embryos but not with each other (a common computational simplification for these type of simulations). All of these bodies are then taken to be distributed across an annulus ranging from an inner radius, $a_{min}$, to an outer radius $a_{max}$, beyond which it is assumed no other solids are present.
In all cases, we assume that embryos are spaced an average of $\Delta$=7.5 mutual Hill radii apart, and thus have initial masses and semi-major axes defined by the relations in @Kokubo2000Icar. For surface density slopes smaller than $p=2$, embryo mass increases with semi-major axis. Following @hansenmurray13, the embryos are placed on initial orbits that are nearly circular, with eccentricities randomly assigned values between 0 and 0.001, and orbital inclinations are randomly assigned from 0 to 2$^{\circ}$. The orientations of the orbits and the initial location of each body on it is determined from a uniform distribution to define the final set of orbital parameters. The initial orbital parameters of the planetesimals are defined in a similar manner, although in their case the individual planetesimal mass is fixed and semimajor axes are chosen so that their distribution follows the surface density profile. For the purposes of calculating the radii of the bodies of interest, we assume a density of $\rho$=3 g cm$^{-3}$. The inner edge of the simulations is set to half the semi-major axis of the innermost body at the start of the simulations, and the timestep is set to 1/20 of the orbital period at the cutoff distance. Bodies that pass interior to this cutoff distance are removed from the simulation.
Each of the simulations we perform is initialized by defining the parameters described above ($\sigma_{o}$, $p$, $f_{e}$, $a_{min}$, and $a_{max}$), and for each set of parameters many individual simulations are run with randomization of the orbital parameters as described above. We refer to each of these as a *Run Set*. The parameter values for all of the runs we consider here are listed in Table \[t:sim\]. Figure \[f:initial\] also shows how the initial conditions compare across the various Run Sets, comparing and contrasting the initial radial distribution of the embryos and their masses. Below we describe the broad categories of simulations that we performed.
![Starting surface density distribution for simulations with different initial conditions. Symbol size is proportional to embryo radius (assuming a constant density). The top panel shows simulations with different inner and outer radii, vertically offset for clarity. The bottom panel shows simulations with different surface density normalization factors and exponents. \[f:initial\]](initcond1-eps-converted-to.pdf "fig:"){width="\linewidth"} ![Starting surface density distribution for simulations with different initial conditions. Symbol size is proportional to embryo radius (assuming a constant density). The top panel shows simulations with different inner and outer radii, vertically offset for clarity. The bottom panel shows simulations with different surface density normalization factors and exponents. \[f:initial\]](initcond2-eps-converted-to.pdf "fig:"){width="\linewidth"}
[l|ccccccccc]{}\[t\] `Gen-HM` & 50.00 & 1.50 & 0.05 & 1.00 & n/a & 1.00 & 45 & 0 & 50\
`Gen-P` & 50.00 & 1.50 & 0.05 & 1.00 & n/a & **0.75 & 51 & 1027 & **8\
`Gen-I` & 50.00 & 1.50 & **0.20 & 1.00 & n/a & 1.00 & 20 & 0 & 50\
`Gen-P-I` & 50.00 & 1.50 & **0.20 & 1.00 & n/a & **0.75 & 23 & 733 & **8\
`Gen-IO` & 50.00 & 1.50 & **0.20 & **4.00 & n/a & 1.00 & 32 & 0 & 50\
`Gen-P-IO` & 50.00 & 1.50 & **0.20 & **4.00 & n/a & **0.75 & 36 & 2055 & **8\
`Gen-O` & 50.00 & 1.50 & 0.05 & **4.00 & n/a & 1.00 & 56 & 0 & 50\
`Gen-O-p1` & 50.00 & **1.00 & 0.05 & **4.00 & n/a & 1.00 & 79 & 0 & 50\
`Gen-O-p2` & 50.00 & **2.00 & 0.05 & **4.00 & n/a & 1.00 & 44 & 0 & 50\
`Gen-O-s22` & **22.50 & 1.50 & 0.05 & **4.00 & n/a & 1.00 & 84 & 0 & 50\
`Gen-O-s10` & **10.00 & 1.50 & 0.05 & **4.00 & n/a & 1.00 & 126 & 0 & 50\
`Gen-O-s5` & **5.00 & 1.50 & 0.05 & **4.00 & n/a & 1.00 & 178 & 0 & 50\
`Gen-M-s50` & **50.00 & 1.50 & 0.05 & **10.00 & **0.05 & 1.00 & 62 & 0 & 50\
`Gen-M-s22` & **22.50 & 1.50 & 0.05 & **10.00 & **0.05 & 1.00 & 93 & 0 & 50\
`Gen-M-s10` & **10.00 & 1.50 & 0.05 & **10.00 & **0.05 & 1.00 & 139 & 0 & 50\
****************************************************************
Embryo-only
-----------
We performed several sets of simulations that consisted only of massive embryos ($f_{e}$=1) based on the initial conditions used in @hansenmurray13 that were found to reproduce the broad properties of planetary systems in terms of planet radii, orbital periods, and multi-planet frequency in the *Kepler* catalog at the time. The nominal case (which we refer to as [`Gen-HM`]{}) is chosen to approximately reproduce their simulations, with surface mass density profile exponent $p$=3/2, normalization factor $\sigma_o = 50 \ \mathrm{g/cm^{2}}$ at 1 au, and embryos distributed between $a_{min}$ of 0.05 au and $a_{max}$ of 1 au. This gives 45 total embryos, with a total mass of 19 M$_{\bigoplus}$ (in comparison, @hansenmurray13 had 32-38 embryos and 20 M$_{\bigoplus}$, likely due to a larger spacing between embryos). In addition to the nominal case, we systematically explored a number of other initial conditions, varying the radial range over which building blocks were distributed ($a_{min}$ - $a_{max}$, simulations [`Gen-I`]{}, [`Gen-IO`]{}, and [`Gen-O`]{}), the normalized surface density $\sigma_o$ at 1 au ([`Gen-O-s22`]{}, [`Gen-O-s10`]{}, [`Gen-O-s5`]{}), and the exponent $p$ of the power law that describes how the mass is distributed with radius ([`Gen-O-p1`]{}, [`Gen-O-p2`]{}. The surface density is a factor of five higher than what is needed to produce the terrestrial planets in the solar system [e.g. @raymond06]. Most simulations were carried out for 10 Myr, though we also ran many for longer periods of time (up to 80 Myr) to explore how planetary system structure varies over these longer timescales. The orbital evolution simulations were performed using the SWIFTER version of the SyMBA integrator [@Duncan1998AJ], which handles close encounters and mergers[^3]. We assume that all collisions result in perfect mergers. Including fragmentation in the simulations would prolong the time-scale for planetary accretion, but generally results in similar planet planetary system properties as most collisional products are re-accreted on the same planet [@chambers13].
Planetesimals and Embryos
-------------------------
We also performed sets of simulations with a population of planetesimals distributed among the embryos ([`Gen-P`]{}, [`Gen-P-I`]{}, and [`Gen-P-IO`]{}). These small bodies have been shown to damp down the inclinations and eccentricities of forming planets through dynamical friction [e.g., @Wetherill1993Icar; @obrien06]. In these cases, we used the same total mass of solid materials as for our embryo-only cases, but distributed it with 75% being present as embryos initially ($f_{emb} = 0.75$) and the remaining 25% initially being contained in planetesimals. In our [`Gen-P`]{} case, which follows our nominal model, the planetesimal mass is distributed as 1000 distinct bodies, giving an individual planetesimal mass that is $\sim$1/20 of the smallest embryo ($\sim$0.005$M_{\oplus}$). The individual planetesimal mass in [`Gen-P-I`]{} and [`Gen-P-IO`]{} is similarly calculated as 1/20 of the mass of the smallest embryos in those simulations, with the number of such bodies varying according to the total mass in the simulation (see Table 1). We again used the SWIFTER version of the SyMBA integrator for these calculations.
Planet migration
----------------
A key issue that can also affect the final architecture of planetary systems is how much of the final stage of planetary assembly occurs while the gas from the protoplanetary disk is still present. To investigate the effects that nebular gas has on the emerging planetary systems, we performed several sets of simulations where we follow the dynamical evolution of embryos while they are embedded in a gaseous protoplanetary disk, leading to radial migration and damping of their eccentricities and inclinations over time. In these cases, we use a modified version of the Mercury integrator [@Chambers1999MNRAS] developed by [@izidoro17]. This code incorporates the 1-D time dependent density and temperature structure model of a viscous disk [@Bitsch2015AA], which was derived from fits to a 2-D numerical hydrodynamical and radiative transfer simulation. The dominant form of migration for embryo-sized bodies is Type-1 migration, which is modeled as in @Paardekooper2010MNRAS [@Paardekooper2011MNRAS], using methods developed by @Cresswell2008AA, @Coleman2014MNRAS, @Fendyke2014MNRAS, and @Papaloizou2000MNRAS. Tidal eccentricity and inclination damping of the embryos is also included [@Papaloizou2000MNRAS; @Tanaka2004ApJ; @Cresswell2006AA; @Cresswell2008AA]. The disk has a lifetime of 5 Myr, and the disk structure is updated every 500 years with a resolution of 0.05 au. The disk dispersal is described by an exponential decrease in disk surface density with a e-folding timescale of 0.01 Myr, see [@izidoro17] for details.
In the disk models, the relevant parameters are the metallicity (dust-to-gas ratio) of 0.01, adiabatic index $\gamma$=1.4 and alpha viscosity $\alpha = \mathrm{5.4 \times 10^{-3}}$. Together, they define the rate of mass and angular momentum transport through the disk, and the resulting temperature profile as a result of accretional heating and stellar irradiation. While the code is able to account for stochastic effects from MRI-driven turbulence, it is not included in the current simulations as [@izidoro17] found that it did not have a significant effect. The inner edge of the gas disk is set to 0.05 au (the same as the inner edge of the initial embryo distribution in [`Gen-HM`]{} and [`Gen-O`]{}), and the surface density near the inner edge is smoothed off using a hyperbolic tangent [@izidoro17 Eq. 4].
The specific setup of the gas disk model is one where planets migrate inward through a continuous disk and get trapped near the inner edge of the gas disk and in orbital resonances with each other. This scenario is consistent with the observed peak in the location of planetary system inner edges at $\sim{}10$ days [@epos1; @carrera19]. We do not explore alternate scenarios for trapping of planets farther out on the disk, for example if the inner disk structure is modified by MHD disk winds [@ogihara15].
{width="\linewidth"}
Architecture of Simulated Planetary Systems
===========================================
Our primary goal in running the suite of simulations described above is to understand how the initial conditions of the final stage of planet formation impact the properties of the planetary systems that formed. The physical framework for the final stages of planetary accretion is that as the initial swarms of planetesimals and embryos around a star begin to gravitationally interact, they perturb one another’s orbits, leading to orbital crossings and collisions. Early on, these collisions are frequent as the bodies begin tightly packed and low in mass, yielding many accretionary events. These events are more frequent close to the star, as the shorter orbital periods there allow for a greater number of orbital crossing events in a given time, increasing the likelihood of accretionary impacts. As bodies grow more massive, they begin to dominate their local annulus, either accreting other bodies or scattering them away from their neighborhood. As bodies begin to dominate their local annulus, their gravitational interactions with neighbors serve to keep these objects on orbits that are more widely spaced and dynamically stable. Collisions still occur as orbital perturbations can grow, but such events typically become rarer as the number of bodies in orbit around the star decreases.
While one might expect systems that begin with similar initial conditions to yield planetary systems with similar properties, it is important to account for the chaotic nature of this stage of evolution. That is, while planetesimals and embryos may begin this stage with identical orbital properties, the outcomes are very sensitive to where along their orbits the bodies begin (something that is randomized between each particular simulation). Thus for each Run Set listed in Table \[t:sim\], we performed multiple simulations in order to provide a statistical measure of the distribution of possible outcomes. In the embryo-only cases, we typically performed 50 realizations of the simulations; the runs with planetesimals were much more computationally demanding, and thus we typically performed 8 simulations for those cases.
[l|cccccc]{}\[t\] `Gen-HM` & $3.7^{+2.1}_{-1.4}$ & $1.2^{+0.94}_{-0.51}$ & $7.5^{+8.4}_{-1.5}$ & $2.7^{+1.1}_{-0.68}$ & $5.2^{+4.6}_{-3.2}$ & $4^{+1}_{-1}$\
`Gen-I` & $4.2^{+1.9}_{-1.4}$ & $0.9^{+0.98}_{-0.32}$ & $49^{+22}_{-7.3}$ & $2.6^{+0.89}_{-0.6}$ & $5^{+3.8}_{-2.7}$ & $2^{+1}_{-0}$\
`Gen-IO` & $8.9^{+2.3}_{-2.6}$ & $0.78^{+0.61}_{-0.21}$ & $78^{+11}_{-17}$ & $4.2^{+0.66}_{-1.4}$ & $14^{+9.7}_{-5.6}$ & $1^{+1}_{-0}$\
`Gen-P` & $2.2^{+1.1}_{-0.59}$ & $1.1^{+0.8}_{-0.41}$ & $6.1^{+0.45}_{-0.69}$ & $2^{+0.27}_{-0.33}$ & $1.8^{+2.3}_{-0.85}$ & $6.5^{+0.5}_{-0.5}$\
`Gen-P-I` & $2.8^{+0.66}_{-1.3}$ & $1^{+1.8}_{-0.28}$ & $38^{+5.3}_{-5.2}$ & $1.8^{+0.46}_{-0.29}$ & $1.1^{+1.3}_{-0.54}$ & $4^{+0}_{-0.89}$\
`Gen-P-IO` & $5.9^{+2.4}_{-3}$ & $1.3^{+1.3}_{-0.76}$ & $65^{+16}_{-16}$ & $3.6^{+0.71}_{-1.2}$ & $5.6^{+3.6}_{-2}$ & $2^{+0}_{-0}$\
`Gen-O` & $7.3^{+3.3}_{-2.4}$ & $1^{+0.5}_{-0.44}$ & $30^{+21}_{-13}$ & $4.4^{+4.7}_{-1.7}$ & $13^{+8.9}_{-6.2}$ & $2^{+1}_{-0.23}$\
`Gen-O-p1` & $7.1^{+4.5}_{-2.7}$ & $1.7^{+1}_{-1}$ & $54^{+49}_{-29}$ & $5.8^{+4.1}_{-3}$ & $16^{+14}_{-8.1}$ & $1^{+1}_{-0}$\
`Gen-O-p2` & $9.5^{+3.6}_{-2.4}$ & $1^{+0.55}_{-0.36}$ & $6.9^{+6.7}_{-1.3}$ & $3.4^{+1.5}_{-0.75}$ & $9.1^{+7.1}_{-5.5}$ & $3^{+1}_{-0}$\
`Gen-O-s22` & $2.2^{+1}_{-0.81}$ & $1.2^{+0.66}_{-0.54}$ & $11^{+13}_{-3.5}$ & $3.1^{+1.2}_{-0.88}$ & $9.2^{+7.5}_{-5.1}$ & $3.5^{+0.5}_{-0.5}$\
`Gen-O-s10` & $1^{+0.46}_{-0.34}$ & $1.3^{+0.6}_{-0.43}$ & $17^{+12}_{-9.4}$ & $3^{+1.3}_{-0.65}$ & $8.6^{+7.9}_{-3.9}$ & $3^{+1}_{-0}$\
`Gen-O-s5` & $0.17^{+0.094}_{-0.054}$ & $1^{+0.77}_{-0.43}$ & $5.5^{+0.5}_{-0.47}$ & $1.6^{+0.22}_{-0.2}$ & $3.4^{+2.7}_{-1.9}$ & $10^{+0.7}_{-1}$\
`Gen-M-s50` & $6.2^{+2.9}_{-2}$ & $1.1^{+0.45}_{-0.54}$ & $2.9^{+1.9}_{-0.92}$ & $1.5^{+0.17}_{-0.17}$ & $0.13^{+0.79}_{-0.12}$ & $10^{+2}_{-2.4}$\
`Gen-M-s22` & $2.3^{+2.6}_{-1.1}$ & $1.1^{+0.96}_{-0.58}$ & $3.3^{+9}_{-1.2}$ & $1.5^{+0.75}_{-0.25}$ & $0.31^{+2.5}_{-0.3}$ & $8.5^{+4.5}_{-4.5}$\
`Gen-M-s10` & $0.99^{+0.89}_{-0.5}$ & $0.99^{+0.72}_{-0.46}$ & $3^{+1.5}_{-0.81}$ & $1.4^{+0.39}_{-0.21}$ & $0.59^{+1.6}_{-0.55}$ & $10^{+7.5}_{-3}$\
{width="0.9\linewidth"}
Figure \[f:individual\] shows a sampling of the realizations produced within each Run Set. While variations in outcomes within a given Run Set are readily seen, it is clear that these differences are small compared to those between the different Run Sets we considered. This indicates that despite the stochastic nature of this stage of evolution, it is possible to identify signatures of the conditions under which planetary accretion took place.
Ultimately, we wish to use the relation between input parameters and realizations of planetary systems to inform the formation histories of the planetary systems found by *Kepler*. In order to do that, we must identify key diagnostics of the *Kepler* systems that can be compared to our models. Given the limited data we have on real systems, we have identified a set of broad, quantifiable parameters that describe the properties of these systems. Specifically, we use the median planet mass ($\overline{M}_p$), the mass ratio of adjacent planets ($\overline{\mathcal{M}}$), the median mutual inclination with respect to the invariable plane of the system, ($\overline{i}$), the median location of the innermost planet in the system, ($\overline{P}_\text{inner}$), the median orbital period ratio of adjacent planets, ($\overline{\mathcal{P}}$), and the median number of planets per system, $\overline{N}$. It is important to note that these parameters focus on the properties of the planetary *system*; that is we focus on the ensemble collection of planets, rather than the specific properties of individual planets, as these will likely inform us about how the planets interacted with one another during planet formation.
In order to ensure we are making the best comparisons between the models and the planetary systems, we focus on those planets in our simulations that are contained inside of 1 au. Because the detection efficiency of *Kepler* drops significantly with distance from the star, few planets are known to exist outside of that distance, making it difficult to make robust and meaningful comparisons. Note that we still allow planets to form in that region in our simulations; we just limit our statistics to those located closer to the star. Table \[t:numbers\] summarizes the median values and 1 $sigma$ ranges for each property for the Run Sets we considered while Figure \[f:tablehist\] provides a visual comparison of the outcomes, providing not only the characteristic values, but also their distributions.
The parameter distributions have mostly distinct shapes between run sets, with the exception of the mass ratio distributions which are more similar across all run sets. In order to quantify the similarities between run sets we calculate the probability that the model distributions are drawn from the same underlying distribution with a KS test. Table \[t:ks\] shows a set of comparions between all run sets and the base model, [`Gen-HM`]{}. The distributions of most parameters ($M_p$, $P_\text{inner}$, $\mathcal{P}$, $i$, and $N$) are distinguishable form the base set at high confidence ($p<0.01$). The exception is the mass ratio between adjacent planets, $\mathcal{M}$, which are statistically indistinguishable from the base set for the majority of run sets. We find the same trend when doing these comparisons between all possible combination of run sets: only the planet mass ratio distributions are often similar between models. None of the run sets are indistinguishable in all six parameter distributions, indicating that the variations in initial conditions indeed result in real differences in planetary system architectures that can be traced by these diagnostics.
[l|cccccc]{}\[t\] `Gen-I` & 0.038 & 0.008 & $<$0.001 & 0.252 & 0.477 & $<$0.001\
`Gen-IO` & $<$0.001 & 0.104 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-P` & $<$0.001 & 0.559 & 0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-P-I` & $<$0.001 & 0.280 & $<$0.001 & $<$0.001 & $<$0.001 & 0.997\
`Gen-P-IO` & 0.001 & 0.763 & $<$0.001 & 0.183 & 0.307 & $<$0.001\
`Gen-O` & $<$0.001 & 0.067 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-O-p1` & $<$0.001 & 0.028 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-O-p2` & $<$0.001 & 0.022 & 0.481 & $<$0.001 & $<$0.001 & 0.055\
`Gen-O-s22` & $<$0.001 & 0.374 & $<$0.001 & 0.025 & $<$0.001 & 0.272\
`Gen-O-s10` & $<$0.001 & 0.132 & $<$0.001 & 0.001 & $<$0.001 & $<$0.001\
`Gen-O-s5` & $<$0.001 & 0.062 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-M-s50` & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-M-s22` & $<$0.001 & 0.005 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
`Gen-M-s10` & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001 & $<$0.001\
In Figure \[f:tablehist\], we also show the approximate values and derived distributions for the *Kepler* exoplanets for reference. As we outline below, comparisons between the model realizations and the *Kepler* systems must be done with caution, and we outline a robust methodology in the next section. Before linking the properties of *Kepler* systems to particular formation histories, it is necessary to understand how the values of various properties are related to the initial conditions of the runs. Here we describe the effects of changes in the initial conditions on the realizations of planetary systems to provide a framework for interpreting the *Kepler* properties.
Effects of Including Planetesimals {#s:model:plts}
----------------------------------
Planetesimals in our models represent the leftover building blocks of planetary embryos, the more massive bodies that are the primary gravitational perturbers during the final stages of planetary accretion. Their incorporation into embryos is not expected to be 100% efficient, and as such, a number of them are expected to be present while planets form. The effects of planetesimals are seen when comparing the realizations from Run Set pairs [`Gen-HM`]{} & [`Gen-P`]{}, [`Gen-I`]{} & [`Gen-P-I`]{}, and [`Gen-IO`]{} & [`Gen-P-IO`]{}; each of these simulation pairs began with the same amount of mass present over the same radial distribution, but the latter had 25% of their mass contained in planetesimals and 75% in embryos, while the former was entirely embryos.
The primary effect of planetesimals is to provide a source of dynamical friction, where their orbits are preferentially excited to high eccentricities and inclinations while embryos and growing planets remain on more circular and roughly coplanar orbits. Inspection of Figures \[f:individual\] and \[f:tablehist\], as well as the numbers reported in Table \[t:numbers\] show that the inclinations of the planets formed in the embryo-only simulations are significantly higher than their planetesimal-containing pairs.
Another consequence of planetesimals being present during planetary accretion is that the number of planets that form in a given system increases. This effect is related to dynamical friction; as the embryos in the system maintain less-excited orbits, their eccentricities and inclinations are damped, limiting the semi-major axis range over which embros can collide. As a result, the feeding zones are narrower, leaving a larger number of bodies behind. This same process is also responsible for the lower masses of the planets in those cases where planetesimals were present: narrower feeding zones means fewer mass-adding events for any given planet (Figure \[f:final\]a).
Effects of Inner Radius {#s:model:inner}
-----------------------
The inner radius of our simulation defines the smallest orbital semi-major axis at which solid materials were available to form planets. In simulations for the accretion of the Solar System, this inner edge is often set at $\sim$0.5 au for computational purposes as closer-in orbits require finer time resolutions to resolve [e.g. @chambers98; @chambers2004; @obrien06; @raymond09]. Thus, these simulations largely focus on the accretion of planets outside of this location. However, given that planetary systems discovered by *Kepler* contain planets located much closer to their parent stars, this requires examining the details of planet formation at these very low orbital periods.
Run Set pairs [`Gen-HM`]{} & [`Gen-I`]{}, [`Gen-HM`]{} & [`Gen-P-I`]{}, and [`Gen-O`]{} & [`Gen-IO`]{} have identical properties to one another but vary the inner edge from 0.05 au (first simulation in pair) to 0.2 au (second). In all cases, the location of the innermost planet correlates with the inner edge of the disk, with those disks extending to 0.05 au yielding innermost planets with lower periods as a result of raw materials being present closer-in for planets to form from (Figure \[f:final\]b). As a result, the number of planets formed within 1 au is larger for smaller inner disk edges.
Variations in planet masses, inclinations, and period ratios are relatively minor among the Run Set pairs (Table \[t:numbers\]). In particular, the masses of planets outside of 0.2 au are not affected by the location of the disk inner edge (Figure \[f:final\]b)
{width="\linewidth"}
Effects of Outer Radius {#s:model:outer}
-----------------------
With most known extrasolar planets, particularly those in the Kepler database, found in low period orbits, many simulations for their formation have focused on the assembly of local materials inside of 1 au [@hansenmurray12; @hansenmurray13; @dawson15; @dawson16]. However, given the decreased detection probability further from the star, it remains uncertain what the distribution of mass at larger distances may be. To investigate the effects of mass at larger distances, we performed simulations where the outer edge of the solids extended to 4 au. The outer edge of where mass may be located is likely set by a combination of the radial extent of the disk as well as how solids migrate prior to being incorporated into planetesimals.
Comparisons of Run Set pairs [`Gen-HM`]{} & [`Gen-O`]{}, [`Gen-P-I`]{} & [`Gen-P-IO`]{}, and [`Gen-I`]{} & [`Gen-IO`]{} show how the extended disk impacts the final realizations. In general, the more extended disk leads to fewer planets forming inside of 1 au, though those planets are more massive than those formed in the more truncated disks (Fig. \[f:final\]c). This is likely due to the greater amount of mass in embryos and planets at larger distances, which can serve to excite the orbits of bodies close in, as well as scatter mass inward to be accreted by the planets growing in this region. The larger excitations among the growing bodies lead to a greater spacing of the planets inside of 1 au, which – in addition to reducing the number of such bodies – leads to increases in the ratio of orbital periods. Further, mutual inclinations increase as the orbits become more excited.
Thus, the amount of material in the outer disk affects the properties of planets formed inside of 1 au, consistent with the study of @ballard16.
Effects of Surface Density Profile {#s:model:sdp}
----------------------------------
The initial surface density profile of solids sets the total mass of embryos and planetesimals initially present in the simulation and how that mass varies with distance from the star. As described above, this is usually defined as a power-law, and thus we examined two variations of how the surface density profile of solids impacted the final realization of planets. Different surface densities may arise as disks have their masses distributed in different ways, or as radial migration of solids leads to the large-scale redistribution of solids prior to planetesimal formation.
In the first case, we considered how the surface density steepness, or exponent of the power-law, impacted the outcome of our models. These effects can be seen by comparing the realizations of [`Gen-O`]{}, [`Gen-O-p1`]{}, and [`Gen-O-p2`]{} where the surface densities profiles varied as $r^{-\frac{3}{2}}$, $r^{-1}$, and $r^{-2}$, respectively. The mass ratios of adjacent planets directly correspond to the slope of the surface density distribution. A slope of $r^{-2}$ leads to roughly equal-mass planets ($\mathcal{M} = 1$), while a shallow slope of $r^{-1}$ leads to the outer planets being more massive ($\mathcal{M} \approx 1.7$).
Comparing the realizations of these different runs, it is seen that more planets, with somewhat higher masses, are produced in the steeper surface density profiles, a result also seen by @raymond05. While the total amount of mass in the simulations is greater in the shallow surface density profiles, as they extend out to 4 au, the amount interior to 1 au, where the surface density is normalized to $50$ g/cm$^2$, is greater in the steep surface profiles, thus providing more raw materials for planets to form at smaller heliocentric distances. The greater mass at larger heliocentric distances in the shallower profiles still impacts the dynamics of the interior planets, exciting them to higher inclinations and resulting in greater radial scattering of the interior planets, leaving the inner-most planet further from the star (Fig. \[f:final\]d).
We also examined how the total mass of solids present in the simulations affected the properties of the final planetary system. Model Runs [`Gen-O`]{}, [`Gen-O-s22`]{}, [`Gen-O-s10`]{}, and [`Gen-O-s5`]{} all had the same surface density steepness (varying as $r^{-\frac{3}{2}}$), but the surface density at 1 au varied, yielding different amounts of total mass to be accreted into the planets. As shown in Table 2 and Figure \[f:final\], the higher surface density runs yielded more massive planets as more material was present to be accreted. The more massive planets, due to their greater gravitational effects, cleared a larger amount of area around their orbits, leading to larger spacing between the planets, and thus fewer of them. As the total mass decreased, the amount of dynamical excitation through the system decreased, with very weak gravitational interactions taking place, leaving a large number of low mass planets behind. Similar outcomes were reported by @kokubo06 and by @raymond07 and @ciesla15 in looking at the formation of planetary systems around stars of different masses: greater numbers of smaller planets were produced around stars that had lower protoplanetary disk masses. The trend of less gravitational excitement of bodies in the lower mass systems also leads to predictable results for the other quantities, with lower values for inclinations, inner planet periods, and period ratios in these same systems.
Effects of Gaseous Disk {#s:model:gas}
-----------------------
The simulations discussed thus far have been performed assuming that the gas from the protoplanetary disk was gone by the time this stage of planetary assembly began. However, as discussed above, some of this growth may have actually occurred early in the evolution of the system while gas was still present. Such early growth is necessary in our Solar System to explain the formation of the gas giants through core accretion [e.g. @pollack96; @hubickyj05] and the possible early formation of Mars relative to the typical lifetime of protoplanetary disks [@dauphas11]. In addition, low bulk densities of mini-Neptunes [@wu13; @hadden14] and the presence of a valley in the planet radius distribution [@fulton17; @owen17] indicate that the *Kepler* planets likely formed in a gaseous environment.
Including disk gas in our simulations would have a dramatic effect on the dynamical evolution of planetary systems as gravitational interactions between embryos and the disk would lead to angular momentum exchange and the inward migration of the embryos. While a single migrating planet would stall at the inner edge of the gas disk, convergent migration of planets leads to inner planets in run sets migrating inside the disk inner edge, for example for run set [`Gen-M-s50`]{} in Fig. \[f:final\]f (see also @carrera19).
In addition to migration, the gas also serves to damp down the eccentricities and inclinations of embryos. This is reflected in the low mutual inclinations of planets in run sets [`Gen-M-s50`]{}, [`Gen-M-s25`]{}, and [`Gen-M-s10`]{} (Fig. \[f:tablehist\].) and the large number of planets that form (as they do not get excited onto crossing orbits). As a result, the masses of individual planets tend to be much lower, but there is a greater number of planets in a given system. Further, the gas-driven migration and damping often drives planets into quasi-stable structures, thus such planets will settle into tightly-packed systems with low period ratios.
![Time-evolution of mutual inclinations in run set [`Gen-M-s22`]{}, which includes gas damping and migration. The disk disperses at $5$ Myr, and system-wide instabilities occur within a few million years for a fraction of systems, after which systems remain fairly stable. \[f:evolution\]](fig5.pdf){width="\linewidth"}
{width="0.9\linewidth"}
While some level of accretion may occur in the presence of gas, protoplanetary disks are estimated to retain significant gas for only 3-5 Myr before the gas is lost via photoevaporation or other processes [e.g. @ercolano17]. Here we assume the gas is present for $\sim$5 Myr, though the gas density does decrease over time (before rapidly disappearing at the end). In the absence of the gas, dynamical interactions and accretion occur as in the previously described cases. In some cases, these interactions are minimal, leaving a tightly-packed, stable system of planets. In other cases, the dynamical interactions are such that the systems go through a period of instability, leading to large increases in mutual inclination. Thus, in our simulations, as in @izidoro17 [@izidoro19], we generally see a dichotomy in this outcome (Fig. \[f:evolution\]), with a bimodal distribution as some systems have very low mutual inclinations while others have significantly elevated values, with few systems with mutual inclinations in between. While this is not readily apparent when looking at the median values in Table \[t:numbers\], it is more obvious when looking at the distributions of run sets [`Gen-M`]{} shown near the bottom rows of Figure \[f:tablehist\], especially for [`Gen-M-s22`]{}. Although most of our simulations are relatively short ( 10 Myr), the result of the bimodal inclination distribution is likely to hold over the lifetimes of the systems: Running the simulations up to 100 Myr did not lead to a large number of these low-inclination systems becoming dynamically unstable and increasing their mutual inclinations.
It is worth noting that the presence of gas can impact the relations that develop compared to those cases where gas is not present. For example, in the higher mass surface density runs with no gas, we find that higher inclinations develop as a result of the gravitational interactions between the planets; those cases with lower surface densities have systems with lower inclinations. The opposite is true when gas is present. This is likely due to accretion occurring more rapidly when surface densities are high; given that a large amount of the gravitational interactions would take place early in the massive cases, the gas is still present to damp down any excitation that develop. As models with lower surface densities would have more protracted planet growth, the gas dissipates before planetary assembly has progressed to a similar level as in a high surface density disk, allowing the planets in the system to become more inclined. Alternatively, more massive planets experience stronger tidal damping from the gas disk.
Summary of different effects
----------------------------
The chaotic nature of planetary accretion means that even with similar starting conditions, the final properties of the emerging planetary systems may differ significantly from one another (Figure \[f:individual\]). However, given a large enough suite of simulations, it is clear that trends and relationships between the final properties and the initial conditions can be recognized (Figs. \[f:tablehist\], \[f:final\]). In this section, we have identified and explained how those trends develop for a set of final planetary system properties that can be inferred from the *Kepler* planetary systems.
Figure \[f:nbody\_summary\] summarizes how changes in individual parameters would lead to changes in the distribution of planetary systems that form. While none of our specific *Run Sets* exactly reproduce the properties of the *Kepler* systems, it is possible to use the trends and relationships that we have outlined here to adjust the parameters and components to provide a better match between the modeled and real systems. However, it is necessary to ensure that a proper comparison between the model realizations and *Kepler* systems is done, requiring a rigorous consideration of observational biases.
{width="\linewidth"}
Statistical Comparison to the Kepler Planet Population
======================================================
In the previous section, we discussed how various accretion scenarios lead to significantly different distributions of planetary system properties. In general, the model systems did reproduce some of the features of the *Kepler* populations, yielding properties that overlapped with what is observed – in particular for the model that includes planetesimals, [`Gen-P`]{}. However, these comparisons must be made with care, as they are done with knowledge of the full properties of the planetary systems that formed, whereas our knowledge of the observed *Kepler* system properties is limited by a number of observational biases. Thus, a detailed comparison between the systems produced in our models must account for these limitations in order to determine which values of the parameters for our synthetic populations would actually be detected (Fig. \[f:arrows\]) and whether the *Kepler* observarions can be used to distinguish between the different accretion scenarios.
To do the forward modeling, we use `epos` version 3.0 [@eposv3], the Exoplanet Population Observation Simulator, whose structure and detailed methodology have been described in @epos1 [@epos2]. In short, `epos` performs a synthetic transit survey by assuming that the planetary populations produced by models make up the real distribution of planets, then performs a simulated transit survey of these systems, assuming that systems are randomly oriented along our sight lines and applies all detection biases of the *Kepler* spacecraft. We then compare the properties of the planets detected in the simulated survey to those of the systems detected by *Kepler*, which are representative of planetary systems around solar-mass stars. This approach, illustrated in Figure \[f:arrows\], allows for an apples-to-apples comparison between simulated planetary systems and the data, without having to parameterize the distribution of planetary system properties.
We use four observational diagnostics of planetary system properties. The first three diagnostics were defined in @epos2, namely the observed frequency of multi-planet systems (§\[s:multi\]), the location of the innermost observed planet in each system (§\[s:inner\]), and the period ratio of adjacent planets (§\[s:ratio\]). The fourth diagnostic is the size ratio of adjacent planets (§\[s:cluster\]), which has recently been shown to be strongly clustered in *Kepler* planetary systems [@hsu19; @sandford19].
An example of this observational comparison for run set [`Gen-HM`]{} is shown in Figure \[f:arrows\]. The five panels in the top row show the distribution of planetary system properties from the planet formation models that are also shown in Fig. \[f:tablehist\]. The bottom panel show the four observational diagnostics calculated by `epos` using a forward modeling approach of the detection biases, compared to the same diagnostic calculated for the exoplanet candidates from *Kepler*.
![Mass-Radius distribution of simulated planetary systems for [`Gen-HM`]{} (blue-ish). The planet radii are drawn using a Monte Carlo simulation from the distribution of @chen17 (dark blue), where the $1\sigma$ dispersion is indicated with the dashed lines. The constant density model assumed in the N-body integrator is shown in gray, which $\rho=3\,g/cm^3$. \[f:MR\] ](fig8.pdf){width="\linewidth"}
An extra step that we perform compared to @epos1 [@epos2] is that the planet radii are calculated from the simulated masses using the planet mass-radius relation from [@chen17]. Figure \[f:MR\] shows this relation and its $1\sigma$ dispersion. For the full forward modeling in the synthetic survey, we draw the radii from a log-normal distribution centered on the best-fit mass-radius distribution using a Monte Carlo simulation. For reference, the constant density of 3 g/cm$^3$ as used in the N-body simulation is also shown, which appears to be a reasonable value for mini-Neptunes as well.
For most runs, especially those where the inner disk edge $a_\text{min}$ is set to $0.05$ au, the subset that is compared to contains the majority of detected planet candidates, and thus a detailed statistical comparison can be made. In general, we find that the run sets with significantly different distributions of planetary system properties (Table \[t:ks\]) are also distinguishable in these observational diagnostics. This demonstrates that while some information on planetary system architectures is inevitably lost due to transit survey biases, different formation scenarios can still be constrained from the observed *Kepler* multi-planet systems.
For certain model runs, however, the overlap with the *Kepler* planet candidates is limited because few planets are formed at short orbital periods and the probability of detection of multiple planets per system is extremely low. In those cases the sub-set of observed planetary systems is too small to allow a meaningful comparison on multi-planet statistics. This is the case for runs with an inner disk edge at $a_\text{min}=0.2$ au ([`Gen-I`]{}, [`Gen-IO`]{}, [`Gen-P-I`]{}, [`Gen-P-IO`]{}) and the run set with a very flat surface density power-law ([`Gen-O-p1`]{}) that need to be compared to *Kepler* planet candidates outside of a $\sim40$-day period. The data in this range contains only a handful of observed 2-planet systems, making them less likely to be relevant to the collection of planetary systems we are interested in. Thus, we omit these simulations from the following analysis.
![Observable multi-planet frequencies for different accretion scenarios. The red dashed line shows the observed frequencies from *Kepler* in the same range of planet radii ($1-3\,R_\oplus$) and orbital periods ($5-300$ days) as the simulated planets. The top panel shows the observable multi-planet frequencies for four different models to highlight the role of additional stirring when enlarging the disk outer edge ([`Gen-O-s22`]{}) compared to the base simulation ([`Gen-HM`]{}) and the role of additional damping by planetesimals ([`Gen-P`]{}) and gas ([`Gen-M-s22`]{}). The bottom panel shows how the multiplicity of planetary systems in the formation model (bars) is completely different from the observable distribution of transiting systems (plus signs) due to non-transiting planets. \[f:epos:multi\]](fig9a.pdf "fig:"){width="\linewidth"} ![Observable multi-planet frequencies for different accretion scenarios. The red dashed line shows the observed frequencies from *Kepler* in the same range of planet radii ($1-3\,R_\oplus$) and orbital periods ($5-300$ days) as the simulated planets. The top panel shows the observable multi-planet frequencies for four different models to highlight the role of additional stirring when enlarging the disk outer edge ([`Gen-O-s22`]{}) compared to the base simulation ([`Gen-HM`]{}) and the role of additional damping by planetesimals ([`Gen-P`]{}) and gas ([`Gen-M-s22`]{}). The bottom panel shows how the multiplicity of planetary systems in the formation model (bars) is completely different from the observable distribution of transiting systems (plus signs) due to non-transiting planets. \[f:epos:multi\]](fig9b.pdf "fig:"){width="\linewidth"}
Number of Planets Observed in Systems {#s:multi}
-------------------------------------
The first diagnostic discussed here is the relative frequency of observed multi-planet systems, i.e., the observed number of stars with $m$ planets. Because most planets in multi-planet systems do not transit, this diagnostic does not always reflect the true multiplicity of systems, $N$. While most exoplanet host stars in the *Kepler* survey only have a single transiting planet detected, statistical studies show that most of these are members of planetary systems with multiple non-transiting planets [e.g. @lissauer11; @zink19]. Instead, the relative frequencies are most sensitive to the mutual inclinations, with a decreased chance of observing more planets per star if the mutual inclinations are higher.
To show what the observable signature of the different planet accretion scenarios would look like, we compare four run sets that all produce planets in the same size range ($1-3\,R_\oplus$) and orbital period range ($5-300$ days) in Figure \[f:epos:multi\]. Run [`Gen-HM`]{} represents an embryo-only scenario, [`Gen-P`]{} a planetesimal scenario, [`Gen-O-s22`]{} a large disk scenario, and [`Gen-M-s22`]{} a migration/damping scenario. The simulated distributions of multi-planet frequencies of these four run sets are significantly different at high confidence ($p<0.01$), showing that these diagnostics can indeed be used to discriminate between the different accretion scenarios shown in Figure \[f:tablehist\]. The observational signature of these runs are then compared to the observed frequencies of multi-planet systems from *Kepler* in the same ranges of planet radii and orbital periods.
The embryo-only simulations of [`Gen-HM`]{} under-predict the frequencies of observed 2, 3, and 4-planet systems compared to *Kepler*, likely because their mutual inclinations are significantly larger than the 2 degrees typically inferred for *Kepler* multi-planet systems (see also Fig. \[f:tablehist\]). While [@hansenmurray13], on which the [`Gen-HM`]{} runs are based, found a good match to the observed ratio of multi-planet systems, they compared their simulated planetary systems to an earlier version of the *Kepler* catalog that had fewer 3- and 4-planet systems. Thus, our model results are consistent with theirs, even though we find no good match to the updated *Kepler* systems.
Models from the [`Gen-O`]{} series in which the outer disk is extended to a semi-major axes of $4$ au, of which [`Gen-O-s22`]{} is shown here, under-predict the number of observable 2 and 3-planet systems compared to the observations by a larger amount than [`Gen-HM`]{}. This is likely due to the increased dynamical stirring from the outer disk mass leading to larger mutual inclinations, which decreases the chance of multiple planets transiting from a given viewing angle (see also @moriarty16). These conclusions extend to models [`Gen-O`]{}, [`Gen-O-p1`]{}, and [`Gen-O-s10`]{} (not shown) which have a similar distribution of planet frequencies and are not distinguishable in this diagnostic. The multi-planet frequencies of model [`Gen-O-p2`]{} are more similar to those of [`Gen-HM`]{} because less mass is added to the outer disk, minimizing its impact on the mutual inclinations and hence the multi-planet frequencies. The model with the lowest surface density ([`Gen-O-s5`]{}) has less dynamical stirring and hence lower mutual inclinations, but a direct comparison with the data is difficult because the small planet sizes yield a very small sample of detected multi-planet systems to compare to.
A better match to the observed multi-planet frequencies is seen in simulations that include planetesimals where dynamical friction reduces the mutual inclinations. [`Gen-P`]{} has a larger number of predicted 2–5 planet systems compared to [`Gen-HM`]{}, and over-predicts the observed frequencies by a small fraction. The ratio of 3/2, 4/3 and 5/4 planet systems is the same as observed, and hence this model would closely match the *Kepler* data if a population of single or highly inclined planets would be included to account for the *Kepler* dichotomy, as described in @johansen12, @hansenmurray13, and @ballard16.
The inclusion of damping by gas leads to mutual inclinations that are so low that the runs over-predict the frequency of multi-planet systems. For example simulation [`Gen-M-s22`]{}, shown in Figure \[f:epos:multi\], which has planets in a similar size range as [`Gen-HM`]{} and [`Gen-P`]{}, over-predicts the number of 3- and 4-planet systems. Moreover, it predicts a large number of 5-8 planet systems be observed which are not present in the *Kepler* data in this size and period range. Simulation run sets that include migration in combination with higher and lower surface densities ([`Gen-M-s50`]{} and [`Gen-M-s10`]{}, respectively) similarly over-predict the number of observable multi-planet systems due to their low mutual inclinations.
Period Ratios of Adjacent Planets {#s:ratio}
---------------------------------
![Period ratio distributions in a simulated survey of the different models (solid lines) compared to the *Kepler* observations (dashed line). These comparisons highlight the role of additional stirring when enlarging the disk outer edge ([`Gen-O-s22`]{}) compared to the base simulation ([`Gen-HM`]{}) and the role of additional damping by planetesimals ([`Gen-P`]{}) and gas ([`Gen-M-s22`]{}). \[f:epos:dP\]](fig10a.pdf "fig:"){width="\linewidth"} ![Period ratio distributions in a simulated survey of the different models (solid lines) compared to the *Kepler* observations (dashed line). These comparisons highlight the role of additional stirring when enlarging the disk outer edge ([`Gen-O-s22`]{}) compared to the base simulation ([`Gen-HM`]{}) and the role of additional damping by planetesimals ([`Gen-P`]{}) and gas ([`Gen-M-s22`]{}). \[f:epos:dP\]](fig10b.pdf "fig:"){width="\linewidth"}
The second diagnostic discussed here is the period ratio distribution of detected planet pairs, which mainly traces the orbital spacing between planets. This period ratio is defined as $\mathcal{P}=P_{k+1}/P_k$ where $k$ enumerates observed planets from inside out. This diagnostic is calculated for planets in the range of $1-3 R_\oplus$ and $5-300$ days. The detected period ratios of four models representative of different accretion scenarios are shown in Figure \[f:epos:dP\] as a cumulative distribution to highlight the excess (or absence) of resonant features. The bottom panel shows how the detected distributions are biased with respect to the distribution predicted by the formation models. The observed distribution from *Kepler* spans a range of period ratios from $\sim$1.3 to almost 10, with a median value at $\mathcal{P} \approx 2.2$.
The period ratio distribution of the different accretion scenarios (Fig. \[f:epos:dP\]) follow a similar trend to that of the multi-planet frequencies: Simulations with the strongest damping (by gas) underpredict the observed period ratios and simulations with stronger dynamical interactions from larger disks progressively over-predict the observed period ratios.
The baseline simulation of [`Gen-HM`]{} (and also [`Gen-O-p2`]{}) has a median period ratio of $\sim3.3$ and the distribution is located at longer period ratios compared to the observations. The amount of dynamical stirring in the embryo-only simulations leads to planets that are spaced too far apart to be consistent with the *Kepler* systems, as also noted by [@hansenmurray13]. The simulations with increased mass in the outer disk ([`Gen-O-s22`]{}, [`Gen-O`]{}, [`Gen-O-p1`]{}) all have similar distributions with even larger period ratios due to the increased dynamical interactions. The simulations with a lower overall surface density ([`Gen-O-s10`]{} and [`Gen-O-s5`]{}) have smaller period ratios, but because these simulations also produce smaller planets, there are not enough observed planets to allow a statistically meaningful comparison with the *Kepler* exoplanet population.
When dynamical friction from planetesimals are included in the simulations, the emerging planets are more tightly packed. For example the predicted period ratio distribution of simulation [`Gen-P`]{} has a mean period ratio of $\mathcal{P} \approx 2.4$, just higher than the observed distribution from *Kepler*. The shape of the distribution is roughly similar to the observed period ratio distribution, and is the closest match to the data of all the simulated systems. We did not explore if a larger number or mass fraction of planetesimals could further reduce the period ratios and provide an even better match to the observed systems.
Simulations including orbital damping by gas and planet migration, on the other hand, result in observable planetary systems that are too compact. For example, simulation [`Gen-M-s22`]{} has a mean period ratio of $\mathcal{P}=1.5$ . The distributions also show strong features at the main orbital resonances (2:1 and 3:2) that are not present in the *Kepler* data. Dynamical instabilities that break these resonances and widen the period ratio distribution are needed to provide a better match to the observed systems, similar to recent results by [@izidoro19].
The solid surface density has a big impact on the observed period ratio distribution of planets formed in a gaseous disk. Simulations with a higher (lower) surface density of solids grow and migrate faster (slower) and therefore experience more (less) damping by gas and less (more) breaking of resonant chains. For example, simulation [`Gen-M-s50`]{} with the highest solid surface density show stronger resonant features and smaller orbital periods and is therefore less consistent with the observed period ratio distribution. On the corollary, the simulation with the lowest surface density, [`Gen-M-s10`]{}, has fewer planets in orbital resonances and a period ratio distribution that is only slightly shortwards of the *Kepler* distribution. While the period ratio distributions of the three migration run sets are statistically distinguishable, they are all inconsistent with the observed distribution from *Kepler*.
![Period of the innermost planet in an observable multi-planet system (solid). The observed distribution from *Kepler* for planets in the same radius range ($1-3~R_\oplus$) is shown in dashed red. These comparisons highlight the role that migration and disk size play in determining the location of the innermost planets. All models utilize the same inner disk edge of 0.05 au. Planet migration moves the innermost planet inwards to the edge of the gas disk which is located at 0.05 au. A large outer disk results in inner planets at larger separations ([`Gen-O`]{}) compared to a $1$ au disk ([`Gen-HM`]{}) unless the outer disk contains less mass ([`Gen-O-p2`]{}). \[f:epos:Pinner\]](fig11a.pdf "fig:"){width="\linewidth"} ![Period of the innermost planet in an observable multi-planet system (solid). The observed distribution from *Kepler* for planets in the same radius range ($1-3~R_\oplus$) is shown in dashed red. These comparisons highlight the role that migration and disk size play in determining the location of the innermost planets. All models utilize the same inner disk edge of 0.05 au. Planet migration moves the innermost planet inwards to the edge of the gas disk which is located at 0.05 au. A large outer disk results in inner planets at larger separations ([`Gen-O`]{}) compared to a $1$ au disk ([`Gen-HM`]{}) unless the outer disk contains less mass ([`Gen-O-p2`]{}). \[f:epos:Pinner\]](fig11b.pdf "fig:"){width="\linewidth"}
Period of the Innermost Planet in Each System {#s:inner}
---------------------------------------------
The third diagnostic discussed here is the location of the innermost detected planet in an observed multi-planet system. For each system with more than one planet detected in the simulated survey, we recorded the location of the innermost detected planet. The distribution of these values is then compared to the locations of the innermost planets in the *Kepler* multi-planet systems. Figure \[f:epos:Pinner\] shows the inner planet locations for four different models in the range $1-3~R_\oplus$. Note that this is a different selection of models than in Figures \[f:epos:multi\] and \[f:epos:dP\], and was chosen to highlight the parameters that most affect the location of the inner planets.
The distributions before and after applying detection biases, shown in the bottom panel of Fig. \[f:epos:Pinner\], differ primarily because two detection biases are at play. First, inner planets at shorter orbital periods have an increased transit probability, which shifts the distribution of detected inner planets towards shorter periods compared to the instrinsic distribution of inner planets. Second, in a fraction of systems the inner planet is not transiting, and the innermost detected planet is located farther out, effectively broadening the distribution. `epos` takes both these biases into account when calculating the detectable distributions.
We find that the period of the observed innermost planet varies significantly even between simulations that have the same inner disk edge but differ in the other initial conditions (Fig. \[f:epos:Pinner\]). The fiducial simulations of [`Gen-HM`]{} and [`Gen-P`]{} predict an observable inner period ratio distribution with roughly the same mean value as found in the *Kepler* data, and are not statistically distinct from each other. However, in both cases the distributions are much narrower than observed. This difference in the widths of the simulated and observed distributions could indicate that a range of inner disk locations, rather than a single value, is needed for this model to match the observed inner period distribution. [@lee17] have suggested that the observed spread in stellar rotation rates of young stars leads to such a range if the inner disk edge is located at the radius corresponding to the stellar co-rotation. This was also suggested by [@mulders15a] based on the stellar-mass (in)dependence of the period-break in the planet occurrence rates.
Increasing the mass available by extending the outer disk (models [`Gen-O`]{}, [`Gen-O-p1`]{}) increases the amount of dynamical interactions between planets in the inner disk and also pushes the location of the innermost planets outward (see Table \[t:numbers\]). The latter is reflected in the simulated observables, where the median inner period of model [`Gen-O`]{} is almost twice that of model [`Gen-HM`]{} (Fig. \[f:epos:Pinner\]). The inner planet orbital period distribution is also wider than that of the fiducial model, indicating that a better fit to the data might be obtained with a larger disk but with a smaller inner disk edge. Simulation [`Gen-O-p2`]{} also has an enlarged outer disk but less mass is added in this case because of the steeper surface density power-law, and this model therefore produces a comparable inner planet orbital period location distribution as run set [`Gen-HM`]{}.
Adding gas to the disk makes planets migrate inward, pushing the observable distribution of innermost planets inward. The median inner period of model [`Gen-M-s22`]{} is more than half that of the baseline model, with a median observable period at 3 days. The simulated distribution is narrower and located at shorter orbital periods as the observed inner planet distribution. A model where the inner edge of the gas disk is located further out at an orbital period of 10 days [e.g. @epos1; @carrera19], and has a dispersion in locations, would likely provide a better match to the observed *Kepler* systems.
Radius Ratios of Adjacent Planets {#s:cluster}
---------------------------------
![Radius ratio distributions in a simulated survey of the different models (solid lines) compared to the *Kepler* observations (dashed line). \[f:epos:dR\]](fig12a.pdf "fig:"){width="\linewidth"} ![Radius ratio distributions in a simulated survey of the different models (solid lines) compared to the *Kepler* observations (dashed line). \[f:epos:dR\]](fig12b.pdf "fig:"){width="\linewidth"}
The fourth diagnostic discussed here is the radius ratio distribution of adjacent detected planets, which mainly traces the clustering in planet sizes. This radius ratio is defined as $\mathcal{R}=R_{k+1}/R_k$ where $k$ enumerates observed planets from inside out. The top panel of Figure \[f:epos:dR\] shows the cumulative distribution of radius ratios for four different models in the simulated survey. Note that this is again a different selection of models than in Figures \[f:epos:multi\], \[f:epos:dP\] and \[f:epos:Pinner\], because the model parameters that affect the distribution of planet radii most are different. The observed distribution from *Kepler* for planets in the range of $1-3\,R_\oplus$ and $5-300$ days spans a range of radius ratios from $\sim$0.5 to almost 3. The distribution is slightly asymmetric with a median at $\mathcal{R} \approx 1.1$.
The bottom panel in (Fig. \[f:epos:dR\]) shows how the size ratio distributions are affected by observation biases. The main detection bias here arises because the minimum detectable planet size increases with orbital period because fewer transits occur at large periods. A planet pair where the outer planet is larger is more likely to be detected than a planet pair where the outer planet is smaller. Thus, the size ratio distribution in the synthetic survey is shifted to larger size ratios than those in the planet formation model.
The radius ratios of the different accretion scenarios, subjected to observational biases, are shown in Figure \[f:epos:dR\]. The simulated distributions typically have a similar shape as the observations in each scenario, indicating that the radius clustering of planet sizes observed with *Kepler* is naturally produced by these simulations. But they differ in how much they are offset toward the larger planet radius ratios. In particular, simulations [`Gen-HM`]{} and [`Gen-P`]{} are offset to larger size ratios compared to the observations.
In a scenario without migration, trends in the size ratio of planets reflect the mass distribution in the initial disk. Assuming the width of a planet’s feeding zone is proportional to its semi-major axis $a$, the surface area of the annulus that a planet accretes from scales as $a^2$. Models with a radial surface density exponent of $p=1.5$ are therefore expected to form planets whose mass increases with semi-major axis as $a^{0.5}$, and are thus expected to have size ratios that are on average larger than $1$. Planets formed in a disk with a radial surface density exponent of $p=2$ are expected to be of similar size at different distances from the stars. We can indeed see that simulation [`Gen-O-p2`]{} does not show a significant offset compared to the observations and is a better match than simulations [`Gen-HM`]{} and [`Gen-P`]{}.
Finally, we find that simulations with migration are almost a perfect fit to the observations, showing both the same offset and dispersion in the simulated planet radius distribution as in the observations. This is perhaps slightly surprising as the underlying surface density profile used in these simulations is the same as in the gas-free simulations. Likely, the inward migration and resonant trapping changes the spacing between forming planets such that the final planetary systems end up closer to equal in mass.
Summary
=======
We have explored the effects that different initial conditions and model configurations have on the final stages of planetary accretion. While the process itself is chaotic, we have identified relationships between the initial conditions and model configurations and the final planetary system properties, including the number of planets, their relative sizes, their orbital spacings, and their mutual inclinations that provide an imprint of their formation histories. Such statistical relationships can be used to inform us about the possible formation pathways of known exoplanetary systems. Doing so, however, requires a detailed assesment of the observational biases to properly reveal the properties of these systems.
In applying our knowledge to the known *Kepler* systems, we find that:
1. Planets within planetary systems are typically similar in size, despite the chaotic nature of the late stages of planetary assembly. In every scenario we explore, whether that includes migration of planets in a gas disk or not, the distribution of planet masses within a system are consistent with the clustering of planet radii observed with *Kepler*.
2. Planet formation did not occur purely through the assembly of planetary embryos as such systems would be too dynamically excited compared to the real systems. Leftover planetesimals that were present during the final stages of accretion would serve as a source of dynamical friction that provide better matches to the multi-planet frequencies and inclinations, though the simulations performed here still yield spacings between detected planets that were larger than observed.
3. Planetary systems where a significant amount of accretion occurs within a gaseous protoplanetary disk, are too dynamically cold to match the observed exoplanet population, with period ratios that are too small and have an overabundance of multi-planet systems. In contrast, systems that undergo dynamical instabilities after the disk has dispersed are dynamically hotter than those that stay stable, and the planets’ mutual inclinations within a system are closer to what is needed to match observed systems. It is, however, difficult to determine a priori which systems will go unstable and which will not.
4. The structure of the outer disk, outside of the region where planets are detected with *Kepler*, influences the properties of planets and planetary systems formed closer to the star, indicating that planet formation is not always a strictly local process. Planets formed in more extended disks are more massive, located farther from their host stars and are less likely to be observed as multi-planet systems.
None of the simulations performed here simultaneously matched all four observational diagnostics of planetary systems, which are the relative frequencies of multi-planet systems, the orbital period ratios of planet pairs, the size ratio of planet pairs, and the locations of the innermost planet. However, the above trends between initial conditions and observables allow us to better estimate the conditions under which these systems formed.
In regards to developing a collection of systems whose innermost planets are similar to those seen in the *Kepler* systems, it is necessary to consider a range of inner disk locations. The embryo-only simulations are consistent with an inner disk edge centered at 0.05 au, but future simulations should consider implementing a range for the inner disk edge rather than a single value. In more extended disks a single inner disk edge location provides a broader detected distribution that is more consistent with the data, though this inner edge has to be located closer in than 0.05 au. In simulations including migration and orbital damping by gas, the inner disk edge needs to be moved outward to $\sim0.1$ au to result in planetary systems consistent with what is seen with *Kepler*.
Dynamical damping by gas in the cases studied here was too efficient, leading to too many observable multi-planets systems and planet orbital period ratios that are too low. Similar findings were found in other studies: population synthesis models with a similar number of interacting seed cores (20-50) also find mutual inclinations of planetary systems that are too low [@epos2]; and pebble-assisted formation and migration models find that not enough systems go unstable after disk dispersal to increase their mutual inclinations to the point of reducing the number of observed planets and increasing the observed period ratios [@izidoro17; @izidoro19]. Less efficient damping by gas may thus be necessary. For example, [@dawson16] and [@MacDonald2020] show that if planetary system assembly is delayed and happens when the disk gas is depleted by a factor 100 compared to the the initial gas mass, the amount of dynamical excitation matches closer to the observed systems. Although the underlying model we use has a gas density that decreases with time, we do not see this effect naturally arising from our simulations. We do see, however, that in disks that form smaller planets a closer match to the observed system properties is found. This is likely a result of these planets growing slower and, therefore, assembling when the gas density is lower or already dissipated, which, in turn, provides weaker or no orbital damping. This could suggest that the final stages of planetary accretion generally took place relatively late in terms of the removal of disk gas.
The general formation conditions outlined here allow us to make predictions about the formation pathways of habitable zone planets, and their potential chemical composition. For example, the likely presence of planetesimals during planetary accretion would provide a means for materials from more distant regions to be delivered to planets that form closer to the Sun. Such a process would be a natural means of delivering water and other volatiles to planets in the Habitable Zone.
The presence of gas might limit such radial mixing of planetesimals, but the gas would also induce type I migration, allowing planetary embryos from the outer disk to drift inwards toward the Habitable Zone. These embryos would also have formed in volatile-rich regions, allowing planets in the habitable zone to contain more water, carbon, or other bio-critical elements than if they were to have formed in situ. Since the amount of orbital damping by gas appears to be limited based on our simulation results, the influence of migration on planet composition is likely also reduced.
Quantifying the impact of different planet formation pathways on planet compositions requires an understanding of both the chemical composition in the disk as well as a model for how the distribution of material at different distances is assembled into planets. As such, we are making all simulated planetary systems discussed in this paper – complete with their accretion history – available in the `Genesis` database at <http://eos-nexus.org/genesis-database/>. The `Genesis` database is part of a larger effort of the Earths in Other Solar Systems (EOS) program aimed at understanding the frequency and formation of earth-like planets with bio-critical ingredients. Such a database will be instrumental in the search for extrasolar planets with atmospheric biosignatures in the Solar neighborhood.
This material is based on work supported by the National Aeronautics and Space Administration under Agreement No. NNX15AD94G for the program Earths in Other Solar Systems. The results reported herein benefitted from collaborations and/or information exchange within NASA‘s Nexus for Exoplanet System Science (NExSS) research coordination network sponsored by NASA‘s Science Mission Directorate We thank Andre Izidoro for providing his modified version of the Mercury integrator.
[^1]: <https://github.com/GijsMulders/epos>
[^2]: <http://eos-nexus.org/genesis-database>
[^3]: http://www.boulder.swri.edu/swifter/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We describe $\cp$ violating observables in resonant $W^\pm$ and $W^\pm$ plus one jet production at the Tevatron. We present simple examples of $\cp$ violating effective operators, consistent with the symmetries of the Standard Model, which would give rise to these observables. We find that $\cp$ violating effects coming from new physics at the $TeV$ scale could in principle be observable at the Tevatron with $10^6$ $W^\pm$ decays.'
---
\#1\#2\#3[[*Phys. Rev.*]{} [**D\#1**]{} \#2 (19\#3)]{} \#1\#2\#3[[*Phys. Lett.*]{} [**\#1B**]{} \#2 (19\#3)]{} \#1\#2\#3[[*Nucl. Phys.*]{} [**B\#1**]{} \#2 (19\#3)]{} \#1\#2\#3[[*Phys. Rev. Lett.*]{} [**\#1**]{} \#2 (19\#3)]{} å 0.0in 0.0in 6.0in 8.75in -1.0in .5in =-0.5in
[**LOOKING FOR $\cp$ VIOLATION IN $W$ PRODUCTION AND DECAY**]{}\
[^1] [**and G. Valencia$^{(b)}$**]{}\
[*$^{(a)}$ Physics Department, Brookhaven National Laboratory, Upton, NY 11973*]{}\
[*$^{(b)}$ Department of Physics and Astronomy, Iowa State University, Ames IA 50011*]{}\
Introduction
============
The origin of $\cp$ violation remains one of the unsolved questions in particle physics. It is therefore very important to search for signals of $\cp$ violation in all the experimentally accessible systems. The Tevatron has now accumulated a sample of about $50 000$ $W^\pm$ events and samples of $10^7$ events should be possible eventually. This makes it timely to think of testing $\cp$ violation in $W^\pm$ production and decay.
In this paper we study the processes $p \pp \ra W^\pm$ and $p \pp \ra
W^\pm \ + 1 {\rm ~jet}$.[^2] Some $\cp$ odd observables in these processes involve the polarization of the $W^\pm$ boson. It is, therefore, convenient to allow the $W^\pm$ to decay leptonically into $\ell^\pm \nu$ pairs and study the complete processes $p \pp \ra \ell^\pm \nu \ + \ {\rm ~0~or~1~jet}$. In this way we recover some information on the direction of the vector boson polarization by measuring the charged lepton momentum. Although it is also possible to construct observables for hadronic $W^\pm$ decays, we will not consider that case in this paper to avoid the complications of hadronization.
In this paper we adopt the strategy of searching for small $\cp$ violating contributions to a dominant process instead of looking for potentially larger effects in rare processes. We therefore limit our discussion to $\cp$ odd observables in the decay chains $$p\pp \ra W^\pm \ + \ {\rm ~0~or~1~jet} \ra \ell^\pm \nu
\ + \ {\rm ~0~or~1~jet}$$ and consequently we study these processes in the narrow width approximation. We concentrate on the dominant parton subprocesses, ignoring for example, top-quark distribution functions in the proton, $\cp$ violation in the parton distribution functions, and $\cp$ violation that occurs only at higher twist.
It is, of course, possible that there are $\cp$ violating effects that vanish within our approximation. An example in $\ell \nu$ production arises from the interference of the one-loop electroweak corrections to the resonant $W^\pm$ Standard Model amplitude with a non-resonant $\cp$ violating four-fermion new physics interaction. This example has been studied by Barbieri [*et. al*]{} [@barbieri].
It should be clear that the $\cp$ odd observables that we discuss are predominantly sensitive to $\cp$ violation in the $u d W$ vertex and in the $W \ell \nu$ vertex and these are the same vertices which are probed in pion decay $\pi \ra \ell \nu$.However, there are several scenarios under which the direct $W$ production and decay processes are more sensitive to $\cp$ violation than the corresponding pion decay. This is true in the examples we discuss for at least one of the following reasons:
- The new physics (at a high energy scale $\Lambda$) that violates $\cp$ contributes to the effective $W(q) f \overline{f^\prime}$ vertex in a way proportional to $(q^2 /\Lambda^2)^n$, $n\geq 1$. Since $q^2 \ll \Lambda^2$, the leading contribution will be for $n=1$. The effects from the new physics on direct $W$ production and decay are therefore enhanced over effects in pion decay by at least a factor $M_W^2 / m_\pi^2 \sim 3.5 \times 10^5$.
- The absorptive phases needed to construct $T$-even observables[^3] are larger in $W$ production and decay. An example is the case where the absorptive phase is due to a rescattering of the final state with electroweak strength. This would be the case if the new physics generates an effective four-fermion interaction that contributes to the $W f \overline{f^\prime}$ vertex at the one-loop level. In this case, the absorptive phase introduces an additional suppression factor of order $\Gamma_W / M_W \sim 3\%$ for direct $W$ production and decay, whereas it introduces a suppression factor of order $G_f m_\pi^2 \sim 10^{-8}$ for pion decays.
- Processes with a sufficient number of independent four-vectors to construct $T$-odd triple products are suppressed at low energy. In direct $W$ production and decay we can look for events with one jet, suppressed by a factor of $\alpha_s$ with respect to the lowest order QCD process. In pion decays we need a three body decay mode with the measurement of a polarization or a four body decay mode. Three body decay modes are four orders of magnitude smaller than $\pi \ra \ell \nu$ and polarization measurements are extremely difficult. Four body decay modes are nine orders of magnitude smaller than $\pi \ra \ell \nu$.
Kaon decay experiments tell us that the $\cp$ violating phases in the Standard Model are extremely small. In addition, $\cp$ odd observables in the standard model vanish in the limit of massless light fermions, and are thus even smaller at higher energies [@gaso]. The Standard Model does not produce a sufficiently large $\cp$ violating signal to be observed in the processes we study [@nacht]. Popular extensions of the standard model in the context of $\cp$ violation include multi-Higgs models. In these models $\cp$ violation is also proportional to fermion masses and thus negligible in processes like $d \uu \ra W^- \ra e^- \nn$ at high energy. We will, therefore, work from the assumption that studies of $\cp$ violation at colliders will only be sensitive to non-Standard Model sources. Furthermore, we will not consider a specific model for $\cp$ violation, but instead we will use an effective Lagrangian approach to parameterize possible $\cp$ violating operators. We will further assume that the origin of these operators lies in the physics that breaks the electroweak symmetry. With a linear realization of the symmetry breaking and a light Higgs, all the operators of dimension $6$ have been given in Refs. [@buchmuller; @burgess]. With a non-linear realization of the symmetry breaking sector, anomalous fermion-gauge-boson couplings have been described in Ref. [@peccei]. Here we consider just a few of these operators to illustrate the physics, but it should be obvious that a similar analysis can be applied to other operators.
There have been a number of studies of $\cp$ violation at colliders that concentrate on effects due to the top-quark or multiple $W$ production Ref. [@rindani]. We concentrate on single $W$ production and complement the work of Refs.[@barbieri; @brand]. We will not discuss detector issues at all, except to make the obvious statement that the detector must be “$\cp$-blind” in order to carry out these studies. We limit ourselves to offer a “proof in principle” that $\cp$ violation could be observed in $W^\pm$ decays at the Tevatron.
$p\pp \ra W^\pm X\ra \ell^\pm \nu X$
====================================
We assume that the proton and anti-proton beams are unpolarized, and that the polarization of the lepton is not measured. In this case, it is only possible to construct $T$-even observables for this reaction. Some $\cp$ odd observables have been listed in Ref. [@nacht]. Here we discuss a few simple observables of this type in the context of a $\cp$ violating four-fermion interaction due to physics beyond the minimal standard model.
Under a $\cp$ transformation, the reaction $p(\vec{p})\pp(-\vec{p}) \ra
\ell^+(\vec{q}) \nu X$ transforms into $p(\vec{p})\pp(-\vec{p})\ra \ell^-(-\vec{q}) \nn \overline{X} $. Here we work in the $p \pp$ center of mass frame and denote by $\overline{X}$ the $\cp$ conjugate of $X$. Also, we have only considered those kinematic variables that can be observed, in this case the momenta of the beam and the lepton. It is conventional to take the $z$-axis as being the direction of the proton beam, and to use rapidity and transverse momentum as the kinematical variables. Recalling that the lepton rapidity is given by (all variables in the $p \pp$ center of mass): $$y_\ell = {1\over 2}\log\biggl({E_\ell + q_{z \ell} \over E_\ell - q_{z \ell}}
\biggr)$$ and the lepton transverse momentum by: $$p_{T\ell} = |\vec{q}\sin\theta|$$ where $\theta$ is the angle between the proton and lepton momenta in the lab frame, one can see that under a $\cp$ transformation $$y_{\ell^-} \goes -y_{\ell^+},\ \ \
p_{T\ell^-} \goes p_{T\ell^+}$$ In terms of these variables we can, therefore, construct $\cp$-odd observables such as: $$\begin{aligned}
\tilde{R}_1 & \equiv &
{{\sigma}^+ - {\sigma}^- \over {\sigma}^+ + {\sigma}^-} \nonumber \\
\tilde{R}_2(y_0) & \equiv &
{ {d {\sigma}^+\over dy_\ell} \mid_{y_\ell=y_0}-
{d {\sigma}^-\over dy_\ell}\mid_{y_\ell= -y_0} \over
{d {\sigma}^+\over dy_\ell}\mid_{y_\ell=y_0} +
{d {\sigma}^-\over dy_\ell}\mid_{y_\ell= -y_0} } \nonumber \\
\tilde{R}_3(p_T) & \equiv &
{ {d {\sigma}^+\over dp_T} -{d {\sigma}^-\over dp_T} \over
{d {\sigma}^+\over dp_T} +{d {\sigma}^-\over dp_T} }
\label{cpobs},\end{aligned}$$ where ${\sigma}^\pm$ refers to $\sigma(p\pp\ra\ell^\pm\nu X)$. Of course, it is also possible to construct $\cp$ odd observables based on more complicated correlations, but we will not pursue that route in this paper.[^4]
To generate the $\cp$ odd observables in Eq. \[cpobs\] it is necessary to have an absorptive phase. We consider $\cp$ violating four-fermion operators, and their one-loop contribution to the $W f \overline{f^\prime}$ amplitudes. The effective four-fermion operators consistent with the symmetries of the Standard Model are listed, for example, in Ref. [@buchmuller; @burgess]. Since we want to interfere the $\cp$ violating amplitude with the lowest order standard model amplitude, we choose a four fermion operator of the form ${\cal O}^{(3)}_{\ell q}$ in the notation of Ref. [@buchmuller]: $${\cal L}_{\cp} = {4 \pi \over \Lambda^2} e^{i\phi}\overline{c}_L\gamma_\mu s_L
\overline{\ell}_L \gamma^\mu \nu_L + {\rm ~h.~c.~}
\label{ffcp}$$ This operator is similar to the one studied in Ref. [@barbieri], chosen so that its interference with the Standard Model is not suppressed by powers of light fermion masses. We keep the same normalization as Ref. [@barbieri] which considered $\ell^\pm \nu$ production away from the $W^\pm$ resonance. We consider the operator Eq. \[ffcp\], instead of a similar one with $\uu d$ quarks (used in Ref. [@barbieri]) for two reasons. First, for the operator with $\uu d$ there is a cancellation between two contributions to $p\pp \ra \ell^\pm \nu$ as discussed in Ref. [@barbieri]. This cancellation is exact for the resonant process that we study here, but it does not occur for the operator with $\overline{c} s$ of Eq. \[ffcp\]. Also, whereas there are several indirect constraints from low energy experiments on the operator with $\uu d$ [@barbieri], analogous constraints on the operator in Eq. \[ffcp\] depend on naturalness assumptions.
We compute the one-loop effects of Eq. \[ffcp\] in the $W f \overline{f^\prime}$ vertex as sketched in Figure \[wff\].
The first diagram in Fig. \[wff\] is just the Standard Model vertex and the second diagram represents the absorptive part of the one-loop contribution from the operator in Eq. \[ffcp\]. This absorptive part contains the $\cp$ violating coupling $\sin\phi$, and can be easily computed using the Cutkosky rule. We obtain an amplitude for $\uu d \ra e^- \nu$: $$\begin{aligned}
{\cal M} (\uu d \ra e^-\nu) &=& \biggl({g \over 2 \sqrt{2}}\biggr)^2
V^\star_{ud} {1\over \hat{s} -M_W^2 + i M_W \Gamma_W}
\biggl[ 1 + {V_{cs} \over 6} \sin\phi {\hat{s} \over \Lambda^2}\biggr]
\nonumber \\ &&
\overline{v}_u \gamma_\mu (1 - \gamma_5)u_d
\uu_e \gamma^\mu (1 - \gamma_5)v_\nu
\label{effver}\end{aligned}$$ We see that the $\cp$ violating contribution to the amplitude is proportional to $\hat{s}=(p_e+p_\nu)^2$ and is thus suppressed in pion decay. The corresponding amplitude for $u \dd \ra e^+ \nu$ is: $$\begin{aligned}
{\cal M} (u \dd \ra e^+\nu) &=& \biggl({g \over 2 \sqrt{2}}\biggr)^2
V_{ud} {1\over \hat{s} -M_W^2 + i M_W \Gamma_W}
\biggl[ 1 - {V^\star_{cs} \over 6} \sin\phi {\hat{s} \over \Lambda^2}\biggr]
\nonumber \\ &&
\overline{v}_d \gamma_\mu (1 - \gamma_5)u_u
\uu_\nu \gamma^\mu (1 - \gamma_5)v_e
\label{effvert}\end{aligned}$$
From these results it is clear that all the differential cross sections can be obtained by multiplying the minimal Standard Model results by an overall factor that depends only on $\hat{s}$. For example: $${d \hat{\sigma}^\pm \over d\ \cos\theta} =
\biggl({d \hat{\sigma}^\pm \over d\ \cos\theta} \biggr)_{SM}
\biggl[1\mp{1\over 3}{\hat{s} \over \Lambda^2}\sin\phi\biggr]
\label{cpfactor}$$
The total hadronic cross-section is obtained as usual: integrating with the parton distribution functions. Within our assumptions, all $\cp$ violation occurs in the parton subprocess and the parton distribution functions satisfy $f_{\dd/\pp}(x)=f_{d/p}(x)\equiv d(x)$ and $f_{u/\pp}(x)=f_{\uu/p}(x)\equiv
{\overline u}(x)$, etc. In this context, the asymmetries in Eq. \[cpobs\] can be trivially computed using the narrow width approximation to replace $\hat{s}$ with $M^2_W$ in the overall factor. They are: $$\tilde{R}_1 = \tilde{R}_2(y_0) = \tilde{R}_3(p_T) \approx
-{1 \over 3}{M_W^2\over \Lambda^2}\sin\phi.
\label{rateas}$$
In order to observe a signal at the one-standard deviation level, we require that the number of events, $N$, be greater than N > [1\_1\^2]{} 200,000 ([1 TeV]{})\^4 [1\^2]{}. \[asy\] The Tevatron is therefore, in principle, capable of observing $\cp$ violation coming from new physics at the $TeV$ scale with a sample of about $10^6$ $W^\pm$ events.
We have argued that it is unlikely that $\cp$ violation in pion decays can place a significant constraint on the strength of this operator. Nevertheless, there are indirect constraints on the strength of the $\cp$ conserving part of operators like Eq. \[ffcp\]. This can be seen by looking at the gauge invariant version of the operator ${\cal O}^{(3)}_{\ell q}$ in the notation of Ref.[@buchmuller]: =[4\^2]{}[Q]{}\_L \^I \_Q\_L [L]{}\_L \^\^I L \[effopb\] where ${\overline Q}=({\overline u},{\overline d})$ and ${\overline L}=({\overline e}, {\overline \nu}$). This contains the term: \^(1-\_5)d \_(1-\_5) which contributes to the decay $K^+\rightarrow \pi^+ \nu
{\overline \nu}$. Using the present experimental bound [@e787], $BR(K^+\rightarrow \pi^+\nu {\overline \nu}) < 10^{-9}$, we obtain a limit, > 70-100 TeV. \[lims\] Assuming that the couplings of operators involving first and second generation fermions are of the same order, this becomes an indirect constraint on the scale appearing in Eq. \[ffcp\].
$p\pp \ra W^\pm \ + \ {\rm ~1~jet}\ra \ell^\pm \nu \ + \ {\rm
{}~1~jet}$
=============================================================
In this process there are several parton subprocesses that contribute at leading order in $\alpha_S$ and there are enough independent four-vectors to give rise to $T$-odd correlations. Ref. [@nacht] has listed several $\cp$ odd observables for this system. We consider a few simple observables generated by a $\cp$ violating effective $udW$ operator that respects the symmetries of the Standard Model.
Working again in the $p\pp$ center of mass frame, a $\cp$ transformation takes the reaction $p(\vec{p})\pp(-\vec{p}) \ra \ell^+(\vec{q}) \nu
{\rm ~jet}(\vec{p}_j) X$ into $p(\vec{p})\pp(-\vec{p}) \ra \ell^-(-\vec{q}) \nn
{\rm ~jet}(-\vec{p}_j)\overline{X}$. In this case, the $\cp$ transformation takes all the particles that form the jet into their respective anti-particles and reverses their momenta. To use jet variables we assume that the algorithm that defines the jet is CP blind in the sense that the probability of finding that a collection of particles with certain momenta forms a jet is the same as the probability of finding that a collection of the respective anti-particles with the momenta reversed forms a jet [@donoghue]. With these definitions, we see that the observable kinematic variables are the rapidity and transverse momentum of the lepton and the jet. By the same arguments of the previous section, we can construct $\cp$ odd observables identical to those in Eq. \[cpobs\]. In this case there are additional distribution asymmetries obtained by replacing $y_\ell$ by $y_{jet}$ and $p_{T\ell}$ by $p_{Tjet}$ in Eq. \[cpobs\]. For example, the same $\cp$ odd interaction of Eq. \[ffcp\] would generate the same asymmetries as in Eq. \[rateas\].
It is also possible to have $T$-odd correlations in this process. The interest of these correlations lies in the fact that they can generate $\cp$ odd observables without requiring additional absorptive phases and thus may test different types of $\cp$ violating physics than the $T$-even asymmetries. To construct a $\cp$ violating observable we have to compare the correlations induced in $W^-$ plus jet production with those induced in $W^+$ plus jet production, in the same way that we constructed $\cp$-odd asymmetries from $T$-even observables.
For the $W$ + 1 jet process there is one $T$-odd correlation that can be observed; in the lab frame it is given by the triple product ${\vec p}_\ell\cdot ({\vec p}_{\rm beam}\times {\vec p}_{\rm jet})$. There are several equivalent ways to use this correlation to construct a $T$-odd observable. The basic idea is to define the plane formed by the beam and jet momenta and count the number of events with the lepton above the plane minus the number of events with the lepton below the plane: $$A^\pm = \sigma^\pm [({\vec p}_{\rm beam} \times
{\vec p}_{\rm jet})\cdot {\vec p}_\ell > 0] -
\sigma^\pm [({\vec p}_{\rm beam} \times {\vec p}_{\rm jet})
\cdot {\vec p}_\ell < 0]
\label{tripobs}$$ where $A^\pm$ refers to the observable for $W^\pm$ events (or $\ell^\pm \nu$ events). A practical way to implement this observable in the calculation (or in the experiment) is to weigh the matrix element squared for a parton subprocess (or to weigh the observed event) by the sign of ${\vec p}_\ell\cdot ({\vec p}_{\rm beam}\times {\vec p}_{\rm jet})$. From the $\cp$ transformation for this reaction, we see that $\cp$ symmetry predicts that $A^+=A^-$. In a manner analogous to the $T$-even observables of Eq. \[cpobs\], it is useful to construct not only the fully integrated asymmetry, but asymmetries for distributions as well. One obvious reason is that the simultaneous study of the different distribution asymmetries provides a handle on the possible $\cp$ odd biases of a detector. Another reason is that it is possible for the integrated asymmetry to vanish while having non-vanishing asymmetries for distributions. Some $T$-odd $\cp$ odd observables are then: $$\begin{aligned}
{R}_1 & \equiv &
{A^+ - A^- \over {\sigma}^+ + {\sigma}^-} \nonumber \\
{R}_2(y_0) & \equiv &
{ {d A^+\over dy} \mid_{y=y_0}-
{d A^-\over dy}\mid_{y= -y_0} \over
{d {\sigma}^+\over dy}\mid_{y=y_0} +
{d {\sigma}^-\over dy}\mid_{y= -y_0} } \nonumber \\
{R}_3(p_T) & \equiv &
{ {d A^+\over dp_T} -{d A^-\over dp_T} \over
{d {\sigma}^+\over dp_T} +{d {\sigma}^-\over dp_T} }
\label{cpobst},\end{aligned}$$ where $y$ and $p_T$ can be the rapidity and transverse momentum of the lepton or the jet (or the $W$). We have chosen to normalize the asymmetries with respect to the respective differential cross-sections. This is because non-zero but $\cp$ conserving $T$-odd correlations arise only through final state interactions and are thus generally small. With our normalization we obtain dimensionless asymmetries that are not misleadingly large.
We now turn our attention to a simple effective interaction that can generate some of these asymmetries. For $W$ plus jet production the only parton subprocesses that can give rise to the $T$-odd correlation are $g d \ra u W^-$, $g \uu \ra \dd W^-$ and $ \uu d \ra g W^-$. These subprocesses receive
contributions from the diagrams shown in Figure \[inig\] and the crossed diagrams. The first two diagrams in this figure correspond to the ones occurring in the standard model with the $W$ vertex replaced by an effective vertex that includes new physics and is represented by the full circle. In general, this vertex will contain derivative couplings and $SU(3)_c$ gauge invariance will require a contact interaction as depicted in the third diagram. Similarly, electromagnetic gauge invariance will require a contact interaction involving a photon. The interaction with the photon does not contribute to the process we study, but is important for $p\pp \ra W \gamma$.
In order to generate a $\cp$-odd triple product correlation, we need interference between the different diagrams that give rise to each parton subprocess. This will only occur if the diagrams have different phases.
An effective $\cp$ violating $W u d$ coupling, can be written down in the non-linear realization of electro-weak symmetry breaking as done in Ref. [@peccei]. In unitary gauge it contains the couplings: $${\cal L} = {g \over 2 \sqrt{2}}\biggl[
\kappa \uu \gamma_\mu (1 - \gamma_5) d W^{+\mu} +
\kappa^\star \dd \gamma_\mu (1-\gamma_5) u W^{-\mu}\biggr]
\label{pzlag}$$ where there is $\cp$ violation if $\kappa$ has an imaginary part. It is easy to see that this operator will not generate a triple product correlation of the type we want because both diagrams[^5] in Figure \[inig\] have the same phase. A study of the diagrams in Figure \[inig\] reveals that it is possible to give them different $\cp$ phases, if the $\cp$ violation is generated by an operator that depends on the momentum carried by the fermions in the $W f \overline{f^\prime}$ coupling. This leads us to consider a higher dimension operator, similar to Eq. \[pzlag\], of the form: $${\cal L}= -{\sqrt{2}\over\Lambda^2}
\biggl[\tilde{\kappa}
\overline{\Psi}_L \overleftarrow{D}_\alpha
\gamma_\mu \Sigma \tau_- \Sigma^\dagger
\overrightarrow{D}^\alpha \Psi_L \Sigma^\mu_+ +
\tilde{\kappa}^\star \overline{\Psi}_L \overleftarrow{D}_\alpha
\gamma_\mu \Sigma \tau_+ \Sigma^\dagger
\overrightarrow{D}^\alpha \Psi_L \Sigma^\mu_- \biggr].
\label{cplagt}$$ The notation is the same as that in Ref. [@peccei]: in unitary gauge $\Sigma=1$ and $\Sigma^\mu_\pm = -{g\over 2} W^{\mu \pm}$. For the processes of interest there will only be one $W$ boson and no $Z$ bosons, so the covariant derivatives refer only to QED and QCD: $$D_\alpha \Psi_L \ra (\partial_\alpha + {i \over 2}g_S\lambda^a G^a_\alpha
+ i e QA_\alpha) \biggl(\begin{array}{c} u \\ d_\theta \end{array}\biggl)_L.
\label{covd}$$ We present some of the Feynman rules from Eq. \[cplagt\] in Figure \[feyn\].
From these Feynman rules it is clear that this operator will give different $\cp$ phases to the diagrams in Fig. \[inig\] and will induce a $\cp$ violating triple product correlation.
For the gluon initiated process $g d \ra u e^- \nu$ we find that the interference of the new interaction with the lowest order standard model amplitude generates the contribution to the spin and color averaged matrix element squared, (within the narrow width approximation): [^6] | [M]{}(g(p\_g)d(p\_d)u(p\_u) e\^-(p\_e) (p\_))|\^2= [C24]{} M(p\_g,p\_d,p\_u,p\_e,p\_) \[interft\] where C=-[(4 )\^4M\_W\_W]{} [\^2 \_ss\_\^4]{} [Im 2 \^2]{} (p\_W\^2-M\_W\^2), $p_W^2=2 p_e\cdot p_\nu$, $\Gamma_W$ is the total $W$ decay width and M(p\_1,p\_2,p\_3,p\_e,p\_)(p\_1,p\_2,p\_3,p\_e) . We ignore interference terms which do not contribute to the $\cp$ violating observables in Eq. \[cpobst\]. For our estimates we use for the matrix element squared for this process the sum of the standard model contribution and Eq. \[interft\]. We take the convention that $p_1,p_2$ are incoming and $p_3,p_e,p_\nu$ are outgoing. The other parton level sub-process amplitudes can be found using crossing symmetry:
A\^[g [d]{}]{}(p\_g,p\_d,p\_u,p\_e,p\_)&=& [C24]{}M(p\_g,p\_d,p\_u,p\_e,p\_)\
A\^[g [u]{}]{}(p\_g,p\_u,p\_d,p\_e,p\_)&=& [C24]{}M(p\_g,-p\_d,-p\_u,p\_e,p\_)\
A\^[g u]{}(p\_g,p\_u,p\_d,p\_e,p\_)&=& [C24]{}M(p\_g,-p\_d,-p\_u,p\_e,p\_)\
A\^[[u]{} d]{}(p\_u,p\_d,p\_g,p\_e,p\_)&=& -[C9]{}M(-p\_g,p\_d,-p\_u,p\_e,p\_)\
A\^[u [d]{}]{}(p\_d,p\_u,p\_g,p\_e,p\_)&=& -[C9]{}M(-p\_g,p\_d,-p\_u,p\_e,p\_).
We have done a numerical simulation to estimate the size of the asymmetries of Eq. \[cpobst\] induced by the interaction Eq. \[cplagt\]. We impose a cut on the jet rapidity that simulates the typical acceptance at the Tevatron $|y_{jet}|< 3.$ For illustration purposes we choose the cut $p_{Tjet}> 30$ GeV to define the jet. Increasing the value of this cut reduces the total number of events but increases the signal to background ratio because the $\cp$ odd contribution is not peaked at low $p_T$.
We present our results for Im $\kappa =1$ and $\Lambda=1$ TeV in Eq. \[cplagt\], and remind the reader that they scale as Im $\kappa/\Lambda^2$.
We find that the fully integrated asymmetry ${R}_1$ vanishes for the interaction Eq. \[cplagt\]. To our knowledge, there is no reason for this asymmetry to vanish in general, so this result is specific to the example we chose.
In Figure \[dady\] we present the distributions of the observable Eq. \[tripobs\] with respect to the lepton rapidity generated by the $\cp$ violating interference Eq. \[interft\]. The corresponding ($\cp$ conserving) distributions for the Standard Model vanish to the order we work (they are non-zero at higher order in QCD[@hagi]). From that figure we can see that although the distributions do not vanish as a function of the lepton rapidity, the integrated observable does vanish. As explained before, a non-zero value of these distributions is not a signal of $\cp$ violation. Rather, the signal of $\cp$ violation is the fact that the distribution for $W^+$ at a given value of $y_e$ is not equal to the distribution for $W^-$ at the value $-y_e$.
In Fig. \[yasym\] we present the asymmetry in the lepton rapidity distribution ${R}_2(y_\ell)$. At the one standard deviation level some $10^6$ $W^\pm$ plus one jet events would be needed to observe this asymmetry. Measurement of this asymmetry for arbitrary values of $y_e$ is complicated by the fact that the acceptance of the detector must be the same for $y_e$ and $-y_e$. Figure \[yasym\] shows that the asymmetry doesn’t necessarily vanish at $y_e=0$, making this a particularly interesting point to search for $\cp$ violation.
We show in Figure \[ptedis\] the distributions of the $A^\pm$ correlations with respect to the lepton transverse momentum. The $\cp$ violating nature of the interaction is reflected in the fact that the two distributions have opposite signs. Normalizing as in Eq.\[cpobst\] we present $R_3(p_{Te})$ in Figure \[ptasym\].
The rise in the asymmetry for large values of $p_{Te}$ is due to the decrease in the standard model distribution $d\sigma/dp_{Te}$ which is peaked around $M_W/2$. This increase in the asymmetry is thus accompanied by a decrease in the total number of events.
Conclusions
===========
We have constructed several $\cp$-odd asymmetries that can be used to search for $\cp$ violation in $W^\pm + (0,1)~{\rm jet}$ events in $p\pp$ colliders. We have estimated the contributions to these asymmetries from some simple $\cp$ violating effective operators that respect the symmetries of the Standard Model. Assuming that the scale of the new physics responsible for these operators is 1 TeV, we find that it is possible to search for $\cp$ violation at the Tevatron with as few as $10^6$ events. Similar observables can be constructed for other processes such as $p\pp \ra W^\pm \gamma$.
[**Acknowledgements**]{} The work of G.V. was supported in part by a DOE OJI award under contract number DEFG0292ER40730. G. V. thanks the theory group at BNL for their hospitality while part of this work was performed. We are grateful to S. Errede, T. Han, J. Hauptman and S. Willenbrock for helpful discussions.
[999]{}
[^1]: This manuscript has been authored under contract number DE-AC02-76CH00016 with the U.S. Department of Energy. Accordingly, the U.S. Government retains a non-exclusive, royalty-free license to publish or reproduce the published form of this contribution, or allow others to do so, for U.S. Government purposes.
[^2]: The Standard Model perturbative amplitudes for $W$ production are given in Ref. [@smprod]. The re-summed amplitudes (valid at small transverse momenta) are given in Ref. [@resumw].
[^3]: By $T$-even (odd) observables we mean those that do not change sign (do change sign) under the naive $T$ operation: inversion of all momentum and spin vectors. Recall that this is not the same as time reversal.
[^4]: The capability of a detector like CDF to study asymmetries in distributions has been demonstrated in the measurement of the charge asymmetry in the production of $W$’s as a function of the $W$ rapidity [@cdf].
[^5]: For this operator there would not be a contact interaction as in the third diagram.
[^6]: $\epsilon(p_1,p_2,p_3,p_4)\equiv \epsilon_{\alpha\beta\gamma\delta}
p_1^\alpha p_2^\beta p_3^\gamma p_4^\delta$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'While the symbol map for the collection of bounded Toeplitz operators is well studied, there has been little work on a symbol map for densely defined Toeplitz operators. In this work a family of candidate symbols, the Sarason Sub-Symbols, is introduced as a means of reproducing the symbol of a densely defined Toeplitz operator. This leads to a partial answer to a question posed by Donald Sarason in 2008. In the bounded case the Toeplitzness of an operator can be classified in terms of its Sarason Sub-Symbols. This justifies the investigation into the application of the Sarason Sub-Symbols on densely defined operators. It is shown that analytic closed densely defined Toeplitz operators are completely determined by their Sarason Sub-Symbols, and it is shown for a broader class of operators that they extend closed densely defined Toeplitz operators (of multiplication type).'
author:
- 'Joel A. Rosenfeld'
bibliography:
- 'toeplitzbibliography.bib'
title: 'The Sarason Sub-Symbol and the Recovery of the Symbol of Densely Defined Toeplitz Operators over the Hardy Space'
---
Introduction
============
The study of bounded Toeplitz operators over the Hardy space $H^2(\mathbb{T})$ is a well developed subject where there are several equivalent definitions of a Toeplitz operator. The simplest definition of a bounded Toeplitz operator is an extension of the definition of a Toeplitz matrix. In this case an operator, $T$, is called a Toeplitz operator if the matrix representation of the operator, with respect to the orthonormal basis $\{ e^{in\theta} \}_{n=0}^\infty$ is constant along the diagonals. Algebraically, this can be represented as $S^* T S = T$. Here $S=M_z$ is the shift operator for the Hardy space. If the coeficients corresponding to each diagonal of the matrix are the Fourier coeficients of a function $\phi \in L^\infty(\mathbb{T})$, then $T = P_{H^2(\mathbb{T})} M_\phi$. Here $P_{H^2(\mathbb{T})}$ is the projection from $L^2(\mathbb{T})\to H^2(\mathbb{T})$, and $M_\phi$ is the bounded multiplication operator from $H^2(\mathbb{T}) \to L^2(\mathbb{T})$ given by $M_\phi f = \phi f$. Finally the converse is true, the bounded operator given by $T_\phi = P_{H^2(\mathbb{T})} M_\phi$ with $\phi \in L^\infty(\mathbb{T})$ satisfies $S^* T S = T$.
When the bounded condition is relaxed to closed and densely defined, the corresponding definitions of Toeplitz operators are no longer equivalent. For instance, if the coefficients of an upper triangular matrix are the coefficients of a Smirnov class function, $\phi \in N^+$, then the operator defined by the closure of this matrix, call it $T$, is densely defined, and the operator is the adjoint of a densely defined multiplication operator (an analytic Toeplitz operator) $M_\phi$. Unlike its bounded counterpart, $T$ can not be represented by a multiplication operator as $P M_{\bar \phi}$ since its domain is strictly larger than the domain of $M_{\bar\phi}$. The operator $T$ does satisfy the following algebraic equations:
1. $D(T)$ is $S$-invariant,
2. $S^* T S = T$, and
3. If $f \in D(T)$ and $f(0)=0$, then $S^*f \in D(T)$.
These can be seen as the densely defined analogue of the algebraic condition for bounded Toeplitz operators. Therefore $T$ satisfies the algebraic conditions for being a Toeplitz operator, but is not a Toeplitz operator in the multiplication sense. However, $T$ is a closed extension of a multiplication type Toeptliz operator. At the close of [@sarason2008] the following problem was posed:
Is it possible to characterize those closed densely defined operators $T$ on $H^2(\mathbb{T})$ with the above three properties? Moreover, is every closed densely defined operator on $H^2(\mathbb{T})$ that satisfies these conditions determined in some sense by a symbol?
This paper aims to address the second half of this question. If a closed densely defined operator, $T$, satisfies the three algebraic conditions above, henceforth a *Sarason-Toeptliz* operator, then is $T$ the extension of an operator of the form $P M_\phi$ where $M_\phi$ is a densely defined multiplication operator from $H^2$ to $L^2$?
For bounded Toeplitz operators the recovery of the symbol of a Toeplitz operator can be achieved through the symbol map on $\mathcal{T}$, the algebra of generated by the collection of Toeplitz operators in $\mathcal{L}(H^2)$. Douglas demonstrated that there is a unique multiplicative mapping, $\phi$ from $\mathcal{T}$ to $L^\infty$ such that $\phi(T_f T_g) = \phi(T_f)\phi(T_g) = fg$ [@Douglas; @axler]. This fact was proven again in [@barriahalmos] by Halmos and Barria using the limits along the diagonals of a Toeplitz matrix in order to find the symbol in $L^\infty$.
The Hardy space can be identified with analytic functions of the disc $\mathbb{D}$ such that the Taylor coefficients of these functions are square summable. By this viewpoint, $H^2$ is a reproducing kernel Hilbert space (RKHS) over $\mathbb{D}$ with the kernel functions $k_w(z) = (1-\bar w z)^{-1}$ for $|w| < 1$.
In the case of bounded Toeplitz operators, the Berezin transform, a tool particular to the study of RKHSs, is sufficient for the recovery of the of $L^\infty$ functions via radial limits of the Berezin transform of a bounded Toeplitz operator [@englis1995]. However, in more general cases the recovery of the symbol of a Sarason-Toeplitz operator is no longer clear. The recovery of the symbol of a densely defined analytic (or a co-analytic) Toeplitz operator with symbol $\phi$ can be accomplished by the use of the Berezin transform. In this case, the adjoint of an analytic Toeplitz operator has the reproducing kernels as eigenvectors, $k_z$, with eigenvalues $\overline{\phi(z)}$ [@sarason2008]. Thus $$\tilde T(z) = (1-|z|^2)\langle k_z,T^* k_z\rangle = (1-|z|^2) \langle k_z, \overline{\phi(z)}k_z \rangle = \phi(z).$$The application of the Berezin transform requires the kernel functions $k_w(z) = (1-\bar w z)^{-1}$ to be in the domain of a operator or in the domain of its adjoint. Thus, the investigation of a new method is justified for the recovery of the symbol of a densely defined Sarason-Toeplitz operator.
We introduce the Sarason Sub-Symbol, which depends on a choice of a function in $D(T)$, as a family of symbol maps for Sarason-Toeplitz operators. In the development, it will be demonstrated that for the bounded case the Sarason Sub-Symbol is unique iff the operator is Toeplitz. Thus the uniqueness of the Sarason Sub-Symbol provides another equivalent definition for a bounded Toeplitz operator. Subsequently it is demonstrated that the Sarason Sub-Symbol for an analytic Toeplitz operator is unique and determines the operator. The rest of the paper is concerned with classes of Toeplitz operators for which the existence of the Sarason Sub-Symbol can be established, and it demonstrates sufficient conditions to show that $T$ is a closed extension of a multiplication type Toeplitz operator.
The Problem of Sarason
======================
For bounded Toeplitz operators, the Sarason Problem has been long settled [@hw; @Douglas]. Indeed, if a Toeplitz operator is bounded, then it can be represented by an $L^\infty$ function. Suarez characterized all closed densely defined operators on $H^2(\mathbb{T})$ that commute with the adjoint of the shift operator [@Suarez], and Sarason gives a different treatment of operators that commute with the shift operator, the so called analytic Toeplitz operators. Both of these collections of operators satisfy the Sarason-Toeplitz condition. In addition, the analytic Toeplitz operators are precisely the operators of multiplication by an function in the Smirnov class, $N^+$ [@sarason2008]. Suarez’s operators are the adjoints of these analytic Toeplitz operators and are called co-analytic Toeplitz operators [@Suarez]. Thus the above classes of Sarason-Toeplitz operators are completely characterized by a symbol.
Analytic and co-analytic Toeplitz operators both satisfy the Sarason conditions. The following propery generalizes this relationship.
If $T$ is a Sarason-Toeplitz operator then so is $T^*$.
$T$ is a closed densely defined operator, which means that $T^*$ is closed and densely defined as well. Thus $D(T^*)$ is nonempty.
To demonstrate that $T^*$ has a shift invariant domain, take $g\in D(T^*)$. By definition this means that $\tilde L(f) = \langle Tf, g \rangle$ is a continuous functional. In order to show that $zg \in D(T^*)$ it must be established that $L(f) = \langle Tf, zg \rangle$ is continuous. Note that $zD(T) \subset D(T)$, and $zD(T)$ has co-dimension 1 in $D(T)$. Thus there exists an $f_0 \in D(T)$ such that $$D(T) = \mathbb{C} \{ f_0 \} \oplus zD(T).$$ The functional $L$ is continouous on $\mathbb{C}\{f_0\}$, since it is finite dimensional. Therefore it suffices to show that $L$ is continuous on $zD(T)$. If $f = zh$ for some $h \in D(T)$, then $$L(f) = L(zh) = \langle Tzh, zg \rangle = \langle Th, g \rangle = \tilde L(h).$$ Thus $L$ is continous on $zD(T)$, since $\tilde L$ is continuous on $D(T)$.
Now suppose that $g \in D(T^*)$ and $g(0) = 0$, and consider the functional $L_2(f) = \langle Tf, S^*g \rangle$ defined for $f \in D(T)$. This functional can be rewritten as $$L_2(f) = \langle S^*TSf, S^*g \rangle = \langle TSf, g \rangle := \tilde L_2(Sf).$$ It follows that $L_2(f)$ is continuous, since $\tilde L_2(Sf)$ is continuous with respect to $f$.
Finally for all $f \in D(T^*)$ and $g \in D(T)$ we have, $$\langle T^*f, g \rangle = \langle f, Tg \rangle = \langle f, S^* T S g \rangle = \langle S^*T^*S f, g \rangle,$$ which yields the second condition.
The Sarason Sub-Symbol
======================
While the Berezin transform can be applied to recover the symbol of densely defined analytic and co-analytic Toeplitz operators, it is not clear if it can be used to recover the symbol of more general densely defined Toeplitz operators. This is because the functions $k_z$ a required to be in the domain of either the operator or the adjoint of the operator for the Berezin transform to be well defined. Instead, in this section the Sarason Sub-Symbol will be introduced as a candidate for the recovery of the symbol of densely defined Sarason-Toeplitz operators.
As a motivating example for the defintion of the Sarason Sub-Symbol, first suppose that $T$ is a bounded Toeplitz operator with symbol $\phi \in L^\infty$. In this case $$a_n = \left\{ \begin{array}{cc}\langle T1, z^n \rangle & n \ge 0\\ \langle Tz^n, 1 \rangle & n < 0 \end{array}\right.$$ are the Fourier coefficients of $\phi$. Thus $\phi$ can be reconstructed as follows $$\phi(e^{i\theta}) = \sum_{n=1}^\infty \langle Tz^n, 1 \rangle e^{-in\theta} + \sum_{n=0}^\infty \langle T1, z^n \rangle e^{in\theta}.$$
While it is not expected that $1 \in D(T)$ in general, given any function $f\in D(T)$ the domain of the densely defined operator $TM_f$ contains the polynomials, since $D(T)$ is shift invariant. The Sarason Sub-Symbol is defined as follows:
Let $T$ be an operator with a shift invariant domain $D(T)$. For $f \in D(T)\setminus \{0\}$ the Sarason Sub-Symbol corresponding to $f$ is given by $R_f = h_f/f$ where $$h_f = \sum_{n=1}^\infty \langle Tfz^n, 1 \rangle e^{-in\theta} + \sum_{n=0}^\infty \langle Tf, z^n \rangle e^{in\theta}$$ where this series is convergent in some sense. The partial Sarason Sub-Symbol corresponding to $f$ is given by $R_{f,N} = h_{f,N}/f$ where $$h_{f,N} = \sum_{n=1}^N \langle Tfz^n, 1 \rangle e^{-in\theta} + \sum_{n=0}^\infty \langle Tf, z^n \rangle e^{in\theta}.$$
Heuristically, if $T$ is a Toeplitz operator associated with multiplication by the symbol $\phi$, then $h_f = \phi \cdot f$. The question of well definedness of the Sarason Sub-Symbol depends on the convergence of the series contained in the definition of $h_f$. When $\phi \in L^\infty$, $h_f = \phi \cdot f$, and is a well defined function in $L^2$. More specifically we can characterize all bounded Toeplitz operators by means of the Sarason Sub-Symbol.
\[bddToep\]Let $V$ be a bounded operator on $H^2$. The operator $V$ is a Toeplitz operator iff the Sarason Sub-Symbol is independent of the choice of $f \in H^2$.
Suppose that $V=T_\phi$ is a Toeplitz operator with symbol $\phi$. $$h_f = \sum_{n=1}^\infty \langle T_\phi f z^n, 1 \rangle_{H^2} e^{-in\theta} + \sum_{n=0}^\infty \langle T_\phi f, z^n \rangle_{H^2} e^{in\theta}$$ $$= \sum_{n=1}^\infty \langle \phi f z^n, 1 \rangle_{L^2} e^{-in\theta} + \sum_{n=0}^\infty \langle \phi f, z^n \rangle_{L^2} e^{in\theta} = \phi \cdot f.$$ Thus $R_f = h_f/f = \phi$ is independent of the choice of $f$.
Now suppose that $V$ is not a Toeplitz operator. This means there is a pair of integers $n,m \in \mathbb{N}$ such that $n < m$ (without loss of generality) and $\langle Vz^n,z^m \rangle \neq \langle V1, z^{m-n} \rangle$. In this case consider the two Sarason Sub-Symbols $$R_1 = \sum_{k=1}^\infty \langle Vz^k, 1\rangle e^{-ik\theta} + \sum_{k=0}^\infty \langle V1, z^k \rangle e^{ik\theta} = \sum_{k=-\infty}^\infty a_k e^{ik\theta}\text{ and}$$ $$R_{z^n}=e^{-in\theta} \left( \sum_{k=1}^\infty \langle Vz^{n+k}, 1\rangle e^{-ik\theta} + \sum_{k=0}^\infty \langle Vz^n, z^k \rangle e^{ik\theta}\right) = e^{-in\theta} \sum_{k=-\infty}^\infty b_k e^{ik\theta}.$$ The difference of the two sub-symbols yields $$R_1 - R_{z^n} = e^{-in\theta} \sum_{k=-\infty}^\infty(a_{k-n}-b_k)e^{ik\theta}.$$ The coeficient $(a_{m-n}-b_m) \neq 0$ by construction. Therefore $R_1 \neq R_{z^n}$.
Thus every bounded Toeplitz operator is characterized by the uniqueness of its Sarason Sub-Symbols. This motivates the investigation into densely defined operators. The following sections investigate the interplay between the Sarason Sub-Symbols and densely defined Sarason-Toeplitz operators.
Analytic Densely Defined Toeplitz Operators
===========================================
Just as in Proposition \[bddToep\], an analytic densely defined Toeplitz operator is completely characterized by a symbol. As shown in [@sarason2008], these operators are precisely the multiplication operators with symbols, $\phi$, in the Smirnov class of functions. That is, each $\phi$ can be written as a ratio of $H^\infty$ functions $b/a$ where $|a(e^{i\theta})|^2 + |b(e^{i\theta})|^2 = 1$ for all $\theta$ and $a$ an outer function. In this setting the Sarason Sub-Symbol is unique.
Given a Sarason-Toeplitz operator $T$, there exists a symbol $\phi \in N^+$ for which $T=M_\phi$ iff $\langle Tzf, 1 \rangle = 0$ for all $f \in D(T)$. Moreover, the Sarason Sub-Symbol is unique.
The forward direction follows since $T_\phi = M_\phi$ for $\phi \in N^+$. This means $TS=ST$, and $\langle Tzf, 1 \rangle = \langle zTf, 1 \rangle = 0$ since $1 \in (zD(T))^\perp$.
In order to establish sufficiency, let $f_1 = \sum_{n=0}^\infty a_n z^n, f_2 = \sum_{n=0}^\infty b_n z^n \in D(T)\setminus \{0\}$. By hypothesis, $h_{f_i} = \sum_{n=0}^\infty \langle Tf_i 1, z^n \rangle z^n = Tf_i \in H^2$ for $i=1,2$.
In order to establish uniqueness of the symbol, $R_{f_1} = R_{f_2}$, consider the function $h_1 f_2 - h_2 f_1 \in L^1(\mathbb{T})$. The Fourier series of $h_1f_2$ and $h_2f_1$ can be computed through convolution. Hence, $$h_1 f_2 = \sum_{n=0}^\infty \left( \sum_{k=0}^n \langle Tf_1, z^{n-k} \rangle b_k \right) z^n = \sum_{n=0}^\infty \left( \sum_{k=0}^n \langle Tz^k f_1, z^n \rangle b_k \right) z^n \text{, and }$$ $$h_2 f_1 = \sum_{n=0}^\infty \left( \sum_{k=0}^n \langle Tz^k f_2, z^n \rangle a_k \right) z^n.$$ The second equality follows since $S^* T S f = Tf$, and $\langle TSf, 1 \rangle = 0$ implies that $TSf(0)=0$ and $S S^* T S f = T S f$. This leads to $$H := h_1f_2 - h_2f_1 = \sum_{n=0}^\infty \left( \sum_{k=0}^n \langle Tz^kf_1, z^{n} \rangle b_k - \sum_{k=0}^n \langle Tz^k f_2, z^n \rangle a_k \right)z^n.$$ In order to establish that each coefficient is in fact zero, consider, for arbtrary $n$, the coefficient of $z^n$: $$\hat H(n) = \sum_{k=0}^n \langle Tz^kf_1, z^{n} \rangle b_k - \sum_{k=0}^n \langle Tz^k f_2, z^n \rangle a_k$$ $$= \left\langle T\left( f_1 \left[ \sum_{k=0}^n b_k z^k - f_2 \right] - f_2 \left[ \sum_{k=0}^n a_k z^k - f_1 \right] \right), z^n \right\rangle.$$ The $H^2$ function inside of $T$ is in fact in the domain of $T$ by the properties of Sarason-Toeplitz operators, and this function has a zero of order greater than $n$ at zero. Denote by $z^{n+1}F_n$, the function in the argument of $T$. By our hypothesis, $$\hat H(n) = \langle Tz^{n+1} F_n, z^n \rangle = \langle TzF_n, 1 \rangle = 0.$$ Therefore $R_{f_1} = R_{f_2}$ for any choice of $f_1, f_2 \in D(T)\setminus \{ 0 \}$, so let $\phi = R_{f_1}$ be the proposed symbol for the Sarason-Toeplitz operator $T$. $h_f = Tf \in H^2$ for each $f \in D(T)$. Further, given any $z \in \mathbb{D}$ there exists $f_z \in D(T)$ such that $f(z) \neq 0$ (this follows from the density of $D(T)$ in $H^2$). Thus $\phi = Tf_z/f_z$ is analytic at $z$ for every point $z \in \mathbb{D}$. Finally note that for each $f \in D(T)$, $M_\phi f = \phi f = ( Tf / f ) f = Tf$. Thus $T=M_\phi$ is a densely defined multiplication operator with an analytic symbol. By [@sarason2008], $\phi \in N^+$.
A Sarason-Toeplitz operator $T$ on $H^2$ is analytic ($ST=TS$) iff $\langle Tzf, 1 \rangle =0$ for all $f \in D(T)$.
Symbols that are ratios of $L^2$ functions and $H^2$ functions
==============================================================
In the case of an analytic densely defined Toeplitz operator with symbol $\phi$ (expressed as $\phi = b/a$ in canonical form) , the domain is given by $D(T)=aH^2$. This means that there is an outer function, in particular $a$, in the domain of $T$. Moreover, since $T=M_\phi$, it is clear that $h_a = \phi a \in H^2$. Therefore, the existence of an outer function $f \in D(T)$ for which $h_f$ is well defined is straighforward in the case of analytic Toeptliz operators.
When we consider a co-analytic Toeptliz operator of the form $M_\phi^*$, its domain is given by $\mathcal{H}(b)$, the de Branges-Rovnyak space corresponding to $b$. $\mathcal{H}(b)$ contains the space $aH^2$ as a subspace. Hence, it also has an outer function in its domain. In particular, if $f=a\cdot p$, where $p$ is a polynomial, then $h_f$ (corresponding to $M_\phi^*$) is in $L^2$. Since $a$ is an outer function, the collection of all such $f$ is dense in $H^2$. Therefore, the set $$D_2(T) = \{ f \in D(T) : h_f \in L^2 \}$$ is dense in $D(T) = D(M_\phi^*) = \mathcal{H}(b)$.
The answer to the question of the nonemptiness (as well as density) of the space $D_2(T)$ is unknown for general Sarason-Toeplitz operators. In this section, the applicability of the Sarason Sub-Symbol is extended to include functions of the form $B/A$ where $B \in L^2$ and $A$ is an $H^2$ outer function.
\[cdd\] Let $\phi$ be a function on the unit circle that can be written as the ratio of an $L^2$ function and an $H^2$ outer function. Let $$D(M_\phi) = \{ f \in H^2 : \phi \cdot f \in L^2\}.$$ The operator $M_\phi : D(M_\phi) \to L^2$ is a closed densely defined operator on $H^2$.
Write $\phi = B/A$ where $B \in L^2$ and $A \in H^2$ is an outer function. Since $B \cdot p \in L^2$ for every polynomial $p(z)$, we see that $A \cdot p \in D(M_\phi)$ for every polynomial $p$. Therefore, $D(M_\phi)$ is dense in $H^2$ by the outer property of $A$.
Now suppose that $\{f_n\} \subset D(M_\phi)$ and $f_n \to f \in H^2$. Suppose further that $M_\phi f_n \to F \in L^2$. Since $f_n \to f$ in the $L^2$ norm, there exists a subsequence, $\{ f_{n_j}\}$, such that $f_{n_j} \to f$ almost everywhere. Since $A$ is an outer function, $A(e^{i\theta}) \neq 0$ for almost every $\theta$. Thus $\phi f_{n_j} \to \phi f$ almost everywhere.
The subsequence $\phi f_{n_j} \to F$ in $L^2$ and so there is a subsequence $\phi f_{n_{j_k}} \to F$ almost everywhere. However, this subsquence also converges almost everywhere to $\phi f$. Thus we may conclude that $\phi f = F$ almost everywhere, which completes the proof.
\[l2symbol\]Let $T$ be a Sarason-Toeplitz operator. If there is an $H^2$ outer function $f \in D(T)$ such that $\sum_{n=1}^\infty \langle Tz^nf, 1 \rangle \bar z^n \in L^2$, then $T$ extends a closed densely defined operator of the form $T_\phi = PM_\phi$ where $\phi=R_f$ is the ratio of an $L^2$ function and an $H^2$ outer function. Moreover, $D_2(T)$ is a dense subset of $D(T)$.
Let $f$ be an $H^2$ outer function in $D(T)$, and let $h_f$ be the corresponding numerator of the Sarason Sub-symbol corresponding to $f$. Express $h_f = \sum_{n=-\infty}^\infty b_n z^n$. By the properties of Sarason-Toeplitz operators, $b_{n-m} = \langle Tfz^m, z^n \rangle$. Now consider the operator $T_{R_f} = P M_{R_f}$ which is closed and densely defined by Lemma \[cdd\].
Since the domain of $T$ is shift invariant, $f \cdot p \in D(T)$ for every polynomial $p$. Moreover, $h_f \cdot p \in L^2$ for every polynomial $p$. It follows that $D_2(T) \subset D(T)$ is dense in $H^2$ since $f$ is an outer function. Define the set $F := \{ f \cdot p : p \text{ is a polynomial } \} \subset D_2(T)$.
Let $p(z) = a_k z^k + \cdots + a_1 z + a_0$ be a polynomial of degree $k \in \mathbb{N}$. The product of $h(z)$ and $p(z)$ can be calculated as follows: $$h(z) \cdot p(z) = \sum_{n=-\infty}^\infty \left( \sum_{m=0}^k b_{n-m}a_m \right) z^n = \sum_{n=-\infty}^\infty \left( \sum_{m=0}^k \langle Ta_m z^m f, z^n \rangle \right) z^n$$ $$=\sum_{n=-\infty}^\infty \langle Tfp, z^n \rangle z^n = w(z) + T(fp)(z).$$ Where $w(z) \in \overline{H^2_0}$ since $h_f \in L^2$. In particular, this means $T_{R_f}(fp) = P(hp) = T(fp)$ for all polynomials $p$. Hence, $T$ agrees with $T_{R_f}$ on a dense domain, and $T$ extends $\left. T_{R_f} \right|_F.$ Finally, by Lemma \[cdd\], $\left. T_{R_f}\right|_F$ is closable, and $\left. T_{R_f} \right|_F \subset T$ implies that $\left. T_{R_f} \right|_F^{**} \subset T^{**} = T$.
The above theorem relies on the ability to find an outer function in $D_2(T)$. Once such a function is found, $T$ is shown to be a closed extension of the corresponding operator $PM_{R_f}$. When such a function does not exist, it can be shown that $T$ is the limit of multiplication type Toeplitz operator on a restricted domain.
Suppose $T$ is a Sarason-Toeplitz operator, let $f \in D(T)$, and define $F=\{f \cdot p : p \text{ is a polynomial } \}$. There exists a sequence of multiplication type Toeplitz operators, $T_{\phi_M} = P M_{\phi_M}$ such that $T_{\phi_M}$ converges to $T$ strongly on all of $F$. Moreover, these operators have a common dense domain.
Let $f \in D(T)$ and let $p(z) = a_k z^k + \cdots + a_1 z + a_0$ be a polynomial of degree $k$. Now, as in Theorem \[l2symbol\], consider the product $h_{f,N}(z)p(z)$: $$h_{f,N}(z)\cdot p(z) = \sum_{n=-N}^\infty \left( \sum_{m=0}^{\min(k,n+N)} b_{n-m} a_m \right) z^n$$ $$= \sum_{n=-N}^{k-N-1} \langle Tf(a_0+\cdots+a_{n+N z^{n+N}}, z^n \rangle z^n + \sum_{n=k-N}^\infty \langle Tfp, z^n\rangle z^n.$$
Therefore, $$T_{R_{f,N}} (fp) (z) = P(h_Np) (z)$$ $$=\sum_{n=0}^{k-N-1} \langle Tf(a_0+\cdots+a_{n+N z^{n+N}}, z^n \rangle z^n + \sum_{n=\min(k-N,0)}^\infty \langle Tfp, z^n\rangle z^n.$$
The left sum is empty for large enough $N$, therefore $\{ T_{R_{f,N}} (fp) \}$ is constant for large enough $N$. This means $T_{R_{f,N}} (fp) \to T(fp)$ as $N \to \infty$.
In order to find a common domain for each of these Toeplitz operators, consider the inner-outer factorization $f = f_i f_o$. The functions, $\tilde h_N = h_{f,N}/f_i \in L^2$ since $f_i$ has modulus 1 on the circle. Thus $\tilde h_N f_o p \in L^2$ for all polynomials $p$. This implies that $$F_0 = \{ f_0 p : p \text{ is a polynomial } \} \subset D(T_{R_{f,N}})$$ for all $N \in \mathbb{N}$.
An Example of a Non-Sarason Toeplitz Operator
=============================================
This section is concerned with demonstrating that a densely defined Toeplitz matrix does not necessarily define a Sarason-Toeplitz operator. In particular, this section will extend an upper triangular Toeplitz matrix and demonstrate that the domain of the extension is not shift invariant. An upper triangular Toeplitz matrix is a matrix of the form $$\left( \begin{array}{cccc}\gamma_0 & \gamma_1 & \gamma_2 & \\0 & \gamma_0 & \gamma_1 & \cdots\\ 0 & 0 & \gamma_0 & \\ & \vdots & & \ddots \end{array}\right).$$ As an operator over $H^2$, this matrix has a natural dense domain, namely the polynomials. The density of the domain does not depend on the sequence $\{ \gamma_n \}_{n\in \mathbb{N}}$. Following Sarason [@sarason2008], this operator may be extended as $$Tf = \sum_{m=0}^\infty \left( \sum_{n=0}^\infty \gamma_n \hat f(n+m)\right) z^m$$ where the domain of $T$ is the collection of functions in $H^2$ for which $Tf \in H^2$.
\[facttoep\]Let $T$ be the extension of an upper triangular matrix given by $$Tf = \sum_{m=0}^\infty \left( \sum_{n=0}^\infty n! \hat f(n+m) \right) z^m.$$ The domain of $T$ is defined to be $D(T) = \{ f \in H^2 : Tf \in H^2 \}.$ Every function $f \in D(T)$ is an entire function and can be written as $f(z) = \sum_{n=0}^\infty a_n \frac{z^n}{n!}$ where $\sum_{n=0}^\infty a_n$ converges.
\[factlemm\]The sequence $\{ c_m = \sum_{n=1}^\infty (n+1)^{-m} \}_{m=2}^\infty$ is an $l^2$ sequence.
Each term of the sequence can be bounded by $$\int_{0}^\infty \frac{1}{(1+x)^m} dx = \frac1{m-1}.$$ Thus, $c_m$ is bounded by an $l^2$ sequence, and so it is also $l^2$.
First suppose that $f \in D(T)$. By definition, the zero-th coefficient of $Tf$ is given by $\sum_{n=0}^\infty n! \hat f(n)$, which must be a convergent series. Declaring $a_n = n! \hat f(n)$, it can be seen that $\sum a_n$ converges. Moreover, since $f(z) = \sum_{n=0}^\infty a_n \frac{z^n}{n!}$, the function $f$ must be an entire function.
For the other direction, suppose that $f(z) = \sum_{n=0}^\infty a_n \frac{z^n}{n!}$ where $\sum a_n$ converges. Define $d_0 = \sum_{n=0}^\infty n! \hat f(n) = \sum_{n=0}^\infty a_n$. Note that since $a_n$ converges so does $\sum_{n=0}^\infty a_n b_n$ for any positive monotonically decreasing sequence $\{ b_n \}$. Thus for each $m = 1, 2, ...$ the series $$d_m = \sum_{n=0}^\infty n! \hat f(n+m) = \sum_{n=0}^\infty \frac{a_{n+m}}{(n+1)(n+2)\cdots(n+m)}$$ converges. This enables us to define $Tf$ formally as $\sum_{m=0}^\infty d_m z^m$.
In order to demonstrate that $d_m$ is in $l^2$, write $d_m$ as follows: $$d_m = \frac{a_m}{m!} + \sum_{n=1}^\infty \frac{a_{n+m}}{(n+1)(n+2)\cdots(n+m)} := s_m + t_m.$$
The sequence $\{ s_m \} \in l^2$ since $a_m \to 0$. The sequence $\{ t_m \}$ is bounded by $\sum_{n=1}^\infty (n+1)^m = c_m$ for sufficiently large $m$, since $|a_m| < 1$ for $m$ sufficiently large. By Lemma \[factlemm\], $t_m$ is in $l^2$. This completes the proof of the theorem.
The domain of the operator given in Theorem \[facttoep\] is not shift invariant.
The function $f(z) = \sum_{n=1}^\infty \frac{(-1)^n}{n} \frac{z^n}{n!} \in D(T)$, since $\sum_{n=1}^\infty (-1)^n/n$ converges. Now consider the function $$zf(z) = \sum_{n=1}^\infty \frac{(-1)^n}{n} \frac{z^{n+1}}{n!} = \sum_{n=2}^\infty \frac{(-1)^{n-1}}{n-1} \frac{z^{n}}{(n-1)!} = \sum_{n=2}^\infty \frac{(-1)^{n-1}n}{n-1} \frac{z^{n}}{n!}.$$ The series $\sum_{n=2}^\infty (-1)^{n-1} \frac{n}{n-1}$ does not converge, which means $zf(z) \not\in D(T)$.
By applying the same techniques used in proving Theorem \[facttoep\], a slightly weaker result can be found when $n!$ is replaced by a sequence of complex numbers $\{ \gamma_n \}$ with the growth condition $|\gamma_{n+1}| > (n+1) |\gamma_n|$.
Let $\{\gamma_n\}$ be a sequence of complex numbers as described above, and define the operator $Tf = \sum_{m=0}^\infty \left( \sum_{n=0}^\infty \gamma_n \hat f (n+m) \right) z^m$ with the domain $D(T)=\{ f \in H^2 : Tf \in H^2 \}$. The operator $T$ is densely defined and functions of the form $f(z) = \sum_{n=0}^\infty a_n \frac{z^n}{\gamma_n}$ where $\sum |a_n| <\infty$ are in its domain.
Conclusion
==========
This paper aims to give a partial answer to a question posed by Donald Sarason in [@sarason2008]. At the end of his paper, Sarason asked if a closed densely defined operator satisfying certain algebraic properties analogous to that of a bounded Toeplitz operator is determined by a symbol in some sense. We call these operators Sarason-Toeplitz operators.
The Sarason sub-symbol was presented as a family of potential symbol maps for Sarason-Toeplitz operators. For bounded and analytic densely defined Toeplitz operators, the Sarason sub-symbol is unique and characterizes the operators. In the case of a coanalytic densely defined Toeplitz operator, the Sarason sub-symbol produces a densely defined multiplication type Toeplitz operator that agrees with the original coanalytic Toeplitz operator on a restricted domain. Finally, these results were extended to a broader class of Sarason-Toeplitz operators, provided their domains contain functions that are ratios of $L^2$ functions and $H^2$ outer funtions.
Section 6 focused on a densely defined coanalytic Toeplitz matrix. The domain of an extension of the matrix was completely classified, and it was shown that the densely defined operator is not of Sarason-Toeplitz type. This demonstrates that the definition of a Toeplitz matrix and that of a Sarason-Toeplitz operator do not coincide.
For general Sarason-Toeplitz operators $T$ the density of the set $D_2(T)$ is unknown. It is also unknown if this set contains any nontrivial elements. This is a question to be addressed in future research, and presumably, if it is to be proven true, the closedness of $T$ must be leveraged.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the first kinematic analysis of the far outer halo globular cluster (GC) population in the Local Group galaxy M31. Our sample contains 53 objects with projected radii of $\sim 20-130$ kpc, of which 44 have no previous spectroscopic information. GCs with projected radii $\gtrsim~30$ kpc are found to exhibit net rotation around the minor axis of M31, in the same sense as the inner GCs, albeit with a smaller amplitude of 7919 km/s. The rotation-corrected velocity dispersion of the full halo GC sample is 10612 km/s, which we observe to decrease with increasing projected radius. We find compelling evidence for kinematic-coherence amongst GCs which project on top of halo substructure, including a clear signature of infall for GCs lying along the North-West stream. Using the tracer mass estimator, we estimate the dynamical mass of M31 within 200 kpc to be $M_{\rm M31} = (1.2-1.5) \pm 0.2 \times 10^{12}M_{\odot}$. This value is highly dependent on the chosen model and assumptions within.'
author:
- 'J. Veljanoski, A.M.N. Ferguson, A.D. Mackey, A.P. Huxor, M.J. Irwin, P. Côté, N.R. Tanvir, E. J. Bernard, S. C. Chapman, R. A. Ibata, M. Fardal, G. F. Lewis, N. F. Martin, A. McConnachie and J. Peñarrubia'
title: Kinematics of Outer Halo Globular Clusters in M31
---
Introduction
============
Globular cluster (GC) systems contain important clues about the assembly history of galaxies [e.g., @West04Nature]. Their kinematics are especially important as different formation channels lead to distinct predictions [e.g. @Forbes97; @BrodieStraderReview]. Moreover, GC kinematics can also be used to model the shape of the gravitational potential and constrain the total mass of the host galaxy [e.g, @Cote01].
Located at a distance of $\sim$780 kpc [@McConnachie05], the Local Group galaxy M31 provides an excellent opportunity to study a rich GC system in unparalleled detail. It has more than 500 confirmed members listed in the Revised Bologna Catalogue [RBC, @Galleti04RBC], most of which lie within a projected radius ([$R_{proj}$]{}) of 30 kpc. In recent years, state-of-the-art wide field surveys [e.g, @Ferguson02; @Ibata07; @McConnachie09] have enabled searches for GCs in M31’s far outer halo. This has led to the discovery of over 90 new halo GCs, extending to [$R_{proj}$]{} $\sim140$ kpc and 3D radii of $\gtrsim$200 kpc (e.g, Huxor et al. 2005; Huxor et al. 2008; Huxor et al 2013, di Tullio Zinn & Zinn 2013). A major step forward in understanding the formation of the outer halo GC system came from the realisation that these objects preferentially lie on stellar streams and other debris features [@Mackey10ApJL; @Mackey13]. Monte Carlo simulations indicate a probability of $\lesssim 1\%$ that such alignments should happen by chance, leading to the conclusion that 80% of the M31 outer halo GCs have been accreted along with their host galaxies, confirming the idea put forward by @SZ78 for the Milky Way. In this Letter, we present the first results of a spectroscopic survey of these outer halo objects, focusing on their global kinematics. A detailed description of the data as well as a full analysis is deferred for a later publication (Veljanoski et al. in prep).
The data
========
Spectra were acquired for 53 GCs spanning [$R_{proj}$]{} $\sim20-130$ kpc, of which 12(6) lie beyond $\gtrsim80(100)$ kpc. This sample is complete down to $g\sim18.5$, and 44 of the clusters had not previously been observed spectroscopically. The data were obtained over 15 nights during 2005-2010 using the ISIS spectrograph on the WHT 4.2m, and the RC spectrograph on the KPNO 4m. ISIS has two detectors that independently sample the blue and red spectral range. We used the R600B and R600R gratings to cover the wavelength range $\sim350-510$ nm with a dispersion of 0.045 nm/pixel and $\sim750-920$ nm with a dispersion of 0.079 nm/pixel respectively. With the RC spectrograph, we used the KPC007 grating with a wavelength coverage of $\sim350-650$ nm and a dispersion of 0.139 nm/pixel. Total integrations were 600-7200 seconds depending on the target brightness. The slit width was 1-2$''$. The signal-to-noise per pixel was $\approx 7-30$ for most targets and 50-70 for the brightest objects.
The data were reduced using standard IRAF[^1] procedures. One-dimensional spectra were extracted with aperture radii of 2-2.5$''$. Heliocentric radial velocities were derived using a chi-squared minimization technique between GC spectra and radial velocity template stars (Veljanoski et al. 2013, in prep). This is analogous to the standard cross-correlation method, and it produces similar results. The method has the advantage that it uses the uncertainties in both the target and template spectra, which helps to eliminate spurious features in the chi-squared function.
The uncertainty in the radial velocity of each cluster is adopted to be the standard deviation of all the independent chi-squared minimizations between the cluster and multiple radial velocity standard stars. Furthermore, as we obtained two independent velocity measurements for the GCs observed with ISIS, these were combined to reduce the uncertainty in the final velocity value. The final median uncertainty of all 53 GC velocity measurements is 12 km/s. Of the 9 GCs in our sample that had published velocities in the RBC, all agree to within one standard deviation. Four of these clusters were found to have more precise RBC velocities compared to our measurements and for these we adopt the RBC values in our subsequent analysis.
As our GC sample spans a large extent on the sky, we converted our heliocentric radial velocities to the Galactocentric frame in order to remove any effects the Solar motion could have on the kinematics. This conversion was done using the relations presented in @Courteau99, with updated values for the Solar motion from @McMillan11 and @Schonrich10. For the purpose of this study, we take the M31 heliocentric velocity to be -301 4 km/s [@Courteau99], which translates to a Galactocentric radial velocity of -109 4 km/s.
Results
=======
The Global Velocity Map
-----------------------
Figure \[fig:velmap\] shows the most recent metal-poor (\[Fe/H\] $\lesssim -1.4$) red giant branch stellar density map from the Pan-Andromeda Archaeological Survey (PAndAS) [@McConnachie09]. Overlaid are the positions of the observed GCs, the colors of which correspond to their measured Galactocentric radial velocities.
Three interesting groups of GCs are indicated on Figure \[fig:velmap\]. The blue rectangle marks 4 GCs that project on the North-West stream. A contiguous velocity gradient is seen along this feature, with the most radially-distant cluster ([$R_{proj}$]{} $\approx$125 kpc) having a velocity of $\approx-183$ km/s and the innermost object ([$R_{proj}$]{} $\approx$67 kpc) having $\approx-380$ km/s. Also marked (red contour) are the GCs that lie on stream D, a feature first identified by @Ibata07. There is no apparent velocity gradient in this case, but the velocities of the GCs observed in the northeastern part of the stream suggest two distinct kinematic groups. In particular, 4 GCs have a mean velocity and dispersion of $-171$ km/s and 50 km/s, while the remaining 3 GCs have mean velocity and dispersion of 71 km/s and 18 km/s. Interestingly, the northeastern part of stream D has been previously identified to be a complicated region where it overlaps stream C [@Ibata07; @Richardson11]. Thus, it seems likely that the two kinematic GC groups we have identified are associated with these different streams, and are moving in the opposite sense around M31, as judged by their mean velocities. Radial velocities for stream stars in this overlapping region are not yet available, but the lone southern GC on stream D has a radial velocity that lies within $\sim20$ km/s of stream stars tentatively identified in a nearby field [@Chapman08]. Interestingly, the And-I dwarf, located on the south end of Stream D, has a Galactocentric velocity of $\sim$-190 km/s, similar to one of the kinematic GC groups. The And-IX dwarf, located on the north end of Stream D with velocity of $\sim$-20km/s, appears uncorrelated with the GC groups.
The yellow region on Figure \[fig:velmap\] marks “association 2” which @Mackey10ApJL identified as a statistical overdensity of GCs not associated with any obvious underlying stellar debris feature. The spread of velocities indicates that not all of the GCs in this region can be members of a kinematically coherent subgroup. However, measurements for the remaining $\sim$60% of the putative “association 2” GCs are required before the presence of a subgroup can be definitely ruled out.
{width="168mm"}
Rotation and Velocity Dispersion
--------------------------------
It has been known for some time that GCs in the inner regions of M31 rotate around the minor optical axis of the galaxy, in the same sense as the disk rotation. @Perrett02 measured a rotational amplitude of the GC system of $\sim140$ km/s, while @Lee08 measured $\sim190$ km/s using a larger sample. Dividing the GCs based on their metalliticy, @Deason11M31rot found more pronounced rotation for the metal-rich (\[Fe/H\] $>$-1) subpopulation. Inspection of Figure \[fig:velmap\] strongly suggests this rotation persists to larger radii, with GCs in the northeast having systematically higher velocities than those in the southwest. Figure \[fig:rot\] shows the Galactocentric radial velocities, corrected for the systemic motion of M31, versus their projected distances along the M31 major axis. GCs belonging to the subgroups identified in Figure \[fig:velmap\] are color-coded. The rotational signature appears to be a property of the bulk population of the outer halo GC sample and is not driven by one or two kinematically-coherent subgroups.
To further investigate the rotational signature, we follow @Cote01 and fit the observed projected Galactocentric radial velocities $v_p$ of the GCs with the function,
$$v_p(\theta) = v_{sys} + Asin(\theta - \theta_{0})
\label{eg:rot}$$
where $\theta$ is the projected position angle, measured east of north, of a cluster relative to M31 center, $\theta_{0}$ is the projected position angle of the GC system rotation axis, $v_{sys}$ is the systemic velocity of the GC system and $A$ is the amplitude of rotation. This approach assumes that the rotation axis of the GC system is perpendicular to the line of sight, and that the intrinsic angular velocity of the system is constant on spherical surfaces. The uncertainties are determined using the numerical bootstrapping technique [@Efron82] and the derived rotational amplitudes are corrected for the inclination of the M31 disk, taken to be 77.5 [@Ferguson02].
We augment our radial velocities with those from the RBC and fit the GC sample as a whole as well as within and beyond 30 kpc. This radius corresponds to a clear break in the GC radial number density profile [@Huxor11] and therefore provides a natural division between the “inner” and “outer" halo. For reference, the outer halo sample consists of the 48 GCs presented here, to which we add a further 2 confirmed GCs from the RBC.
-------------------------- ---------- -------------- --------------------- ----------
A $\theta_{0}$ Velocity Dispersion $N_{GC}$
\[km/s\] \[degrees\] \[km/s\]
All GCs 133 11 124 4 115 5 595
[$R_{proj}$]{}<30 kpc 137 10 124 4 114 5 545
[$R_{proj}$]{}>30 kpc 79 19 123 27 106 12 50
-------------------------- ---------- -------------- --------------------- ----------
: Derived Rotational Properties for M31 Halo GCs[]{data-label="tab:1"}
The systemic velocity of the GC system was set to the M31 Galactocentric systemic velocity. The results of this fitting are displayed in Table \[tab:1\]. The results remain unchanged when $v_{sys}$ in Equation \[eq:rot\] is left to vary as a free parameter, or when the mean velocity of the GC system is used. Given the position angle of the M31 major axis is 38, the derived rotation axis is consistent with the minor axis of M31, and is essentially indistinguishable from the rotation axis of the inner halo GCs.
![Galactocentric radial velocity, corrected for the M31 systemic motion, vs. projected radius along the M31 major axis. The black and colored squares mark the velocities of the GCs presented in this work, with different colors marking GC subgroups identified on Figure \[fig:velmap\]. RBC values are shown as small grey circles. The open squares correspond to the mean velocities of the GCs with [$R_{proj}$]{}>30 kpc calculated in 20 kpc bins. The $x$ error bars represent the bin size, while the $y$ error bars indicate the standard deviation of the mean. The green solid lines correspond to our measured amplitude for the outer GCs corrected for inclination.[]{data-label="fig:rot"}](fig2.ps){width="86mm"}
Finally, we derived the rotation-corrected velocity dispersion of the M31 GC sample using the biweight scale of @Beers90. Figure \[fig:vproj\] shows the rotation-corrected Galactocentric radial velocities (corrected for the systemic motion of M31) versus [$R_{proj}$]{} of the M31 GCs. It reveals a decreasing velocity dispersion with increasing distance from the M31 center, varying from $\sim122$ km/s at 60 kpc to $\sim57$ km/s at 120 kpc. For reference, the metal-poor field star velocity dispersion is $\sim$98 km/s at 60 kpc (measured), and $\sim$44 km/s at 120 kpc (extrapolated) [@Chapman06].
![Rotation-corrected Galactocentric radial velocity, corrected for the M31 systemic motion, vs. [$R_{proj}$]{} from the M31 center. Symbols are as in Figure \[fig:velmap\]. The open squares correspond to the mean velocities of the halo clusters in 20 kpc bins. The $x$ error bars mark the bin size, while the $y$ error bars represent the rotation-corrected velocity dispersion. The dotted line marks the 30 kpc radius.[]{data-label="fig:vproj"}](fig3.ps){width="86mm"}
An M31 Mass Estimate
--------------------
Assuming the halo GC system is spherically symmetric, we can estimate the mass of M31 by solving the Jeans equation [@BT87]. Because we have found evidence for rotation in the M31 halo GC system, the Jeans equation is separated into a rotating and a non-rotating component. The total mass is obtained by summing the mass supported by pressure $M_p$ and the rotationally-supported mass $M_{r}$. The rotational component is determined via:
$$M_{r} = \frac{R_{max}v^{2}_{max}}{G}
\label{eq:rot}$$
where $R_{max}$ is the projected radius of the outermost GC in our sample, $v_{max}$ is the rotational amplitude of the outer GC population and $G$ is the gravitational constant.
To determine $M_{p}$ we use the solution of the non-rotating Jeans equation proposed by @Evans03, known as the tracer mass estimator (TME):
$$M_{p} = \frac{C}{GN}\sum\limits_{i=1}^{N}\left(v_i-v_{sys}\right)^{2}R_i
\label{eq:tme}$$
where $R_i$ is the projected radius from the center of M31 for a given cluster, $v_i$ is the radial velocity of the GC with the rotational component removed and $N$ is the total number of clusters in our sample. The constant $C$ depends on the shape of the potential, the radial distribution of the tracer objects and the anisotropy in the system. For a spherical, isotropic system, it has the following form:
$$C = {4(\alpha\!+\!\gamma) \over \pi}
{4\!-\!\alpha\!-\!\gamma\over 3\!-\!\gamma}
{1\!-\!(r_{\rm in}/r_{\rm out})^{3-\gamma} \over 1\!-\!(r_{\rm in}/r_{\rm out})^{4\!-\!\alpha\!-\!\gamma}}.
\label{eq:C}$$
In equation (\[eq:C\]), $r_{\rm in}$ and $r_{\rm out}$ correspond to the smallest and largest 3D radii of the halo GCs respectively. In this case, we assume $r_{\rm in} = 30$ kpc, while $r_{\rm out}$ is set to 200 kpc assuming MGC1 is the most remote GC [@Mackey10MNRAS]. The constant $\alpha$ is related to the underlying gravitational field, which is assumed to be scale-free, at least between $r_{\rm in}$ and $r_{\rm out}$. For an isothermal halo potential where the system has a flat rotation curve at large radii, $\alpha$ is zero. If we assume a NFW dark matter profile [@NFW], $\alpha \approx 0.55$ [@Watkins10]. The $\gamma$ parameter is the slope of the GC volume density distribution. We calculate this using the surface density distribution of all 84 GCs that have projected radii larger than 30 kpc (Huxor et al. 2013, in prep; Mackey et al. 2013, in prep) and find $\gamma \sim 3.34 $. It is also important to note that even though the TME uses a GC sample in a shell around the center of M31, it calculates the total mass enclosed by the furthest cluster in that sample.
Using the above method, and setting $\alpha$ to zero, we calculate the total mass of M31 to be $M_{\rm M31}=1.5\pm0.2\times10^{12}
M_{\odot}$, where the pressure component is $M_{\rm p}=1.3\pm0.2
\times10^{12}M_{\odot}$ and the rotation contribution is $M_{\rm
r}=2\pm1\times10^{11}M_{\odot}$. Assuming $\alpha = 0.55$, we find the M31 mass to be $M_{\rm M31}=1.2\pm0.2\times10^{12}
M_{\odot}$, where $M_{\rm p}=1.0\pm0.2\times10^{12}M_{\odot}$ and $M_{\rm r}=2\pm1\times10^{11}M_{\odot}$. For reference, applying the TME in a single step (ignoring rotation), gives $M_{\rm M31}=1.8\pm0.2\times10^{12}M_{\odot}$ and $M_{\rm M31}=1.3\pm0.2\times10^{12}M_{\odot}$ for $\alpha$ values of 0 and 0.55 respectively. The quoted errors incorporate the statistical uncertainties only, and in reality they are much larger. In our mass calculations, we assume isotropic orbits, a steady state for our tracer population and a power-law form for the potential. This is the simplest approach we can take, although the presence of the substructure in the spatial distribution of the outer halo GCs suggests that it may not be correct. Nonetheless, studies suggest the presence of substructure in the tracer population will bias results only at the 20% level [e.g., @Yencho06; @Deason12]. To explicitly test the assumption of steady state, we recalculate the M31 mass excluding GCs which lie along the North-West stream and stream D. We find the total mass of M31 decreases by $0.3 \times 10^{12}
M_{\odot}$ for both values of $\alpha$, with the formal statistical errors remaining unchanged.
Discussion
==========
Analysis of the radial velocities of M31 outer halo GCs strongly supports our earlier finding that many of these objects have been accreted [@Mackey10ApJL]. Clear kinematic correlations are seen amongst subgroups of GCs which lie on top of stellar debris features, and, in the case of the North-West stream, an unambiguous signature of radial infall is observed. Interestingly, recent work has also began to detect phase-space substructure in large samples of GCs around distant galaxies [e.g. @Strader11; @Blom12; @Romanowsky12]. However it is only within the Local Group that we can attempt the obvious next step of comparing GC velocities with those of underlying stream stars.
A surprising discovery is the high degree of rotation in the M31 outer halo GC population, which is in the same sense as the inner halo population as well as the main stellar disk. While it is common to find GCs rapidly rotating in inner regions of galaxies, strong rotation beyond a few tens of kpc seems to be a rare occurrence, at least amongst early type galaxies [e.g. @Woodley10; @Strader11; @Blom12; @Pota13]. It is natural to speculate on how such coherent motion could arise if the outer GC population were largely accreted from numerous dwarf galaxy hosts. One possibility is that the donor dwarf galaxies have been accreted from a few preferred directions on the sky and hence have aligned angular momenta, as seen in some recent cosmological simulations [@Libeskind05; @Libeskind11; @Lovell11]. Such a scenario has also been suggested as the origin of the planar alignments of dwarf galaxies seen in both the Milky Way and M31 [e.g, @Metz07; @ibata13]. Curiously, the plane of dwarf galaxies reported in M31 rotates in the same sense as the outer halo GC population, although the rotation axis of that plane appears to be inclined by $\sim 45\deg$ to the minor axis. It is also possible that the bulk of the M31 outer halo GC population was accreted in a single event involving a moderate mass satellite. Support for this idea could come from the M31 thick disk which rotates in the same sense, albeit somewhat faster, than the halo GC population [@collins11]. However, a rather massive satellite would be required to bring in a population of several tens of GCs, raising the question of how the M31 disk could survive such an encounter. Furthermore, it would seem difficult to explain the spatial correlation between outer halo GCs and the numerous tidal streams in this case. In a different scenario, numerical modelling [@Bekki10M31] suggests that a past major merger between M31 and another disk galaxy could give rise to the rapid rotation of the resulting GC system, including the rotation in the halo population.
Despite being our closest massive galaxy, it is remarkable that we are still unable to measure the total mass of M31 with good precision. Indeed, there is still debate as to whether M31 or the Milky Way is more massive [@Watkins10]. @Evans03 used GC kinematics to find a M31 mass of $1.2\times10^{12}
M_{\odot}$ out to a deprojected 3D distance of $\sim$100 kpc, while @Galleti06 and @Lee08 used expanded samples to calculate masses of $1.9-2.4
\times10^{12}M_{\odot}$ within the same radial range. Several authors have attempted to determine the M31 mass via the motions of its dwarf satellite galaxies [e.g. @Cote00; @Evans03]. The most recent such measurement is presented by @Watkins10 who estimated $1.4\pm0.4\times10^{12}
M_{\odot}$ using 23 satellites out to 300 kpc, assuming isotropy. Our estimate of $M_{\rm M31}=1.2-1.5\pm0.2\times10^{12}M_{\odot}$ within a 3D radius of 200 kpc agrees well with this value but suffers from the similar systematic uncertainties due to model assumptions. Notably, @vdMarel12 use the velocity vector of M31 with respect to the Milky Way in combination with the timing argument to derive a total mass for the Local Group $M_{LG} = (4.39
\pm1.63)\times10^{12} M_{\odot}$, which would push the M31 mass higher than any estimate thus far using dynamical tracers.
ADM is supported by ARC grant DP1093431. APH was partially supported by Sonderforschungsbereich SFB 881 “The Milky Way System" of the German Research Foundation. The WHT is operated on the island of La Palma by the Isaac Newton Group in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias.
[*Facilities:*]{} .
Beers, T. C., Flynn, K., & Gebhardt, K. 1990, , 100, 32
Bekki, K. 2010, , 401, L58
Binney, J., & Tremaine, S. 1987, Princeton, NJ, Princeton University Press, 1987, 747 p.,
Blom, C., Forbes, D. A., Brodie, J. P., et al. 2012, , 426, 1959
Brodie, J. P., & Strader, J. 2006, , 44, 193
Chapman S. C., Ibata R., Lewis G. F., Ferguson A. M. N., Irwin M., McConnachie A., Tanvir N., 2006, ApJ, 653, 255
Chapman, S. C., Ibata, R., Irwin, M., et al. 2008, , 390, 1437
Collins M. L. M., et al., 2011, MNRAS, 413, 1548
C[ô]{}t[é]{}, P., Mateo, M., Sargent, W. L. W., & Olszewski, E. W. 2000, , 537, L91
C[ô]{}t[é]{}, P., McLaughlin, D. E., Hanes, D. A., et al. 2001, , 559, 828
Courteau, S., & van den Bergh, S. 1999, , 118, 337
Deason, A. J., Belokurov, V., Evans, N. W., et al. 2012, , 425, 2840
Deason, A. J., Belokurov, V., & Evans, N. W. 2011, , 411, 1480
Efron, B. 1982, CBMS-NSF Regional Conference Series in Applied Mathematics, Philadelphia: Society for Industrial and Applied Mathematics (SIAM), 1982,
Evans, N. W., Wilkinson, M. I., Perrett, K. M., & Bridges, T. J. 2003, , 583, 752
Ferguson, A. M. N., Irwin, M. J., Ibata, R. A., Lewis, G. F., & Tanvir, N. R. 2002, , 124, 1452
Forbes, D. A., Brodie, J. P., & Grillmair, C. J. 1997, , 113, 1652
Galleti, S., Federici, L., Bellazzini, M., Fusi Pecci, F., & Macrina, S. 2004, , 416, 917
Galleti, S., Federici, L., Bellazzini, M., Buzzoni, A., & Fusi Pecci, F. 2006, , 456, 985
Huxor, A. P., Tanvir, N. R., Irwin, M. J., et al. 2005, , 360, 1007
Huxor, A. P., Tanvir, N. R., Ferguson, A. M. N., et al. 2008, , 385, 1989
Huxor, A. P., Ferguson, A. M. N., Tanvir, N. R., et al. 2011, , 414, 770
Ibata, R., Martin, N. F., Irwin, M., et al. 2007, , 671, 1591
Ibata, R. A., Lewis, G. F., Conn, A. R., et al. 2013, , 493, 62
Lee, M. G., Hwang, H. S., Kim, S. C., et al. 2008, , 674, 886
Libeskind, N. I., Frenk, C. S., Cole, S., et al. 2005, , 363, 146
Libeskind, N. I., Knebe, A., Hoffman, Y., et al. 2011, , 411, 1525
Lovell, M. R., Eke, V. R., Frenk, C. S., & Jenkins, A. 2011, , 413, 3013
Mackey, A. D., Ferguson, A. M. N., Irwin, M. J., et al. 2010a, , 401, 533
Mackey, A. D., Huxor, A. P., Ferguson, A. M. N., et al. 2010b, , 717, L11
Mackey, A. D., Huxor, A. P., Ferguson, A. M. N., et al. 2013a, , 429, 281
van der Marel, R. P., Fardal, M., Besla, G., et al. 2012, , 753, 8
McConnachie, A. W., Irwin, M. J., Ferguson, A. M. N., et al. 2005, , 356, 979
McConnachie, A. W., Irwin, M. J., Ibata, R. A., et al. 2009, , 461, 66
McMillan, P. J. 2011, , 414, 2446
Metz, M., Kroupa, P., & Jerjen, H. 2007, , 374, 1125
Navarro, J. F., Frenk, C. S., & White, S. D. M. 1996, , 462, 563
Perrett, K. M., Bridges, T. J., Hanes, D. A., et al. 2002, , 123, 2490
Pota, V., Forbes, D. A., Romanowsky, A. J., et al. 2013, , 428, 389
Richardson J. C., et al., 2011, ApJ, 732, 76
Romanowsky A. J., Strader J., Brodie J. P., Mihos J. C., Spitler L. R., Forbes D. A., Foster C., Arnold J. A., 2012, ApJ, 748, 29
Searle, L., & Zinn, R. 1978, , 225, 357
Sch[ö]{}nrich, R., Binney, J., & Dehnen, W. 2010, , 403, 1829
Strader, J., Romanowsky, A. J., Brodie, J. P., et al. 2011, , 197, 33
Watkins, L. L., Evans, N. W., & An, J. H. 2010, , 406, 264
West, M. J., C[ô]{}t[é]{}, P., Marzke, R. O., & Jord[á]{}n, A. 2004, , 427, 31
Woodley, K. A., G[ó]{}mez, M., Harris, W. E., Geisler, D., & Harris, G. L. H. 2010, , 139, 1871
Yencho B. M., Johnston K. V., Bullock J. S., Rhode K. L., 2006, ApJ, 643, 154
di Tullio Zinn, G., & Zinn, R. 2013, , 145, 50
[^1]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The dispersion relation of an electromagnetic wave in an unmagnetized neutral plasma is well known to be $\omega^{2}=\omega_{p}^{2}+c^{2}k^{2}$. A modified dispersion relation is presented taking into account the ion restoring force in the transverse direction.'
address: 'Department of Electrical Engineering, IIT Madras, Chennai - 600036.'
author:
- Kushal Shah
title: Dispersion relation of an electromagnetic wave in unmagnetized cold plasma
---
Introduction
============
It is well known that the dispersion relation of a plane electromagnetic wave in an unmagnetized neutral cold plasma is given by $\omega^{2}=\omega_{p}^{2}+c^{2}k^{2}$. To derive this, it has been assumed that the electron oscillations in the transverse direction are of the same amplitude for all the electrons on a plane perpendicular to a given point on the axis. Though this assumption is true in theory, but questionable in practice. In any real experiment, the electrons will oscillate with varying amplitude in the transverse direction. This could happen due to boundary effects or many other reasons. This effect can be small or large depending on the problem at hand. For large amplitude variations, the above dispersion relation will certainly breakdown. The question at hand is: Does the above dispersion relation hold in the case of arbitrarily small differences in the amplitude of single electron oscillations in the transverse direction? Before we answer this, let us consider the nature of langmuir oscillations.
The nature of langmuir oscillations is very peculiar. For a 1d neutral cold plasma, the dispersion relation for langmuir waves is $\omega=\omega_{p}$. This dispersion relation does not seem to depend on $k$, but there is indeed a dependence. The correct way of writing the dispersion relation for a langmuir wave is:$$\begin{aligned}
\omega & = & 0\qquad k=0\\
& & \omega_{p}\qquad k\ne0\end{aligned}$$ This is a very peculiar dispersion relation because $\omega$ is not a smooth function of $k$. There is an abrupt jump at $k=0$. This dispersion relation also holds in the nonlinear limit. The only thing to make sure is that the wave amplitude should not be large enough to cause wave-breaking.
As discussed before, when a plane electromagnetic wave propagates in a plasma, we derive its dispersion relation assuming that there is no variation in the transverse direction. In any practical system, this is certainly not true. The assumption made is valid for systems where the scale length of transverse variation is much larger compared to the scale length of longitudinal variation. But no matter how small the transverse variation may be, it will always be non-zero. And the derivation of langmuir waves shows that all that we need to have electrostatic oscillations at $\omega=\omega_{p}$ is a non-zero $k$. This $k$ can be as small as we like, but as long as it is $>0$, we will certainly have electrostatic oscillations. This may sound trivial, but has very important implications.
Thus, when an electromagnetic wave travels in a plasma, there will always be electrostatic oscillations along the transverse direction. And, as we will see, these electrostatic oscillations do modify the dispersion relation of the electromagnetic wave.
We first derive the known dispersion relation for em waves in a plasma. Then, we add the electrostatic restoring force and show how it modifies the em wave dispersion relation.
Usual Derivation
================
This derivation is given in any book on plasma physics[@nicholson]. I am reproducing it here for the benefit of the reader.
The momentum equation is:$$mn\left[\frac{\partial\vec{v}}{\partial t}+\vec{v}\cdot\vec{\nabla}\vec{v}\right]=qn\left[\vec{E}+\vec{v}\times\vec{B}\right]$$ And the two Maxwell’s equations required are:$$\begin{aligned}
\vec{\nabla}\times\vec{B} & = & \mu_{0}\vec{J}+\mu_{0}\epsilon_{0}\frac{\partial\vec{E}}{\partial t}\\
\vec{\nabla}\times\vec{E} & = & -\frac{\partial\vec{B}}{\partial t}\end{aligned}$$ Under the linear approximation, we drop the nonlinear terms in the momentum equation. Then the momentum equation becomes,$$\frac{\partial\vec{v}}{\partial t}=\frac{q}{m}\vec{E}$$ $$\Rightarrow\frac{\partial\vec{J}}{\partial t}=\frac{q^{2}n}{m}\vec{E}$$ Differentiating the first of the Maxwell’s equation listed w.r.t. time and substituting appropriately, we get,$$-\vec{\nabla}\times\left(\vec{\nabla}\times\vec{E}\right)=\frac{q^{2}n}{m}\vec{E}+\frac{1}{c^{2}}\frac{\partial^{2}\vec{E}}{\partial t^{2}}$$ Now, since $\vec{\nabla}\cdot\vec{E}=0$, we get,$$\vec{\nabla}^{2}\vec{E}-\frac{1}{c^{2}}\frac{\partial^{2}\vec{E}}{\partial t^{2}}=\frac{w_{p}^{2}}{c^{2}}\vec{E}$$ Now, on fourier transforming in space and time, we get the usual dispersion relation:$$\omega^{2}=\omega_{p}^{2}+c^{2}k^{2}$$
A twist in the tale
===================
As explained in the introduction, as the electrons oscillate perpendicular to the direction of propagation of the wave, they will experience a restoring force due to ions due to infinitesimally small differences in the amplitude of electron motion in the transverse direction. And as long as wave breaking does not happen, this restoring force will be proportional to the displacement of the particle. This is true even in the nonlinear regime, as was shown by Dawson [@dawsom].
We assume that the em wave is polarized along the $\hat{x}$ direction and travels along the $\hat{z}$ direction. Thus, the correct momentum equation for the electrons should be,$$\frac{\partial^{2}x}{\partial t^{2}}+w_{p}^{2}x=\frac{q}{m}\vec{E}=\frac{q}{m}E_{0}\cos\left(\omega t-kz\right)$$ $$\Rightarrow x=-\frac{q}{m}\frac{E_{0}}{w^{2}-w_{p}^{2}}\cos\left(\omega t-kz\right)$$ $$\Rightarrow qnv=\frac{q^{2}n}{m}\frac{wE_{0}}{w^{2}-w_{p}^{2}}\sin\left(\omega t-kz\right)$$ $$\Rightarrow\frac{\partial J}{\partial t}=\frac{q^{2}n}{m}\frac{w^{2}E_{0}}{w^{2}-w_{p}^{2}}\cos\left(\omega t-kz\right)$$ The Maxwell’s equation was,$$-\vec{\nabla}\times\left(\vec{\nabla}\times\vec{E}\right)=\mu_{0}\frac{\partial\vec{J}}{\partial t}+\frac{1}{c^{2}}\frac{\partial^{2}\vec{E}}{\partial t^{2}}$$ $$\Rightarrow\omega^{2}=\frac{\omega_{p}^{2}\omega^{2}}{\omega^{2}-\omega_{p}^{2}}+c^{2}k^{2}$$ If $\omega\gg\omega_{p}$, then the above expression reduces to the conventional dispersion relation, $\omega^{2}=\omega_{p}^{2}+c^{2}k^{2}$.
The above expression gives,$$\omega^{4}-2\omega_{p}^{2}\omega^{2}-c^{2}k^{2}\omega^{2}+c^{2}k^{2}\omega_{p}^{2}=0$$ $$\begin{aligned}
\Rightarrow\omega^{2} & = & \frac{\left(2\omega_{p}^{2}+c^{2}k^{2}\right)\pm\sqrt{\left(2\omega_{p}^{2}+c^{2}k^{2}\right)^{2}-4c^{2}k^{2}\omega_{p}^{2}}}{2}\\
& = & \frac{\left(2\omega_{p}^{2}+c^{2}k^{2}\right)\pm\sqrt{4\omega_{p}^{4}+c^{4}k^{4}}}{2}\\
& = & \left(\omega_{p}^{2}+\frac{c^{2}k^{2}}{2}\right)\pm\sqrt{\omega_{p}^{4}+\frac{c^{4}k^{4}}{4}}\end{aligned}$$ For $ck\ll\omega_{p}$, we have,$$\begin{aligned}
\omega^{2} & = & \left(\omega_{p}^{2}+\frac{c^{2}k^{2}}{2}\right)\pm\omega_{p}^{2}\sqrt{1+\frac{c^{4}k^{4}}{4\omega_{p}^{4}}}\\
& \approx & \left(\omega_{p}^{2}+\frac{c^{2}k^{2}}{2}\right)\pm\omega_{p}^{2}\left(1+\frac{c^{4}k^{4}}{8\omega_{p}^{4}}\right)\end{aligned}$$ If $ck\gg\omega_{p}$, we have,$$\begin{aligned}
\omega^{2} & = & \left(\omega_{p}^{2}+\frac{c^{2}k^{2}}{2}\right)\pm\frac{c^{2}k^{2}}{2}\sqrt{1+\frac{4\omega_{p}^{4}}{c^{4}k^{4}}}\\
& \approx & \left(\omega_{p}^{2}+\frac{c^{2}k^{2}}{2}\right)\pm\frac{c^{2}k^{2}}{2}\left(1+\frac{2\omega_{p}^{4}}{c^{4}k^{4}}\right)\end{aligned}$$ As can be seem from the above expressions, $\omega<\omega_{p}$ is an allowed solution of the dispersion relation. We can get this by taking the -ve sign in the above expressions. This seems to an anomaly since it is conventionally believed that for $\omega<\omega_{P}$ , the em wave should not be able to penetrate in the plasma. There are two possibilities: Either the em wave with $\omega<\omega_{P}$ does actually penetrate into the plasma or there is some reason for us to discard the -ve sign in the above expression. This question can be answered only by more careful experiments in the future.
One could argue saying that the above field is not purely electromagnetic since we have also included an electrostatic component in the transverse direction. It does not really matter by what name we call it. All that matters is that the dispersion relation of a “initially” transverse wave in a plasma seems to be given by $$\omega^{2}=\frac{\omega_{p}^{2}\omega^{2}}{\omega^{2}-\omega_{p}^{2}}+c^{2}k^{2}$$
I would be glad to receive comments and views on this article through email. I am open to both sides of the arguments. If you can find a flaw in the above arguments and prove that this wrong, you are welcome. And if you have evidence or a strong reason to say that the above could be right, then of course, you are most welcome.
[99]{} D R. Nicholson, Introduction to Plasma Theory, (Wiley, New York, 1983) J. M. Dawson, Phys. Rev. 113, 383 (1959)
|
{
"pile_set_name": "ArXiv"
}
|
---
address: ' School of Physics, Georgia Institute of Technology, Atlanta, Georgia 30332-0430 '
author:
- Constantine Yannouleas and Uzi Landman
date: January 1997
---
[Comment on “Density functional theory study of some structural\
and energetic properties of small lithium clusters” \[J. Chem. Phys. 105, 9933 (1996)\]]{}
The experimentally determined size-evolutionary patterns (SEPs) of simple metal clusters pertaining to ionization potentials (IPs), electron affinities, and monomer separation energies have attracted much attention in the past ten years [@dehe; @brec], since they may reflect the electronic shell structure of clusters (the major features of SEPs are associated with steps at magic numbers, but often there is in addition substantial fine structure, such as odd-even alternations).
Due to the required computational effort, first-principles (FP) theoretical studies (which incorporate the ionic geometry) of such SEPs are usually limited to cluster sizes of the order of ten atoms. [@bona; @barn] Recently, however, systematic theoretical investigations of such SEPs have been performed for a broad range of cluster sizes using the jellium-related Shell-Correction-Method (SCM) approach. [@yann1; @yann2; @yann3; @yann4]
In a recent article, Gardet [*et al.*]{} [@gard] have studied the IPs of small lithium clusters Li$_N$ (with $N \leq 12$) using a FP - density functional theory (DFT) and found reasonable agreement between theory and experiment. Furthermore, through a comparison of their results with those obtained from Kohn-Sham local-density-approximation (KS-LDA) calculations on a [*spherical*]{} jellium background, [@ekar; @pusk] they concluded that the jellium model is inadequate for a proper description of the IPs of such systems.
The purpose of this comment is to clarify that, while modelling the ions in a cluster via a uniform jellium background is certainly an approximation, the above conclusion of Gardet [*et al.*]{} pertains to limitations introduced through neglect of deviations from spherical symmetry, rather than to the jellium approximation itself. Indeed, in a series of papers, [@yann2; @yann3; @yann4] we have demonstrated that consideration of triaxial shape deformations drastically improves the agreement between the jellium approximation and experiment for all instances of the aforementioned SEPs and for sizes up to 100 atoms, as well as for a variety of metal species (namely, alkali metals, such and Na and K, and noble metals, such as Cu and Ag).
To further elucidate the importance of shape deformations, we display in Fig. 1 the IPs of small Li$_N$ clusters (in the same size-range as with Ref. ) calculated [@note] with our SCM for three different families of shapes of the jellium background \[namely, spherical, spheroidal (axially symmetric), and ellipsoidal (triaxial)\], and compare them to the experimental measurements. [@broy]
Fig. 1 reveals that spherical shapes (top panel) exhibit a characteristic sawtoothed profile, well known from previous spherical KS-LDA studies [@ekar] and similar to the curve [@vezi] labeled jellium-LDA in Fig. 12 of Ref. . Apart from major-shell closures, this sawtoothed profile describes the data rather poorly (notice in particular the absence of fine structure between major shell closures at $N=2$, 8 and 20).
The spheroidal model (middle panel) exhibits substantial improvement in describing the experimental trend. Furthermore, the ellipsoidal case (bottom panel) improves the agreement between the SCM results and experiment even further, in particular in the size range $11 \leq N \leq 14$. The essential improvement introduced by the deformations over the spherical case concerns the very good description of the subshell closure at $N=14$ and of odd-even alternations between major shell closures.
In summary, we have illustrated once again that, within the jellium approximation, deformed cluster shapes provide an adequate description of the observed systematic size dependence of the properties of simple metal clusters and should necessarily be employed in comparisons with other theoretical approaches.
This research is supported by the US Department of Energy (Grant No. FG05-86ER-45234). Studies were performed at the Georgia Institute of Technology Center for Computational Materials Science.
W.A. de Heer, Rev. Mod. Phys. [**65**]{}, 611 (1993). see articles in [*Nuclear Aspects of Simple Metal Clusters*]{}, edited by C. Bréchignac and Ph. Cahuzac, Comments At. Mol. Phys. [**31**]{} (1995) nos. 3-6. V. Bonačić-Koutecký, P. Fantucci, and J. Koutecký, Chem. Rev. [**91**]{}, 1035 (1991). R.N. Barnett, U. Landman, and G. Rajagopal, Phys. Rev. Lett. [**67**]{}, 3058 (1991). C. Yannouleas and U. Landman, Phys. Rev. B [**48**]{}, 8376 (1993); Chem. Phys. Lett. [**210**]{}, 437 (1993). C. Yannouleas and U. Landman, Phys. Rev. B [**51**]{}, 1902 (1995). C. Yannouleas and U. Landman, in [*Large Clusters of Atoms and Molecules*]{}, edited by T.P. Martin, NATO ASI Series E: Applied Sciences – Vol. 313 (Kluwer, Dordrecht, 1996), p. 131. C. Yannouleas and U. Landman, Phys. Rev. Lett. [**78**]{}, 1424 (1997). G. Gardet, F. Rogemond, and H. Chermette, J. Chem. Phys. [**105**]{}, 9933 (1996). W. Ekardt, Phys. Rev. B [**29**]{}, 1558 (1984). M.J. Puska, R.M. Nieminen, and M. Manninen, Phys. Rev. B [**31**]{}, 3486 (1985).
In the case of Li$_N$ clusters, the parameters entering in the SCM calculation are $U_0=-0.035$, $r_s=3.25$ a.u., $t=2.22$ a.u., $\delta_0=\delta_2=0$, $W=2.31$ eV, $\alpha_v=-7.985$ eV, $\alpha_s=0.629$ eV, and $\alpha_c=0.125$ eV (see Ref. for their meaning). The liquid-drop parameters $\alpha_v$, $\alpha_s$, and $\alpha_c$ were specified through a fit to extended-Thomas-Fermi total energy calculations for spherical clusters in conjunction with a [*stabilized-jellium*]{}-LDA energy functional \[see J.P. Perdew, H.Q. Tran, and E.D. Smith, Phys. Rev. B [**42**]{}, 11 627 (1990) and C. Yannouleas and U. Landman, Chem. Phys. Lett. [**217**]{}, 175 (1994)\]. The need to use the stabilized-jellium-LDA energy functional arises from the fact that the usual jellium-LDA fails to describe the surface energy of metals with high densities (like the case of Li). The mass of the delocalized valence electrons was taken equal to the free-electron mass. Ph. Dugourd [*et al.*]{}, Chem. Phys. Lett. [**197**]{}, 433 (1992). B. Vezin-Pintar, Thesis, Université Lyon-I, 1994.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Takeshi <span style="font-variant:small-caps;">Matsumura</span>$^1$[^1], Hideaki <span style="font-variant:small-caps;">Ishida</span>$^1$, Taku J. <span style="font-variant:small-caps;">Sato</span>$^2$, Kenichi <span style="font-variant:small-caps;">Katoh</span>$^3$, Yuzuru <span style="font-variant:small-caps;">Niide</span>$^3$ and Akira <span style="font-variant:small-caps;">Ochiai</span>$^4$'
title: Spin Fluctuation and Crystal Field Excitation of a Heavy Fermion Compound YbAgGe studied by Inelastic Neutron Scattering
---
A ternary rare-earth compound YbAgGe has been attracting interest as a new Yb-based heavy-fermion material. [@Katoh04; @Umeo04; @Budko04; @Morosan04; @Budko04b] One major aspect is the coexistence of Kondo effect and geometrical frustration in the *kagomé*-lattice-like triangular coordination of Yb atoms in ZiNiAl-type crystal structure (hexagonal space group $P\bar{6}2m$). Heavy electron mass is observed which reaches about 400 mJ/mol K$^2$ at 2 K followed by successive magnetic phase transitions at quite low temperatures of 0.8 K and 0.65 K. [@Katoh04; @Umeo04] The two step transition is considered to be a characteristic of the geometrical frustration as observed in isostructural YbPtIn and YbRhSn. [@Trovarelli00; @Kaczorowski00] Another aspect of importance is the magnetic field induced non-Fermi-liquid (NFL) behavior which is typically manifested in the $-\ln T$ dependence of magnetic specific heat $C_{\text{mag}}/T$ in a wide temperature range from above 10 K to below 1 K. [@Budko04] YbAgGe offers a good platform to study a crossover from magnetically ordered state with small moments to a Fermi-liquid (FL) state through a quantum critical point (QCP), as well as recently discovered NFL compound YbRh$_2$Si$_2$. [@Trovarelli00b]
Characteristics of Kondo effect and heavy fermion are typically manifested in the magnetic specific heat. [@Katoh04] At low temperatures $C_{\text{mag}}/T$ increases with decreasing temperature in proportion to $-\ln T$ from about 15 K to 3 K. Katoh *et al.* analyzed the data with a single-ion Kondo model of $J=1/2$ for $T_{\text{K}}=24$ K. While the magnetic susceptibility at high temperatures follows Curie-Weiss behavior of well localized moment, $\chi(T)$ for $H\parallel a$ exhibits a weak maximum around 4 K, which suggests disappearance of the local moment. Although this is reminiscent of isostructural heavy fermion compound YbCuAl, [@Murani85] YbAgGe is different in that a well defined crystal field (CF) anomaly is observed; $C_{\text{mag}}(T)$ exhibits a Schottky-type peak at around 60 K, which indicates the first excited CF level at around 110 K. [@Katoh04] Our final goal in YbAgGe will be to understand the whole process from the local-moment formation at zero field to the FL state at high fields through the QCP where the NFL state is observed. For this purpose, it is of essential importance to study correlation and fluctuation of the magnetic moments by inelastic neutron scattering (INS). The aim of the present study, as a starting subject, is to investigate spin fluctuation associated with the Kondo effect and to determine the CF level scheme, using polycrystalline sample at zero magnetic field.
INS experiment was performed using the LAM-D spectrometer at the KENS pulsed neutron source. LAM-D is an inverted geometry spectrometer using crystal analyzer mirrors. [@Inoue93] The incident neutrons have a Maxwellian distribution of energy at room temperature and the scattered neutrons with an energy of 4.59 meV are selected by the (002) reflection of the pyrolytic graphite analyzer-crystals before being counted by the detector. The energy and momentum transfer are analyzed by the time of flight and the scattering angle. Two analyzer mirrors cover a range of low scattering angle around $\phi=35^{\circ}$; the other two cover high scattering angle around $\phi=85^{\circ}$ which we use to estimate the contribution from phonon scattering and to check the scattering-vector dependence of magnetic scattering.
A collection of small pieces of polycrystals with total mass 14.2 g was sealed in an aluminium container with helium exchange gas and attached to a liquid-helium cryostat. The details of the sample preparation and characterization are described in Ref. . The scattering intensity was put on an absolute scale in mb/meV/sr/Yb by comparing the intensity with that of the incoherent scattering from a vanadium standard sample. Contribution of phonon scattering was estimated from the data for $\phi=35^{\circ}$ and $\phi=85^{\circ}$ at high temperatures above 100 K; at low temperatures it was estimated from the high temperature results by calculating the Bose factor.
Figure \[fig1\] shows the magnetic part of the neutron scattering function $S_{\text{mag}}(\phi,\omega)$ averaged for $\phi=35^{\circ}$ and $85^{\circ}$. Two features are clearly observed; one is a quasielastic (QE) scattering around $\hbar\omega=0$ meV and the other a magnetic excitation around 12 meV. Both of them are much broader than the energy resolution. The 12 meV excitation can be ascribed to the crystal field (CF) excitation of the Yb$^{3+}$ ions. A scenario that the first excited doublet is located at 12 meV above the ground doublet is consistent with the Schottky-type specific heat anomaly around 60 K reported in Ref. . No other significant peak structures were observed up to 60 meV.
![Magnetic part of the neutron scattering function of YbAgGe per Yb ion. The data for 20 K, 50 K, and 100 K, are shifted by 15, 30, and 45, respectively. Solid lines are the fits with Lorentzian spectral functions. The dashed line is the measured spectrum of a vanadium standard sample normalized to the intensity of the incoherent scattering of the YbAgGe sample, representing the resolution function at 0 meV. []{data-label="fig1"}](Espect1.eps){width="7.5cm"}
We have analyzed the magnetic excitation spectra with $$S_{\text{mag}}(\phi,\omega)=\frac{\hbar\omega}{1-\exp (-\hbar\omega/k_{\text{B}}T)}
\sum_{i=1}^{2} f_{i} P_{i}(\hbar\omega; \Delta_i)\;,
\label{eq:1}$$ where $P_i$ represents a spectral function of the $i$th excitation with an energy transfer $\hbar\omega=\Delta_i$ and $f_i$ the strength of the excitation. Lorentzian spectral function was assumed here. Fitting results are shown by the solid lines in Fig. \[fig1\]. With respect to the low energy excitation, the spectra above 20 K can be fitted well by the Lorentzian centered at $\hbar\omega=0$ meV, which indicates that the excitation is quasielastic. Although there remains a possibility that the excitation is inelastic with $\Delta$ less than 1 meV, it was not possible to determine within our energy resolution. The excitation at 12 meV can also be fitted well by the Lorentzian at all temperatures.
The temperature dependences of the spectral strengths $f_i$ roughly exhibit the behavior of the bulk magnetic susceptibility. The strength of the QE scattering decreases with increasing temperature in accordance with the Curie term, while that of the 12 meV excitation do not vary with temperature very much since this corresponds to the Van-Vleck term. The decrease in the observed intensity in Fig. \[fig1\] is due to the temperature effect in eq. (\[eq:1\]). The sum of the two spectral strengths at 1.5 K becomes $\sim 160$ mb, which corresponds to the static susceptibility of 0.042 emu/mol and is consistent with the bulk susceptibility reported in Ref. .
Figure \[fig2\] shows the temperature dependence of the half width $\Gamma$ of the Lorentzian spectral functions. The widths of both QE scattering and CF excitation become broad with increasing temperature. Although the CF excitation has broader width than the QE peak, this is due to the broader energy resolution at 12 meV than at 0 meV; the two peaks have the same intrinsic widths. Even if we take into account the convolution effect by the resolution function, the intrinsic widths of the QE peak will be almost the same as those in Fig. \[fig2\].
![Temperature dependence of the half width of the crystal field excitation and the quasielastic scattering. Solid and dashed lines represent fits with $\Gamma_{0}+A\sqrt{T}$. []{data-label="fig2"}](gamma.eps){width="7.5cm"}
Here, let us consider the CF level scheme. Since the eightfold multiplet of Yb$^{3+}$ ($J=7/2$) splits into four Kramers doublets, three excitations are expected at the lowest temperature. However, only one excitation was observed experimentally. It is necessary to look for a level scheme in which only the excitation from the ground doublet to the first excited doublet has a significant transition strength and the other excitations has negligibly small probability.
Since the local symmetry of the Yb site is $C_{2v}$ when viewed from the $a$ axis, all the CF parameters $B_{kq}$ with even $k$ and $q$ up to six appear. Then, we first estimated the CF parameters from the electrostatic field of point charges in which we assumed negative charges for Ag and Ge, and a positive charge for Yb. As a result, we found that the $B_{22}$ term becomes the main source of the crystal field and the $B_{20}$ term modifies it slightly; higher order terms with $k=4$ and 6 are negligibly small. This is a straightforward result of the crystal structure of YbAgGe. Looking from the $c$ axis at the Yb ion at ($x$,0,0) on the $3f$ site and the surrounding Ag and Ge atoms, we notice that the Yb ion is in a strong $B_{22}$ type ($=x^2-y^2$ type) crystal field. [@Gibson96] The $a$ axis becomes the principal axis of the CF anisotropy for this Yb ion. The anisotropy along the $c$ axis, which is a natural result of the hexagonal structure, induces the $B_{20}$ term. Since there is not an inversion symmetry, CF parameters of odd rank also exist; but these do not work as an actual crystal field, and the lowest order CF parameters are $B_{22}$ and $B_{20}$.
Following the above observation, we have neglected the 4th and 6th order terms and calculated the CF eigenfunctions only with $B_{20}$ and $B_{22}$. First, with $B_{22}$ only, we found that only the transition strength between the ground and 1st excited doublets has a significantly large value. Next, we modified the eigenfunctions by slightly including $B_{20}$ so that the magnetization curve and the temperature dependence of the magnetic susceptibility can be reproduced as well as possible. Our model of the CF parameters are $B_{22}=-14.77$ K and $B_{20}=1.477$ K, which give 1st, 2nd, and 3rd excited levels at 140 K ($=12$ meV), 228 K, and 323 K, respectively.
The matrix elements of $J_{\perp}$, the perpendicular component of $\mib{J}$ to the scattering vector, for thus obtained eigenfunctions are listed in Table \[tabl1\]. It is noted that the direction of the principal axis of the CF anisotropy for the three Yb ions at ($x$,0,0), (0,$x$,0), and ($-x$,$-x$,0), are different from each other by $120^{\circ}$. In addition, the matrix elements were averaged over the direction of the scattering vector to estimate the value for the polycrystalline sample. It can be seen from this table that the matrix element for the $(0,1)$ transition, which represents the transition between the ground state $|0\rangle$ and the 1st excited state $|1\rangle$, is much larger than those for the $(0,2)$ and $(0,3)$ transitions. Then, at low temperatures where only the ground state is populated, it is expected that there appear only one sizable excitation for $(0,1)$. When the temperature is elevated so that the 1st excited level is populated, the excitation for the $(1,2)$ transition is also expected.
$|0\rangle$ $|1\rangle$ $|2\rangle$ $|3\rangle$
-------------- ------------- ------------- ------------- -------------
$\langle 0|$ 20.17 3.99 0.11 0.02
$\langle 1|$ 9.61 6.70 0.45
$\langle 2|$ 7.32 4.63
$\langle 3|$ 15.13
: Polycrystalline average of the squared matrix elements $|\langle i | J_{\perp} | j \rangle |^2$ for the crystal field eigenstates of Yb$^{3+}$ described in the text. []{data-label="tabl1"}
----------- --------- --------- --------- --------- --------- ---------
$(i,j)$ $(0,1)$ $(0,2)$ $(0,3)$ $(1,2)$ $(1,3)$ $(2,3)$
$E$ (meV) 12.1 19.7 27.8 7.6 15.8 8.1
$T$ (K)
1.5 K 31.2 0.51 0.055 0.00 0.00 0.00
50 K 27.3 0.47 0.051 3.89 0.15 0.44
100 K 17.0 0.33 0.038 8.6 0.40 2.40
----------- --------- --------- --------- --------- --------- ---------
: Polycrystalline average of the transition strengths of the $|i\rangle \rightarrow |j\rangle$ excitation at 1.5 K, 50 K, and 100 K. []{data-label="tabl2"}
Calculated transition strengths at three different temperatures are listed in Table \[tabl2\] in unit of mb. Corresponding energy positions are also listed. The calculated result at 1.5 K is consistent with the experimental result that only one excitation is observed, and it can be concluded that the excitation corresponds to the $(0,1)$ transition. At elevated temperatures above 100 K, another excitation corresponding to the $(1,2)$ transition is expected to appear around 7.6 meV with its intensity about half as large as that of the $(0,1)$ transition. If this peak was observed experimentally, we could determine the energy level of the 2nd excited state; however, it was difficult to identify this excitation in the background of phonon scattering which dominates the observed intensity at high temperatures. More careful subtraction of phonon scattering using a reference sample is necessary to identify this peak.
The experimental results and the analysis of INS support the CF model in which the $B_{22}$ term is dominant. Figure \[fig3\] demonstrates the calculated magnetization and magnetic susceptibility for the CF parameters used in the analysis: $B_{22}=-14.77$ K and $B_{20}=1.477$ K. Molecular field is not considered; it just shifts $1/\chi(T)$ along the vertical axis and changes the scale of magnetic field in the $M(H)$ curves. Although the effective moment at high temperatures and the magnetic anisotropy along $a$ and $c$ axes are consistent with the experiment, $1/\chi(T)$ along the $c$ axis at low temperatures deviates from the experimental results. Since the calculated ground state has strong magnetic anisotropy within the $ab$ plane, $\chi(T)$ for $H\parallel c$ levels off below temperatures where the 1st excited state is depopulated. However, the experimental data follows the Curie-Weiss law down to the lowest temperature. Slight modification of $B_{20}$ cannot alter this behavior. Some mechanism which induce magnetic moments along the $c$ axis is necessary, which could be a hybridization with the conduction band or a ferromagnetic exchange in the Yb chain along the $c$ axis. Concerning $H\parallel a$, $1/\chi(T)$ in calculation is not far from the experimental results if we include a molecular field. Calculated magnetization exhibits a saturation moment of about 2.5 $\mu_{\text{B}}$, which also is roughly consistent with the experimental result that 2.0 $\mu_{\text{B}}$ is induced at 15 T. [@Morosan04]
These CF parameters correspond to the set No. 12 in Table I of Ref. which can explain the bulk properties and the INS results of an isostructural ErNiAl. [@Javorsky01] CF excitations in PrNiAl, NdNiAl, and ErCuAl, have also been measured, but the CF parameters have not been deduced. [@Javorsky02] It should be noted, in our analysis, that the sign of $B_{22}$ could not be determined since the calculated results little depend on the sign of $B_{22}$ in the present analysis.
![Calculated magnetization curve at 5 K and temperature dependence of the inverse magnetic susceptibility for the crystal field model of YbAgGe described in the text. Experimental data in Ref. is shown by the circles ($H\parallel a$) and squares ($H\parallel c$). []{data-label="fig3"}](calc.eps){width="7.5cm"}
With respect to the spectral line shape, it is basically described by a Lorentzian in the whole temperature range of this study. The spectra in Fig. \[fig1\] may be regarded as single-site excitations. The QE line width follow roughly a $\Gamma_0+A\sqrt{T}$ law with intrinsic $\Gamma_0=0.9$ meV. The CF excitation is also broadened to the same width and exhibits the same temperature dependence. This behavior is commonly observed in Kondo systems in which the $4f$ electrons are almost localized; the width at $T=0$ indicates a single-site spin fluctuation and is regarded as a measure of $T_{\text{K}}$. [@HollandMoritz94; @Maekawa85] The value of $\Gamma_0=0.9$ meV and $\Gamma (300\text{K})\sim 3$ meV is slightly larger than the widths of YbBe$_{13}$, YbAuCu$_4$, YbPdCu$_4$, and YbPtBi, which are in the Kondo regime with well localized moments, [@HollandMoritz94; @Walter85; @Severing90; @Robinson95] but much smaller than those of YbCuAl and YbAgCu$_4$ in the intermediate valence regime, [@Murani85; @HollandMoritz94; @Severing90] Then, from the aspect of spin fluctuation, YbAgGe can be categorized in the Kondo regime and has a relatively high characteristic temperature $T_{\text{K}}$.
The transition temperatures of 0.8 K and 0.65 K in YbAgGe are several times lower than the Néel temperatures of other isostructural heavy-fermion Yb-compounds with lower $T_{\text{K}}$: $T_{\text{N}}=$3.4 K and 1.4 K in YbPtIn, 1.9 K and 1.7 K in YbRhSn, 3.5 K in YbPtSn, and 2.9 K in YbNiAl. [@Trovarelli00; @Kaczorowski00; @Schank95; @Diehl95] In view of the fact that the Néel temperatures of RAgGe for R=Tb-Yb roughly scale with the de Gennes factors, [@Morosan04] it is considered that the RKKY interaction gives $T_{\text{N}}$ of a few Kelvin for the Yb compounds with ZiNiAl-type structure. Since a distinctive magnetic structure of frustrated moments has been observed in YbNiAl which orders at 2.9 K, [@Ehlers97] it is supposed that the low transition temperature in YbAgGe is due to the Kondo effect rather than the geometrical frustration.
At the lowest temperature of 1.5 K, the QE line shape deviates from the Lorentzian as can be seen in Fig. \[fig1\]. This could be ascribed to emergence of spin correlation leading to a scattering-vector dependence of the QE intensity as observed, e.g., in CeCu$_6$ and CeRu$_2$Si$_2$. [@RossatMignod88; @Appeli86] This point will be an important subject in future experiments using a single crystalline sample.
In conclusion, we have measured inelastic neutron scattering spectra of a new heavy fermion compound YbAgGe for a polycrystalline sample at zero magnetic field. The spectra exhibit a quasielastic scattering and a crystal field excitation at 12 meV, both of which are broadened because of the spin fluctuation associated with Kondo effect. The temperature dependence of the line width follows a typical behavior of $\Gamma_0+A\sqrt{T}$ observed in Kondo systems. The intrinsic width of 0.9 meV at 1.5 K and 3 meV at 300 K is relatively high in comparison with other Yb-based Kondo systems. Crystal field parameter was also deduced from the analysis of the observed excitations. Model calculation assuming the $B_{22}$ term as the main component successfully explains the excitation spectra. This is related with the ZiNiAl-type structure in which the Yb site is in a strong field of anisotropy.
[99]{} K. Katoh, Y. Mano, K. Nakano, G. Terui, Y. Niide and A. Ochiai: J. Magn. Magn. Mater. **268** (2004) 212. K. Umeo, K. Yamane, Y. Muro, K. Katoh, Y . Niide, A. Ochiai, T. Morie, T. Sakakibara and T. Takabatake: J. Phys. Soc. Jpn. **73** (2004) 537. S. L. Bud’ko, E. Morosan and P. C. Canfield: Phys. Rev. B **69** (2004) 014415. E. Morosan, S. L. Bud’ko, P. C. Canfield, M. S. Torikachvili and A. H. Lacerda: J. Magn. Magn. Mater. **277** (2004) 298. S. L. Bud’ko, E. Morosan and P. C. Canfield: cond-mat/0406435. O. Trovarelli, C. Geibel, R. Cardoso, S. Mederle, R. Borth, B. Buschinger, F. M. Grosche, Y. Grin, G. Sparn and F. Steglich: Phys. Rev. B **61** (2000) 9467. D. Kaczorowski, B. Andraka, R. Pietri, T. Cichorek and V. I. Zaremba: Phys. Rev. B **61** (2000) 15255. O. Trovarelli, C. Geibel, S. Mederle, C. Langhammer, F. M. Grosche, P. Gegenwart, M. Lang, G. Sparn and F. Steglich: Phys. Rev. Lett **85** (2000) 626. A. P. Murani, W. C. M. Mattens, F. R. de Boer and G. H. Lander: Phys. Rev. B **31** (1985) 52. K. Inoue, T. Kanaya, Y. Kiyanagi, K. Shibata, K. Kaji, S. Ikeda, H. Iwasa and Y. Izumi: Nucl. Inst. Meth. A **327** (1993) 433. B. Gibson, R. Pöttgen, R. K. Kremer, A. Simon and K. R. A. Ziebeck: J. Alloys and Compounds **216** (1996) 34. P. Javorský, M. Diviš, H. Sugawara, H. Sato and H. Mutka: Phys. Rev. B **65** (2001) 014404. P. Javorský, H. Mutka and H. Nakotte: Appl. Phys. A **74** (2002) S658. E. Holland-Moritz and G. H. Lander: *Handbook on the Physics and Chemistry of Rare Earths*, ed. K. A. Gschneidner, Jr., L. Eyring, G. H. Lander and G. R. Choppin (Elsevier Science, B. V., 1994) Vol. 19, p. 1. S. Maekawa, S. Takahashi, S. Kashiba and M. Tachiki: J. Phys. Soc. Jpn. **54** (1985) 1955. U. Walter, Z. Fisk and E. Holland-Moritz: J. Magn. Magn. Mater. **47& 48** (1985) 159. A. Severing, A. P. Murani, J. D. Thompson, Z. Fisk and C.-K. Loong: Phys. Rev. B **41** (1990) 1739. R. A. Robinson, M. Kohgi, T. Osakabe, F. Trouw, J. W. Lynn, P. C. Canfield, J. D. Thompson, Z. Fisk and W. P. Beyermann: Phys. Rev. Lett. **75** (1995) 1194. C. Schank, G. Olesch, J. Köhler, U. Tegel, U. Klinger, J. Diehl, S. Klimm, G. Sparn, S. Horn, C. Geibel and F. Steglich: J. Magn. Magn. Mater. **140-145** (1995) 1237. J. Diehl, H. Davideit, S. Klimm, U. Tegel, C. Geibel, F. Steglich and S. Horn: Physica B **206& 207** (1995) 344. G. Ehlers, C. Geibel, F. Steglich and H. Maletta: Z. Phys. B **104** (1997) 393. J. Rossat-Mignod, L. P. Regnault, J. L. Jacoud, C. Vettier, P. Lejay, J. Flouquet, E. Walker, D. Jaccard and A. Amato: J. Magn. Magn. Mater. **76& 77** (1988) 376. G. Appeli, H. Yoshizawa, Y. Endoh, E. Bucher, J. Hufnagl, Y. Onuki and T. Komatsubara: Phys. Rev. Lett. **57** (1986) 122.
[^1]: E-mail address: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using data from gene expression databases on various organisms and tissues, including yeast, nematodes, human normal and cancer tissues, and embryonic stem cells, we found that the abundances of expressed genes exhibit a power-law distribution with an exponent close to -1, i.e., they obey Zipf’s law. Furthermore, by simulations of a simple model with an intra-cellular reaction network, we found that Zipf’s law of chemical abundance is a universal feature of cells where such a network optimizes the efficiency and faithfulness of self-reproduction. These findings provide novel insights into the nature of the organization of reaction dynamics in living cells.'
author:
- Chikara Furusawa
- Kunihiko Kaneko
title: 'Zipf’s Law in Gene Expression'
---
In a cell, an enormous number of organized chemical reactions are required to maintain its living state. Although enumeration of detailed cellular processes and the construction of complicated models is important for a complete description of cellular behavior, it is also necessary to search for universal laws with regard to the intra-cellular reactions common to all living systems, and then to unravel the logic of life leading to such universal features. For example, scale-free networks have recently been discussed as a universal property of some biochemical reaction networks within existing organisms [@metabolic; @lethal]. These studies, however, only focused on the properties of the network topologies, while the reaction dynamics of the networks were not discussed. Here, we report a universal property of the reaction dynamics that occurs within cells, namely a power-law distribution of the abundance of expressed genes with an exponent close to -1, i.e. a power-law distribution that obeys Zipf’s law [@Zipf]. By using an abstract model of a cell with simple reaction dynamics, we show that this power-law behavior in the chemical abundances generally appears when the reaction dynamics leads to a faithful and efficient self-reproduction of a cell. These findings provide insights into the nature of the organization of complex reaction dynamics in living cells.
In order to investigate possible universal properties of the reaction dynamics, we examined the distributions of the abundances of expressed genes (that are approximately equal to the abundances of the corresponding proteins) in 6 organisms and more than 40 tissues based on data publicly available from SAGE (Serial Analysis of Gene Expression) databases [@SAGE1; @SAGE2; @SAGE3]. SAGE allows the number of copies of any given mRNA to be quantitatively evaluated by determining the abundances of the short sequence tags which uniquely identify it [@SAGE_method].
In Fig.1, we show the rank-ordered frequency distributions of the expressed genes, where the ordinate indicates the frequency of the observed sequence tags (i.e. the population ratio of the corresponding mRNA to the total mRNA), and the abscissa shows the rank determined from this frequency. As shown, the distributions follow a power-law with an exponent close to -1 (Zipf’s law). We observed this power-law distribution for all the available samples, including 18 human normal tissues, human cancer tissues, mouse (including embryonic stem cells), rat, nematode ([*C. elegans*]{}), and yeast ([*S.cerevisiae*]{}) cells. All the data over 40 samples (except for 2 plant data) show the power-law distributions with the exponent in the range from $-1 \sim -0.86$. Even though there are some factors which may bias the results of the SAGE experiments, such as sequencing errors and non-uniqueness of tag sequences, it seems rather unlikely that the distribution is an artifact of the experimental procedure.
The abundance of each protein is the result of a complex network of chemical reactions that is influenced by possibly a large number of factors including other proteins and genes. Then, why is Zipf’s law universally observed, and what class of reaction dynamics will show the observed power-law distribution? Because the power-law distribution applies to a wide range of existing organisms, it is expected that it appears as a general feature of the reaction dynamics of cellular systems.
In order to investigate the above questions, we adopt a simple model of cellular dynamics that captures only its basic features. It consists of intra-cellular catalytic reaction networks that transform nutrient chemicals into proteins. By studying a class of simple models with these features, we clarify the conditions under which the reaction dynamics leads to a power-law distribution of the chemical abundances.
Of course, real intra-cellular processes are much more complicated, but if the mechanism is universal, the power-law should be valid regardless of how complicated the actual processes are. Hence it is relevant to study as simple as possible a model when trying to understand a universal law in real data.
Consider a cell consisting of a variety of chemicals. The internal state of the cell can be represented by a set of numbers $(n_1,n_2,\cdots ,n_k)$, where $n_i$ is the number of molecules of the chemical species $i$ with $i$ ranging from $i=1$ to $k$. For the internal chemical reaction dynamics, we chose a catalytic network among these $k$ chemical species, where each reaction from some chemical $i$ to some other chemical $j$ is assumed to be catalyzed by a third chemical $\ell$, i.e. $(i + \ell \rightarrow j + \ell)$ . The rate of increase of $n_j$ (and decrease of $n_i$) through this reaction is given by $\epsilon n_i n_{\ell}/N^2$, where $\epsilon$ is the coefficient for the chemical reaction. For simplicity all the reaction coefficients were chosen to be equal [@hetero], and the connection paths of this catalytic network were chosen randomly such that the probability of any two chemicals $i$ and $j$ to be connected is given by the connection rate $\rho$ [@stem].
Some resources (nutrients) are supplied from the environment by diffusion through the membrane (with a diffusion coefficient $D$), to ensure the growth of a cell. Through the calaytic reactions, these nutrients[@note1] are transformed into other chemicals. Some of these chemicals may penetrate [@hetero] the membrane and diffuse out while others will not. With the synthesis of the unpenetrable chemicals that do not diffuse out, the total number of chemicals $N= \sum_i n_i$ in a cell can increase, and accordingly the cell volume will increase. We study how this cell growth is sustained by dividing a cell into two when the volume is larger than some threshold. For simplicity the division is assumed to occur when the total number of molecules $N= \sum_i n_i$ in a cell exceeds a given threshold $N_{max}$. Chosen randomly, the mother cell’s molecules are evenly split among the two daughter cells.
In our numerical simulations, we randomly pick up a pair of molecules in a cell, and transform them according to the reaction network. In the same way, diffusion through the membrane is also computed by randomly choosing molecules inside the cell and nutrients in the environment. In the case with $ N \gg k$ (i.e. continuous limit), the reaction dynamics is represented by the following rate equation:
$$\begin{aligned}
dn_i/dt = \sum_{j,\ell}Con(j ,i ,\ell) \;\epsilon \;n_j \;n_{\ell} /N^2~~~~~~~~~ \nonumber \\
- \sum_{j',{\ell}'}Con(i ,j' ,{\ell}') \;\epsilon \;n_i \; n_{{\ell}'} /N^2
+ D \sigma_i (\overline{n_i}/V - n_i/N), \nonumber\end{aligned}$$
where $Con(i,j,\ell)$ is 1 if there is a reaction $i+\ell \rightarrow j + \ell$, and 0 otherwise, whereas $\sigma_i$ takes 1 if the chemical $i$ is penetrable, and 0 otherwise. The third term describes the transport of chemicals through the membrane, where $\overline{n_i}$ is a constant, representing the number of the [*i*]{}-th chemical species in the environment and $V$ denotes the volume of the environment in units of the initial cell size. The number $\overline{n_i}$ is nonzero only for the nutrient chemicals.
If the total number of molecules $N_{max}$ is larger than the number of chemical species $k$, the population ratios $\{ n_i/N \}$ are generally fixed, since the daughter cells inherit the chemical compositions of their mother cells. For $k>N_{max}$ [@large_K], the population ratios do not settle down and can change from generation to generation. In both cases, depending on the membrane diffusion coefficient $D$, the intra-cellular reaction dynamics can be classified into the three classes [@rho].
First, there is a critical value $D = D_c$ beyond which the cell cannot grow continuously. When $D > D_c$, the flow of nutrients from the environment is so fast that the internal reactions transforming them into chemicals sustaining ‘metabolism’ cannot keep up. In this case all the molecules in the cell will finally be substituted by the nutrient chemicals and the cell stops growing since the nutrients alone cannot catalyze any reactions to generate unpenetrable chemicals. Continuous cellular growth and successive divisions are possible only for $D \leq D_c$. When the diffusion coefficient $D$ is sufficiently small, the internal reactions progress faster than the flow of nutrients from the environment, and all the existing chemical species have small numbers of approximately the same level. A stable reaction network organization is obtained only at the intermediate diffusion coefficient below $D_c$, where some chemical species have much larger number of molecules than others.
The rank-ordered number distributions of chemical species in our model are plotted in Fig.2, where the ordinate indicates the number of molecules $n_i$ and abscissa shows the rank determined by $n_i$. As shown in the figure, the slope in the rank-ordered number distribution increases with an increase of the diffusion coefficient $D$. We found that at the critical point $D = D_c$, the distribution converges to a power-law with an exponent -1.
The power-law distribution at this critical point is maintained by a hierarchical organization of catalytic reactions, where the synthesis of higher ranking chemicals is catalyzed by lower ranking chemicals. For example, major chemical species (with e.g. $n_i>1000$) are directly synthesized from nutrients and catalyzed by chemicals that are slightly less abundant (e.g. $n_i \sim 200$). The latter chemicals are mostly synthesized from nutrients (or other major chemicals), and catalyzed by chemicals that are much less abundant. In turn these chemicals are catalyzed by chemicals that are even less abundant, and this hierarchy of catalytic reactions continues until it reaches the minor chemical species (with e.g. $n_i < 5$) [@hierarchy].
Based on this catalytic hierarchy, the observed exponent -1 can be explained using a mean field approximation. First, we replace the concentration $n_i/N$ of each chemical $i$, except the nutrient chemicals, by a single average concentration (mean field) $x$, while the concentrations of nutrient chemicals $S$ is given by the average concentration $S=1- k^{*}x$, where $k^{*}$ is the number of non-nutrient chemical species. From this mean field equation, we obtain $S=\frac{DS_0}{D+\epsilon
\rho}$ with $S_0=\sum_j \overline{n_j}/{V}$. With linear stability analysis, the solution with $S \neq 1$ is stable if $D< \frac{\epsilon \rho}{S_0-1} \equiv D_c$. Indeed, this critical value does not differ much from numerical observation.
Next, we study how the concentrations of non-nutrient chemicals differentiate. Suppose that chemicals $\{ i_0 \}$ are synthesized directly from nutrients through catalyzation by chemicals ${j}$. As the next step of the mean-field approximation we assume the concentrations of the chemicals $\{ i_0 \}$ are larger than the others. Now we represent the dynamics by two mean-field concentrations; the concentration of $\{ i_0 \}$ chemicals, $x_0$, and the concentration of the others, $x_1$. The solution with $x_0 \neq x_1$ satisfies $x_0 \approx x_1/\rho$ at the critical point $D_c$. Since the fraction of the $\{ i_0 \}$ chemicals among the non-nutrient chemicals is $\rho$, the relative abundance of the chemicals $\{ i_0
\}$ is inversely proportional to this fraction. Similarly, one can compute the relative abundances of the chemicals of the next layer synthesized from $i_0$. At $D \approx D_c$, this hierarchy of the catalytic network is continued. In general a given layer of the hierarchy is defined by the chemicals whose synthesis from the nutrients is catalyzed by the layer one step down in the hierarchy. The abundance of chemical species in a given layer is $1/\rho$ times larger than chemicals in the layer one step down. Then, in the same way as this hierarchical organization of chemicals, the increase of chemical abundances and the decrease of number of chemical species are given by factors of $1/\rho$ and $\rho$, respectively. This is the reason for the emergence of power-law with an exponent -1 in the rank-ordered distribution [@details].
In general, as the flow of nutrients from the environment increases, the hierarchical catalyzation network pops up from random reaction networks. This hierarchy continues until it covers all chemicals, at $D \rightarrow D_c-0$. Hence, the emergence of a power-law distribution of chemical abundances near the critical point is quite general, and does not rely on the details of our model, such as the network configuration or the kinetic rules of the reactions. Instead it is a universal property of a cell with an intra-cellular reaction network to grow, by taking in nutrients, at the critical state, as has been confirmed from simulations of a variety of models.
There are two reasons to assume that such a critical state of the reaction dynamics is adopted in existing cellular systems. First, as shown in Fig.3, the growth speed of a cell is maximal at $D = D_c$. This suggests that a cell whose reaction dynamics are in the critical state should be selected by natural selection. Second, at the critical point, the similarity of chemical compositions between the mother and daughter cell is maximal as shown in Fig.3. Indeed, for $k>N$, the chemical compositions differ significantly from generation to generation when $D \ll D_c$. When $D \approx D_c$, several semi-stable states with distinct chemical compositions appear. Daughter cells in the semi-stable states inherit chemical compositions that are nearly identical to their mother cells over many generations, until fluctuations in molecule numbers induce a transition to another semi-stable state. This means that the most faithful transfer of the information determining a cell’s intra-cellular state is at the critical state. (Inheritance of chemical compositions is also discussed in [@Lancet] in connection with the origin of reproducing cells). In this state, cells of specific chemical compositions are reproduced and can also ’evolve’ into other states. For these reasons, it is natural to conclude that evolution favors a critical state [@SOC] for the reaction dynamics.
Last, we investigated the relationship between the abundance of a chemical species and the number of reaction paths connected with it. By comparing the SAGE data and the protein-protein interaction data in yeast ([*S.cerevisiae*]{}) [@yeast_data1; @yeast_data2], obtained by systematic two-hybrid analysis, we found that there is a significant negative correlation between the abundance of any given mRNA and the number of protein-protein interaction links that the corresponding protein takes part in ($p<0.01$; determined by randomization test). In our model simulations, this negative correlation between the abundance of chemical species and the number of possible catalytic paths of the chemical is also found. In this sense, chemicals minor in abundance can play a relatively important role in the control of the behavior of a cell[@minority]. In the future it will be important to study this kind of interplay in the context of evolution since the evolution of reaction networks has only been discussed in the context of network topology [@metabolic; @lethal].
We would like to thank Tetsuya Yomo and Lars Martin Jakt for stimulating discussions and Frederick H. Willeboordse and Adam Ponzi for critical reading of the manuscript. Grant-in-Aids for Scientific Research from the Ministry of Education, Science and Culture of Japan (11CE2006).
[999]{}
H. Jeong, et al., [*Nature*]{} [**407**]{}, 651 (2000).
H. Jeong, S. P. Mason, A.-L. Barabási, [*Nature*]{} [**411**]{}, 41 (2001).
G. K. Zipf, [*Human Behavior and the Principle of Least Effort*]{} (Addison-Wesley, Cambridge, 1949).
A.E. Lash et al.,[*Genome Research*]{} [**10**]{}(7), 1051 (2000).
V.E. Velculescu et al., [*Cell*]{} [**88**]{}, 243 (1997): SAGE Data is available from http://www.sagenet.org/
S. J. Jones et al., [*Genome Res.*]{} [**11**]{}(8), 1346 (2001): SAGE Data is available from http://elegans.bcgsc.bc.ca/SAGE/
V.E. Velculescu, L. Zhang, B. Vogelstein, K. W. Kinzler, [*Science*]{} [**270**]{}, 484 (1995).
Even if the reaction coefficient and diffusion coefficient of penetrating chemicals are not identical but distributed, the results reported here are obtained.
K. Kaneko, T. Yomo, [*Bull. Math. Biol*]{} [**59**]{}, 139 (1997); C. Furusawa, K. Kaneko, [*Phys. Rev. Lett.*]{} [**84**]{}, 6130 (2000); [*Jour. Theor. Biol.*]{} [**209**]{}, 395 (2001); J. D. Farmer, S. A. Kauffman, and N. H. Packard, [*Physica D*]{} [**22D**]{}, 50 (1986). In contrast to these studies, auto-catalytic reactions are not relevant to the present study.
The nutrient chemicals have no catalytic activity in order to prevent the occurrence of catalytic reactions in the environment.
Note that, in the case $k>N_{max}$ the number of some chemical species $n_i$ is 0, while a sub-population of chemical species sustains the intra-cellular dynamics.
These three classes of intra-cellular dynamics also appear when changing the connection rate $\rho$. There is a critical value $\rho = \rho_c$, where in the case $\rho < \rho_c$ the cell stops growing. The power-law distribution of chemical abundances with an exponent -1 appears at $\rho = \rho_c$.
In the case depicted in Fig.2, a hierarchical organization of catalytic reactions with $5 \sim 6$ layers is observed at the critical point.
Within a given layer, a further hierarchy exists, which again leads to the Zipf rank distribution. For details, see C. Furusawa and K. Kaneko, to be published.
P. Bak and K. Sneppen, [*Phys. Rev. Lett.*]{} [**71**]{}, 4083 (1993); P. Bak, [*How Nature works*]{}(Springer, New York, 1996)
D. Segré, B. Danfna, D. Lancet, [*Proc. Natl. Acad. Sci. USA*]{} [**97**]{}(8), 4112 (2000).
P. Uewtz et al., [*Nature*]{} [**403**]{}, 623 (2000).
I. Xenarios et al., [*Nucleic Acids Res.*]{} [**28**]{}, 289 (2000).
K. Kaneko, T. Yomo, [*Jour. Theor. Biol.*]{} [**312**]{}, 563 (2002).
![ Rank-ordered frequency distributions of expressed genes. (a), Human liver; (b), kidney; (c), Human colorectal cancer (caco2); (d), Mouse embryonic stem cells; (e),[*C. elegans*]{}; (f) yeast ([*Saccharomyces cerevisiae*]{}). The exponent $\alpha$ of the power law is in the range from $-1 \sim -0.86$ for all the samples inspected, except for two plant data ([seedlings of *Arabidopsis thaliana*]{} and the trunk of [*Pinus taeda*]{}), whose exponents are approximately $-0.63$. ](./fig1.eps){width="11cm" height="15cm"}
{width="11cm" height="9cm"}
![ The growth speed of a cell and the similarity between the chemical compositions of the mother and daughter cells, plotted as a function of the diffusion coefficient $D$. The growth speed is measured as the inverse of the time for a cell to divide. The degree of similarity between two different states $m$ (mother) and $d$ (daughter) is measured as the scalar product of k-dimensional vectors $ H({\bf n}_m,{\bf n}_d) = ({\bf n}_m/|{\bf n}_m|)\cdot ({\bf n}_d/|{\bf n}_d|)$, where ${\bf n}=(n_1,n_2,...,n_k)$ represents the chemical composition of a cell and $|{\bf n}|$ is the norm of ${\bf n}$ [@Lancet]. Both the growth speed and the similarity are averaged over 500 cell divisions. Note that the case $H = 1$ indicates an identical chemical composition between the mother and daughter cells. ](./fig3.eps){width="11cm" height="9cm"}
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Johannes Hendriks
- 'Nicholas O’Dell'
- Adrian Wills
- Anton Tremsin
- Christopher Wensrich
- Takenao Shinohara
bibliography:
- 'References.bib'
title: 'Bayesian Non-parametric Bragg-edge Fitting for Neutron Transmission Strain Imaging'
---
Introduction and Motivation {#sec:introduction}
===========================
Energy resolved neutron transmission techniques can provide high-resolution images of strain within polycrystalline samples [@santisteban02; @tremsin12] by analysing features known as Bragg-edges. These strain images can be used to study the residual strain, and hence stress, within engineering components. Residual stresses are those that remain after the applied load is removed, for example, due to heat treatment, or plastic deformation [@withers2001bresidual]. The presence of residual stress can have a significant and unintended impact on a component’s mechanical performance — in particular its fatigue life.
Modern microchannel plate detectors (MCP) [@tremsin11], in combination with a spallation neutron source, can measure the transmission of neutrons as a function of wavelength over a 512 by 512 grid of $\SI{55}{\micro\meter}$ pixels. Strain is estimated from these measurements by analysing the relative shift in the location of Bragg-edges. Bragg-edges are a sudden increase in the transmission as a function of wavelength. This sudden increase occurs when the diffraction angle $2\theta$ reaches $180^\circ$ beyond which no further coherent scattering can occur. The location of these edges, $\lambda_{hkl}$, is given by Bragg’s law [@bragg1913reflection], $\lambda_{hkl}=2d_{hkl}\sin\theta$, and is related to strain through $$\label{eq:relative_strain}
\langle \epsilon \rangle = \frac{d_{hkl} - d_0}{d_0},$$ where $d_{hkl}$ is the lattice spacing and $d_0$ is the equivalent lattice spacing in a stress-free sample.
Several methods exist for estimating strain by analysing a single Bragg-edge in the measured transmission — presented by @santisteban2001time, @tremsin2016investigation, and @ramadhan2018neutron. These methods are reviewed in Section \[sec:existing\_approaches\]. Given a measured transmission these methods produce:
1. An estimate of the strain,
2. A quantification of their certainty in this estimate; this could be predicted confidence limits or standard deviation.
The measurements of the transmission are corrupted by noise. It is commonplace to average over several pixels either using a running average or by grouping pixels together into macro pixels. However, this averaging results in unwanted smoothing and should be kept to a minimum. Therefore, in order to confidently use these methods to study residual stress and make engineering decisions, it is important that they are accurate even when the data is noisy. Further, it is important that the quantification of certainty provided by the methods is reliable, allowing the user to make appropriate decisions about the validity of the estimated strains. Additionally, if the methods are capable of providing accurate estimates of strain even from noisier data, then less time is required for data acquisition.
The ability of these methods to produce accurate estimates of strain along with reliable confidence estimates from noisy data is particularly important for strain tomography and for dynamic measurement strain. Dynamic measurements of strain during in-situ material loading have measurement acquisition times limited by the process itself and these short acquisition times lead to noisy data [@mostafavi2017dynamic]. Strain tomography is analogous to computed medical tomography and reconstructs the full triaxial strain field within a sample from a set of strain images. Several methods for strain tomography have been developed and an overview can be found in @hendriksThesis2020. Methods for neutron strain tomography take as inputs a large number of strain images requiring many thousands of Bragg-edges to be analysed. Therefore, in order for strain tomography methods to be accurate, it is imperative that the methods used to produce these strain images are accurate and produce reliable estimates of confidence, which the strain tomography methods can take into account [@hendriks2019tomographic]. Additionally, as many of the strain tomography methods use either least-squares [@gregg2018tomographic] or an assumption of Gaussian noise [@jidling2018probabilistic], it would be beneficial if the errors given by the methods for Bragg-edge analysis are Gaussian.
The contribution of this paper is twofold. First, it provides a numerical simulation study of the accuracy and reliability of existing methods when applied to data with varying levels of noise. This analysis will show that the more recent approach presented by @ramadhan2018neutron on average produces smaller errors than the approaches by @santisteban2001time and @tremsin2016investigation. However, this is at the cost of very inaccurate quantification of certainty. Second, a novel non-parametric Bayesian method for estimating strain from measurements of a Bragg-edge is outlined. This method accurately quantifies the confidence in the produced strain estimates, whilst still giving competitively small errors. The existing methods and our proposed approach are also demonstrated on two sets of experimental data.
Problem Statement {#sec:problem_statement}
=================
This paper is concerned with methods for estimating the elastic normal strain ${\langle\epsilon\rangle}$ from noisy data collected of the transmission $\text{Tr}(\lambda) = \sfrac{I(\lambda)}{I_0(\lambda)}$ over a region containing a single Bragg-edge.
The data set contains $n$ measurements of the recorded transmission at specified wavelengths $\lambda_i$ which take the form $$\label{eq:noiseModel}
y_i = \text{Tr}(\lambda_i) + e_i$$ where $e$ is additive noise corrupting the measurement. The complete set of measurements will be denoted $y_{1:n} = \{y_1,\dots,y_n\}$.
Taking a Bayesian viewpoint, we wish to determine the distribution of ${\langle\epsilon\rangle}$ conditioned on this set of measurements $p({\langle\epsilon\rangle}| y_{1:n})$. An illustration of the problem we are solving is given in Figure \[fig:illustrative\_example\]. Since it may not be feasible to determine the full distribution, the methods should provide a reliable estimate of the mean and a suitable measure of certainty, such as confidence interval or standard deviation. These estimates of the mean and certainty should be both accurate and robust in the presence of varying levels of noise. That is, given noisy data the method should not produce estimates that are outliers and the provided measure of certainty should reflect the distribution of error between the estimated strain and the true strain.
\
In the following sections, we will assess the accuracy and robustness of existing methods, described in Section \[sec:existing\_approaches\], and compare these to our proposed approach presented in Section \[sec:our\_approach\]. In order to make this assessment, it is important to determine the distribution of noise corrupting the measurements $p(e|\vartheta)$ and an analysis of this noise is undertaken in Section \[sec:relative\_transmission\_intensity\_noise\_analysis\]. The results of this analysis are used in Section \[sec:simulation\_demonstration\_and\_error\_analysis\], which undertakes simulated random trials to numerically evaluate the accuracy and robustness of the methods. In Section \[sec:experimental\_data\_demonstration\] our proposed approach is applied to two sets of experimental data and compared to results from the existing approaches.
Transmission Intensity Noise Analysis {#sec:relative_transmission_intensity_noise_analysis}
=====================================
In this section, we undertake an analysis to determine a suitable model for the noise corrupting our measurements of transmission intensity, $p(e \vert \vartheta)$. For instance, the noise could have a Gaussian distribution, a student’s t-distribution, or something more complex and it is useful to know which. Understanding the way in which noise affects our measurements is necessary for understanding the limiting assumptions of both our method and existing methods and helpful for generating realistic simulation data.
This analysis is performed using two sets of neutron transmission data collected during strain tomography experiments; the first is data from an experiment described in @hendriks2019tomographic and the second set is from an experiment described Section \[sub:strain\_imaging\_example\]. These data sets were suitable for this analysis as each contains a large number of measurements — more than 75 measurements of the transmission spectrum were made of a sample for 2 hours for each experiment — enabling the distribution of noise to be accurately determined. Due to the short projection times the signal to noise ratio for individual 55 pixels is generally poor, and, therefore, averaging the data over groups of pixels to create macro pixels was performed to improve this.
In reality, noise enters the measurements through discrete errors in neutron counts at a given pixel on the detector for both the open-beam intensity, $I_0(\lambda)$, and projection intensity, $I(\lambda)$ and manifests itself when the quotient of these two quantities is calculated to produce the *measurement* of transmission. However, we determine that the noise can be modelled as an additive term on the transmission, according to $$y_i = \text{Tr}(\lambda_i) + e_i \tag{\ref{eq:noiseModel} revisited}$$
To investigate the properties of $p(e \vert \vartheta)$ we have chosen to obtain samples of $e_i$ by utilising an appropriate model as *ground truth*. Decaying exponential functions, $\exp[ -(a_0 + b_0 \lambda)]$, have been shown to model transmission near a Bragg-edge very well [@santisteban2001time], therefore we have utilised decaying exponential functions fitted to the data using least squares sufficiently far from the edge location where the model fits well and calculating the difference. The exponential model fits well past the last Bragg-edge due to the absence of coherent scattering and fits well before the Bragg-edge provided that the sample does not exhibit significant preferred crystal orientation (texture). These assumptions are valid for our data since we are only considering data immediately left and right of the last Bragg-edge as the data sets used are of samples with minimal grain texture.
By inspection of the data, the amount of noise varies with the amount of attenuation. Regions of the data with more attenuation have far less noise than regions of low attenuation, which can be seen in Figure \[fig:GP\_Tr\_fit\]. Therefore by *binning* the data according to ranges of transmission, we can see that the distribution of error conditioned on the true transmission $p(e_i \vert \operatorname{Tr}(\lambda_i))$, is very close to a zero-mean Gaussian, shown in Figure \[fig:histbinTR\].
\[fig:Hist1\]
\[fig:Hist1\]
\
\[fig:Hist1\]
\[fig:Hist1\]
Following on from the result in Figure \[fig:histbinTR\], we can use a Gaussian distribution to model the noise and establish a relationship between transmission ratio and the variance of our Gaussian model, this relationship is shown in Figure \[fig:sigVsTr\] for three macro-pixel sizes. From this we can establish that variance appears to be a linear function of the transmission ratio, we can therefore write $$p\left(y_i \vert \text{Tr}(\lambda_i)\right) \sim \mathcal{N}\left( \text{Tr}(\lambda_i),\sigma_i^2\right)$$ where, $$\sigma_i^2 = a + b\left[\text{Tr}(\lambda_i)\right].$$ Importantly, the form of the distribution is not affected by how many pixels are combined to create a measurement, only that the variance of the noise decreases as more pixels are included in the measurement.
\[fig:JPARC2019trend\]
\[fig:JPARC2020trend\]
Several pertinent conclusions can be drawn from these results which have implications for both our proposed method and existing ones. As the distributions are incredibly close to Gaussian, we can conclude that that the use of a Gaussian likelihood model is appropriate. This also validates the choice of a least-squares cost function proposed by some existing approaches described in Section \[sec:existing\_approaches\]. Additionally, we can conclude that due to the absence of a bias, the use of the exponential functions to locally model attenuation far from the edge is appropriate.
Existing Approaches {#sec:existing_approaches}
===================
This section provides a brief overview of existing approaches.
The predominant approaches used for determining strain images from neutron Bragg-edge data are presented by @santisteban2001time [@santisteban02b], and @tremsin2016investigation [@tremsin12]. Both of these approaches fit a parametric model to the recorded transmission to determine $\lambda_{hkl}$ from which strain is determined using Equation \[eq:relative\_strain\]. Recently, an alternative approach that uses cross-correlation to estimate $\Delta_{hkl} = \lambda_{hkl}-\lambda_0$ has been presented by @ramadhan2019characterization [@ramadhan2018neutron]. For readability, we will refer to these methods as the Santisteban method, the Tremsin method, and the cross-correlation method, respectively.
There are also methods available that analyse the full spectrum (containing multiple Bragg-edges) [@vogel2000tof; @sato2011rietveld]. Ideally, such a full spectrum analysis is used for analysis of micro-structure in the sample, including strain. Such analysis has only been demonstrated for a small number of measured spectra as it requires substantial computing power to fit a large number of parameters and some *a priori* information about the micro-structure. Hence, the analysis of the entire spectrum for strain imaging and strain tomography seems to be impractical at the present time where hundreds or thousands of transmission spectrum may need to be analysed. Additionally, single edge fitting is implemented not only because of acceptable computing time, but because different Bragg edges can have a different response to a specific stress induced on the sample.
The Santisteban method models the recorded transmission by $$\label{eq:santistban_model}
\begin{split}
\text{Tr}(\lambda) =& B(\lambda,\lambda_{hkl},\sigma_B,\tau)\exp(-a_0-b_0\lambda)\\ &+ (1-B(\lambda,\lambda_{hkl},\sigma_B,\tau))\exp(-a_0-b_0\lambda)\\&\hspace{20mm}\exp( - a_{hkl} - b_{hkl}\lambda),
\end{split}$$ where the exponential terms provide good models for the data either side of the edge and the edge shape $B$ is chosen as the integral of the Kropff model [@kropff1982bragg] given by $$\label{eq:kropff_model}
\begin{split}
B(\lambda,\lambda_{hkl},\sigma_B,\tau) =& \frac{1}{2}\Bigg[\text{erfc}\left(-\frac{\lambda-\lambda_{hkl}}{\sqrt{2}\sigma_B}\right)\\ &- \exp\left(-\frac{\lambda-\lambda_{hkl}}{\tau}+\frac{\sigma_B^2}{2\tau^2}\right)\\&\hspace{9mm}\text{erfc}\left(-\frac{\lambda-\lambda_{hkl}}{\sqrt{2}\sigma_B}+\frac{\sigma_B}{\tau}\right)\Bigg].
\end{split}$$ The model is fit to the data using a least-squares method, such as the Matlab function `lsqcurvefit`. To help avoid local minima that exist due to the non-linear nature of the model, @santisteban2001time suggest fitting the model in three stages. First, $a_0$ and $b_0$ are estimated by fitting $\exp(-a_0-b_0\lambda)$ to the far side of the edge where $B(\lambda,\lambda_{hkl},\sigma_B,\tau)=1$. Then, $a_{hkl}$ and $b_{hkl}$ are estimated by fitting $\exp(-a_0-b_0\lambda-a_{hkl}-b_{hkl}\lambda)$ to the far left of the edge where $B(\lambda,\lambda_{hkl},\sigma_B,\tau)=0$. Finally, the values of $\lambda_{hkl}$, $\sigma$, and $\tau$ are estimated by fitting the full model to data spanning the Bragg-edge. In each stage, the previously estimated parameters are held constant.
Additionally, @santisteban2001time mentions that an alternative, more complex, model for the Bragg-edge presented by @vogel2000tof can be used. This models the Bragg-edge by the integral of a Gaussian convoluted with two back-to-back exponentials with different rise and decay rates; $$\label{eq:vogel_model}
B(\lambda,p)\hspace{-1mm} =\hspace{-1mm}\frac{1}{2}\text{erfc}(w) - \frac{\beta\exp(u)\text{erfc}(y)\hspace{-1mm} -\hspace{-1mm} \alpha\exp(v)\text{erfc}(z)}{2(\alpha+\beta)} ,$$ where $p = \{\lambda_{hkl},\sigma_B,\alpha,\beta\}$ and $$\begin{split}
\delta =& \lambda_{hkl}-\lambda, \quad
w = \frac{\delta}{\sqrt{2}\sigma_B}, \\
u =& \frac{\alpha}{2}(\alpha\sigma_B^2 + 2\delta), \quad
v = \frac{\beta}{2}(\beta\sigma_B^2 - 2\delta), \\
y =& \frac{\alpha\sigma_B^2+\delta}{\sqrt{2}\sigma_B}, \quad
z = \frac{\beta\sigma_B^2-\delta}{\sqrt{2}\sigma_B}.
\end{split}$$ In the authors’ experience, this more complex model provides a better fit for some edge shapes. However, it becomes more difficult to avoid local minima while fitting.
The Tremsin method fits a five parameter model to the region around the Bragg-edge. This method does not use exponential models for the attenuation either side of the edge. Instead, it adds additional parameters to the Kropff Bragg-edge model given in Equation \[eq:kropff\_model\]: edge height, $h$, and base height $b$. This gives the model as $$B(\lambda,\lambda_{hkl},\sigma_B,\tau,b,h) = h*B(\lambda,\lambda_{hkl},\sigma_B,\tau,b,h) + b$$ This simple model is easy to implement and fit using non-linear least squares. However, this model has been found to less adequately describe the slope trends on the left-hand side of the edge for certain data sets [@ramadhan2018neutron]. Despite this, in the authors experience it can give good results provided that it is fit over a region of the transmission tightly cropped around the Bragg-edge.
Since both the Santisteban and Tremsin method fit a parametric model to the edge shape, the performance of both of these methods may suffer if the Bragg-edge to be fit is of a shape that cannot be adequately represented by these models. This could occur, for instance, in samples with significant preferred crystal orientation (texture), which distorts the edge shape. This was a significant motivating factor behind the development of the cross-correlation method by @ramadhan2018neutron and also motivates our use of a non-parametric edge shape in Section \[sec:our\_approach\].
The cross-correlation method proceeds as follows. First, smoothed numerical derivatives of the recorded transmission for both the stressed and stress-free sample are computed. Second, cross-correlation is performed on the two derivatives which determines the correlation between the edge shapes as a function of displacement $\Delta=\lambda-\lambda_0$. third, a pseudo Voigt function is fit to the peak-shaped correlation curve; $$\begin{split}
V(\Delta,\Delta_{hkl},A,\mu,w_l,w_g) =& y_0 + A\Bigg[ \frac{\mu 2\pi w_l}{4(x-\Delta)^2+w_l^2} \\
&+(1-\mu)\frac{\sqrt{4\log(2)}}{\sqrt{\pi}w_g} \\
&\exp\Bigg(-\frac{4\log(2)(\Delta-\Delta_{hkl})^2}{w_g^2}\Bigg)\Bigg].
\end{split}$$ Once $\Delta_{hkl}$ is estimated, an estimate of strain can be determined given knowledge of $\lambda_{0}$ according to ${\langle\epsilon\rangle}=\sfrac{\Delta_{hkl}}{\lambda_0}$.
The use of cross-correlation between the $\lambda_0$ and $\lambda_{hkl}$ data allows this method to be applied to a wide variety of Bragg-edge shapes. However, it is unclear if the $\lambda_0$ Bragg-edge and the $\lambda_{hkl}$ Bragg-edge having significantly different shapes would bias the strain estimates. Additionally, since differentiation magnifies noise present in data, a method for producing smoothed derivatives, such as the Savitzky-Golay method [@gorry1990general] is required. To achieve good results for different noise levels this requires manual tuning of the fit window and polynomial order parameters.
Each of these methods requires a parametric model to be fit. In @santisteban2001time and @tremsin2016investigation it is specified that non-linear least squares is used. Given that the analysis in Section \[sec:relative\_transmission\_intensity\_noise\_analysis\] determined the noise to have a Gaussian distribution we can conclude that the choice of a least-squares fitting method is appropriate. This is because least-squares is equivalent to finding a maximum likelihood estimate with a Gaussian noise model [@hendriksThesis2020]. As such, these methods can be interpreted as providing a maximum likelihood estimate of strain; $${\langle\epsilon\rangle}_{ML} = \operatorname*{argmax}_{{\langle\epsilon\rangle}} p(y_{1:n},{\langle\epsilon\rangle}),$$ which is equivalent to the maximum *a posteriori* estimate given a uniform prior for $p({\langle\epsilon\rangle})$.
Having obtained an estimate using least squares, it is also possible quantify the certainty of this estimate. A typical approach is to determine the Fisher information matrix, which can be estimated around the solution using Taylor’s theorem, from which confidence intervals or covariance can be calculated [@geyer2007fisher]. However, due to the non-linear nature of the models this estimate of the confidence interval may not always be accurate, particularly if the true conditional distribution $p({\langle\epsilon\rangle}| y_{1:n})$ is non-Gaussian or in extreme cases multi-modal. Additionally, as with most non-linear optimisation problems, there is no guarantee that a global minima is found. If instead a local minima is found, then both the estimate of strain and its confidence interval could be wrong.
Proposed Bayesian Approach {#sec:our_approach}
==========================
In this section, we propose a non-parametric Bayesian approach for estimating strain from time-of-flight neutron transmission Bragg-edge measurements. The approach models the Bragg-edge shape using a Gaussian process. This non-parametric model is beneficial as it is capable of fitting a wide range of edge shapes. However, it becomes more challenging to determine strain from these fits as there is no longer a model parameter that represents $\lambda_{hkl}$. This is overcome, by using the maximum of the derivative which can be exactly computed for a Gaussian process.
Our approach is presented as follows. A brief introduction to Gaussian process regression is presented in Section \[sub:an\_introduction\_to\_gaussian\_processes\]. Section \[sub:gaussian\_process\_bragg\_edge\_fitting\] describes a procedure to fit the transmission data using a Gaussian process model for the edge shape. A method is outlined in Section \[sub:strain\_measurements\] for determining the strain from this non-parametric estimate of the edge shape. Lastly, Gaussian processes have a number of hyperparameters; for example, the length-scale which controls the characteristic smoothness of the function, and Section \[sub:hyperparameter\_optimisation\] outlines how to optimise these parameters.
An Introduction to Gaussian Processes Regression {#sub:an_introduction_to_gaussian_processes}
------------------------------------------------
This section provides a brief introduction to Gaussian process regression, a more thorough discourse is given by @rasmussen2006gaussian.
Gaussian process regression is a non-parametric Bayesian method for fitting spatially correlated functions. A Gaussian process (GP) is a non-parametric Gaussian distribution of functions that is fully defined by a mean function $m:\mathbb{R}\to\mathbb{R}$ and covariance function $k:\mathbb{R}\times\mathbb{R}\to\mathbb{R}$. For clarity, we restrict the input space to be scalar.
The covariance function $k$ determines the correlation between any two function values $f(x)$ and $f(x')$ given by $k(x,x')$. As such, the choice of covariance function determines the characteristics of functions belonging to a Gaussian process, such as the degree of smoothness and differentiability of these functions. Two common covariance functions are the squared-exponential, $k_{SE}$, and Matérn, $k_M$, given by $$\begin{split}
k_{SE}(x,x') &= \sigma_f\exp\left(-\frac{1}{2l^2}(x-x')^2\right), \\
k_{M}(x,x') &= \sigma_f\frac{2^{1-\nu}}{\Gamma(\nu)}\left(\frac{\sqrt{2\nu}r}{l}\right)^\nu I_\alpha\left(\frac{\sqrt{2\nu}r}{l}\right),
\end{split}$$ where $r=|x-x'|$ is the distance between input locations, $l$ is the length-scale, $\sigma_f$ is the prior variance, $\nu$ is the degrees of freedom of the Matérn covariance function, and $I_\alpha$ is the modified Bessel function [@abramowitz1964handbook]. With both these covariance functions, the correlation between two function values is proportional to the distance between their input locations scaled by the length-scale. Hence, a larger length-scale will impose a greater degree of smoothness on the functions. The length-scale and prior variance are commonly referred to as hyperparameters and their selection and the selection of covariance function is discussed in Section \[sub:hyperparameter\_optimisation\].
For a function belonging to a GP distribution, any finite set of function values at locations have a multivariate Gaussian distribution. Hence, an unknown function value $f_*$ at $x_*$ and a measurement of the function $$\label{eq:GP_measurement}
y_i=A_if(x_i)+e,$$ where $A$ is a linear mapping and e is Gaussian noise with standard deviation $\sigma$, are *a priori* multivariate Gaussian[^1]; $$\setlength\arraycolsep{1pt} {\begin{bmatrix}y_i \\ f_*\end{bmatrix}} \sim \mathcal{N}\left({\begin{bmatrix}A_i m(x_i) \\ m(x_*)
\end{bmatrix}},{\begin{bmatrix}A_ik(x_i,x_i)A_i^{{\mathsf{T}}}\hspace{-1.1mm} +\hspace{-0.6mm} \sigma^2 & A_ik(x_i,x_*) \\
k(x_*,x_i)A_i^{{\mathsf{T}}} & k(x_*,x_*)
\end{bmatrix}}\right)$$ Without loss of generality, we can assume the *a priori* mean function to be zero [@rasmussen2006gaussian]. Given that we know the measurement value, we can update our knowledge of $f_*$ using standard Gaussian conditioning $$\label{eq:GP_posterior}
\begin{split}
f_*| y_i \sim \mathcal{N}\left(\mu_{f_*|y_i},\Sigma_{f_*|y_i}\right)
\end{split}$$ where $$\begin{split}
\mu_{f_*|y_i} &= k(x_*,x_i)A_i^{{\mathsf{T}}}(A_ik(x_i,x_i)A_i^{{\mathsf{T}}}+\sigma^2)^{-1}y_i \\
\Sigma_{f_*|y_i} &= k(x_*,x_*) - k(x*,x_i)A_i^{{\mathsf{T}}}\\&\hspace{10mm}(A_ik(x_i,x_i)A_i^{{\mathsf{T}}}+\sigma^2)^{-1}A_ik(x_i,x_*)
\end{split}$$
Note that this is trivially extended to conditioning a set of function values on a set of measurements. This means that Gaussian process regression can easily be used to condition an estimate of the unknown functions over a range of input locations based on a set of measurements.
Gaussian process Bragg-edge fitting {#sub:gaussian_process_bragg_edge_fitting}
-----------------------------------
Here, a method for fitting Bragg-edges using Gaussian process regression is presented. To do this, the transmission intensity $\text{Tr}(\lambda) = \sfrac{I(\lambda)}{I_0(\lambda)}$ is modelled similarly to the approach in @santisteban2001time, using two decaying exponentials and an edge shape function, see Equation .
Our approach differs in that we use a Gaussian process to model the edge shape $B(\lambda)$ rather than a parametric model. Rearranging Equation we can write $$\begin{split}
\text{Tr}(\lambda) - \gamma_1(\lambda) = (\gamma_2(\lambda) - \gamma_1(\lambda))B(\lambda),
\end{split}$$ where $\gamma_1(\lambda) = \exp(-(a_0+b_0\lambda+a_{hkl}+b_{hkl}\lambda))$ and $\gamma_2(\lambda) = \exp(-(a_0+b_0\lambda))$. If we now let $\bar{y}_i = y_i - \gamma_1(\lambda_i)$ and $A(\lambda) = (\gamma_2(\lambda_i)-\gamma_1(\lambda_i))$ then our measurements are of the same form as Equation \[eq:GP\_measurement\], and Equation \[eq:GP\_posterior\] can be applied to estimate the unknown function $B$.
The Bragg-edge can be fit using the following procedure
1. The values of $a_0$ and $b_0$ are determined by fitting $\exp(-(a_0 + b_0\lambda))$ to the recorded transmission to the far right of the edge.
2. Then, $a_{hkl}$ and $b_{hkl}$ are determined by fitting $\exp(-(a_0 + b_0\lambda - a_{hkl}-b_{hkl}\lambda))$ to the recorded transmission to the far left of the edge.
3. Lastly, Gaussian process regression is used to estimate $B(\lambda)$.
An example of using this approach to fit a recorded transmission belonging to the data set described in Section \[sub:strain\_imaging\_example\] is shown in Figure \[fig:GP\_Tr\_fit\].
Strain measurements {#sub:strain_measurements}
-------------------
Having fit the transmission data using a Gaussian process model for the edge shape we now wish to determine the distribution of strain values, $p(\langle\epsilon\rangle | Y)$. This distribution will provide us with both the expected value and the confidence in this value.
Using a non-parametric model for the edge shape has the advantage that it is not constrained to a particular shape. However, it presents the challenge that this edge shape model does not have a $d_{hkl}$ parameter and so Equation \[eq:relative\_strain\] cannot be directly applied to determine the strain. Therefore, our approach determines the strain from the relative shift in the maximum gradient of the edge shape; $$\langle \epsilon \rangle = \frac{\zeta - \zeta_0}{\zeta_0}$$ where $$\label{eq:lambda_equation}
\zeta = \operatorname*{argmax}_{\bar{\lambda}} \frac{\partial}{\partial\lambda}B(\lambda)\Big|_{\lambda=\bar{\lambda}}.$$ That is, $\zeta$ is the wavelength for which the gradient of the edge shape is maximised. Similarly, $\zeta_0$ is computed in the same manner from the transmission recorded of a stress-free sample.
An advantage of using Gaussian process regression to model the edge shape is that the required derivative can be computed directly, avoiding the use of numerical differentiation. This is possible as differentiation is a linear operator and Gaussian processes are closed under linear operators [@rasmussen2006gaussian; @wahlstrom2015modeling]. As a consequence, the gradient of the edge shape, $g=\frac{\partial B(\lambda)}{\partial\lambda}$, at $\lambda_*$ and a measurement, $\bar{y}_i=Tr(\lambda_i)-\gamma_1(\lambda_i)$, at $\lambda_i$ are *a priori* multivariate Gaussian; $$\setlength\arraycolsep{1pt} \begin{split}
{\begin{bmatrix}\bar{y}_i \\ g_*\end{bmatrix}} &\sim \mathcal{N} \left({\begin{bmatrix}0 \\ 0
\end{bmatrix}},{\begin{bmatrix}Ak(x_i,x_i)A^{{\mathsf{T}}} + \sigma^2 & k_*^{{\mathsf{T}}} \\ k_* & k_{**}\end{bmatrix}}
\right), \\
k_* &= \frac{\partial}{\partial \lambda}k(\lambda,\lambda_i)A(\lambda_i)^{{\mathsf{T}}}\Big|_{\lambda=\lambda_*}, \\
k_{**} &= \frac{\partial^2}{\partial \lambda\partial\lambda'}k(\lambda,\lambda')\Big|_{\substack{\lambda=\lambda_*\\\lambda'=\lambda_*}}, \\
\end{split}$$ where the prior mean was assumed zero. Therefore the gradient of the edge shape conditioned on a measurement, $p(g|\bar{y}_i)$, will be Gaussian with mean $\mu_{g|\bar{y}_i}$ and covariance $\Sigma_{g|\bar{y}_i}$ according to $$\begin{split}
\mu_{g|\bar{y}_i} &= k_*(Ak(x_i,x_i)A^{{\mathsf{T}}} + \sigma^2 )^{-1}\bar{y}_i, \\
\Sigma_{g|\bar{y}_i} &= k_{**} - k_*(Ak(x_i,x_i)A^{{\mathsf{T}}} + \sigma^2 )^{-1}k_*^{\mathsf{T}}.
\end{split}$$ This is easily extended to compute $p(g|\bar{y}_{1:n})$ from which $p(\zeta|\bar{y}_{1:n})$ can be determined by solving $$p(\zeta | \bar{y}_{1:n}) = \operatorname*{argmax}_{\bar{\lambda}}g(\lambda)p(g(\lambda);g|\bar{y}_{1:n}).$$ Since this is analytically intractable Monte Carlo sampling can be used to draw samples from $p(\zeta| \bar{y}_{1:n})$, $p(\zeta_0 | \bar{y}_{1:n})$ and $p(\langle\epsilon\rangle|y_{1:n})$ according to; $$\label{eq:eps_samples}
\begin{split}
\zeta^{(i)} &= \operatorname*{argmax}_{\bar\lambda} \mathcal{G}^{(i)}(\lambda), \quad \mathcal{G}^{(i)} \sim \mathcal{N}(\mu_{g|\bar{y}_{1:n}} | \Sigma_{g|\bar{y}_{1:n}}),
\\
\zeta_0^{(i)} &= \operatorname*{argmax}_{\bar\lambda} \mathcal{G}_0^{(i)}(\lambda), \quad \mathcal{G}_0^{(i)} \sim \mathcal{N}(\mu_{g_0|\bar{y}_{1:n}} | \Sigma_{g_0|\bar{y}_{1:n}}),
\\
\langle\epsilon\rangle^{(i)} &= \frac{\zeta^(i)-\zeta_0^{(i)}}{\zeta_0^{(i)}}
\end{split}$$ The samples of the edge shape gradient $\mathcal{G}^{(i)}$ and stress-free edge shape gradient $\mathcal{G}_0^{(i)}$ are easy to compute given that they have Gaussian distributions. Given $N$ samples $\langle\epsilon\rangle$ the sample mean and variance of $p(\langle\epsilon\rangle|\bar{y}_{1:n})$ is given by $$\label{eq:sample_mean_var}
\begin{split}
\mu_{\langle\epsilon\rangle|y_{1:n}} &= \frac{1}{N}\sum_i^N \langle\epsilon\rangle^{(i)}, \\
\Sigma_{\langle\epsilon\rangle|y_{1:n}} &= \frac{1}{N}\sum_i^N(\langle\epsilon\rangle^{(i)} - \mu_{\langle\epsilon\rangle|\bar{y}_{1:n}})^2.
\end{split}$$ Figure \[fig:eps\_hist\] shows a histogram of $\langle\epsilon\rangle^{(i)}$ samples corresponding to the Bragg-edge data shown in Figure \[fig:GP\_Tr\_fit\]. The Gaussian probability density function (pdf) with mean and variance given by Equation is also plotted. The Gaussian pdf very closely matches the histogram indicating that, for this data, the mean and variance are sufficient to describe the distribution $p({\langle\epsilon\rangle}|y_{1:n})$.
Before proceeding, a few pertinent remarks can be made. First, $\lambda$ should not be used as an estimate of the lattice spacing $\lambda_{hkl}$. This is apparent by considering the Kropff model [@kropff1982bragg] for the instrument resolution function with non-zero asymmetry. In this case, the value for $\lambda_{hkl}$ will be offset from the peak value of the derivative. This offset is not a problem when calculating strain provided that the asymmetry is relatively consistent between the stress-free Bragg-edge and the Bragg-edge used to calculate strain since the offset will cancel.
Second, although the Monte Carlo approximation of $p({\langle\epsilon\rangle}|\bar{y}_{1:n})$ shown in Figure \[fig:eps\_hist\] is very close to Gaussian this may not always be the case. For example, when the ratio of the edge height to noise is low, $p({\langle\epsilon\rangle}|\bar{y}_{1:n})$ is less likely to be Gaussian and in extreme cases may even be multimodal. If the result is bimodal or multimodal then the sample mean and variance may not be a good representation.
Covariance Function and Hyperparameter Optimisation {#sub:hyperparameter_optimisation}
---------------------------------------------------
Typically, a covariance function has several hyperparameters, $\theta$. For instance, both the squared-exponential and Matén covariance functions have hyperparameters $\theta = \{\sigma_f,l\}$, where $\sigma_f$ encodes the prior uncertainty and $l$ governs the characteristic smoothness of functions belonging to the GP. The covariance function and its hyperparameters can be selected by maximising the marginal log-likelihood [@rasmussen2006gaussian]. For each covariance function considered the hyperparameters are chosen according to $$\begin{split}
\theta_* &= \operatorname*{argmax}_\theta \ \log p(y_{1:n} | \theta) \\
&= \operatorname*{argmax}_\theta \ -\frac{1}{2}\left[\log\det(K_y) + Y^{{\mathsf{T}}}K_y^{-1}Y\right]
\end{split}$$ where $Y = {\begin{bmatrix}y_1 & \hdots & y_n\end{bmatrix}}^{{\mathsf{T}}}$, $K_y = K+I\sigma^2$ and $$\setlength\arraycolsep{1pt}
\begin{split}
K = {\begin{bmatrix}A(\lambda_1)k(\lambda_1,\lambda_1)A(\lambda_1)^{{\mathsf{T}}} & \hdots & A(\lambda_1)k(\lambda_1,\lambda_n)A(\lambda_n)^{{\mathsf{T}}} \\
\vdots & \ddots & \vdots \\
A(\lambda_n)k(\lambda_n,\lambda_1)A(\lambda_1)^{{\mathsf{T}}} & \hdots & A(\lambda_n)k(\lambda_n,\lambda_n)A(\lambda_n)^{{\mathsf{T}}}\end{bmatrix}}
\end{split}$$ After optimising the hyperparameters for each covariance function, the covariance function that yielded the highest marginal log-likelihood is chosen.
This optimisation can be performed using a gradient-based method, such as the BFGS algorithm presented by @wright1999numerical with the gradients given by @rasmussen2006gaussian as $$\frac{\partial \log p(y_{1:n}|\theta)}{\partial\theta_i} = \frac{1}{2}Y^{{\mathsf{T}}}K_y^{-1}\frac{\partial K}{\partial \theta_i}K_y^{-1}Y -\frac{1}{2}\text{tr}\left(K_y^{-1}\frac{\partial K}{\partial \theta_i}\right).$$
Numerical Simulation Study and Error Analysis {#sec:simulation_demonstration_and_error_analysis}
=============================================
This section provides a numerical analysis of the accuracy and robustness of the existing approaches and our proposed approach using simulated random trials. Measurements of the transmission were simulated using Equation with the Bragg-edge shape given by Equation as this provides a good approximation near the region of a Bragg-edge [@santisteban2001time]. These measurements were corrupted with noise according to the noise model determined in Section \[sec:relative\_transmission\_intensity\_noise\_analysis\]. To each simulated set of measurements, representing data of a single Bragg-edge, the existing approaches and our proposed approach are applied to estimate the strain and the error between the estimate and the true strain calculated. In the calculation of strain, the true $\lambda_0$ value was used for the Santisteban and Tremsin methods, and the true stress-free edge profile was used for the cross-correlation method and our approach.
Three different noise levels are investigated; standard deviations given by the analysis with 24 by 24 pixel binning (which will be indicated by $\sigma_{24\times24}$), these standard deviation reduced by a factor of 10 to represent high-quality Bragg-edge data with a high edge height to noise ratio, and these standard deviation increased by a factor of 10 to represent low-quality Bragg-edge data with a low edge height to noise ratio. For each noise level, random trials were conducted, broken into 100 groups. For each group the parameters of the edge shape were randomly chosen selected from the ranges $\sigma_B=[\num{4.7e-3},\num{1.4e-2}]$, and $\tau=[0,\num{1.3e-2}]$. The results are summarised in Table \[tab:simulation\_analysis\].
Ideally we would like the methods to produce the following:
- A low mean error as this means the results are unbiased.
- A low mean magnitude of the errors.
- A low standard deviation of the errors, as this would indicate that there is not a large spread of errors.
- For the mean of the methods predictions of standard deviation (std) to be close to the calculated standard deviation of the errors, as this indicates the methods can provide an accurate estimate of confidence in the results.
- Lastly, for the maximum error to be within a couple of standard deviations of zero, otherwise it indicates the presence of outliers.
The results presented in Table \[tab:simulation\_analysis\] indicate both the cross-correlation method and our proposed approach achieve good mean absolute error, mean error, and standard deviation of error. Significantly, applying these methods to the noisiest data ($10\sigma_{24\times24}$) sets yields results almost as accurate as the Santisteban and Tremsin methods applied to the data sets with the standard noisy level ($\sigma_{24\times24}$). For the cross-correlation procedure this required manually tuning the smoothing parameters for each noise level — this could be challenging to do well for a large set of strain images with varying Bragg-edge height to noise ratios.
When well-tuned, the cross-correlation method gives slightly lower mean absolute error, mean error, and standard deviation of error than our proposed approach. However, our proposed approach yields a significantly more accurate prediction of standard deviation, and a lower maximum error on all but the lowest noise simulations. This higher maximum error for the cross-correlation method is indicative that the method occasionally produces outlying estimates of strain.
It is our hypothesis that the outliers are due to the need to perform a numerical derivative. A significant challenge when applying this approach is determining the fit window and polynomial order that produce the lowest bias, mean absolute errors, and maximum error. In general, it was found that increasing the fit window decreased the standard deviation of the errors and reduced the likelihood of outliers, but after a certain point would increase the bias. Since the results here are from a simulated numerical study, the authors had the true values of strain to aid in tuning these parameters; it is unclear what the best way to tune these parameters is in general. A secondary challenge is choosing a good starting guess for the parameters when fitting the Voigt function, since a bad starting guess could result in the optimisation procedure finding a local rather than global minima.
Additionally, the standard deviations predicted by the cross-correlation method are, for the most part, far too low. This is likely due to the fact that the method has three stages (derivatives are taken, cross-correlation is performed, a parametric function is fit) whereas the prediction of standard deviation is based only on the final fitting stage. As a consequence, the user may be given the impression that the results are accurate even if this is not the case. For example, the average predicted standard deviation is only marginally higher for the standard noise simulations ($12.76\mu\text{strain}$) than for the low noise simulations ($11.01\mu\text{strain}$), whereas the mean error is twice as large and the maximum error is far greater. In contrast, our proposed approach gives predictions of the standard deviation that are relatively accurate for the standard noise and high noise simulations.
The error distributions calculated for the $\sigma_{24\times24}$ noise level are shown in Figure \[fig:error\_dists\] along with the mean predicted confidence interval given by each method. Confidence intervals have been calculated from the predicted standard deviation using a Gaussian assumption, whereby $\mu\pm2\sigma$ gives the $95\%$ region. These histograms show that the error distributions from the Santisteban, and Tremsin have heavy tails. The histogram does not show the outliers, maximum error magnitudes, of the cross-correlation method as they do not occur at high enough density to be clearly displayed.. Further, a significant number of the errors for the Santisteban and Tremsin methods are outside the average predicted $95\%$ confidence intervals, while for the cross-correlation method the prediction of this region is on average far too small — to the extent that zero-error is right on the edge of the $95\%$ confidence interval. In contrast, our proposed approach has a roughly Gaussian distribution for the errors with a reasonably accurate $95\%$ confidence interval.
\
Experimental Data Demonstration {#sec:experimental_data_demonstration}
===============================
The existing methods and our proposed approach are demonstrated on two sets of experimentally collected neutron transmission Bragg-edge data. The first set is of a AlSiCP metal matrix composite plate and was previously used to demonstrate the cross-correlation method. For this data set, there are corresponding diffraction strain measurements allowing for a quantitative comparison to be made. The second data set is part of a strain tomography data set and qualitatively compares the ability of the methods to produce strain images that can be used for strain tomography.
Data from Sample with Significant Texture {#sub:data_from_cross_correlation_paper}
-----------------------------------------
The proposed approach is demonstrated on experimental neutron transmission Bragg-edge data described in detail by @ramadhan2018neutron and compared to results from the existing approaches and diffraction strain measurements @ramadhan2019characterization. This data is taken of an AlSiCP metal matrix composite (MMC) plate composed of an AL 2124 matrix and pure silicon with dimensions of $\SI{15}{\milli\meter}$ in $z$ and $\SI{35}{\milli\meter}$ in $x$ and $y$. This data is an appropriate experimental test set as it was previously used to compare the existing approaches for strain estimation.
Data was collected using a microchannel plate (MCP) detector [@tremsin11] which has a 512 by 512 array of $\SI{55}{\micro\meter}$ by $\SI{55}{\micro\meter}$ pixels, a five hour measurement time, and the samples $x$ dimension aligned with the neutron beam. The data was then averaged over regions of $\SI{1}{\milli\meter}$ in the $z$-direction and $\SI{20}{\milli\meter}$ in the $y$-direction, for which there is minimal strain variation. After averaging, the Bragg-edge data has minimal noise.
The strain is estimated by applying the methods to the Aluminium $(111)$ Bragg-edge and the results are shown in Figure \[fig:cross\_corr\_data\_comp\]; the estimated strains and the one standard deviation error bars are shown. As in @ramadhan2018neutron, it was found that applying the approach by @tremsin2016investigation yielded very similar but slightly worse results than the Santisteban function and so they are not shown.
![Strain estimates from Bragg-edge neutron transmission data of a AlSiCP MMC plate. The estimated strain values and the one standard deviation confidence intervals are shown. Also shown for comparison are results from neutron diffraction [@ramadhan2019characterization].[]{data-label="fig:cross_corr_data_comp"}](ramadhan_data_comparison-eps-converted-to){width="1.0\linewidth"}
Calculating the mean absolute difference between the diffraction measurements and the estimated strains from transmission data gives; $94.42\pm{17.28}\,\mu\text{Strain}$ for the cross-correlation method, $105.63\pm30.77\,\mu\text{Strain}$ for our approach, and $145.73\pm19.2216\,\mu\text{Strain}$ for the approach by @santisteban2001time. The one standard deviation range on the mean difference has contributions from both the predicted standard deviation of estimated strain from neutron transmission and diffraction, as neutron diffraction are also a measurement and not ground truth. Given that both the cross-correlation and our proposed approach have a mean absolute difference one standard deviation smaller than the Santisteban function, we can tentatively conclude that they perform better on this data. With the cross correlation performing marginally better.
This is expected as the Al 2124 metal matrix is reported to have significant preferred crystal orientation (texture), which distorts the Al $(111)$ edge shape [@ramadhan2018neutron]. As a result, the parametric edge shapes used by @santisteban2001time and @tremsin2016investigation are not a good model for the edge shape. Hence, the use of the cross-correlation method and our proposed non-parametric approach are beneficial.
Strain Imaging Example {#sub:strain_imaging_example}
----------------------
The proposed approach is demonstrated on experimental neutron transmission Bragg-edge data collected on the RADEN energy-resolved-neutron-imaging instrument at J-PARC [@shinohara2015commissioning; @shinohara2016final], and compared with existing methods. The sample was manufactured primarily for triaxial strain tomography, it is EN26 steel (medium carbon, low alloy) consisting of a $17 \times 17 \times 17~\si{\milli\metre}$ steel cube with a precision ground hole of diameter 12 along the diagonal into which an EN26 steel plug was fit with an interference of $40\pm2$ interference fit, i.e., shrink fit — shown in Figure \[fig:samle\]. This sample was manufactured to provide a reference for three dimensional strain tomography. Strain images were measured using a micro-channel plate detector ($512~\times~512$ pixels, 55 per pixel) [@tremsin11] at a distance of 17.9 from the source. At the time of the experiment (February 2020) the source power was 500 .
The data-set was collected for tomographic reconstruction of a strain field, for which many projections are collected in a limited amount of beam-time and as little averaging over pixels as possible is desired to obtain a high resolution strain reconstruction. Both of these factors contribute to noise levels in the data, motivates Bragg-edge fitting methods which perform well in the presence of noise and provide an accurate measure of certainty which can be used by strain tomography methods to weight the importance of different strain measurements.
Figure \[fig:FEA\] shows two simulated strain images generated using finite element results. The first projection is aligned with the face of the cube and the second is aligned with the axis of the plug. Lacking secondary experimental data for this sample, e.g., diffraction, the FEA results will provide a point of reference for each method’s performance.
[L[0.02]{} L[0.2]{} L[0.01]{} L[0.18]{} L[0.01]{} L[0.19]{} L[0.01]{} L[0.2]{} L[0.01]{} ]{} & @santisteban2001time &&@tremsin2016investigation&&@ramadhan2018neutron&&Proposed method&\
Strain is estimated by applying the methods to the (110) Bragg-edge and results are shown in Figure \[fig:StrainImageGrid\]. It is apparent that both the cross-correlation method and our proposed approach perform substantially better in the presence of higher levels of noise than the other methods. Recalling how pixel grouping affects measurement noise from Section \[sec:relative\_transmission\_intensity\_noise\_analysis\], this can be observed in the degree of agreement between each strain image produced by the same method. Both our approach and the cross-correlation method experience good agreement between the 24-by-24, 18-by-18 and 13-by-13 pixel groupings. While both of the other methods experience significantly many more outliers, indicated by the black and dark red pixels which saturate the colour scale, with the introduction of higher noise levels.
We can also assess the quality of each image by observing its *smoothness*, giving a rough indication of how noisy each image is. Large unexpected jumps in strain between neighbouring pixels give an indication of a *noisy* image and smooth strain images give an indication that the image is *less noisy*. These unexpected jumps in strain are even present in the strain images produced by the Santisteban and Tremsin method using the relatively noise free 24-by-24 pixel grouping, where-as for both our proposed approach and the cross correlation method this *noise* is not present in the image until significantly more noise is present in the data, e.g., the 13-by-13 pixel grouping.
Table \[tab:AverageCI\] shows the mean predicted standard deviation for each of the Bragg-edges fitted to produce Figure \[fig:StrainImageGrid\]. As each ray has the same irradiated length in the shown projection, and therefore very similar noise levels, the average confidence interval should give a good indication of each methods confidence interval for the whole image. These results reflect the ones from Section \[sec:simulation\_demonstration\_and\_error\_analysis\], the cross-correlation method is consistently overconfident in its results, where-as our method’s predicted confidence interval reflects the addition of noise introduced by averaging across fewer pixels.
A second strain imaging projection shown in Figure \[fig:projection2\] depicts the results of both the cross-correlation method and our proposed approach. In this projection the beam passes through varying amounts of material, ranging from 29 in the centre of the image to zero at the perimeter, for this reason the signal to noise ratio is better in the centre and worse near the edges, this should be reflected in the standard deviation predicted by each method. Although the strain image produced by the cross correlation method appears to be slightly less noisy the associated predicted standard deviation is over confident, and has virtually no variation across the image. In contrast, the standard deviation predicted by our proposed approach increases from the centre of the image to the edge. The ability to accurately estimate the confidence interval for Bragg-edge fitting is imperative for some strain tomography methods [@hendriks2019tomographic].
[L[0.01]{} L[0.42]{} L[0.42]{} L[0.02]{} ]{} &@ramadhan2018neutron & Proposed method&\
Conclusion and Future Work {#sec:conclusion}
==========================
In this paper we have focussed on the problem of estimating strain from energy resolved neutron transmission Bragg-edge data. Methods for estimating strain from this data can be used to generate strain images allowing the residual stress and strain inside a sample to be studied. They can also be used as an intermediate step in strain tomography. For both these applications, the methods used should accurately estimate the strain and give a reliable measure of certainty even when the data is noisy.
A novel Bayesian non-parametric approach is proposed for estimating strain from this data. As part of this approach, the standard deviation of the strain can be accurately predicted using a Monte Carlo method. Additionally, the non-parametric nature of this approach is beneficial when the Bragg-edge shape is distorted, such as when the the sample has significant preferred crystal orientation.
A numerical simulation study and two experimental data sets were used to compare the existing approaches and our proposed approach. The results show that both our proposed approach and the cross-correlation method provide more accurate results than the Tremsin and Santisteban method when the data is noisy. The numerical simulation study indicates that the cross-correlation approach gives, on average, slightly lower error magnitudes than our proposed approach. However, this is at the cost of larger maximum errors and inaccurate predictions of standard deviation.
The main challenge in applying the cross correlation method is the need to manually tune the smoothing for the numerical derivatives. This numerical derivative is possibly the cause of any outliers produced by this method.
A path forward would be to combine parts of our proposed approach and the cross-correlation method. Our proposed approach provides a method for computing exact, rather than numerical, derivatives and through hyperparameter optimisation removes the need for manual tuning. Once the derivatives are calculated using our approach, the remainder of the cross-correlation method could be applied. Additionally, this would potentially allow accurate prediction of the standard deviation by using Monte Carlo sampling of the derivatives and pushing these samples through the rest of the procedure.
Acknowledgements {#sec:acknowledgements .unnumbered}
================
This work was supported by the Australian Research Council through a Discovery Project Grant No. DP170102324. Access to the RADEN instrument was made possible through the user-access program of J-PARC (J-PARC Long Term Proposal No. 2017L0101).
[^1]: The notation $\mathcal{N}(\mu,\Sigma)$ is used to denote a multivariate Gaussian with mean $\mu$ and covariance $\Sigma$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We associate to a finite digraph $D$ a lattice polytope $P_D$ whose vertices are the rows of the Laplacian matrix of $D$. This generalizes a construction introduced by Braun and the third author. As a consequence of the Matrix-Tree Theorem, we show that the normalized volume of $P_D$ equals the complexity of $D$, and $P_D$ contains the origin in its relative interior if and only if $D$ is strongly connected. Interesting connections with other families of simplices are established and then used to describe reflexivity, $h^*$-polynomial, and integer decomposition property of $P_D$ in these cases. We extend Braun and Meyer’s study of cycles by considering cycle digraphs. In this setting we characterize reflexivity and show there are only four non-trivial reflexive Laplacian simplices having the integer decomposition property.'
address:
- |
Department of Mathematics\
Stockholm University\
SE-$106$ $91$ Stockholm\
Sweden
- 'Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Suita, Osaka 565-0871, Japan'
- |
Department of Mathematics\
University of Kentucky\
Lexington, KY 40506–0027
- 'Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Suita, Osaka 565-0871, Japan'
author:
- Gabriele Balletti
- Takayuki Hibi
- Marie Meyer
- Akiyoshi Tsuchiya
title: Laplacian simplices associated to digraphs
---
Introduction
============
The use of linear algebra to study properties of graphs is an established technique in combinatorics and consequently has led to the development of the so called *spectral graph theory*. There is an extensive literature on algebraic aspects of spectral graph theory and on how combinatorial properties are encoded in characteristic polynomials, eigenvalues and eigenvectors of adjacency or Laplacian matrices of graphs (see [@Chu97] for a survey). It is tempting to take a step forward and associate a polytope $P_G$ to any graph $G$, by interpreting the rows of a matrix encoding the data of $G$ as the vertices of $P_G$. This is the case of the *edge polytope* [@OH98; @Vil98], the convex hull of the rows of the unsigned vertex-edge incidence matrix of a graph, whose geometric and combinatorial properties have been extensively studied in the last two decades (see e.g. [@MHNH11; @TZ14]), and used to build counterexamples [@OH99]. Recently the third author and Braun [@BM17] took a similar direction by associating to any graph $G$ the simplex $T_G$ (called the *Laplacian simplex*) whose vertices are the rows of the Laplacian matrix of $G$. They established basic properties of $T_G$ and study reflexivity, the integer decomposition property, and unimodality of the Ehrhart $h^*$-vectors of $T_G$ for some special classes of graph.
Our contribution is to provide a more general setting for the investigation on Laplacian simplices. We do this by allowing $G$ to be a directed multigraph. In this way the objects studied in [@BM17] can be seen as a special case of our setting (see Remark \[rmk:BM\]). We have reasons to believe this generalization is the correct direction to take. Indeed, in the undirected and simple case, the origin of a Laplacian simplex coincides with the barycenter of its vertices, which is an uncommon property for a lattice simplex. In our setting, it is clarified (Proposition \[prop:dependence\] and Corollary \[cor:0ininterior\]) that this happens only for special digraphs, i.e. they need to be strongly connected and have the same number of spanning trees converging to each vertex. Moreover, in the original settings the volume of a Laplacian simplex associated to a graph with $n$ vertices equals $n$ times the number of spanning trees of the graph $G$. Extending to digraphs, it turns out (Proposition \[prop:volume\]) that the factor $n$ appears because in the undirected case each vertex has the same number of spanning trees converging to it.
Main results and organization of the paper {#main-results-and-organization-of-the-paper .unnumbered}
------------------------------------------
In Section \[sec:2\] we set notation, basic definitions and prove the first important properties of Laplacian simplices in this new settings. In particular, we prove that the Laplacian simplex $P_D$ associated to a digraph $D$ with $n$ vertices, satisfies the following properties.
1. $P_D$ is a $(n-1)$-simplex if and only if $D$ has positive complexity, i.e. if $D$ has at least a spanning converging tree (Proposition \[prop:characterization\_rank\]).
Now assume that $D$ has positive complexity.
2. The numbers of spanning trees converging to each of the vertices of $D$ encode the barycentric coordinates of the origin with respect to the vertices of $P_D$ (Proposition \[prop:dependence\]).
3. $P_D$ contains the origin ${\textnormal{\textbf{0}}}$, which is in the strict relative interior of $P_D$ if and only if $D$ is strongly connected (Corollary \[cor:0ininterior\]).
4. The normalized volume of $P_D$ equals the total complexity of $D$, i.e. the total number of spanning converging trees (Proposition \[prop:volume\]).
In Section \[sec:3\], moreover we prove that, under some assumptions on $D$, $P_D$ is equivalent to the simplex associated to a weighted projective space (Proposition \[prop:wps\]). Under even more restrictive assumptions, $P_D$ is equivalent to one of the $\Delta_{(1,{\textnormal{\textbf{q}}})}$ simplices described in [@BDS16]. In such cases, we use this equivalence to characterize reflexivity (Corollary \[cor:refl\]), describe the Ehrhart $h^*$-polynomial, and the integer decomposition property of $P_D$ in terms of the spanning converging trees of $D$.
In Section \[sec:4\] we use these descriptions to extend the study of cycle graphs of Braun–Meyer [@BM17] to *cycle digraphs*, i.e. strongly connected simple digraphs whose underlying graph is a cycle (see Definition \[def:cycle\]). We prove the following results.
1. In Proposition \[prop:simple\] we prove that a Laplacian simplex associated to a simple digraph has at most one interior lattice point. In particular, Laplacian simplices associated to cycle digraphs have exactly one interior point.
2. In Theorem \[thm:terminal\] we prove that the Laplacian simplices $P_D$ associated to a cycle digraph is terminal Fano, i.e. it contains no lattice points other than the origin and its vertices, unless $D$ is one of the following six digraphs.
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1) edge \[bend left=20\] (2);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1) edge \[bend left=20\] (3);
3. In Theorem \[thm:ref\_dir\_cycles\] we characterize the reflexivity of $P_D$ in terms of combinatorial properties of the cycle digraph $D$.
4. In Theorem \[thm\_IDP\] we prove that a reflexive Laplacian simplex $P_D$ has the integer decomposition property if and only if $D$ is the oriented cycle $1 \to 2 \to \cdots \to n-1 \to n \to 1$ for any $n$, or one of the following four exceptional digraphs.
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1) edge \[bend left=20\] (2);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
5. In Theorem \[thm:non\_unim\] we construct a family of reflexive Laplacian simplices with symmetric but non-unimodal $h^*$-vector $(1,\ldots,1,2,\ldots,2,1,\ldots,1,2,\ldots,2,1,\ldots,1,2,\ldots,2,1,\ldots,1)$.
In the final Section \[sec:underlying\] we try to understand how the structure of the underlying simple and undirected graph of a digraph affects the reflexivity of its Laplacian polytope. In particular we show that there is a graph which is not the underlying graph of any simple directed graph whose Laplacian simplex is reflexive. A more general version of this problem (Question \[q3\]) remains open.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank Benjamin Braun for helpful comments. The project started during the “Workshop on Convex Polytopes for Graduate Students” at Osaka University. The first author is partially supported by the Vetenskapsr[å]{}det grant NT:2014-3991. The fourth author is partially supported by Grant-in-Aid for JSPS Fellows 16J01549.
The Laplacian Polytope Construction and properties of Laplacian simplices {#sec:2}
=========================================================================
The Laplacian of a digraph
--------------------------
Let $D$ be a finite directed graph (*digraph*) on the vertex set $V(D)=[n]$, where $[n] \coloneqq \{ 1 , \ldots , n \}$. Let $E(D)$ be the set of the *directed edges* of $D$. A directed edge $e = (i,j) \in E(D)$ points from a vertex $i$ (called the *tail* of $e$) to another vertex $j$ (called the *head* of $e$). Multiple directed edges between vertices are allowed, and we denote by $a_{i,j}$ the number of directed edges having tail on the vertex $i$ and head on the vertex $j$ of $D$, with $i,j\in [n]$ and $i \neq j$. Since loops do not affect the Laplacian matrix, we assume $D$ to be without loops, and thus $a_{i,i}=0$ for all $i \in [n]$. The number of edges with vertex $i$ as a tail is called the *outdegree* of $i$ and is denoted by ${\mathrm{outdeg}}(i)$, while the number of edges with vertex $i$ as a head is called the *indegree* of $i$ and is denoted by ${\mathrm{indeg}}(i)$. We call $D$ *strongly connected* if it contains a directed path from $i$ to $j$ for every pair of vertices $i,j \in [n]$ and *weakly connected* if there exists a path (not necessarily directed) between $i$ and $j$ for every pair of vertices $i, j \in [n]$. In this paper we assume $D$ has no isolated vertices, i.e. vertices with indegree and outdegree equal to zero. A *converging tree* is a weakly connected digraph having one vertex with outdegree zero, called the *root* of the tree, while all other vertices have outdegree one. We say that a subgraph $D'$ of $D$ is *spanning* if the vertex set of $D'$ is $[n]$. All the data of $D$ can be encoded in the *adjacency matrix* of $D$, that is, the $n \times n$ matrix $A(D) \coloneqq (a_{i,j})_{1 \leq i,j \leq n}$. We define the *outdegree matrix* of $D$ to be $O(D) \coloneqq (d_{i,j})_{1 \leq i,j \leq n}$, the $n \times n$ matrix with $d_{i,j}={\mathrm{outdeg}}(i)$, if $i=j$, and $d_{i,j}=0$ otherwise. We define the *Laplacian matrix* of $D$ to be the matrix $L(D) \coloneqq O(D) - A(D)$.
Observe the sum of the entries of each row of $L(D)$ is zero. Thus the rank of the Laplacian matrix is never maximal, i.e. $$\label{eq:rank}
\operatorname{rk}(L(D)) \leq n-1.$$ A combinatorial interpretation for having equality in is given by the Matrix-Tree Theorem, which is presented here in its generalized version for digraphs. The interpretation is given in terms of spanning converging trees of $D$. For any $i \in [n]$, we denote by $c_i$ the number of spanning trees which converge to $i$, i.e. the converging trees of $D$ with $n$ vertices having $i$ as the root. We denote by $c(D)$ the total number of converging spanning trees of $D$, i.e. $c(D) \coloneqq \sum_{i = 1}^n c_i$. The number $c(D)$ is usually referred to as the *complexity* of the digraph $D$.
\[thm:MTT\] Let $D$ be a digraph without loops on the vertex set $[n]$. Let $i,j \in [n]$, and $L(D)_{i,j}$ the matrix obtained from $L(D)$ by removing its $i$-th row and $j$-th column. Then the determinant of $L(D)_{i,j}$ equals, up to a change of sign, the number of spanning trees of $D$ converging to $i$, i.e. $$(-1)^{i+j} \det L(D)_{i,j} = \det L(D)_{i,i} = c_i.$$ In particular the complexity of $D$ is $$c(D)=\sum_{i=1}^n \det L(D)_{i,i}.$$
The Laplacian polytope associated to a digraph
----------------------------------------------
Let $D$ be a digraph on the vertex set $[n]$. To $D$ we associate a convex polytope in ${{\mathbb R}}^n$ having vertices in the integer lattice ${{\mathbb Z}}^n$. We call the *Laplacian polytope* associated to $D$ the polytope $P_D:=\operatorname{conv}({\{{\textnormal{\textbf{v}}}_1, \ldots, {\textnormal{\textbf{v}}}_n\}}) \subseteq {{\mathbb R}}^n$, where ${\textnormal{\textbf{v}}}_i$ is the $i$-th row of the Laplacian matrix of $D$. The polytope $P_D$ is not full dimensional; since the sum of the entries in each row of $L(D)$ vanishes, $P_D$ is contained in the hyperplane $H \coloneqq \{{\textnormal{\textbf{x}}}=(x_1,\ldots,x_n) : \sum_{i=1}^n x_i = 0 \}$ of ${{\mathbb R}}^n$. In particular, the dimension of the Laplacian polytope, $\dim(P_D)$, equals the rank of the Laplacian matrix $L(D)$. When the rank of $L(D)$ is equal to $n-1$, then $P_D$ is a simplex, called the *Laplacian simplex* associated to $D$.
\[rmk:BM\] The Laplacian simplex in this context is a generalization of the Laplacian simplex introduced by Braun-Meyer in [@BM17]. For a connected simple graph $G$, they define the simplex $T_G$ as the convex hull of the rows of the graph Laplacian matrix of $G$. The Laplacian $L(G)$ of $G$ can be interpreted as the Laplacian of a digraph $D_G$, and thus the resulting simplices are equal, that is, $T_G = P_{D_G}$.
Two lattice polytopes $P,Q\subset {{\mathbb R}}^n$ are said to be *unimodularly equivalent* if there exists an affine lattice automorphism $\varphi\in \textrm{GL}_n({{\mathbb Z}})\ltimes{{\mathbb Z}}^n$ of ${{\mathbb Z}}^n$ such that $\varphi_{{\mathbb R}}(P)=Q$. Sometimes it is convenient to work with full dimensional lattice polytopes, i.e. lattice polytopes embedded in a space of their same dimension. Given a Laplacian simplex $P_D$, one can easily get a full dimensional unimodularly equivalent copy of $P_D$ by considering the lattice polytope defined as the convex hull of the rows of $L(D)$ with one column deleted. An example of this can be observed in Example \[ex:first\_example\].
\[ex:first\_example\] Let $D$ be the following digraph with its Laplacian matrix $L(D)$.
[c]{}
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (2);
$L(D)=
\begin{pmatrix}
1 & -1 & 0 \\
0 & 1 & -1\\
-1& -1 & 2
\end{pmatrix}$
Note that $L(D)$ has rank two, which means $P_D$ is a two dimensional simplex in ${{\mathbb R}}^3$. Full dimensional unimodularly equivalent copies of $P_D$ can be obtained by deleting any of the columns of $L(D)$ and considering the convex hull of the rows as in Figure \[fig:ex1\].
(-1,0)–(1,-1)–(-1,2)–cycle; (-1,0)–(1,-1)–(-1,2)–cycle; (-1,0) circle (0.25em) node\[below\]; (-1,1) circle (0.25em) node\[below\]; (-1,2) circle (0.25em) node\[below\]; (0,0) circle (0.25em) node\[below\]; (1,-1) circle (0.25em) node\[below\];
(1,0)–(0,-1)–(-1,2)–cycle; (1,0)–(0,-1)–(-1,2)–cycle; (-1,2) circle (0.25em) node\[below\]; (0,-1) circle (0.25em) node\[below\]; (1,0) circle (0.25em) node\[below\]; (0,0) circle (0.25em) node\[below\]; (0,1) circle (0.25em) node\[below\];
(1,-1)–(0,1)–(-1,-1)–cycle; (1,-1)–(0,1)–(-1,-1)–cycle; (-1,-1) circle (0.25em) node\[below\]; (0,-1) circle (0.25em) node\[below\]; (1,-1) circle (0.25em) node\[below\]; (0,0) circle (0.25em) node\[below\]; (0,1) circle (0.25em) node\[below\];
Properties of Laplacian simplices
---------------------------------
From the Matrix-Tree Theorem (Theorem \[thm:MTT\]) the following characterization can be immediately obtained.
\[prop:characterization\_rank\] Let $D$ be a digraph on $n$ vertices. The following are equivalent:
1. $D$ has positive complexity $c(D)$;
2. $\operatorname{rk}( L(D) ) = n-1$;
3. $P_D$ is an $(n-1)$-simplex.
Following the work of Braun-Meyer [@BM17], we focus our attention to the case in which a digraph $D$ on $n$ vertices defines an $(n-1)$-simplex. Proposition \[prop:characterization\_rank\] asserts we will always assume the digraph $D$ has positive complexity. As another consequence of Theorem \[thm:MTT\], we deduce the following result.
\[prop:dependence\] Let $D$ a digraph with positive complexity. Then the numbers of spanning converging trees $c_1,\ldots,c_n$ of $D$ encode the unique linear dependence among the vertices ${\textnormal{\textbf{v}}}_1, \ldots, {\textnormal{\textbf{v}}}_n$ of $P_D$, i.e. $$\sum_{i=1}^n c_i {\textnormal{\textbf{v}}}_i = {\textnormal{\textbf{0}}}.$$
Since the determinant of $L(D)$ is zero, the Laplace expansion along the $j$-th column of $L(D)$ yields $\sum_{i=1}^n (-1)^{i+j}\det{L(D)_{i,j}}v_{i,j} = 0$, where $L(D)_{i,j}$ is the matrix of $L(D)$ obtained by removing the $i$-th row and $j$-th column of $L(D)$, and $v_{i,j}$ is the $j$-th entry of ${\textnormal{\textbf{v}}}_i$. By Theorem \[thm:MTT\], $\det{L(D)_{i,i}} = c_i$.
\[cor:0ininterior\] Let $D$ be a digraph on $n$ vertices having positive complexity. Then ${\textnormal{\textbf{0}}}\in P_D$. Moreover ${\textnormal{\textbf{0}}}$ is an interior point of $P_D$ if and only if $D$ is strongly connected.
The first statement is a direct consequence of Proposition \[prop:dependence\]. For the second it is enough to note that $D$ is strongly connected if and only if each vertex has at least one spanning converging tree.
In this setting we prove a formula for the normalized volume of $P_D$. If a polytope $P$ is $n$-dimensional, its *normalized volume* $\operatorname{Vol}(P)$ is defined to be $n!$ times the Euclidean volume of $P$.
\[prop:volume\] Let $D$ be a digraph with positive complexity. Then its normalized volume equals the complexity of $D$, i.e. $$\operatorname{Vol}(P_D) = c(D),$$
In this case $P_D$ is a $(n-1)$-simplex by Proposition \[prop:characterization\_rank\]. For $i=1,\ldots,n$, we denote by $F_i$ the facet of $P_D$ not containing the vertex $i$. By Proposition \[prop:dependence\], ${\textnormal{\textbf{0}}}\in P_D$, so we can triangulate $P_D$ with the $S_i \coloneqq \operatorname{conv}(\{{\textnormal{\textbf{0}}}\} \cup F_i)$. In particular $$\operatorname{Vol}(P_D) = \sum_{i=1}^n \operatorname{Vol}(S_i),$$ where we consider $\operatorname{Vol}(S_i)=0$ if $S_i$ is not of dimension $n-1$. Let $S'_i$ be the unimodular copy of $S_i$ obtained as the convex hull of the rows of $L(D)_{i,i}$, the matrix obtained from $L(D)$ by removing the $i$-th row and $i$-th column. $$\operatorname{Vol}(P_D) = \sum_{i=1}^n \operatorname{Vol}(S_i) = \sum_{i=1}^n \operatorname{Vol}(S'_i) = \sum_{i=1}^n \det L(D)_{i,i} = \sum_{i=1}^n c_i,$$ where the last equality follows from Theorem \[thm:MTT\].
Connections with other families of simplices {#sec:3}
============================================
Laplacian simplices associated to strongly connected digraphs have interesting intersections with the study of weighted projective space arising from algebraic geometry and with the study of other families of simplices. We use these connections to describe properties of Laplacian simplices with particular attention to reflexivity, the integer decomposition property, and $h^*$-vectors of lattice polytopes. For the convenience of the reader, the next two subsections are a quick introduction to these topics.
Weighted projective spaces
--------------------------
Given positive integers $\lambda_1,\ldots,\lambda_n$ which are coprime, i.e. such that $\gcd \{ \lambda_1 , \ldots , \lambda_n \} = 1$, we define the polynomial algebra $S(\lambda_1 , \ldots , \lambda_n) \coloneqq {{\mathbb C}}[x_1,\ldots,x_n]$ graded by $\deg x_i \coloneqq \lambda_i$. A *weighted projective space* with weights $\lambda_1 , \ldots , \lambda_n$ is the projective variety ${{\mathbb P}}(\lambda_1 , \ldots , \lambda_n) \coloneqq \textrm{Proj}(S(\lambda_1 , \ldots , \lambda_n))$. Since ${{\mathbb P}}(\lambda_1 , \ldots , \lambda_n)$ is a toric variety, it corresponds to a fan $\Delta$ that can be characterized as follows. Let ${\textnormal{\textbf{v}}}_1,\ldots,{\textnormal{\textbf{v}}}_n$ be primitive lattice points which generate the lattice and satisfy $\sum_{i=1}^n \lambda_i {\textnormal{\textbf{v}}}_i = {\textnormal{\textbf{0}}}$, where $\gcd \{ \lambda_1 , \ldots , \lambda_n \} = 1$. Then, up to isomorphism, the fan $\Delta$ is the fan whose rays are generated by the ${\textnormal{\textbf{v}}}_i$. Note that the fan $\Delta$ identifies uniquely the simplex $S_\Delta \coloneqq \operatorname{conv}(\{ {\textnormal{\textbf{v}}}_1, \ldots, {\textnormal{\textbf{v}}}_n\})$. With an abuse of terminology, we say that a simplex is the weighted projective space ${{\mathbb P}}(\lambda_1 , \ldots , \lambda_n)$ if it is unimodularly equivalent to the simplex $S_\Delta$. For details we refer the reader to [@Dol82; @Ian00].
Ehrhart Theory, reflexivity and integer decomposition properties of lattice polytopes {#ssec:ehrhart_theory}
-------------------------------------------------------------------------------------
For a proper introduction to Ehrhart Theory and related topics, we refer to the textbook [@BR15]. A classical result by Ehrhart states that the number of lattice points in integer dilations of a $d$-dimensional lattice polytope $P\subset {{\mathbb R}}^d$ behaves polynomially. In terms of generating series this translates into the equality $$1 + \sum_{n\ge1} | nP \cap {{\mathbb Z}}^d | z^n = \frac{h_d^*z^d + \cdots + h_1^*z + 1}{(1-z)^{d+1}},$$ where $h^*(z) = h_d^*z^d + \cdots + h_1^*z + 1$ is a polynomial with non-negative integer coefficients, called the *$h^*$-polynomial* of $P$. This is an important invariant as it preserves much information about $P$. For example, the following relations are well known: $$h^*_1=|P \cap {{\mathbb Z}}^d|-d-1, \qquad h^*_d=|P^\circ \cap {{\mathbb Z}}^d|, \qquad 1 + \sum_{i=1}^d h^*_i=\operatorname{Vol}(P),$$ where $P^\circ$ denotes the relative interior of $P$. The $h^*$-polynomial of $P$ is often identified with the vector of its coefficient $(1,h^*_1,\ldots,h^*_d)$, called the *$h^*$-vector* of $P$. We call a vector $(x_0,x_1,\ldots,x_d)$ *unimodal* if there exists a $1 \leq j \leq d$ such that $x_i \leq x_{i+1}$ for all $0 \leq i < j$ and $x_k \geq x_{k+1}$ for all $j \leq k < d$. An important open problem in Ehrhart Theory is to understand under which conditions $h^*$-vectors are unimodal (see [@Bra16] for a survey).
The *dual polytope* $P^*$ of a full dimensional rational polytope $P$ which contains the origin is the polytope $P^* \coloneqq \{ {\textnormal{\textbf{x}}}\in {{\mathbb R}}^d \mid {\textnormal{\textbf{x}}}\cdot {\textnormal{\textbf{y}}}\le 1 \mbox{ for all } {\textnormal{\textbf{y}}}\in P \}$. If $P$ is a lattice polytope, we call it *reflexive* if its dual $P^*$ is again a lattice polytope. We extend the definition of reflexive to all the lattice polytopes which are unimodular equivalent to $P$. Reflexive polytopes were first introduced in [@Bat94]. A well known result of the second author [@Hib92] characterizes reflexive polytopes as lattice polytopes having a *symmetric* $h^*$-vector, i.e. such that $h^*_i = h^*_{d-i}$ for $0 \leq i \leq \lfloor \frac{d}{2} \rfloor$.
We say that a lattice polytope $P$ has the *integer decomposition property*, if, for every positive integer $n$ and for all ${\textnormal{\textbf{x}}}\in nP \cap {{\mathbb Z}}^d$ there exist ${\textnormal{\textbf{x}}}_1,\ldots,{\textnormal{\textbf{x}}}_n \in P \cap {{\mathbb Z}}^d$ such that ${\textnormal{\textbf{x}}}= {\textnormal{\textbf{x}}}_1 + \ldots + {\textnormal{\textbf{x}}}_n$. A polytope having the integer decomposition property is often called *IDP*.
Many efforts have been made to find sufficient conditions for unimodality. It has been conjectured by Stanley [@MR1110850] that a standard graded Cohen-Macaulay integral domain has unimodal $h$-vector. Although this has proven to be wrong in general [@MR1310575], in the context of lattice polytope this can be translated in the following question which remains open: does an IDP polytope always have unimodal $h^*$-vector? A weaker statement of this question has also been suggested by Ohsugi and the second author [@OH06], who conjectured that being reflexive and IDP is a sufficient condition for a lattice polytope to have unimodular $h^*$-vector.
Connections with weighted projective spaces and $\Delta_{(1,{\textnormal{\textbf{q}}})}$-simplices
--------------------------------------------------------------------------------------------------
We now relate Laplacian polytopes to weighted projective spaces.
\[prop:wps\] Let $D$ be a strongly connected digraph such that $\gcd\{c_1 , \ldots, c_n \}=1$. Then $P_D$ is equivalent to the weighted projective space ${{\mathbb P}}(c_1 , \ldots, c_n)$.
By Corollary \[cor:0ininterior\], $\sum_{i=1}^n \lambda_i {\textnormal{\textbf{v}}}_i = {\textnormal{\textbf{0}}}$; so we just need to prove that the vertices of $P_D$ span the lattice. Let $L$ be the lattice spanned by all the vertices, and $L_i$ the lattice spanned by all the vertices ${\textnormal{\textbf{v}}}_j$ such that $j \neq i$. Then we have the following inclusions of subgroups of ${{\mathbb Z}}^d$: $L_i \subset L \subset {{\mathbb Z}}^n$. In particular for all $i$, $|{{\mathbb Z}}^n : L| |L : L_i|=|{{\mathbb Z}}^n : L_i|=c_i$, which implies that $ L = {{\mathbb Z}}^n$.
In [@Con02; @Kas13] characterizations for properties of weighted projective spaces are given in terms of their weights and are used to perform classifications. We use these results to translate properties of $D$ to properties of $P_D$. Motivated by the open questions mentioned in the previous subsection, we focus on reflexivity, integer decomposition property, and description of the $h^*$-polynomial.
We use the following result of Conrads, presented below in a slightly weaker form.
\[prop:conrads\] Let $S = \operatorname{conv}({\textnormal{\textbf{v}}}_1,\ldots,{\textnormal{\textbf{v}}}_n)$ be an $(n-1)$-simplex such that $\sum_{i=1}^n q_i {\textnormal{\textbf{v}}}_i = {\textnormal{\textbf{0}}}$ for some positive integers $q_1,\ldots,q_n$ satisfying $\gcd(q_1,\ldots,q_n)=1$. Then $S$ is reflexive if and only if $$\label{eq:conrads}
q_i \quad \text{divides the total weight} \quad |Q|=\sum\limits_{j=1}^{n} q_j \quad for \quad i=1,\ldots,n.$$
From this we can derive the following corollary.
\[cor:refl\] Let $D$ be a strongly connected digraph such that $\gcd\{c_1 , \ldots, c_n \}=1$. Then $P_D$ is reflexive if and only if $c_i | c(D)$ for all $i$.
Proposition \[prop:conrads\] is also used by Braun–Davis–Solus [@BDS16] to define an interesting class of reflexive simplices. In particular they are interested in studying the integer decomposition property and unimodality of the $h^*$-vectors of such simplices constructed the following way. Let ${\textnormal{\textbf{q}}}=(q_1,\ldots,q_n)$ be an nondecreasing sequence of positive integers satisfying the condition $q_j | (1+\sum_{i \neq j}q_i )$ for all $j=1,\ldots,n$. For such a vector ${\textnormal{\textbf{q}}}$, the simplex $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is defined as $$\Delta_{(1,{\textnormal{\textbf{q}}})}:=\operatorname{conv}\left\{ {\textnormal{\textbf{e}}}_1,{\textnormal{\textbf{e}}}_2,\ldots,{\textnormal{\textbf{e}}}_n,-\sum_{i=1}^n q_i {\textnormal{\textbf{e}}}_i \right\},$$ where $e_i\in {{\mathbb R}}^n$ is the $i$-th standard basis vector. By Proposition \[prop:conrads\], $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is a reflexive simplex. Note that $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is equivalent to the weighted projective space with weights $(1,q_1,\ldots,q_n)$.
The next proposition shows that the simplices $\Delta_{(1,{\textnormal{\textbf{q}}})}$ are a subfamily of Laplacian simplices arising from special star-shaped strongly connected digraphs.
\[prop:Delta\_1q\] Let ${\textnormal{\textbf{q}}}=(q_1,q_2,\ldots,q_n)$ be any nondecreasing sequence of positive integers. Then there is a strongly connected digraph $D$ such that $P_D \cong {{\mathbb P}}(1,q_1,\ldots,q_n)$. In particular, if ${\textnormal{\textbf{q}}}$ satisfies the condition $q_j | (1+\sum_{i \neq j}q_i )$ for all $j=1,\ldots,n$, then $P_D \cong \Delta_{(1,{\textnormal{\textbf{q}}})}$.
As in Figure \[fig:star\_shaped\], we define $D$ as the star-shaped digraph on the vertices $1,\ldots,n+1$ such that
1. for $i=1,\ldots,n$ there are $q_i$ many edges directed from $1$ to $i+1$;
2. for $i=1,\ldots,n$ there is one edge directed from $i+1$ to $1$.
It is easy to verify that $c_1=1$ and, for $i \geq 2$, $c_i = q_i$. Proposition \[prop:wps\] concludes the proof.
at (0:0) (1) [$1$]{}; at (72:2) (2) [$2$]{}; at (0:2) (3) [$3$]{}; at (288:2) (4) [$\cdots$]{}; at (216:2) (5) [$n$]{}; at (144:2) (6) [$n+1$]{};
\(1) edge \[bend left=20\] node\[xshift=.3cm,yshift=-.3cm\] [$q_1$]{} (2) edge \[bend left=20\] node\[xshift=.0cm,yshift=-.3cm\] [$q_2$]{} (3) edge \[bend left=20\] node\[xshift=-.2cm,yshift=-.2cm\] [$\cdots$]{} (4) edge \[bend left=20\] node\[xshift=-.2cm,yshift=.3cm\] [$q_{n-1}$]{} (5) edge \[bend left=20\] node\[xshift=.4cm,yshift=.1cm\] [$q_{n}$]{} (6) (2) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1) (4) edge \[bend left=20\] (1) (5) edge \[bend left=20\] (1) (6) edge \[bend left=20\] (1);
In [@BDS16] an explicit formula for the $h^*$-polynomial of the simplices $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is given. We remark that such formula can be also extracted from [@Kas13], where it is proved in the more general setting of weighted projective spaces; however, the formulation given in [@BDS16] perfectly fits our needs.
\[thm:hstarq\] The $h^*$-polynomial of $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is $$h^*(\Delta_{(1,{\textnormal{\textbf{q}}})};z) = \sum_{b=0}^{q_1+\cdots +q_n}z^{w(b)}$$ where $$w(b):=b-\sum_{i=1}^n\left\lfloor\frac{q_ib}{1+q_1+\cdots +q_n} \right\rfloor \, .$$
Finally, in [@BDS16], necessary conditions for a $\Delta_{(1,{\textnormal{\textbf{q}}})}$ simplex to be IDP are given.
\[lem:BDS\_IDP\] If $\Delta_{(1,{\textnormal{\textbf{q}}})}$ is IDP, then for all $j=1,2,\ldots,n$ $$\frac{1}{q_j}+\sum_{i\neq j}\left\{\frac{q_i}{q_j}\right\}=1.$$
Laplacian simplices associated to cycle digraphs {#sec:4}
================================================
We now want to extend the study of Braun–Meyer on simplices associated to cycle graphs. They show that the Laplacian simplex associated to a cycle is reflexive if and only if the cycle has odd length $n$; in that case it has a unimodal $h^*$-vector and fails to be IDP for $n \ge 5$ [@BM17 Section 5]. We generalize their study by extending the notion of cycle graphs to cycle digraphs. A natural way to do it, would be to consider digraphs whose underlying simple graphs are cycle graphs. Here by underlying simple graph $G_D$ of a digraph $D$ we mean the simple undirected graph on the vertex set $V(G_D) \coloneqq V(D)$ such that the edge $\{i,j\}$ is in $E(G_D)$ if and only if there is at least one directed edge between $i$ and $j$ in $D$ (in either of the two directions). But since we are interested in reflexivity, we know, by Corollary \[cor:0ininterior\], that $D$ has to be strongly connected, therefore $D$ needs to contain a cycle entirely oriented in one of the two possible directions. This generalization of cycle graphs will be made clear later (Definition \[def:cycle\]). Moreover, in order to ensure the presence of no more than one interior point, we will assume that $D$ is a simple digraph. In the next subsection we show that this is always true for any simple graph.
Laplacian simplices associated to simple digraphs
-------------------------------------------------
In this section we focus on *simple* digraphs, where by simple we mean there is at most one directed edge from $i$ to $j$, for any pair of vertices $i,j \in [n]$, $i \neq j$. Note that the presence of both a directed edge from $i$ to $j$ and one from $j$ to $i$ is allowed. As in the previous section, we restrict our attention to those digraphs having positive complexity. Note this case still generalizes the work of Braun–Meyer [@BM17] (see Remark \[rmk:BM\]) and defines polytopes with at most one interior point. Indeed we prove that all the Laplacian polytopes of a simple directed graph on $n$ vertices are subpolytopes of $P_{K_n}$, the Laplacian simplex associated to the complete simple digraph. Observe $P_{K_n}$ is equivalent to the $n$-th dilation of an $(n-1)$-dimensional unimodular simplex, and therefore it has exactly one interior lattice point.
\[prop:simple\] Let $D$ be a simple digraph on $n$ vertices. Then $P_D$ is a subpolytope of $P_{K_n}$. In particular, if $D$ is strongly connected, then $P_D$ has exactly one interior lattice point.
Corollary \[cor:0ininterior\] implies that $P_D$ has at least one interior lattice point, so the second statement follows directly from the first one. In order to prove the first part, we show that any vertex ${\textnormal{\textbf{u}}}$ of $P_D$ is in $P_{K_n}$. Up to a relabeling of the vertices, we can assume that ${\textnormal{\textbf{u}}}=(a,-1,\ldots,-1,0,\ldots,0)$, where $a$ equals the number entries of $u$ which are equal to $-1$. We know that the Laplacian $L(K_n)$ is $$L(K_n) =
\begin{bmatrix}
n-1 & -1 & \dots & -1 \\
-1 & n-1 & \dots & -1 \\
\vdots & \vdots & \ddots & \vdots \\
-1 & -1 & \dots & n-1
\end{bmatrix}.$$ We denote by ${\textnormal{\textbf{v}}}_i$ the $i$-th row of $L(K_n)$, as well as the corresponding vertex of $P_{K_n}$. It is then enough to prove that ${\textnormal{\textbf{u}}}$ can be written as a convex combination of the vertices of $K_n$, i.e. that ${\textnormal{\textbf{u}}}= \sum_{i=0}^n \lambda_i {\textnormal{\textbf{v}}}_i$, with $0 \leq \lambda_i \leq 1$ and $\sum_{i=0}^n \lambda_i = 1$. This can be done with the following choice of barycentric coordinates: $$\lambda_i =
\begin{cases}
\frac{a+1}{n}, & \mbox{if } i = 1 \\
0, & \mbox{if } 2 \leq i \leq a+1 \\
\frac{1}{n}, & \mbox{if } a+2 \leq i \leq n
\end{cases}.$$ This proves that $P_D$ is a subpolytope of $P_{K_n}$.
Lattice simplices associated to generalized cycles
--------------------------------------------------
In [@BM17] the authors study the Laplacian simplex associated to the undirected cycle graph $C_n$, proving the following result.
\[thm:ref\_cycles\] For $n \ge 3$, the simplex $T_{C_n}$ is reflexive if and only if $n$ is odd.
The rest of this section is aimed to generalize their result to the case of directed cycles. Note that in order to have reflexivity (or, in particular, to have one interior lattice point) we need the digraph to be strongly connected (Corollary \[cor:0ininterior\]). Therefore, all cycles we consider will always contain a cycle entirely oriented in one of the two possible directions and some additional edges directed in the opposite direction. Informally speaking, we define a cycle digraph to have all the edges pointing clockwise and some edges pointing counterclockwise.
\[def:cycle\] Let $n \geq 3$. We say that a digraph $D$ on the vertex set $[n]$ is a *cycle digraph* if, up to a relabeling of the vertices, $E(D) = \overrightarrow{E}(D) \cup \overleftarrow{E}(D)$, where $$\begin{aligned}
\overrightarrow{E}(D) &= \{(1,2),(2,3),\ldots,(n-1,n),(n,1)\},\\
\overleftarrow{E}(D) &\subseteq \{(n,n-1),(n-1,n-2),\ldots,(2,1),(1,n)\}.\end{aligned}$$ If such a relabeling exists, $D$ is completely determined by $\overleftarrow{E}(D)$, and we denote it by $D=C_n^S$, where $S \subseteq [n]$ is the set of the tails of the directed edges in $\overleftarrow{E}(D)$. As an example see Figure \[fig:dir\_cycle\].
at (72:1) (1) [1]{}; at (0:1) (2) [2]{}; at (288:1) (3) [3]{}; at (216:1) (4) [4]{}; at (144:1) (5) [5]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (5) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (5) (5) edge \[bend left=20\] (1);
We first prove that, for most of the directed cycles, the associated Laplacian simplex has no other lattice points other than the origin and the vertices. Borrowing some terminology from the algebraic geometers, we call *terminal Fano* a simplex with this property.
\[thm:terminal\] Let $D$ be any cycle digraph. $P_{D}$ is terminal Fano if and only if $D$ is not, up to a relabeling of the vertices, one of the following six exceptional directed cycles.
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1) edge \[bend left=20\] (2);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1) edge \[bend left=20\] (3);
We prove that, for $n \geq 5$, $P_{C_n^S}$ is terminal Fano for all $S \subseteq [n]$. The lower dimensional cases are checked individually, leading to the six exceptional cases above. For $1 \leq i \leq n$, we set ${\textnormal{\textbf{v}}}_i=a_{i-1}{\textnormal{\textbf{e}}}_{i-1}+b_{i}{\textnormal{\textbf{e}}}_i-{\textnormal{\textbf{e}}}_{i+1}$ with $a_{j} \in \{-1,0\}$ and $b_{j}=1-a_{j} \in \{1,2\}$, where $a_0=a_n$, ${\textnormal{\textbf{e}}}_0={\textnormal{\textbf{e}}}_n$ and ${\textnormal{\textbf{e}}}_{n+1}={\textnormal{\textbf{e}}}_{1}$. Assume that $P_{C_n^S}$ is not terminal Fano. Let ${\textnormal{\textbf{x}}}=(x_1,\ldots,x_n)$ be a lattice point of $P_{C_n^S} \setminus (\{{\textnormal{\textbf{v}}}_1,\ldots,{\textnormal{\textbf{v}}}_n, {\bf 0} \})$ and set ${\textnormal{\textbf{x}}}= \sum_{i=1}^{n}\lambda_i {\textnormal{\textbf{v}}}_i$ with $0 \leq \lambda_1,\ldots,\lambda_n <1$ and $\lambda_1+\cdots+\lambda_n=1$. Then one has $x_i=-\lambda_{i-1}+b_{i}\lambda_i+a_{i+1}\lambda_{i+1} \in \{-1,0,1\}$ for each $i$, where $\lambda_0=\lambda_n$, $\lambda_{n+1}=\lambda_1$ and $a_{n+1}=a_1$. If $x_2=-1$, then we obtain $a_{3}=-1$, $0<\lambda_{1},\lambda_{3} <1$ and $\lambda_j=0$ for any $j \neq 1,3$. Since $x_{3}=2\lambda_{3}$, it follows that $\lambda_1=\lambda_3=1/2$ and $a_{1}=-1$. However, $x_4=-1/2$, a contradiction. Hence we have $x_i \in \{0,1\}$ for each $i$. Since ${\textnormal{\textbf{x}}}\neq {\textnormal{\textbf{0}}}$, we can assume that $x_3=1$. Then one has $b_3=2$ and $\lambda_3 \geq 1/2$. If $b_2=1$, then we obtain $\lambda_2=\lambda_3=1/2$. However, $x_3=1/2$, a contradiction. Hence one has $b_2=b_4=2$. Moreover, it follows from $\lambda_3 \geq 1/2$ that $x_2=x_4=0$. Then it follows that $\lambda_1=3a_5\lambda_5+4\lambda_4-2 \leq 4 \lambda_4-2 \leq 0$. Hence we obtain $\lambda_3=\lambda_4=1/2$. However, $x_3=1/2$, a contradiction. Therefore, $P_{C_n^S}$ is terminal Fano.
Now we characterize reflexivity for Laplacian simplices $P_{C_n^S}$, extending Theorem \[thm:ref\_cycles\] by Braun–Meyer.
\[thm:ref\_dir\_cycles\] The Laplacian simplex $P_{C_n^S}$ associated to a cycle digraph $C_n^S$, is reflexive if and only if one of the following conditions is satisfied:
1. $S = \varnothing$, or
2. $S = [n]$ and $n=2$, or
3. $S = [n]$ and $n$ is odd, or
4. $\varnothing \subsetneq S \subsetneq [n]$, such that $k | c(D)$ for each integer $1 \leq k \leq K+1$, where $K$ is the longest chain of consecutive edges pointing counterclockwise, i.e. $$K \coloneqq \max \{ j \mid \{a+1,\ldots,a+j\} \subseteq S, \mbox{ for some } a \in [n]\},$$ where, since $S \subsetneq [n]$, we have assumed without loss of generality, that $1 \notin S$.
If $S$ satisfies $(1)$ or $(2)$, then it trivial to check that $P_{C_n^S}$ is reflexive. If $S$ satisfies $(3)$, then $P_{C_n^S}$ is reflexive by Theorem \[thm:ref\_cycles\]. Suppose now that $S$ satisfies $(4)$. In particular we have assumed that $1 \notin S$. This implies vertex $n$ has exactly one spanning converging tree, i.e. $c_{n}=1$. As usual, $c_i$ denotes, the number of spanning trees which converge to vertex $i$. Then $\gcd(c_1, \ldots, c_n) = 1$, and $P_{C_n^S}$ is a weighted projective space by Proposition \[prop:wps\]. For each vertex $i$ we denote by $K_i$ the length of the longest chain of consecutive edges pointing counterclockwise ending in $i$, i.e. $$K_i \coloneqq \max \{ j \mid \{i+1,\ldots,i+j\} \subseteq S\}, \qquad \mbox{for $1 \leq i \leq n$ }$$ in particular $K_n=0$ and $K= \max \{K_i \mid i \in [n] \}$. Given $i \in [n]$, note that there are exactly $K_i + 1$ spanning trees converging to $i$. There are $K_i$, having edge set $$\{ (j,j-1),\ldots,(i+1,i),(j+1,j+2),\ldots,(n-1,n),(n,1),\ldots,(i-1,i) \},$$ for all $j \in \{i+1,\ldots,i+K_i\}$, plus an additional “clockwise tree” with edges $$\{(i+1,i+2),\ldots,(n-1,n),(n,1),\ldots,(i-1,i) \}.$$ By Corollary \[cor:refl\], $P_{C_n^S}$ is reflexive if and only if $c_i|c(D)$, for all $i \in [n]$. We conclude by noting that if $c_i > 1$ for some $i \in [n]$, then $c_{i+1}=c_i-1$, in particular $\{c_i \mid i \in [n] \} = \{ 1, \ldots, K+1\}$.
We now have all the tools to completely characterize all the reflexive IDP simplices arising from cycle digraphs.
\[thm\_IDP\] Let $C_n^S$ be a cycle digraph on $n$ vertices such that $P_{C_n^S}$ is reflexive. Then $P_{C_n^S}$ possesses the integer decomposition property if and only if $D$ satisfies one of the following conditions:
1. $S=\varnothing$, or
2. $D$ is, up to a relabeling of the vertices, one of the following directed cycles.
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1);
at (60:1) (1) [$1$]{}; at (300:1) (2) [$2$]{}; at (180:1) (3) [$3$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (3) (2) edge \[bend left=20\] (3) edge \[bend left=20\] (1) (3) edge \[bend left=20\] (1) edge \[bend left=20\] (2);
at (60:1) (1) [$1$]{}; at (330:1) (2) [$2$]{}; at (240:1) (3) [$3$]{}; at (150:1) (4) [$4$]{};
\(1) edge \[bend left=20\] (2) edge \[bend left=20\] (4) (2) edge \[bend left=20\] (3) (3) edge \[bend left=20\] (4) edge \[bend left=20\] (2) (4) edge \[bend left=20\] (1);
If $S=\varnothing$ then $C_n^S$ is always a reflexive IDP simplex. If $S=[n]$, from Theorem \[thm:ref\_cycles\], $P_{C_n^{[n]}}$ is reflexive if and only if $n$ is odd. In this case it is known that $P_{C_n^{[n]}}$ is IDP if and only if $n=3$ [@BM17 Corollary 5.11]. Now, assume that $\varnothing \neq S \neq [n]$ and $P_{C_n^S}$ is IDP. We use the same notation introduced in Theorem \[thm:ref\_dir\_cycles\]. Then we can assume that $c_1=1,c_2=K+1, c_3 = K, \ldots,c_{K+1}=2$. Set ${\textnormal{\textbf{q}}}=(c_2,\ldots,c_n)$. It follows that $P_{C_n^S}$ is unimodularly equivalent to $\Delta_{(1,{\textnormal{\textbf{q}}})}$. By Lemma \[lem:BDS\_IDP\], we know that, for each $2 \leq j \leq n$ $$\label{eq:IDP}
\frac{1}{c_j}+\sum_{i\neq j}\left\{\frac{c_i}{c_j}\right\}=1.$$ But, if $K \geq 3$, by we get $$\dfrac{1}{K+1}+\sum_{i=3}^{n}\left\{\dfrac{c_i}{K+1}\right\} \geq \dfrac{1}{K+1}+\dfrac{K-1}{K+1}+\dfrac{K}{K+1}>1,$$ so $K$ must satisfy $1 \leq K \leq 2$. By applying in these cases one gets $n \leq 4$. By checking all the cycle digraph having up to four vertices, we get the exceptions represented above.
As an application of the tools developed in this section, we build a special family of cycle digraphs whose Laplacian polytopes are reflexive and have non unimodal $h^*$-vectors.
\[thm:non\_unim\] Let $\alpha, \beta, k \in {{\mathbb Z}}_{>0}$ such that $\alpha \le \beta \le k-1$ and $\alpha + \beta \le k+1$. Let $D=C_n^S$ be the cycle digraph, with $n \coloneqq 6(k+1) - 2 \alpha - \beta$, and $S \coloneqq S_1 \cup S_2 \cup S_3$, with $$\begin{aligned}
& S_1 \coloneqq \{1 + 3h \mid 0 \leq h \leq \alpha - 1\}, \\
& S_2 \coloneqq \{2 + 3h \mid 0 \leq h \leq \alpha - 1\}, \\
& S_3 \coloneqq \{ 3\alpha + 1 + 2h \mid 0 \leq h \leq \beta-\alpha - 1\}.\end{aligned}$$ Then $P_{D}$ is a reflexive simplex of dimension $6(k+1) - 2 \alpha - \beta - 1$ with symmetric and nonunimodal $h^*$-vector $$(
\underbrace{1,\ldots,1}_{2(k+1)-\alpha} ,
\underbrace{2,\ldots,2}_{\alpha} ,
\underbrace{1,\ldots,1}_{(k+1)-\alpha-\beta} ,
\underbrace{2,\ldots,2}_{\beta} ,
\underbrace{1,\ldots,1}_{(k+1)-\alpha-\beta} ,
\underbrace{2,\ldots,2}_{\alpha} ,
\underbrace{1,\ldots,1}_{2(k+1)-\alpha} )$$
An example of the digraph in the statement is represented in Figure \[fig:non\_unim\]. The digraph has no more than two consecutive vertices with outdegree two, so the number of spanning trees converging to each of the vertices of $D$ is at most three. Specifically, $$c_i =
\begin{cases}
3, & \mbox{if } i \in S_1,\\
2, & \mbox{if } i \in S_2 \cup S_3,\\
1, & \mbox{if } i \in [n] \setminus S.
\end{cases}$$ Since each $c_i$ divides $c(D) = \sum_{i = 1 }^n c_i = 6(k+1)$, then $P_D$ is reflexive by Theorem \[thm:ref\_dir\_cycles\]. Now we use Theorem \[thm:hstarq\] to describe its $h^*$-polynomial. In particular $$h^*(P_D,z)=\sum_{b=0}^{c(D)-1} x^{w(b)}, \quad \text{with} \; w(b)=b-\sum_{i=1}^{n} \left \lfloor \frac{c_i b}{6(k+1)} \right \rfloor.$$ In our case this becomes $$w(b) = b - \alpha \left \lfloor \frac{b}{2(k+1)}\right \rfloor - \beta \left \lfloor \frac{b}{3(k+1)} \right \rfloor,$$ which yields $$w(b) = \left\{\begin{alignedat}{5}
&b, && \mbox{if } \quad &0 &\leq b \leq 2(k+1)-1, \\
&b-\alpha, && \mbox{if } &2(k+1) &\leq b \leq 3(k+1)-1, \\
&b-\alpha-\beta, && \mbox{if } &3(k+1) &\leq b \leq 4(k+1)-1, \\
&b-2\alpha-\beta, \quad && \mbox{if } &4(k+1) &\leq b \leq 6(k+1)-1.
\end{alignedat}\right.$$ From this we deduce the $i$-th coefficient of the $h^*$-polynomial: $$h^*_i =
\begin{cases}
2, & \mbox{if } \qquad
\begin{cases}
2(k+1)-\alpha \leq i \leq 2(k+1)-1, \,\text{or}\\
3(k+1)-\alpha -\beta \leq i \leq 3(k+1)-\alpha-1,\,\text{or}\\
4(k+1)-2\alpha -\beta \leq i \leq 4(k+1)-\alpha-\beta-1; \,\\
\end{cases} \\
1, & \mbox{otherwise.} \\
\end{cases}$$
at (96:3) (1) [$1$]{}; at (72:3) (2) [$2$]{}; at (48:3) (3) [$3$]{}; at (24:3) (4) [$4$]{}; at (0:3) (5) [$5$]{}; at (336:3) (6) [$6$]{}; at (312:3) (7) [$7$]{}; at (288:3) (8) [$8$]{}; at (264:3) (9) [$9$]{}; at (240:3) (10) [$10$]{}; at (216:3) (11) [$11$]{}; at (192:3) (12) [$12$]{}; at (168:3) (13) [$13$]{}; at (144:3) (14) [$14$]{}; at (120:3) (15) [$15$]{};
\(1) edge \[bend left=10\] (2) edge \[bend left=10\] (15) (2) edge \[bend left=10\] (3) edge \[bend left=10\] (1) (3) edge \[bend left=10\] (4) (4) edge \[bend left=10\] (5) (5) edge \[bend left=10\] (6) (6) edge \[bend left=10\] (7) (7) edge \[bend left=10\] (8) (8) edge \[bend left=10\] (9) (9) edge \[bend left=10\] (10) (10) edge \[bend left=10\] (11) (11) edge \[bend left=10\] (12) (12) edge \[bend left=10\] (13) (13) edge \[bend left=10\] (14) (14) edge \[bend left=10\] (15) (15) edge \[bend left=10\] (1);
Further Questions {#sec:underlying}
=================
Note that, in the case of undirected cycles studied by Braun-Meyer [@BM17], the reflexivity is influenced by the number of vertices of the graph (Theorem \[thm:ref\_cycles\]). On the other hand, when passing to the undirected case we discussed in Section \[sec:4\], it is clear (from Theorem \[thm:ref\_dir\_cycles\]) that one can build reflexive Laplacian simplices starting from cycles of any length. This can be done by orienting a cycle in one of the two directions.
It is natural to wonder how the structure of the underlying simple graph $G_D$ of a digraph $D$ plays a role in determining the reflexivity of $P_D$ (see Section \[sec:4\] for the definition of underlying simple graph). We formalize this problem with the three questions below. We define an *oriented graph* to be a simple digraph $D$ such that if there is an edge pointing from $i$ to $j$, then there is no edge pointing from $j$ to $i$.
\[q1\] For any simple graph $G$ on $[n]$, does there exist an oriented graph $D$ on $[n]$ such that $G_D=G$ and $P_D$ is a reflexive $(n-1)$-simplex?
\[q2\] For any simple graph $G$ on $[n]$, does there exist a digraph $D$ on $[n]$ such that $G_D=G$ and $P_D$ is a reflexive $(n-1)$-simplex?
\[q3\] For any simple graph $G$ on $[n]$, does there exist a simple digraph $D$ on $[n]$ such that $G_D=G$ and $P_D$ is a reflexive $(n-1)$-simplex?
We have been able to answer negatively the first two questions with the following examples. Let $G_1$ be the following graph.
at (90:1) (1) [$1$]{}; at (180:1) (2) [$2$]{}; at (270:1) (5) [$5$]{}; at (0:1) (4) [$4$]{}; at (0:0) (3) [$3$]{}; at (180:2) (6) [$G_1=$]{};
\(1) edge \[bend right=0\] (2) edge \[bend left=0\] (3) edge \[bend left=0\] (4) (5) edge \[bend left=0\] (2) edge \[bend left=0\] (3) edge \[bend left=0\] (4);
Assume that $D_1$ is an orientation of $G_1$ such that $P_{D_1}$ is a reflexive $4$-simplex. Since $D_1$ is strongly connected, we may assume, without loss of generality, that the edges $(5,3),(3,1),(1,2),(2,5)$ are in $E(D_1)$. It follows that either $(1,4),(4,5)$ or $(5,4),(4,1)$ are in $E(D_1)$. In both cases, $P_{D_1}$ is not reflexive. In particular there are graphs that cannot be oriented to produce reflexive Laplacian simplices. Hence Question \[q1\] is not true in general. However, $G_1$ is not a counterexample of Question \[q2\]. Indeed, let $D_1'$ be the following simple digraph.
at (90:1) (1) [$1$]{}; at (180:1) (2) [$2$]{}; at (270:1) (5) [$5$]{}; at (0:1) (4) [$4$]{}; at (0:0) (3) [$3$]{}; at (180:2) (6) [$D_1'=$]{};
\(1) edge \[bend right=0\] (2) edge \[bend left=20\] (3) edge \[bend left=0\] (4) (2) edge \[bend left=20\] (5) (3) edge \[bend left=20\] (1) (4) edge \[bend left=20\] (5) (5) edge \[bend left=20\] (2) edge \[bend left=0\] (3) edge \[bend left=20\] (4);
Note that its underlying simple graph is still $G_1$, but $P_{D_1'}$ is reflexive. However, let $G_2$ be the following graph.
at (120:1) (1) [$1$]{}; at (60:1) (2) [$2$]{}; at (180:1) (3) [$3$]{}; at (0:1) (4) [$4$]{}; at (240:1) (5) [$5$]{}; at (300:1) (6) [$6$]{}; at (180:2) (7) [$G_2=$]{};
\(1) edge \[bend right=0\] (3) (2) edge \[bend right=0\] (4) (3) edge \[bend left=0\] (5) edge \[bend left=0\] (4) edge \[bend left=0\] (6) (4) edge \[bend left=0\] (5) edge \[bend left=0\] (6) (5) edge \[bend left=0\] (6);
Note that there are finitely many possible directed simple graphs having it as underlying graph. A computer-assisted check shows that none of them produces a reflexive Laplacian simplex. Hence Question \[q2\] is also not true in general. However, $G_2$ is not a counterexample of Question \[q3\]. Indeed, let $D'_2$ be the following digraph (the label on an edge from $i$ to $j$, if present, represents the total number of edges from $i$ to $j$).
at (120:1) (1) [$1$]{}; at (60:1) (2) [$2$]{}; at (180:1) (3) [$3$]{}; at (0:1) (4) [$4$]{}; at (240:1) (5) [$5$]{}; at (300:1) (6) [$6$]{}; at (180:2) (7) [$D_2'=$]{};
\(1) edge \[bend right=0\] (3) (2) edge \[bend right=0\] (4) (3) edge \[bend right=30\] node\[xshift=-.35cm,yshift=-.1cm\] [${\tiny 3}$]{} (5) edge \[bend left=20\] node\[xshift=-.0cm,yshift=-.2cm\] [${\tiny 3}$]{} (4) edge \[bend left=10\] node\[xshift=-.3cm,yshift=-.2cm\] [${\tiny 3}$]{} (6) (4) edge \[bend left=0\] node\[xshift=-.0cm,yshift=.2cm\] [${\tiny 3}$]{} (3) edge \[bend left=10\] node\[xshift=-.0cm,yshift=.3cm\] [${\tiny 3}$]{} (5) edge \[bend left=30\] node\[xshift=-.1cm,yshift=.3cm\] [${\tiny 3}$]{} (6) (5) edge \[bend left=0\] node\[xshift=.2cm,yshift=.2cm\] [${\tiny 3}$]{} (3) edge \[bend left=10\] node\[xshift=.2cm,yshift=-.2cm\] [${\tiny 3}$]{} (4) edge \[bend left=0\] node\[xshift=.0cm,yshift=-.2cm\] [${\tiny 3}$]{} (6) (6) edge \[bend left=10\] node\[xshift=.0cm,yshift=.3cm\] [${\tiny 3}$]{} (3) edge \[bend left=0\] node\[xshift=.1cm,yshift=-.3cm\] [${\tiny 3}$]{} (4) edge \[bend left=20\] node\[xshift=.0cm,yshift=.1cm\] [${\tiny 3}$]{} (5);
Then $P_{D'_2}$ is reflexive simplex. Question \[q3\] remains open.
[10]{}
Victor V. Batyrev. Dual polyhedra and mirror symmetry for [C]{}alabi-[Y]{}au hypersurfaces in toric varieties. , 3(3):493–535, 1994.
Matthias Beck and Sinai Robins. . Undergraduate Texts in Mathematics. Springer, New York, second edition, 2015. Integer-point enumeration in polyhedra, With illustrations by David Austin.
Benjamin Braun. Unimodality problems in [E]{}hrhart theory. In [*Recent trends in combinatorics*]{}, volume 159 of [*IMA Vol. Math. Appl.*]{}, pages 687–711. Springer, \[Cham\], 2016.
Benjamin Braun, Robert Davis, and Liam Solus. Detecting the integer decomposition property and ehrhart unimodality in reflexive simplices, 2016. Preprint [[`arXiv:1608.01614`](http://arxiv.org/abs/1608.01614)]{}.
Benjamin Braun and Marie Meyer. Laplacian simplices, 2017. Preprint [[`arXiv:1706.07085`](http://arxiv.org/abs/1706.07085)]{}.
Francesco Brenti. Log-concave and unimodal sequences in algebra, combinatorics, and geometry: an update. In [*Jerusalem combinatorics ’93*]{}, volume 178 of [*Contemp. Math.*]{}, pages 71–89. Amer. Math. Soc., Providence, RI, 1994.
Fan R. K. Chung. , volume 92 of [*CBMS Regional Conference Series in Mathematics*]{}. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1997.
Heinke Conrads. Weighted projective spaces and reflexive simplices. , 107(2):215–227, 2002.
Igor Dolgachev. Weighted projective varieties. In [*Group actions and vector fields ([V]{}ancouver, [B]{}.[C]{}., 1981)*]{}, volume 956 of [*Lecture Notes in Math.*]{}, pages 34–71. Springer, Berlin, 1982.
Takayuki Hibi. Dual polytopes of rational convex polytopes. , 12(2):237–240, 1992.
A. R. Iano-Fletcher. Working with weighted complete intersections. In [*Explicit birational geometry of 3-folds*]{}, volume 281 of [ *London Math. Soc. Lecture Note Ser.*]{}, pages 101–173. Cambridge Univ. Press, Cambridge, 2000.
Alexander Kasprzyk. Classifying terminal weighted projective space, 2013. Preprint [[`arXiv:1304.3029`](http://arxiv.org/abs/1304.3029)]{}.
Tetsushi Matsui, Akihiro Higashitani, Yuuki Nagazawa, Hidefumi Ohsugi, and Takayuki Hibi. Roots of [E]{}hrhart polynomials arising from graphs. , 34(4):721–749, 2011.
H. Ohsugi and T. Hibi. A normal [$(0,1)$]{}-polytope none of whose regular triangulations is unimodular. , 21(2):201–204, 1999.
Hidefumi Ohsugi and Takayuki Hibi. Normal polytopes arising from finite graphs. , 207(2):409–426, 1998.
Hidefumi Ohsugi and Takayuki Hibi. Special simplices and [G]{}orenstein toric rings. , 113(4):718–725, 2006.
Richard P. Stanley. Log-concave and unimodal sequences in algebra, combinatorics, and geometry. In [*Graph theory and its applications: [E]{}ast and [W]{}est ([J]{}inan, 1986)*]{}, volume 576 of [*Ann. New York Acad. Sci.*]{}, pages 500–535. New York Acad. Sci., New York, 1989.
Richard P. Stanley. , volume 62 of [ *Cambridge Studies in Advanced Mathematics*]{}. Cambridge University Press, Cambridge, 1999. With a foreword by Gian-Carlo Rota and appendix 1 by Sergey Fomin.
Tuan Tran and Günter M. Ziegler. Extremal edge polytopes. , 21(2):Paper 2.57, 16, 2014.
Rafael H. Villarreal. On the equations of the edge cone of a graph and some applications. , 97(3):309–317, 1998.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report the observation of the $A_{\rm g}$-symmetric one-magnon Raman peak in the magnon Bose-Einstein condensation phase of TlCuCl$_{3}$. Its Raman shift traces the one-magnon energy at the magnetic $\Gamma$ point, and its intensity is proportional to the squared transverse magnetization. The appearance of the one-magnon Raman scattering originates from the exchange magnon Raman process and reflects the change of the magnetic-state symmetry. Using the bond-operator representation, we theoretically clarify the Raman selection rules, being consistent with the experimental results.'
author:
- Haruhiko Kuroe
- Kouhei Kusakabe
- Akira Oosawa
- Tomoyuki Sekine
- Fumiko Yamada
- Hidekazu Tanaka
- Masashige Matsumoto
title: ' Magnetic field-induced one-magnon Raman scattering in the magnon Bose-Einstein condensation phase of TlCuCl$_{3}$ '
---
\[Introduction\]Introduction
============================
Currently, many physicists are examining the Bose-Einstein condensation (BEC) of atoms in ultracooled dilute gases, and in particular, the BEC of magnons. The latter, which is the magnetic-field induced quantum phase transition to the magnon BEC phase, has been reported in $S = 1/2$ antiferromagnets with a spin gap, such as KCuCl$_{3}$, TlCuCl$_{3}$,[@Oosawa1999; @PhysRevLett.84.5868; @Tanaka2001; @Yamada2008] BaCuSi$_{2}$O$_{6}$,[@Jaime2004] and Pb$_{2}$V$_{3}$O$_{9}$.[@Pb2V3O9] The change of the magnon dispersion relation in TlCuCl$_{3}$ through the magnon BEC phase transition at $H_{\rm c} \sim 6$ T has been observed by inelastic neutron scattering [@Ruegg2003] and has been explained using the bond-operator representation.[@Matsumoto2004] One of the characteristic features of the magnon BEC phase is the formation of massless excitation, i.e., the Goldstone mode at the magnetic $\Gamma$ point, indicating the spontaneous breaking of the continuous symmetry. However the details of the magnon excitations, especially their symmetries, have not yet been established. Raman scattering is a powerful tool to study phase transitions. Because the magnon Raman process is sensitive to the symmetries of the ground and excited states,[@Fleury1968] Raman-scattering measurement above $H_{\rm c}$ presents great potential to study the change of the ground and excited states through the magnon BEC phase transition.
This paper reports the observation of one-magnon Raman scattering originating from changes of the ground and excited states through the magnon BEC phase transition. This study focused on TlCuCl$_{3}$ where the magnon excitations and magnetic parameters below and above $H_{\rm c}$ have been studied in detail.[@Ruegg2003; @Matsumoto2004] First, we show our experimental results above $H_{\rm c}$. We then construct the microscopic theory of one-magnon Raman scattering in the exchange magnon Raman process using the bond-operator representation, which can explain the experimental results clearly. Based on our results, the Raman selection rule will be clarified.
Experiments
===========
Single crystals of TlCuCl$_{3}$ were prepared by the vertical Bridgman method.[@Oosawa1999] The 5145-Å line of Ar$^+$-ion laser polarized along the (201) axis was incident on the (010) cleavage surface. We set the samples in the cryostat under the dried N$_{2}$ or He gas atmosphere, because the sample was easily damaged by moisture in air. We placed the microscope in the vacuum chamber of superconducting magnet in order to collect the scattered light effectively. This enabled us to select good surface positions of crystals and the effects of the direct scattering in the low-energy region were avoided. Magnetic fields of up to 10 T were applied nearly parallel to the $(010)$ axis. The effect of a weak component of magnetic field along the (201) axis due to the experimental setting is negligible because the effect of the anisotropic $g$ tensor along these directions is small.[@Glazkov2004]
Results
=======
![ Polarization characteristics of Raman spectra at 9 T. The inset shows the integrated Raman intensity of the phonon peak at 137 cm$^{-1}$. The arrows indicates the Raman peaks coming from the $B_{\rm g}$ phonons. The unpolarized Raman spectrum at 0 T is also shown. []{data-label="Polarization"}](Polarization.eps){width="45.00000%"}
Figure \[Polarization\] compares the low-temperature Raman spectrum at 9 T (above $H_{\rm c}$ of TlCuCl$_{3}$) to that at 0 T. At 0 T, we observed several sharp phonon peaks superimposed on the two-magnon Raman band extending from 11 cm$^{-1}$, i.e., twice the energy of the magnetic gap,[@Ruegg2003] to about 150 cm$^{-1}$. This spectrum is consistent with the results of refs. and . At 9 T, we observed the new Raman peak with a Lorentzian lineshape, called P1, at 20 cm$^{-1}$. No other significant change was observed. P1 excitation has the $A_{\rm g}$ symmetry, which is obtained with the following procedures. For the incident laser polarized along the (201) axis, ${\bm E}_{\rm in} // (201)$, we measured the scattered light with polarization ${\bm E}_{\rm sc}$ which is rotated from ${\bm E}_{\rm in}$ with an angle $\theta$. The Raman intensities are normalized so that the 137-cm$^{-1}$ $A_{\rm g}$-symmetric phonon peaks in each spectrum, for which the $\theta$-dependence is shown in the inset of Fig. \[Polarization\], have the same intensity. The Raman intensity from the quasiparticles with $A_{\rm g}$ symmetry, including P1, is $\theta$-independent in this plot while those with $B_{\rm g}$ symmetry indicated by arrows increased with increasing $\theta$, as shown in Fig. \[Polarization\].
Figure \[RamanSpectra\] shows the detailed magnetic-field dependence of Raman spectra at 1.9 K. At 0 T, the 25- and 32-cm$^{-1}$ phonon Raman peaks, called P2 and P3, respectively, are superimposed on the two-magnon Raman band starting at 11 cm$^{-1}$. P2 and P3 have the $A_{\rm g}$ symmetry as well as P1, as shown in Fig. \[Polarization\]. The Raman spectra below 5 T are magnetic-field independent. Above 7 T, we clearly observed that the frequency and intensity of P1 strongly depended on the applied magnetic field, as denoted by the hatched areas in Fig. \[RamanSpectra\], of which the details will be explained later. Around 6 T, the increase of the Rayleigh scattering around 0 cm$^{-1}$ suggests the quasielastic (or critical) light scattering reflecting the large magnetic specific heat around $H_{\rm c}$ as observed in several antiferromagnets or spin-Peierls system.[@FePS3; @Kuroe1997; @Halley1978] To discuss the quasielastic light scattering quantitatively, Raman-scattering measurements in the anti-Stokes region are necessary.
![ Magnetic-field dependence of Raman spectra in TlCuCl$_{3}$ at 1.9 K. The fitting curves are superimposed on the experimental data above 7 T. The details of the fitting curve (solid curves) together with the background generated by the two-magnon Raman band (dashed lines) are given in the text. The hatched areas show the component of the Raman intensity generated by P1. []{data-label="RamanSpectra"}](RamanSpectra.eps){width="45.00000%"}
The lineshapes of P1, P2, and P3 are well described by three Lorentzian curves superimposed on the background: $$\displaystyle
I(\omega) =
\sum_{i = 1}^{3}
\frac{(n + 1)k_{i}^{2} \omega \Gamma_{i}}{( \omega^{2} - \omega_{i}^{2} )^{2}
+ (\omega \Gamma_{i})^{2}}
+ {\rm background}
\ ,
\label{Lorentzian}$$ where $k_{i}$, $\hbar\omega_{i}$, and $\Gamma_{i}$ indicate the Raman coupling coefficient, the energy, and the halfwidth of P$i$, respectively. Here, the Bose factor $(n + 1)$ can be treated as unity because the temperature is much lower than the energies of quasiparticles. The background generated by the two-magnon Raman band peaking around 50 cm$^{-1}$ was assumed to be a linear function, as shown by the dashed lines in Fig. \[RamanSpectra\]. The calculated curves reproduced the observed data well. We show the Raman intensity generated by P1 (the term related to the subscript $i$ = 1 in eq. (\[Lorentzian\])) on the linear background as the hatched area in Fig. \[RamanSpectra\]. Around $H_{\rm c}$, we could not distinguish P1 from the two-magnon Raman band because of its weak intensity.
Figure \[RamanParms\](a) shows the peak energies $\hbar\omega_{1}$ and $\hbar\omega_{2}$ as functions of magnetic field together with the calculated one-magnon energy $E_{g\alpha}({\bm Q})$ ($\alpha = -,0,+$) with the wavevector $\bm{Q} = (0,0,2\pi)$, [@PhysRevLett.89.077203; @Matsumoto2004] where the magnetic gap is closed.[@Tanaka2001; @Oosawa2002] One can see that $\hbar \omega_{1}$ below 10 T agrees with $E_{g-}({\bm Q})$ within experimental accuracy. Figure \[RamanParms\](b) shows the squared Raman coupling coefficient $k^{2} = k_{1}^{2}/k_{3}^{2}$, which is proportional to the integrated Raman intensity of P1. Here, we normalized $k_{1}^{2}$, which is proportional to the area hatched in Fig. \[RamanSpectra\], by $k_{3}^{2}$ to correct errors due to the small deviations of optic alignment. The errors of $k^{2}$ are similar to the symbol size in Fig. \[RamanParms\](b). For comparison, we show the magnetic-field dependences of squared transverse magnetization $M_{xy}^{2}$ (ref. ) and squared longitudinal magnetization $M_{z}^{2}$ (ref. ) together with their calculated values.[@Matsumoto2004] One can see that the magnetic-field dependence of Raman intensity is well scaled to the former.
![ (a) Magnetic-field dependence of one-magnon (filled circles) and phonon (empty circles) energies together with that of one-magnon energies calculated in refs. and . (b) Magnetic-field dependence of squared Raman coupling coefficient (filled circles). Squared transverse magnetization $M_{xy}^{2}$ (empty squares, ref. ) and squared longitudinal magnetization $M_{z}^{2}$ (filled squares, ref. ) are also plotted with the calculated values for them (a solid and a dashed curves for $M_{xy}^{2}$ and $M_{z}^{2}$, respectively).[@Matsumoto2004] []{data-label="RamanParms"}](RamanParms.eps){width="45.00000%"}
\[Discussion\]Discussion
========================
First, we discuss the origin of P1. Judging from the polarization characteristics, the scattering process of P1 comes from the exchange magnon Raman scattering, as proposed by Fleury and Loudon,[@Fleury1968] which usually creates the broad two-magnon Raman band with $A_{\rm g}$ symmetry reported at zero magnetic field.[@Choi2003; @Kusakabe2007] Because the two-magnon Raman band at 9 T is almost the same as that at 0 T, as seen in Fig. \[Polarization\], we do not need to consider the increase of the magnon-magnon interaction which may cause the two-magnon Raman band with a nearly Lorentzian lineshape [@Weber1989; @PhysRevB.42.4842] as well as the formation of the two-magnon bound state.[@Sekine1998] P1 does not originate from a three-magnon process, where the thermally excited triplets play an essential role.[@PhysRevLett.79.5138; @Choi2005PRB] The three-magnon Raman intensity, which is proportional to $n$ at low temperatures, should be negligibly small at 1.9 K. Because P1 has a Lorentzian lineshape, the one-magnon Raman scattering can be considered as the origin of P1. Hereafter, we consider the detail of the magnon Raman process using the bond-operator representation and clarify that the one-magnon Raman scattering from the exchange magnon Raman process becomes possible in the magnon BEC phase.
The effective Raman operator ${\cal R}$ in the above-mentioned exchange magnon Raman process has a form given by the isotropic Heisenberg-type exchange interaction between spins $\bm{S}_{i}$ and $\bm{S}_{j}$: $${\cal R} =
\sum_{i,j} {\cal R}_{i,j} =
\sum_{i,j} F_{i,j}
(\hat{\bm E}_{\rm in} \cdot \hat{\bm r}_{ij})
(\hat{\bm E}_{\rm sc} \cdot \hat{\bm r}_{ij})
\bm{S}_{i} \cdot \bm{S}_{j} \ ,
\label{FleuryLoudon}$$ where $\bm{r}_{ij}$ indicates the position vector between $\bm{S}_{i}$ and $\bm{S}_{j}$ and the sum runs over all the interacting spin pairs. Here, $\hat{\bm r} = {\bm r}/|{\bm r}|$, $\hat{\bm E} = {\bm E}/|{\bm E}|$, and the coefficient $F_{i,j}$ depends on the pathway of interaction between $\bm{S}_{i}$ and $\bm{S}_{j}$. ${\cal R}$ depends on the experimental setting through the $(\hat{\bm E}_{\rm in} \!\! \cdot \! \hat{\bm r}_{ij})
(\hat{\bm E}_{\rm sc} \!\! \cdot \! \hat{\bm r}_{ij})$ term. The matrix element of ${\cal R}$ between the initial state $| {\rm i} \rangle$ and the final one $| {\rm f} \rangle$ is called Raman tensor. The magnon Raman intensity is given as $${\cal I}^{\rm (in,sc)} (\omega) \propto
\displaystyle
|\bm{E}_{\rm in}|^{2}
\sum_{| {\rm f} \rangle}
\left|
\langle {\rm i} | {\cal R} | {\rm f} \rangle
\right|^{2}
\delta(\omega - \omega_{\rm if})
\ ,$$ where $\hbar \omega_{\rm if}$ is the excitation energy between the states $|{\rm i}\rangle$ and $|{\rm f}\rangle$.
We note here that ${\cal R}$ is [*always*]{} written using the pure singlet operator $s_{\bm{k}}$ and the triplet operators $t_{\bm{k}\alpha}$ $(\alpha = -,0,+)$, which annihilate the triplets with $S^{z} = \alpha$. We need to rewrite ${\cal R}$ using the creation and annihilation operators of eigenstates in the magnon BEC phase. As discussed by Matsumoto [*et al.*]{}, [@PhysRevLett.89.077203; @Matsumoto2004] the following transformed operators based on the bond-operator representation characterize the magnon excitations in the magnon BEC phase: $$\begin{array}{l}
\displaystyle
s_{\bm{k}} \! = \! u a_{\bm{k}} \! - \! v b_{\bm{k}+\bm{Q}+},
\\
\displaystyle
t_{\bm{k}+} \! = \! vf a_{\bm{k}-\bm{Q}} \! + \! uf b_{\bm{k}+} \! - \! g b_{\bm{k}-} ,
\\
\displaystyle
t_{\bm{k}0} \! = \! b_{\bm{k}0},
\\
\displaystyle
t_{\bm{k}-} \! = \! vg a_{\bm{k}-\bm{Q}} \! + \! ug b_{\bm{k}+} \! + \! f b_{\bm{k}-}
,
\label{mixedrepresentation}
\end{array}$$ where the momentum-independent real-number parameters $u$, $v$, $f$, and $g$ satisfy $f^{2}+g^{2} = u^{2}+v^{2} = 1$. Below $H_{\rm c}$, $v = 0$ and one can obtain the simple relations $a_{\bm{k}} = s_{\bm{k}}$ and $b_{\bm{k}\alpha} = t_{\bm{k}\alpha}$. Above $H_{\rm c}$, $v \neq 0$ and the operators $a_{\bm{k}-\bm{Q}}$ and $b_{\bm{k}\pm}$ are linearly combined. The mixing of $b_{\bm{k}\pm}$ and $b^{\dagger}_{-\bm{k}\pm}$ is treated using the Bogoliubov transformation, as will be shown in detail later. It should be noted that the ground state does not include the $b_{\bm{k}0}$ state, indicating that the $E_{g0}({\bm Q})$ mode is not Raman-active.
We consider the Raman operator ${\cal R}_{\bm d}$ associated with the intradimer interaction, which can be written in the reciprocal lattice space as $$\begin{array}{l}
{\cal R}_{\bm d}
=
\displaystyle
F_{\bm d}
(\hat{\bm{E}_{\rm in}} \!\! \cdot \! \hat{\bm{d}})
(\hat{\bm{E}_{\rm sc}} \!\! \cdot \! \hat{\bm{d}})
\!\!
\sum_{\bm{k}}
\!\!
\left( \!\!
- \frac{3}{4} % - \mu
s_{\bm{k}}^{\dagger} s_{\bm{k}}
+ \sum_{\alpha} \!
\frac{1}{4}% + \alpha h %- \mu
t_{\bm{k}\alpha}^{\dagger}t_{\bm{k}\alpha}
\!\!
\right)
\\
=
\displaystyle
F_{\bm d}
(\hat{\bm{E}_{\rm in}} \!\! \cdot \! \hat{\bm{d}})
(\hat{\bm{E}_{\rm sc}} \!\! \cdot \! \hat{\bm{d}})
\! \left[ \!
\left( \! \frac{1}{4}-u^{2} \! \right) \! \bar{a}^{2}
\! + \! uv\bar{a}(b_{\bm{Q}+}+b_{\bm{Q}+}^{\dagger})
\right.
\\
\displaystyle
+
\left.
\sum_{{\bm k}}
\left\{
\!\! \left( \! \frac{1}{4} - v^{2} \! \right) \!\! b_{\bm{k}+}^{\dagger}b_{\bm{k}+}
+ \frac{1}{4} b_{\bm{k}-}^{\dagger}b_{\bm{k}-}
+ \frac{1}{4} b_{\bm{k}0}^{\dagger}b_{\bm{k}0}
\right\}
\! \right]
\ ,
\end{array}
\label{Rintra}$$ where ${\bm d}$ indicates the position vector between two spins forming a dimer, which is almost parallel to the $(201)$ direction, and $F_{\bm d}$ originates from $F_{i,j}$. Here, we used the facts that the operator $a_{\bm{k}}$ can be treated as a uniformly condensed mean-field parameter $\bar{a}\delta_{\bm{k},0}$ (ref. ) and $b_{\bm{-Q}+}^{\dagger} = b_{\bm{Q}+}^{\dagger}$ at the magnetic $\Gamma$ point. The term $uv\bar{a}(b_{\bm{Q}+} + b_{\bm{Q}+}^{\dagger})$ in eq. (\[Rintra\]) gives the momentum selection rule of one-magnon Raman scattering. The parameter $uv$ indicates the appearance of one-magnon Raman scattering only above $H_{\rm c}$ and the Raman intensity proportional to $M_{xy}^{2}$ (see eq. (5) of ref. ), which is consistent with the observation, as shown in Fig. \[RamanParms\](b). The terms $\bar{a}^{2}$ and $b_{{\bm k}\alpha}^{\dagger}b_{{\bm k}\alpha}$ in eq. (\[Rintra\]) do not give one-magnon Raman scattering.
The quadratic terms of the magnetic Hamiltonian ${\cal H}_{\pm}$ in ref. can be diagonalized using the $\alpha_{\bm k}^{\pm}$ bosonic operators which annihilate the $E_{g\pm}({\bm k})$ modes, respectively. The ground state in the magnon BEC phase is the vacuum state for the $\alpha_{\bm k}^{\pm}$ operators. Using the bosonic commutation relations of $\alpha_{\bm k}^{\pm}$, we obtain the inverse Bogoliubov transformation as $$\left( \!\!
\begin{array}{c}
b_{\bm k-} \\
b_{\bm k+} \\
b_{-{\bm k}-}^{\dagger} \\
b_{-{\bm k}+}^{\dagger}
\end{array}
\!\! \right)
=
\left(
\!\!
\begin{array}{cccc}
\! u_{-{\bm k}-}^{-*} &
\!\! u_{-{\bm k}-}^{+*} &
\!\!\!-v_{{\bm k}-}^{-} &
\!\!-v_{{\bm k}-}^{+} \\
\! u_{-{\bm k}+}^{-*} &
\!\! u_{-{\bm k}+}^{+*} &
\!\!\!-v_{{\bm k}+}^{-} &
\!\!-v_{{\bm k}+}^{+} \\
\!-v_{-{\bm k}-}^{-*} &
\!\!-v_{-{\bm k}-}^{+*} &
\!\!\! u_{{\bm k}-}^{-} &
\!\! u_{{\bm k}-}^{+} \\
\!-v_{-{\bm k}+}^{-*} &
\!\!-v_{-{\bm k}+}^{+*} &
\!\!\! u_{{\bm k}+}^{-} &
\!\! u_{{\bm k}+}^{+} \\
\end{array}
\!\!
\right)
\!\!\!
\left(
\!\!
\begin{array}{c}
\alpha_{\bm k}^{-}\\
\alpha_{\bm k}^{+}\\
\alpha_{-{\bm k}}^{-\dagger}\\
\alpha_{-{\bm k}}^{+\dagger}
\end{array}
\!\!
\right)
\ .
\label{InvTransformation}$$
![Changes of spin polarizations for the amplitude and phase modes, corresponding to the ionic-vibration patterns in phonons. The solid and dashed lines indicate the dimers at the corner and center of the chemical unit cell denoted by parallelograms, respectively. Only the Cu$^{2+}$ sites are shown with the symbols of inversion centers and screw axes. The direction of the transverse magnetization $x$ in the magnon BEC phase is shown by a bold arrow with $M_{xy}$.[@Tanaka2001] The direction $y$ is perpendicular to $x$ and $b$.[]{data-label="Struct"}](LatStruct.eps){width="45.00000%"}
The one-magnon term in eq. (\[Rintra\]) can be rewritten as $$b_{\bm Q+} + b_{{\bm Q}+}^{\dagger}
\!=\!
(u_{Q+}^{-}-v_{Q+}^{-})^{\!*}\!\alpha_{\bm Q}^{-}
+ (u_{Q+}^{+}-v_{Q+}^{+})^{\!*}\!\alpha_{\bm Q}^{+}
+ {\rm H. c.} \ ,$$ indicating that both the $E_{g+}({\bm Q})$ and $E_{g-}({\bm Q})$ modes are symmetry-allowed. These modes can be described as the mixed amplitude and phase modes which are related to the spin correlation functions along the $x$ and $y$ directions in Fig. \[Struct\], respectively. The amplitude mode changes the amplitudes of $M_{xy}$ without changing their directions whereas the phase mode is described as uniform rotations of $M_{xy}$. Both of these modes have $A_{\rm g}$ symmetry, i.e., these are intrinsically Raman-active because the continuous [*rotational*]{} symmetry is broken above $H_{\rm c}$. This is one of the most distinguishing characteristics of the magnetic excitations in the magnon BEC phase. In systems of density waves,[@Gruner1994] the Goldstone mode (phase mode), which corresponds to the continuous [*translational*]{} operation, is IR-active.
Let us show that the phase mode does not give the finite Raman intensity. When the phase mode is the Goldstone mode, i.e., for $E_{g+}({\bm Q}) = 0$, ${\cal H}_{\pm}$ can be diagonalized by $$\left(
\!\!\!
\begin{array}{cccc}
\epsilon_{{\bm Q}+} &-\Delta_{{\bm Q}+} & \epsilon_{{\bm Q}\pm} &-\Delta_{{\bm Q}\pm} \\
\Delta_{{\bm Q}+} &-\epsilon_{{\bm Q}+} & \Delta_{{\bm Q}\pm} &-\epsilon_{{\bm Q}\pm} \\
\epsilon_{{\bm Q}\pm} &-\Delta_{{\bm Q}\pm} & \epsilon_{{\bm Q}-} &-\Delta_{{\bm Q}-} \\
\Delta_{{\bm Q}\pm} &-\epsilon_{{\bm Q}\pm} & \Delta_{{\bm Q}-} &-\epsilon_{{\bm Q}-}
\end{array}
\!\!\!
\right)
\!\!\!
\left( \!\!
\begin{array}{c}
u_{{\bm Q}+}^{+}\\
v_{{\bm Q}+}^{+}\\
u_{{\bm Q}-}^{+}\\
v_{{\bm Q}-}^{+}\\
\end{array}
\!\!\! \right) = 0,
\label{HamiltonianMatrix}$$ where the definitions of $\epsilon_{{\bm Q}\beta}$ and $\Delta_{{\bm Q}\beta}$ ($\beta = +, -, \pm$) are given in ref. . We can reduce eq. (\[HamiltonianMatrix\]) to the form of $$\left(
\!\!\!
\begin{array}{cc}
\epsilon_{{\bm Q}+} + \Delta_{{\bm Q}+} & \epsilon_{{\bm Q}\pm} + \Delta_{{\bm Q}\pm} \\
\epsilon_{{\bm Q}\pm} + \Delta_{{\bm Q}\pm} & \epsilon_{{\bm Q}-} + \Delta_{{\bm Q}-}
\end{array}
\!\!\!
\right)
\!\!\!
\left( \!\!
\begin{array}{c}
u_{{\bm Q}+}^{+} -
v_{{\bm Q}+}^{+}\\
u_{{\bm Q}-}^{+} -
v_{{\bm Q}-}^{+}
\end{array}
\!\!\! \right) = 0.
\label{HamiltonianMatrixSmall}$$ Because the matrix in eq. (\[HamiltonianMatrixSmall\]) is invertible, we find that $u_{{\bm Q}+}^{+} - v_{{\bm Q}+}^{+} = 0$, i.e., the Raman intensity for the phase mode is zero although this mode is symmetry-allowed. This result indicates that only the spin correlation function along $x$ is detectable with a factor $M_{xy}^{2}$ by the first-order Raman scattering.
In case of TlCuCl$_{3}$, the anisotropic exchange interaction gives a small magnetic gap $E_{g+}({\bm Q}) \approx 1.7$ cm$^{-1}$.[@Glazkov2004] In this case, the $E_{g+}({\bm Q})$ mode is the mixed amplitude and phase modes and it gives a finite one-magnon Raman intensity. In our measurements, however, we could not detect it because of the strong direct scattering at 0 cm$^{-1}$. It is worthwhile to consider the Raman scattering from the $E_{g+}({\bm Q})$ because this mode is thermally populated at 1.9 K. The transition to the ground state is the one-magnon anti-Stokes Raman scattering, which cannot be detected in our measurements, as stated above. The transition to the $E_{g-}({\bm Q})$ mode is obtained from the terms $\alpha_{\bm k}^{-\dagger}\alpha_{\bm k}^{+}$ in Raman tensor. Substituting $\alpha_{\bm k}^{\pm}$ and $\alpha_{\bm k}^{\pm\dagger}$ in eq. (\[InvTransformation\]) for $b_{\bm k\pm}$ and $b_{\bm k\pm}^{\dagger}$ in eq. (\[Rintra\]), ${\cal R}_{\bm d}$ contains the terms $$\begin{array}{l}
\displaystyle
\sum_{\bm k}
\left\{
\!\frac{1}{4}\!\left(
u_{-{\bm k}+}^{+*}\!u_{-{\bm k}+}^{-}\!\!+\!
v_{-{\bm k}+}^{+*}\!v_{-{\bm k}+}^{-}\!\!+\!
u_{-{\bm k}-}^{+*}\!u_{-{\bm k}-}^{-}\!\!+\!
v_{-{\bm k}-}^{+*}\!v_{-{\bm k}-}^{-}
\right)
\right.
\\
\displaystyle
\ \ \ \ \ \ \
\biggl.
\!-v^{2}\!\left(
u_{-{\bm k}+}^{+*}\!u_{-{\bm k}+}^{-}\!\!+\!
v_{-{\bm k}+}^{+*}\!v_{-{\bm k}+}^{-}
\right)
\!\biggr\}
\alpha_{\bm k}^{-\dagger}\alpha_{\bm k}^{+}
\ .
\end{array}$$ This result indicates that the transition from the $E_{g+}({\bm Q})$ state to the $E_{g-}({\bm Q})$ one may be detected as a part of the two-magnon Raman band and its intensity is expected to have no drastic change, at least below 10 T, because $v^{2}$ below 10 T is very small.[@Matsumoto2004; @PhysRevLett.89.077203] Actually, the profile and intensity of the two-magnon Raman band at 9 T are almost similar to those at 0 T, as shown in Fig. \[Polarization\].
When we consider one-magnon Raman scattering caused by the interdimer interaction, we can substitute the expectation value for one of the spin operators in eq. (\[FleuryLoudon\]): $${\cal R} =
\sum_{i,j} F_{i,j}
(\hat{\bm{E}_{\rm in}} \cdot \hat{\bm{r}_{ij}})
(\hat{\bm{E}_{\rm sc}} \cdot \hat{\bm{r}_{ij}})
\bm{S}_{i} \cdot \langle \bm{S}_{j} \rangle \ .
\label{FleuryLoudonMF}$$ The Raman tensor from the interdimer interaction also contains the terms $(b_{\bm{Q}+} + b_{\bm{Q}+}^{\dagger})$, suggesting that the one-magnon Raman scattering from the $E_{g-}(\bm{Q})$ mode can be detected above $H_{\rm c}$. Because $\langle S \rangle = M_{xy}$, one can expect that the Raman intensity is also proportional to $M_{xy}^{2}$. The precise analytic form of the Raman tensor coming from the interdimer interaction has been established in our recent letter in case of the pressure-induced magnon BEC phase at zero magnetic field. [@Matsumoto2008] The magnetic field-induced magnon BEC case will be published elsewhere. [@MatsumotoFuture]
We point out that our theory for the appearance of the one-magnon Raman scattering is applicable to the pressure-induced magnon BEC phase transition in TlCuCl$_{3}$. [@Matsumoto2004; @Goto2004; @Oosawa2003] In case of pressure-induced magnon BEC phase transition, the pure amplitude mode, which is the longitudinal spin-wave mode coupled only with the spin correlation function along the $x$ direction in Fig. \[Struct\], is expected to be observed.[@Matsumoto2007; @Matsumoto2008]
The two-magnon Raman band is also interesting as well as the appearance of the one-magnon Raman peak which is the main purpose of this paper. In case of two-magnon Raman scattering, both ${\cal R}_{\bm d}$ and the Raman tensors generated from the interdimer interactions, which create the magnon pair with the zero total momentum, play an essential role. At present, it was difficult to calculate the lineshape of the two-magnon Raman band because the values of $F_{i,j}$, which the two-magnon Raman spectrum is sensitive to, cannot be obtained directly.
Conclusion
==========
We have assigned the origin of the Raman peak appearing above $H_{\rm c}$ in TlCuCl$_{3}$ to one-magnon Raman scattering, which comes from the exchange magnon Raman process. This is based on (1) the Lorentzian lineshape of the peak, (2) its Raman shift tracing $E_{g-}({\bm Q})$, (3) its polarization characteristics, i.e. this one-magnon Raman scattering is $A_{\rm g}$-symmetric as well as the second-order magnetic Raman scattering, and (4) the observation that the peak’s Raman intensity is proportional to $M_{xy}^{2}$. Using the bond-operator representation, we calculated the Raman intensity to clarify the Raman selection rule of one-magnon Raman scattering in the exchange magnon Raman process. The intensity of the one-magnon Raman scattering is related to the spin correlation function along the direction of $M_{xy}$, i.e., the $x$ direction in Fig. 4. And therefore the $E_{g\pm}({\bm Q})$ modes with the finite excitation energies are $A_{\rm g}$-symmetric and Raman-active. In the isotropic limit, the Goldstone mode for $E_{g+}({\bm Q}) = 0$, which is related to the spin correlation function along the $y$ direction, is $A_{\rm g}$-symmetric but has no Raman intensity. The $E_{g0}({\bm Q})$ mode and the magnetic excitation at the chemical $\Gamma$ point are Raman-inactive. The change of the ground and excited states through the magnon BEC phase transition can be detected via the appearance of a new one-magnon Raman peak from the $E_{g-}({\bm Q})$ mode in the magnon BEC phase.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Meng-Li Shih$^{12}$\
[[email protected]]{}
- |
Shih-Yang Su$^{1}$\
[[email protected]]{}
- |
Johannes Kopf$^3$\
[[email protected]]{}
- |
Jia-Bin Huang$^1$\
[[email protected]]{}
- '$^{1}$Virginia Tech'
- '$^{2}$National Tsing Hua University'
- '$^{3}$Facebook'
- '<https://shihmengli.github.io/3D-Photo-Inpainting>'
bibliography:
- 'main.bib'
title: '3D Photography using Context-aware Layered Depth Inpainting'
---
This project is supported in part by NSF (\#1755785) and MOST-108-2634-F-007-006 and MOST-109-2634-F-007-016.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the sample of long Gamma-ray bursts (GRBs) detected by [*Swift*]{}-BAT before June 2007, we measure the cumulative distribution of the peak photon fluxes ($\log N$–$\log P$) of the [*Swift*]{} bursts. Compared with the [BATSE]{} sample, we find that the two distributions are consistent after correcting the band pass difference, suggesting that the two instruments are sampling the same population of bursts. We also compare the $\log N$–$\log P$ distributions for sub-samples of the [*Swift*]{} bursts, and find evidence for a deficit (99.75% confident) of dark bursts without optical counterparts at high peak flux levels, suggesting different redshift or $\gamma$-ray luminosity distributions for these bursts. The consistency between the $\log N$–$\log P$ distributions for the optically detected bursts with and without redshift measurements indicates that the current sample of the [*Swift*]{} bursts with redshift measurements, although selected heterogeneously, represents a fair sample of the non-dark bursts. We calculate the luminosity functions of this sample in two redshift bins ($z<1$ and $z\ge1$), and find a broken power-law is needed to fit the low redshift bin, where $dN/dL \propto L^{-1.27\pm0.06}$ for high luminosities ($L_{peak} > 5\times10^{48}{\hbox{erg~s$^{-1}$}}$) and $dN/dL \propto L^{-2.3\pm0.3}$ at for low luminosities, confirming the results of several studies for a population of low luminosity GRBs.'
author:
- Xinyu Dai
title: 'Intensity Distribution and Luminosity Function of the *Swift* Gamma-Ray Bursts'
---
Introduction
============
The cumulative distribution of source intensities ($\log N$–$\log S$) is a useful tool for studying the source populations, especially when the redshifts of the sources are not measured. In the Gamma-ray burst (GRB) field, since the peak photon flux of the GRB is directly related to the detection threshold, the cumulative distribution of peak photon flux ($\log N$–$\log P$) is used in many studies. A number of these studies have made use of the [BATSE]{} sample of more than 2,000 GRBs. This sample was first used to demonstrate the cosmological nature of the bursts (e.g., Fenimore et al. 1993; Pendleton et al. 1996), and then used to constrain the GRB populations (e.g., Kommers et al. 2000; Lin et al. 2004; Guetta et al. 2005; Dai & Zhang 2005), though many assumed that GRB rate follows the star formation rate in the analyses. Since the launch of [*Swift*]{} (Gehrels et al. 2004), the sample of the [*Swift*]{}-BAT bursts has reached to 237 in June 2007 (Sakamoto et al. 2008). With the sample size, it is now possible to compare the [*Swift*]{} and [BATSE]{} samples to test whether the two instruments sample the same population of bursts. It is possible that they are different since the two instruments have different band passes and sensitivities. For example, Band (2006) showed that [*Swift*]{} is more sensitive to soft bursts. Moreover, there is still a large number of bursts without redshift measurements, e.g., dark bursts, and the $\log N$–$\log P$ distribution provides a means of studying their source population besides using pseudo-redshifts derived from spectral or timing properties (e.g., Norris 2002).
The fraction of the [*Swift*]{} bursts with redshift measurements has increased significantly compared to the [BATSE]{} bursts. Although selected heterogeneously, it is tempting to measure the luminosity function of the [*Swift*]{} bursts using this sample (e.g., Liang et al. 2007). Besides issues with redshift selection effects, the [*Swift*]{} trigger efficiency has not been well studied, which presents an additional difficulty. We show that by studying the $\log N$–$\log P$ distributions of the [*Swift*]{} and [BATSE]{} bursts, we can justify the usage of the heterogeneous redshift sample and set detection thresholds for measuring luminosity functions. We present the luminosity functions using the heterogeneous redshift sample, where we adopt a cosmology of $H_0 = 70~\rm{km~s^{-1}~Mpc^{-1}}$, $\Omega_{\rm m} = 0.3$, and $\Omega_{\Lambda}= 0.7$.
The [*Swift*]{} Burst Sample
============================
We use the [*Swift*]{}-BAT GRB catalog published in Sakomoto et al. (2008). The sample consists of 237 [*Swift*]{}bursts detected before 2007 June 16. All the bursts are triggered by the BAT instrument on board [*Swift*]{}. The [*Swift*]{}-BAT catalog contains a number of basic properties of the bursts such as the burst duration, spectral index, and peak photon flux in several bands. There are 229 bursts with peak photon flux estimates in the 15–150 keV band, with a minimum value of 0.23 . Of these 229 bursts, 210 bursts can be identified as long bursts and 15 as short bursts, where we divide the sample at $T_{90} = 2$ s. In this paper, we focus on these 210 long bursts. To compare with the [BATSE]{} burst sample, we use the long [BATSE]{} burst sample from Kommers et al. (2000), which consists of 2176 long GRBs from both online and off-line searches.
The $\log N$–$\log P$ Distribution
==================================
We plot the $\log N$–$\log P$ distributions of the [*Swift*]{} and [BATSE]{} bursts in Figure \[fig:np\]. We use the $>50$% coded field of view of $\Omega_{{\emph{Swift}}}=1.4$ str for the BAT and a flight time of $T_{{\emph{Swift}}}=2.49$ yr until 2007 June 16 to calculate the burst rate for the [*Swift*]{} sample. For the [BATSE]{} sample, we adopt the values from Kommers et al. (2000) with $T_{{BATSE}} = 1.33\times10^{8}$ s and a mean field of view of $\Omega_{{BATSE}} = 0.67*4\pi$ str. We also correct for the band pass differences between the BAT (15–150 keV) and [BATSE]{} (50–300 keV) instruments, where we use the spectral fits provided by Sakamoto et al. (2008) for the [*Swift*]{} sample. Using the simple power-law fits to the BAT spectra, which fit well for most [*Swift*]{} bursts (Sakamoto et al. 2008), we find a mean relation, $N_{50-300} = 0.4*N_{15-150}$, which we use to correct the photon flux for the [BATSE]{} bursts. Figure \[fig:np\] shows that the two distributions are quite consistent, and only small discrepancies with $\sim1\sigma$ significance exist at high photon flux levels at $f_{peak} > 8~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$. We perform a Kolmogorov-Smirnov (K-S) test for the two samples, and find that the null probability that the two samples are drawn from the same distribution is 69% (76% for the sub-samples with $f_{peak} > 8~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$). This result indicates that the BAT and [BATSE]{} instruments sample the same population of GRBs, even though they have different band passes. We also convert the BAT photon flux to the [BATSE]{} band individually for each burst using the simple power-law fits of Sakamoto et al. (2008) to reduce the uncertainty introduced by scatter of the spectral indices around their mean, and compare the two distributions in the [BATSE]{} band. We again find that the two distributions are not significantly different. For some bursts, the spectra are better fit by a cut-off power-law (Sakamoto et al. 2008), which introduces additional uncertainties in the flux conversion; however they are generally negligible compared to the uncertainty introduced by scatter of the photon indices. We note that the detection limits of the two samples also match each other after the band pass correction.
Comparisons between [*Swift*]{} Sub-Samples
-------------------------------------------
Next, we compare the $\log N$–$\log P$ distributions for sub-samples of the [*Swift*]{} bursts. First, we compare the first half and the second half of the [*Swift*]{} bursts separated by time, finding that the two samples are consistent, suggesting that there is no significant degradation in the BAT detector. Next, we compare the distributions for bursts with or without redshift measurements or optical detections (Figure \[fig:snp\]). There are 68 bursts with redshift measurements and optical afterglow detections, 48 with optical afterglow detections but without redshift measurements, and 94 without optical afterglow detections. The last category of bursts with no optical afterglow detections is the dark bursts, though another definition uses an optical to X-ray threshold ratio to define this sample (Jakobsson et al. 2004). In this paper, we use the simple definition of no optical detection for dark bursts. In most cases, we find no significant difference between the sub-samples based on the K-S test probabilities (45% between samples with and without redshift measurements, 34% between samples with redshift measurements and dark bursts, and 24% between samples with redshift measurements and without redshift measurements but with optical detections).
However, Figure \[fig:snp\] shows that there is an indication for a deficit of dark bursts at high peak flux levels compared to the bursts with optical afterglow detections. We test it by including a flux filter and select bursts with $f_{peak} > 5~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$, and find that the K-S probability that the dark bursts and bursts with optical detections are drawn from the same population is 4%. It seems that the deficit is more pronounced in the flux level at $f_{peak} > 10~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$. Since the number of bursts detected in this regime is too small and that the K-S test is no longer applicable, we use a simple Poisson argument instead. The null model is that the dark bursts follow the same distribution as the optically detected bursts, which predicts that the number of the dark bursts detected at $f_{peak} > 10~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$ is 8.2$\pm$0.8. This model prediction and the associated error-bar are obtained by aligning the normalizations of the two distributions at low flux levels. Since we only detect one dark burst with $f_{peak} > 10~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$, the Poisson probability of detecting no more than one burst is 0.0025 for an expected value of 8.2. If we use the lower end of the model prediction (7.4), the corresponding probability is 0.0051. We argue that there is evidence that the dark bursts do not follow the $\log N$–$\log P$ distribution of the optically detected bursts. It can be either a deficit of dark bursts at high peak flux levels or an over abundance of dark bursts at low flux levels. Future analysis including the whole [*Swift*]{} sample is needed to confirm this results.
Since the ground-based optical follow-up observations of GRBs can be potentially biased by several reasons, such as the scheduling issues or unfavorable burst sky positions, we also test by limiting the dark burst sample using those bursts observed by UVOT but without UVOT or other optical detections (UVOT dark bursts). This provides a more uniform selection because UVOT routinely observes the GRB fields after the burst triggers. In addition, this excludes some of the bursts with unfavorable sky positions to follow-up in the optical bands, such as those located close to the Sun. We also limit the sample by excluding bursts with large Galactic extinctions, $A_{V, Gal} > 1.5$ mag, where the dark burst fraction is significantly larger compared to that for the total population. However, since the UVOT flux limit is generally shallower than those of the ground-based observations using mid-sized telescopes, we still classify the dark bursts as those without optical detections. We find the $\log N$–$\log P$ distribution of the UVOT dark burst sample follows that of the general dark burst sample (or no optical counterparts sample, Figure \[fig:snp\]). This is because only a small fraction of bursts have unfavorable burst positions that cannot be observed by the ground-based telescopes. In addition, a number of optical telescopes are used to observe GRB afterglows world wide, which collectively reduces the scheduling problems such as the whether conditions in individual optical sites. Because of the shallowness of the UVOT flux limit, it is still possible that both of the dark burst (no optical counterpart) samples contain a fraction of normal, non-dark bursts, because they are not observed promptly by the mid or large size ground-based telescopes. However, this population should not contribute to the difference between the $\log N$–$\log P$ distributions between dark and normal bursts.
Luminosity Function
===================
The bursts with redshift measurements account for 59% of the bursts with optical afterglow detections. Figure \[fig:snp\] shows that the $\log N$–$\log P$ distributions between the optically detected bursts with and without redshift measurements are quite similar. This suggests that the current redshift sample, although obtained heterogeneously, represents a fair sample of the bursts with optical afterglow detections. We compute the luminosity function (LF) of this sample using the $1/V_{max}$ method (Schmidt 1968; Avni & Bahcall 1980), where we include an additional factor of $1/(1+z)$ in the differential volume element to account for the cosmological time dilation effect. We calculate the k-corrections using the power-law fit to the BAT spectra from Sakomoto et al. (2008), and use a detection limit of $f_{peak} = 0.25~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$, the smallest value in the redshift sample (GRB 060218), to calculate the $V_{max}$ values. Since the $\log N$–$\log P$ distribution of the [*Swift*]{} sample matches that of the [BATSE]{} sample when approaching the detection limits (Figure \[fig:np\]), we use the trigger efficiency analysis of Kommers et al. (2000) to estimate the efficiency of the [*Swift*]{} triggers, and then assign weights to the [*Swift*]{} bursts with a maximum weight of 2. Using a maximum weight is a conservative approach to avoid huge corrections at low peak flux levels, where the uncertainties of the efficiency analysis could be large.
We show the total LF and those in redshift bins of $z<1$ and $z\ge1$ in Figure \[fig:lf\], where we choose a bin size of 0.5 dex. We find that the LF for the lower redshift bin is consistent with the total LF, and they cannot be fit well by a single power-law model with $\chi^2/dof =$ 4.6 and 2.8, respectively, for the total and $z<1$ LFs. We add another power-law component at the low luminosity end, and fit the $z<1$ LF. We find $dN/dL \propto L^{-2.3\pm0.3}$ at the low luminosity end and $dN/dL \propto L^{-1.27\pm0.06}$ at the high luminosity end ($L_{peak} > 5\times10^{48}{\hbox{erg~s$^{-1}$}}$) with $\chi^2/dof =$1.3. The complete LF for the $z<1$ bin is in Equation \[eqn:lf\], $$\frac{dN}{dL} = (1.7\pm0.2) {\rm h_{70}^3 Gpc^{-3} Yr^{-1}} \left[ \left(\frac{L}{5\times10^{48}{\rm erg~s^{-1}}}\right)^{-2.3\pm0.3}+\left(\frac{L}{5\times10^{48}{\rm erg~s^{-1}}}\right)^{-1.27\pm0.06} \right]\label{eqn:lf}$$ Comparing the LFs in the two redshift bins, the two low luminosity data points for the $z\ge1$ LF at $L_{peak} = 10^{50-51}{\hbox{erg~s$^{-1}$}}$ are below the $z<1$ LF. Since the star formation rate drops significantly below $z=1$ (e.g., Hopkins & Beacom 2006), if the GRB rate follows the star formation rate, we expect that the LF for the $z\ge1$ bin should have a higher normalization than that for the $z<1$ LF. It is possible that the two low luminosity data points of the $z\ge1$ LF are incomplete because they are close to the [*Swift*]{} flux limit (e.g., Kistler et al. 2008), even though we have partially modeled the [*Swift*]{} trigger incompleteness by using the [BATSE]{} trigger efficiency model (Kommers et al. 2000). If we neglect these two low luminosity data points, the remaining two data points of the $z\ge1$ LF are not lower than the $z<1$ LF. Considering the uncertainties, the results do not present a significant challenge to the hypothesis that the GRB rate follows the star formation rate. We note that the LFs do not include the contribution from the dark bursts. We need to multiply by a factor of 1.8 to include them; however, this may not be accurate since the dark bursts may have a different redshift or $\gamma$-ray luminosity distribution (§3.1). In addition, if the majority of the dark bursts are at high redshifts (§5), the normalization of the $z\ge1$ LF will be significantly increased compared to the $z<1$ LF, which can raise the GRB rate to be consistent with the star formation rate, or even evolving faster as suggested by some recent studies (e.g., Salvaterra & Chinarini 2007; Kistler et al. 2008; Salvaterra et al. 2009).
Discussion
==========
We compare the $\log N$–$\log P$ distributions for the [*Swift*]{} bursts and [BATSE]{} bursts, and find that they are consistent after correcting the band pass differences as suggested by the K-S test results. This shows that the two instruments sample the same population of bursts, although they have different band passes. We can also compare the normalizations of the distributions directly, since the field of view and the observing time of the two samples are measured, and we find that they are consistent. Indirect comparison between the [*Swift*]{} and [BATSE]{} samples has been performed in previous studies (Salvaterra & Chincarini 2007; Virgili et al. 2009), where the two samples are found to be consistent with same theoretical models. Given that [*Swift*]{} is more sensitive than [BATSE]{} for the soft bursts (Band 2006), it is puzzling that the two distributions are consistent. Currently, we are comparing the on-board triggered [*Swift*]{} sample with the [BATSE]{} off-line search sample. It is possible that the difference will become significant when comparing both the off-line search samples.
We also compare the $\log N$–$\log P$ distributions for the sub-samples of the [*Swift*]{} bursts with or without optical afterglow detections or redshift measurements. We find them to be broadly consistent. However, at the high peak flux regime ($f_{peak} > 5~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$), we find an indication that there is a deficit (96% confident) of “dark bursts” (bursts without optical counterparts). This deficit is more evident at $f_{peak} > 10~{\hbox{photon~cm$^{-2}$~s$^{-1}$}}$, where we find the effect is significant at the 99.75% confidence level.
A comparison between the $\log N$–$\log P$ distributions for dark bursts and optically bright bursts can place important constraints on the origin of the dark bursts. There are several models proposed for the dark bursts. They can be bursts with intrinsic normal optical-to-$\gamma$-ray ratios but with large intrinsic or foreground optical extinction (e.g., Taylor et al. 1998; Djorgovski et al. 2001; Fynbo et al. 2001), have intrinsic faint optical afterglows, i.e., low optical-to-$\gamma$-ray ratios, (e.g., Groot et al. 1998a; Frail et al. 1999), are normal bursts but exist in extremely high redshifts where the Lyman break lands in the optical (e.g., Groot et al. 1998b; Fynbo et al. 2001), or a combination of several origins mentioned above. A different $\gamma$-ray luminosity distribution or redshift distribution is needed to interpret the difference in the $\log N$–$\log P$ distributions. Since the $\gamma$-ray/hard X-ray flux is not sensitive to the ISM absorption, the hypotheses with optical extinction or intrinsic optical faintness predict that the $\log N$–$\log P$ distributions should have a similar shape between dark and optically detected bursts. For the high redshift scenario, it is possible for the dark bursts to have a different shape in the $\log N$–$\log P$ distribution from the normal, non-dark bursts. Therefore, the deficit of the dark bursts at high flux levels suggests that a significant fraction of dark burst is from extreme high redshifts. In addition, it is possible that the bulk of the dark bursts are selected due to their low optical luminosity, and the deficit of dark bursts at high peak photon fluxes merely reflects the detection threshold of optical observations under the assumption of a constant optical-to-$\gamma$-ray ratio.
We measure the luminosity function of the [*Swift*]{} bursts using the sample with redshift measurements. We find the luminosity function can be fit by a broken power-law with $dN/dL \propto L^{-2.3\pm0.3}$ at the low luminosity end, and $dN/dL \propto L^{-1.27\pm0.06}$ at the high luminosity end. We compare this result with previous measurements (Schmidt 2001; Norris 2002; Stern et al. 2002; Firmani et al. 2004; Guetta et al. 2005; Liang et al. 2007; Salvaterra & Chincarini 2007) and find a best match with the result from Liang et al. (2007). The requirement for an additional component at the low luminosity end confirms the existence of the population of low luminosity GRBs claimed by several studies (e.g., Cobb et al. 2006; Piran et al. 2006; Soderberg et al. 2006; Guetta & Della Valle 2007; Liang et al. 2007; Virgili et al. 2009). Although we model it as a power-law, it could be the tail of a Gaussian component. At the high luminosity range, the measured slope $dN/dL \propto L^{-1.27\pm0.06}$ is close to the prediction of the “quasi-universal Gaussian jet” ($dN/dL \propto L^{-1}$, Zhang et al. 2004). At the very high luminosity end ($L_{peak} > 10^{51}{\hbox{erg~s$^{-1}$}}$), the “quasi-universal Gaussian jet” predicts $dN/dL \propto L^{-2}$ (Lloyd-Ronning et al. 2004; Dai & Zhang 2004), which cannot be tested with the current data.
Unlike many of the previous studies, we do not make any assumption on the GRB rate, since there are recent studies suggesting that the GRB rate does not follow the star formation rate (e.g., Stanek et al. 2006). Instead, we measure the average GRB luminosity function in a large redshift bin. We compare the LFs in two luminosity bins. The $z\ge1$ LF shows a drop at two low luminosity data points. This result can be affected by the uncertainties in the [*Swift*]{} trigger efficiency close to the [*Swift*]{}detection limit. If we neglect these two data points, the LFs for the two redshift bins are consistent. However, the large measurement uncertainties in the $z<1$ LF make it difficult to test whether the GRB rate follows the star formation rate.
We thank B. Zhang, C. S. Kochanek, R. Salvaterra, and the anonymous referee for helpful discussion.
Avni, Y., & Bahcall, J. N. 1980, , 235, 694
Band, D. L. 2006, , 644, 378
Cobb, B. E., Bailyn, C. D., et al. 2006, , 645, L113
Dai, X., & Zhang, B. 2005, , 621, 875
Djorgovski, S. G., Frail, D. A., Kulkarni, S. R., et al. 2001, , 562, 654
Fenimore, E. E., et al. 1993, , 366, 40
Firmani, C., Avila-Reese, V., et al. 2004, , 611, 1033
Frail, D. A., et al. 1999, , 525, L81
Fynbo, J. U., et al. 2001, , 369, 373
Gehrels, N., et al. 2004, , 611, 1005
Groot, P. J., et al. 1998a, , 502, L123
Groot, P. J., et al. 1998b, , 493, L27
Guetta, D., Granot, J., & Begelman, M. C. 2005, , 622, 482
Guetta, D., & Della Valle, M. 2007, , 657, L73
Hopkins, A. M., & Beacom, J. F. 2006, , 651, 142
Jakobsson, P., Hjorth, J., Fynbo, J. P. U., et al. 2004, , 617, L21
Kommers, J. M., [et al. ]{}2000, ApJ, 533, 696
Kistler, M. D., Y[ü]{}ksel, H., et al. 2008, , 673, L119
Liang, E., Zhang, B., Virgili, F., & Dai, Z. G. 2007, , 662, 1111
Lin, J. R., Zhang, S. N., & Li, T. P. 2004, ApJ, 605, 819
Lloyd-Ronning, N. M., Dai, X., & Zhang, B. 2004, ApJ, 601, 371
Norris, J. P. 2002, ApJ, 579, 386
Pendleton, G. N., et al. 1996, , 464, 606
Pian, E., et al. 2006, , 442, 1011
Sakamoto, T., et al. 2008, , 175, 179
Salvaterra, R., & Chincarini, G. 2007, , 656, L49
Salvaterra, R., Guidorzi, C., et al. 2009, , in press, arXiv:0805.4104
Schmidt, M. 1968, , 151, 393
Schmidt, M. 2001, ApJ, 552, 36
Soderberg, A. M., et al. 2006, , 442, 1014
Stanek, K. Z., et al. 2006, Acta Astronomica, 56, 333
Stern, B. E., Tikhomirova, Y., & Svensson, R. 2002, ApJ, 573, 75
Taylor, G. B., Frail, D. A., Kulkarni, S. R., et al. 1998, , 502, L115
Virgili, F. J., Liang, E.-W., & Zhang, B. 2009, , 392, 91
Zhang, B., Dai, X., et al. 2004, ApJ, 601, L119
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A construction of a thermal neutron testing detector with a thin \[ZnS(Ag)+$^6$LiF\] scintillator is described. Results of an investigation of sources of the detector pulse origin and the pulse features in a ground and underground conditions are presented. Measurements of the scintillator own background, registration efficiency and a neutron flux at different objects of the BNO INR RAS were performed. The results are compared with the ones measured by the $^3$He proportional counter.'
author:
- |
V.V. Alekseenko$^{a}$, I.R. Barabanov$^{a}$, R.A. Etezov$^{a}$, Yu.M. Gavrilyuk$^{a}$,\
A.M. Gangapshev$^{a}$, A.M. Gezhaev$^{a}$, V.V. Kazalov$^{a}$, A.Kh. Khokonov$^{b}$,\
V.V. Kuzminov$^{a}$, S.I. Panasenko$^{c}$, S.S. Ratkevich$^{c}$
title: |
Characteristics of a thermal neutrons scintillation detector with\
the \[ZnS(Ag)+$^6$LiF\] at different conditions of measurements
---
\[intro\]Introduction
=====================
Large area scintillation detector of the thermal neutrons on a base of a thin \[ZnS(Ag)+ $^6$LiF\] scintillator has been put into operation at the Baksan Neutrino Observatory of the INR RAS at the last time [@p1; @p2]. The neutron registration is occurred as a result of $^6$Li($n,\alpha$)$^3$H+4786 keV reaction. The cross section is 945 b [@p3]. A kinetic energy of the reaction products ($E_\alpha=2051$ keV, $E_{\rm H}=2735$ keV) converted at a light flash by the scintillator. The scintillator components enter into the composition as an alloy. Spectrometric characteristics of such detectors are known not enough. The aim of the work was to investigate and refinement these parameters.
Detector construction
=====================
A schematic view of the test detector and its electronics are sown on the Fig. \[fig:schematic\_view\].
{width="4.25"}
The detector is assembled in a rectangular case with $30\times30\times50$ cm$^3$ sizes made from galvanized iron of the 0.7 mm thickness. The covers of the top and bottom are detachable. A dividing plate with a central 150 mm diameter hole aimed to install photomultiplier FEU-173 (PMT) mounted in the middle of the case. A charge sensitive preamplifier (CSP) with $\sim100$ $\mu$s self-discharge time is installed on the wall inside of the upper part of the case. Pulses from the PMT’s anode resistor (4.8 M$\Omega$) go to the CSP input and further at the input of the LAn-10M5 digital oscilloscope card and are recorded into the personal computer (PC) memory. A sampling frequency was equal to 6.25 MHz. The flat flexible plate with the \[ZnS(Ag)+$^6$LiF\] scintillator is placed on the floor of the bottom section. It contains of the white sheet of plastic film ($207\times295$ mm$^2$ ) with a sticky side covered by grains of the scintillator with an average thickness of $\sim0.1$ mm [@p2]. The sheet is laminated by the lavsan films. A surface of the bottom section covered with a reflecting mylar film for a better light collection. Densities of the pure ZnS(Ag) and $^6$LiF are equal to 4.09 g$\cdot$cm$^{-3}$ [@p4] and $\sim2.64$ g$\cdot$cm$^{-3}$ [@p7]. A density of the mixture in the proportion $1:3$ is $\sim3$ g$\cdot$cm$^{-3}$.
Results of measurements
=======================
The two types of working pulses shown on the Fig. \[fig:pulses\] ($a,b$) were observed in a measurement with the detector at the ground laboratory. The first type ($a$) has a front time duration $\tau_f = 16-25$ $\mu$s which corresponds to the own de-excitation time of the fine-grained scintillator $\tau_s= 8-10$ $\mu$s ($\tau_f = 3 \tau_s$). The CSP integrates an input current and a pulse maximum is shapes at the point where the current charging velocity turns equal to the preamplifier own charge decay velocity. The second type (*b*) of a pulse has a front time duration $\tau_f = 0.8 \tau_s$.
![\[fig:pulses\] The two types of pulses with test the detector. “*a*” - the long pulse rise time, “*b*” - the short pulse rise time. ](fig2.eps){width="2.1in"}
Its shape is similar to the noise pulses which occur after an irradiation of the scintillator and PMT by an external light during of an adjustment work. Intensity and amplitudes of the noise pulses in the range of interest are fall down to zero after some hours. A residual part of the type (*b*) pulses is not a noise because of its intensity is proportional to the intensity of cosmic rays and decreases with moving of the detector on a deeper underground. It is known that a part of a photomultiplier own noise pulses could be created by the charged particles from an outer radioactive background and cosmic rays and also by charged particles appeared in decays of radioactive residual impurity isotopes in the PMT construction materials [@p6]. A separate investigation was done for an alignment of a type ($b$) appearing mechanism. The two possibilities were examined. The first one is a direct generation of the primary electrons from the photocathode or dynode by the cosmic rays. The second one is an appearing of a photoelectron from the photocathode as a result of absorption of a Cerenkov radiation created by a charged particle in a glass of the entrance window of the PMT. For the last case, the cosmic rays coming from directions around the vertical can generate in the window a Cerenkov radiation directed outside. This light will return into the PMT after the reflection from the bottom sell walls. The PMT entrance window was covered by the black paper to check this assumption. The pulses with a short front only remained in the spectrum. Their intensity is changed with the cosmic rays intensity at a moving of the detector from the second floor of a laboratory building to the ground one. A conclusion that the short pulses are generated by cosmic rays in the photocathode or dynode has been made.
A set of measurements was made with the detector at the ground and underground conditions of the BNO INR RAS. They are:
1. Deep Underground Low Background Laboratory (DULB-4900). An unshielded place located at the underground hall under the mountain thickness equal to the 4900 m of the water equivalent (m w.e.). First spectrum (*a*) was measured without any shielding materials and the second one (*b*) was measured with the $0.1\times100\times100$ cm$^3$ Cd sheet (absorber) placed under the detector;
2. DULB-4900 low background compartment with the walls made of 25 cm polyethylene $+0.1$ cm Cd$+15$ cm Pb. A spectrum (*a*) was measured;
3. “2” + thermal neutron source;
4. Low background underground laboratory “KAPRIZ” at the 1000 m w.e. The spectra (*a*) and (*b*) were measured;
5. Low background underground laboratory “NIKA” at the 660 m w.e. The spectra (*a*) and (*b*) were measured;
6. Underground hall of the “CARPET-2” set-up muon detector at the 5 m w.e. The spectra (*a*) and (b) were measured;
7. Ground building “ELLING” of the “CARPET-2” set-up. The spectra (*a*) and (*b*) were measured;
8. The open soil. The spectra (*a*) and (*b*) were measured;
9. The fourth floor (room 404) of the four-storey laboratory building (LAB). The spectrum (*a*) was measured;
10. The river side of the second floor (room 204) of the four-storey laboratory building. The spectra (a) and (b) were measured;
11. The valley side of the second floor (room 211) of the four-storey laboratory building. The spectrum (*a*) was measured;
12. The ground floor of the four-storey laboratory building. The spectrum (*a*) was measured.
The objects are placed in a list in an accordance of a thickness of the cosmic rays absorber above the installation in a sequence of 4900 m w.e.$\rightarrow$ 0 m w.e.$\rightarrow$1.3 m w.e. A brief description of the objects is given in [@p11].
A spectrum “a” was preliminary measured in the point (10) at 63.12 h to obtain a general imagine about statistical characteristics of pulses. The spectrum is shown on Fig. \[fig:Pulse\_amplitude\_spectra\],
![\[fig:Pulse\_amplitude\_spectra\] Pulse amplitude spectra of the detector in the point (10) collected at 63.12 h): “$a$” – a total spectrum, “$b$” – a spectrum of pulses with a short front, “$c$” – a spectrum of pulses with a long front. ](fig3_v1.eps){width="2.65in"}
spectrum ($a$). A distribution of the pulse front durations (0.2-0.8 region of the total amplitude normalized for 1 hour) is shown on the Fig. \[fig:Front\_duration\].
![\[fig:Front\_duration\] Front duration distribution of pulses from the spectrum ($a$) from the Fig. \[fig:Pulse\_amplitude\_spectra\] normalized for 1 hour. ](fig4_v1b.eps){width="2.5in"}
The peaks in the regions of 0.5-1.6 $\mu$s and 3.2-13 $\mu$s are visible. The last one is correspond to the pulses from the scintillator. A selection of pulses correspondent to this marked regions allows one to separate the spectrum ($a$) on the Fig. \[fig:Pulse\_amplitude\_spectra\] at fast \[spectrum ($b$)\] and slow \[spectrum ($c$)\] components. It is seen that the fast pulses contribute a main part at the low amplitudes. The more detailed analysis of the fast pulse shapes shows that the pulses have different decay shapes. Such difference could be explained by small variable contributions of the scintillation light to the PMT Cherenkov light. A proportion of the two components depends on a number of particles in an event and a quantity of tracks crossing the two light generators. A thin scintillator has a low sensitivity to the cosmic rays and electrons. Amplitudes of pulses of the events with the cosmic rays particles crossing of the scintillator are small and lie below the registration threshold.
Significant number of extraneous pulses could presence in the data sets measured with a low count rate in the underground conditions. Pickups from a periodic switching of industrial equipment and furnishings are the main component of such pulses. They have shapes differ considerably from the useful ones. The extraneous pulses could be excluded from the spectra by using of the discrimination on a base of a pulse shape analysis.
The detector own background should be known for the low count rate measurements. A measurement at the point (2) was done for this purpose. The spectrum normalized at 100 h is shown on the Fig. \[fig:background\_pulse\_amplitude\] \[spectrum ($a$)\].
![\[fig:background\_pulse\_amplitude\] The detector own background pulse amplitude spectrum “$a$” and the spectrum of a neutron calibration source “$b$”. ](fig5_v1.eps){width="2.5in"}
The corresponding distribution of the front durations is shown on the Fig. \[fig:Front\_duration\_distributions\] (curve “$a$”).
![\[fig:Front\_duration\_distributions\] Front duration distributions of pulses from the spectrum ($a$) (curve “$a$”) and spectrum ($b$) (curve “$b$”) from the Fig. \[fig:background\_pulse\_amplitude\] normalized for 1 hour. ](fig6_v1.eps){width="2.5in"}
It is seen from a comparison of the spectra on the Fig. \[fig:background\_pulse\_amplitude\] that the background spectrum has a longer energy extension than the neutron one. Its front distribution is shifted to the shorter times simultaneously. A conclusion could be done that the detector background created by strongly ionized particles with the energy larger than the energy of the products of the neutron reactions. The $\alpha$-particles from decays of the $^{232}$Th and $^{238}$U natural long lived radioactive isotopes and its daughters contained as micro impurities in the scintillator could be possible sources of the background. The $^{210}$Po ($T_{1/2}=138.4$ d, $E_\alpha=5.3$ MeV) generated in a decay chain of the $^{210}$Pb ($\beta^-$ - decay, $T_{1/2}=21.8$ y) could be another background source. The last isotope was born charged in the air in the radon decay chain and deposited at the scintillator charged surfaces during its preparation. This source could be essentially suppressed by using of the special protecting arrangements against the radon and its daughters penetration into the gas environment of the scintillator production area. The $\alpha$-particles which were born outside of the scintillator plate will be absorbed in the covering lavsan film and will not give any noticeable effect.
Estimated path lengths of $\alpha$-particles with the energies of 2051 keV, 4800 keV and 5300 keV in the scintillator with a density of 3.0 g$\cdot$cm$^{-3}$ are equal to 10.6 $\mu$m, 28.6 $\mu$m and 32.1 $\mu$m [@p8]. The one of the 2735 keV triton is 56.3 $\mu$m. The energy spectrum extension will be longer due to the energy difference if the main source of the background is not the neutron reactions products but the $\alpha$-particles from the $^{210}$Po decays. A difference could increase additionally if a specific light yield of the ZnS(Ag) decreased at lower particle energy with an ionization density rise as it occurs for the other nonorganic scintillators [@p4]. The light output for the two reaction products with the 4800 keV sum energy will be less than the one for the $\alpha$-particle with the same energy. The lesser front duration of the background pulses could be explained by a location of the $\alpha$-source on a surface of the scintillator grains. Trajectories of $\alpha$-particles will be directed into the grains in this case and the main part of energy will be released inside the ones. It is mentioned in the [@p4] that a scintillator based on the dispersed ZnS(Ag) has a long afterglow. It could be explained possibly by a relative increasing of long-lived excitation traps number located on grain surfaces. A particle absorbed in the grain surface layer will give a longer deexcitation time in this case. Vertexes of neutron reactions are distributed uniformly in a $^6$LiF component volume. The reaction products should exit from this materials and fall into the surface layer of the ZnS(Ag) grain to produce a scintillation. An ionization density increases with a particle energy decreasing in accordance with the energy loss dependence [@NIST]. A considerable part of the energy will be released in the surface layers of the two adjacent grains giving longer pulse front duration in comparison with a particle absorbed inside the grain. The features of the detector background distributions mentioned above could be an indirect evidence of a background source grains surface location.
Count rates of the detector at 1 hour for the data integrated above the third channel of the spectra “*b*” and “*c*” Fig. \[fig:Pulse\_amplitude\_spectra\] are presented in the Table 1 for the all objects cited in the list.
[l c c r c r c r c]{} & & &\
No.& & & & & & Thermal neutron flux density,\
& & & & Short pulse rise time & & Long short pulse rise time & & (s$^{-1}\cdot$cm$^{-2}$)\
1a & & DULB-4900 & & 0.13$\pm$0.03 & & 21.3$\pm$0.4 & & (2.6$\pm$0.4)$\times10^{-5}$\
1b & & -//- + (Cd) & & 0.09$\pm$0.03 & & 18.7$\pm$0.5 & & (1.2$\pm$0.4)$\times10^{-5}$\
2 & & DULB-4900 & & 0.10$\pm$0.02 & & **16.4$\pm$0.3** & & $\leq3.8\times10^{-7}$ (90% C.L.)\
& & (low background) & & & & & & ($^3$He prop. counter)\
3 & & -//- + (n-source) & & 0.2$\pm$0.1 & & 80$\pm$2 & & (3.4$\pm$0.4)$\times10^{-4}$\
4a & & KAPRIZ & & 0.13$\pm$0.03 & & 16.7$\pm$0.3 & & $\leq5.9\times10^{-6}$ (90% C.L.)\
4b & & -//- + (Cd) & & 0.13$\pm$0.03 & & 16.5$\pm$0.4 & & $\leq5.9\times10^{-6}$ (90% C.L.)\
5a & & NIKA & & 0.19$\pm$0.04 & & 17.8$\pm$0.4 & & (7.5$\pm$3.1)$\times10^{-6}$\
5b & & -//- + (Cd) & & 0.14$\pm$0.03 & & 16.9$\pm$0.3 & & (3.4$\pm$3.1)$\times10^{-6}$\
6a & & $\mu$-detector & & 8.4$\pm$0.5 & & 64$\pm$1 & & (2.8$\pm$0.3)$\times10^{-4}$\
6b & & -//- + (Cd) & & 8.2$\pm$0.4 & & 43$\pm$1 & & (1.4$\pm$0.2)$\times10^{-4}$\
7a & & ELLING & & 23$\pm$1 & & 1415$\pm$8 & & (7.5$\pm$0.6)$\times10^{-3}$\
7b & & -//- + (Cd) & & 22$\pm$3 & & 730$\pm$16 & & (3.8$\pm$0.4)$\times10^{-3}$\
8a & & Open soil & & 28$\pm$3 & & 1704$\pm$31 & & (9.0$\pm$0.8)$\times10^{-3}$\
8b & & -//- + (Cd) & & 27$\pm$3 & & 702$\pm$14 & & (3.7$\pm$0.3)$\times10^{-3}$\
9 & & LAB, 404 & & 22$\pm$1 & & 1439$\pm$9 & & (7.6$\pm$0.6)$\times10^{-3}$\
10a& & LAB, 204 & & 16$\pm$2 & & 866$\pm$14 & & (4.5$\pm$0.4)$\times10^{-3}$\
10b& & -//- + (Cd) & & 16$\pm$1 & & 482$\pm$6 & & (2.5$\pm$0.2)$\times10^{-3}$\
11 & & LAB, 211 & & 19.1$\pm$0.9 & & 672$\pm$5 & & (3.5$\pm$0.3)$\times10^{-3}$\
12 & & LAB, ground & & 6.5$\pm$0.5 & & 240$\pm$3 & & (1.2$\pm$0.1)$\times10^{-3}$\
Values of a thermal neutron flux $F$ in the all examined points could be obtained from the neutron count rates. A flux $F$ of particles according to a definition is a ratio of a particles number $\Delta N$ falling on a given surface at a $\Delta t$ time interval to this interval: $F=\Delta N/\Delta t$. The measured count rate **n** connected with $F$ by the $ \textbf{n}=\varepsilon \times F$ relation where $\varepsilon$ is a neutron registration efficiency of the scintillation plate. A neutron flux density parameter $\phi$ is used usually to a lightening of a comparison of results obtained with the different geometry detectors. The $\phi$ according to a definition is a ratio of a particles flux $dF_S$ penetrated into the volume of an elementary sphere to the area of it’s central cross section $dS: \phi=dF_S/dS$. The $dF_S$ value could be determined using a specific neutron flux falling on an elementary square of the scintillator as $dF_S=4\times F/S$ where $S$ is an area of the scintillator plane and the coefficient 4 is equal to the ratio of the sphere surface area to the area of the sphere cross section. The $\phi$ is equal to $\phi=4n/(2S\varepsilon)$ as a result.
It is seems impossible to calculate $\varepsilon$-value because of uncertainties of a composition and a structure of a scintillator layer. This value was obtained experimentally from a comparison of the $n_1$ count rate of the described detector and $n_{1(2)}$ count rate of the detector with an additional similar passive scintillator plate (2) put under the active plate (1). The plate (2) was light intercepted. The detector in the measurements was shielded by the 1 mm cadmium foil for the thermal neutron coming on from the upper hemisphere to shape a single-sided neutron flux. A count rate of the standard detector is equal to $n_1=\varepsilon_1 \times F$ and the one for the modified detector is $n_{1(2)}=\varepsilon_{1(2)} \times (F-\varepsilon_2 \times F)$ where $\varepsilon_1$ is a plate (1) absorption efficiency of a thermal neutron flux $F$, $\varepsilon_{1(2)}$ is the absorption efficiency of thermal neutron flux plate (1) in the presence of isolation from the light of the plate (2) absorb some of the flux $F$ with efficiency $\varepsilon_2$. A value of efficiency is an integral characteristic of a process of absorption for the neutrons coming at different angles and depends on a path passed by a neutron in the scintillator. An angular distribution of neutrons after passing the one scintillator layer is pulled in the direction normal because of emptive absorption of particles coming at odd angles. An absorption will be lower for the passed neutrons and $\varepsilon_{1(2)}$ will be lower than $\varepsilon_1$ ($\varepsilon_{1(2)} \leq \varepsilon_1$). The count rates are specified by the expressions $n_2=\varepsilon_2 \times F$ and $n_{2(1)}=\varepsilon_2(1) \times (F-\varepsilon_1 \times F)$ in a case when the plate (2) is used as the active one. Five unknown variables $\varepsilon_1$, $\varepsilon_{1(2)}$, $\varepsilon_2$, $\varepsilon_{2(1)}$ and $F$ are in four obtained equations. One needs to measure additionally a total count rate $n_{[1+2]}=n_{[2+1]}$ for the case when the both plate used in the active mode to determine precisely all five values. Such measurement is possible if the plates emit scintillation light into the both hemispheres in the detector having two PMTs. The task could be solved with the reviewed detector if the plate adsorbs neutron not strongly. It could be taken that $\varepsilon_{1(2)}\approx \varepsilon_1$ and $\varepsilon_{2(1)} \approx \varepsilon_2$ in this case. Simple conversions give the expressions $\varepsilon_1=[n_1-n_{1(2)}]/n_2$ and $\varepsilon_2=[n_1-n_{1(2)}]/n_1$.
The measurements were done in the point (9) of the objects list. The result $\varepsilon_1= \varepsilon_2=0.14 \pm 0.01_{\rm{stat.}} \pm 0.02_{\rm{syst.}} $ was obtained. A difference between $\varepsilon_1$ and $\varepsilon_{1(2)}$ is calculated for the two homogeneous plate having $\varepsilon=0.14$ at a registration of isotropic neutron flux coming from the one side uses as a systematic uncertainty.
A background count rate measured in the point (2) was subtracted from the data in a process of a determination of a thermal neutron flux density in the each other point. The obtained $\phi$-values are resulted in the last column of the Table \[T1\]. A limit of a thermal neutron flux density in the point (2) was obtained on a base of measurements with a CH-04 neutron proportional counter with $^3$He [@p13].\
**Discussion of results**
A specific own background of the scintillator plate was found to be to (2.69$\pm$0.05) h$^{-1}\times$(100 cm)$^{-2}$ in the point (2). This value is comparable with a value of a surface $\alpha$-activity of the commercial copper and steel samples which is $\sim(0.5-1.0)$ h$^{-1}\times$(100 cm)$^{-2}$ [@p9]. A surface $\alpha$-activity of the silicon semiconductor samples could reach $\sim 0.1$ h$^{-1}\times$(100 cm)$^{-2}$. It is seems possible to prepare a scintillator plate with similar surface $\alpha$-activity by using of a specially selected low background \[ZnS(Ag)+$^6$LiF\] material and a clean technology for the plate preparation. A sensitivity of such scintillator detector (SD) for the thermal neutrons would be comparable with a sensitivity of the $^3$He-proportional counter. A present ratio of the sensitivities is $\sim 16$ as it is seen from a comparison of the data for the (2) and (4) points of the Table \[T1\]. A difference of the front time duration distribution of background and neutron pulses could be useful for the plate sensitivity improving also. The obtained results are in a good agreement with the one measured by the $^3$He-proportional counter in the Ref. [@p10].
A value of the own background defines a sensitivity of the neutron measurement in the deep underground conditions as it seen from the Table \[T1\]. A count rate of neutrons in the DULB-4900 (point (1a)) is equal to (21.32-16.4)/6.11=(0.8$\pm$0.08) h$^{-1}\times$(100 cm)$^{-2}$ with the effect to background ratio equal to $\sim 0.3$. The neutrons are born in the rock mainly due to the ($\alpha,n$)-reactions with the light elements. Walls of the “KAPRIZ” laboratory are covered with a 30 cm layer of a low background concrete made on a base of a dunite crushed rock. The concrete is decreased considerably a neutron flux from the rock. (The dunite concrete was used in the construction of the “NIKA” laboratory too). A comparison of neutron fluxes in the “DULB-4900” and “KAPRIZ” measured with CH-04 proportional counter shows that the concrete decreases a neutron flux at $\sim 5.2$ times. One can estimate an expected neutron effect in the “KAPRIZ” using of this coefficient and SD count rate in the point (1a) as (0.15$\pm$0.02) h$^{-1}\times$(100 cm)$^{-2}$. A calculation with the data from the Table \[T1\] gives a value (0.05$\pm$0.08) h$^{-1}\times$(100 cm)$^{-2}$ which is not contradicts to the estimated one. The last value was used to obtain a limit for the neutron flux density in the points (1a and 1b) at 90% C.L. as $2 \times (0.05+1.64 \times 0.08)/3600/100/0.17 =5.9 \times 10^{-6}$ cm$^{-2}$c${-1}$.
The own SD background enters a minor deposit into the detector count rate in the measurements at the ground and shallow underground points where reactions of cosmic rays with element nuclei of the environment are the main source of the neutron.
A comparison of the SD count rates with and without a cadmium absorber shows that the absorber decreased a neutron flux at $\sim 1.9$ times under a ceiling and at $\sim 2.4$ times on the open place. Thus, a ratio of neutron fluxes from the soil and from the atmosphere on the open place is equal to $\sim 1.4$.
A SD count rate of the short front pulses in the underground conditions does not depend practically on a value of the external $\gamma$-quanta background level as it seen from the Table \[T1\]. This noise component could be born directly in the PMT by densely ionizing particles. The pulses could appear at the photocathode as a result of direct generation of electrons by $\alpha$-particles of the window surface $\alpha$-activity. A count rate of such pulses connected with the muon intensity and increased proportionally with its growth.\
**Conclusions**
Measurements of some working characteristics of a thermal neutron scintillator detector prepared on a base of a thin 216$\times$304 mm$^2$ plate of a fine-grained \[ZnS(Ag)+$^6$LiF\] - scintillator made at the BNO INR RAS. An analysis of shapes of charge pulses recorded by means of a digital oscilloscope has showed that background pulses with a short front created by the cosmic rays directly in the photomultiplier are presented among the working pulses in a low energy region of spectra. Measurements of the detector thermal neutron count rate were made at the ground and underground objects of the BNO INR RAS at different shielding for the cosmic rays. The inherent background of the detector created by $\alpha$-particles from decays of inner radioactive admixture with a surface $\alpha$-activity at level of (2.69$\pm$0.05) h$^{-1}\times$(100 cm$^{-2}$)$^{-1}$ was measured. A ratio of an effect to a background was equal to $\sim 0.3$ at the underground conditions. The neutron pulses have a shorter front than the background ones. It could be used for a discrimination of the background pulses at $\sim 2$ times with an insignificant rejection of the neutron events at the low neutron flux measurements.
**Acknowledgement**
The work was carried out in part with the financial support of the Federal Objective Program of the Ministry of Education and Science of the Russian Federation “Research and Development in the 2007-2013 years on the Priority Directions of the Scientific and Technological Complex of the Russia” under the contract No. 16.518.11.7072 and the “Russian Foundation for Basic Research” under the grant No. 14-22-03059.
We are thankful to Yu.V. Stenkin for providing the samples of the scintillator and numerous useful critical comments.
[99]{}
V.V. Alekseenko, Yu.M. Gavrilyuk, D.M. Gromushkin, at al., “Correlation of Variations in the Thermal Neutron Flux from the Earth’s Crust with the Moon’s Phases and with Seismic Activity” // Physics of the Solid Earth, 2009, Vol. 45, No. 8, p.709.
V.V. Alekseenko, D.D. Dzhappuev, V.A. Kozyarivsky, at al., “Analysis of Variations in the Thermal Neutron Flux at an Altitude of 1700 m above Sea Level” // Bulletin of the Russian Academy of Sciences: Physics. 2007, Vol. 71, No. 7, p. 1047
A.I. Abramov, Yu.A. Kazansky and E.S. Matusevich, [Bases of the Experimental Methods of Nuclear Physics]{}, Moscow: Energoatomizdat, 1985.
Moscow: Atomizdat 1977.
. [$^6$LiF/ZnS:Ag Phosphor/Scintillator Data Sheet 39- ND - iss1.doc]{}.
M.E. Globus, B.V. Grinev, [Inorganic scintillators (new and traditional materials).]{} Kharkiv: Akta, 2001.
O.F. Nemets, Yu.V. Ghofman, [Handbook of Nuclear Physics.]{} Kiev: Naukova Dumka. 1975.
M.J. Berger, J.S. Coursey, M.A. Zucker and J. Chang, [Stopping-Power and Range Tables for Electrons, Protons, and Helium Ions.]{} [NIST, Physical Measurement Laboratory,]{} [http://www.nist.gov/pml/data/star/]{}
E.L. Kovalchuk, V.V. Kuzminov, A.A. Pomansky, [“Surface alfa activity of different materials”.]{} Proc. of the Int. Conf. “The Natural Radiation Environment III”, Houston, Texas, April 23-28, 1978, V.1, 1980, P.673.
V.V. Alekseenko, I.R.Barabanov, R.A. Etezov, et al., [“Results of measurements of an environment neutron background at BNO INR RAS objects with the helium proportional counter”]{} // arXiv: 1510.05109 \[physics.ins-det\].
V.V. Kuzminov., [“The Baksan Neutrino Observatory”.]{} // Eur. Phys. J. Plus 127 (2012) 113. doi: 10.1140/epjp/i2012-12113-0; http://www.inr.ac.ru/
Moscow, Energoatomizdat, 1995.
// Physics of Atomic Nuclei, V.73, No.9, 2010, P.1482.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Claudia de Rham,$^{a}$ Lavinia Heisenberg$^{a,b}$ and Raquel H. Ribeiro$^{a}$'
bibliography:
- 'references.bib'
title: Quantum Corrections in Massive Gravity
---
Introduction
============
Over the past century a considerable amount of effort has been devoted to understanding gravity. With the observational evidence of the recent accelerated expansion of the Universe [@Perlmutter:1997zf; @Riess:1998cb; @Tonry:2003zg], this quest has become more urgent because it might be an indication that we do not understand gravity at the largest distance scales. Another striking puzzle in cosmology is the cosmological constant problem [@Weinberg:1988cp], for which the discrepancy between field theory predictions and observations is of many orders of magnitude.
A similar disparity arises in the standard model of particle physics, the hierarchy problem. This is the problem of why the Higgs mass is so small relative to the Planck/unification scales. These hierarchies are puzzling as they do not seem to be protected without the help of new physics. [*Technically natural*]{} tunings on the other hand are less of an issue and are common within the standard model. For example, the electron mass, $m_{\rm e}$, is hierarchically smaller than the electroweak scale. In the limit $m_{\rm e}\to0$ there is an enhancement of the symmetry of the system, due to the recovery of chiral symmetry. The existence of this symmetry in the massless limit is enough to protect the electron mass from receiving large quantum corrections thanks to the *’t Hooft naturalness* argument [@tHooft:1979bh; @Dimopoulos:1979es]. Therefore quantum corrections will only give rise to a renormalization of the electron mass proportional to $m_{\rm e}$ itself, and thus the hierarchy between the electron mass and the electroweak scale is technically natural.
In the case of the cosmological constant, $\Lambda_{\rm CC}$, there is no symmetry recovered in the limit $\Lambda_{\rm CC}\to 0$. This is because the (anti-)de Sitter and Poincaré groups have the same number of generators. Its tuning to a small scale compared to the Higgs mass for instance is therefore unnatural in the ’t Hooft sense. In the context of General Relativity (GR) it has been very hard to address this problem, which has motivated the search for local modifications of gravity in the infra-red. One possibility is to introduce a technically natural small scale which could account for the late time acceleration of the Universe. This can be achieved by giving the mediator of gravitational interactions, the graviton, a small mass of order of the Hubble constant today.
However, such a modification usually comes with a number of pathologies [@ArkaniHamed:2002sp; @Creminelli:2005qk], most notably the presence of the so-called Boulware–Deser ghost, [@Boulware:1973my]. In recent years we have seen considerable progress in formulating a well-defined, non-linear theory of massive gravity, which has no propagating ghost degrees of freedom [@Nibbelink:2006sz; @Gabadadze:2009ja; @deRham:2010gu; @deRham:2012az]. Such a theory has been proposed in Refs. [@deRham:2010ik; @deRham:2010kj], which extends the Fierz–Pauli action [@Fierz:1939ix; @VanNieuwenhuizen:1973fi] and relies on a very specific structure of a 2-parameter family interaction potential. The proof of the absence of ghosts has now been generalized to a multitude of languages and formalisms, see for instance Refs. [@Hassan:2011hr; @deRham:2011rn; @Hassan:2011ea; @deRham:2011qq; @Mirbabayi:2011aa; @Golovnev:2011aa; @Hassan:2012qv; @Hinterbichler:2012cn; @Kluson:2012wf; @Deffayet:2012zc]. Despite these recent developments, the stability of massive gravity against quantum corrections remains an open question, especially when including interactions with other massive matter fields. This question is not only tied to the smallness of the graviton mass when taking quantum corrections into account; it also raises the issue of whether the quantum corrections are capable of detuning/destabilizing the structure of the potential of massive gravity. Since it is the special nature of this potential which prevents the presence of the Boulware–Deser ghost, a detuning would reintroduce the ghost at some energy scale.
It is therefore crucial to address the question of the quantum stability of massive gravity, and we shall take some first steps in this direction. We will also clarify the role played by the Vainshtein mechanism [@Vainshtein:1972sx; @Deffayet:2001uk; @Berezhiani:2013dw; @Berezhiani:2013dca] at the level of the quantum corrections, given its importance in enhancing the strongly coupled interactions and curing the classical discontinuity problem [@Kogan:2000uy; @Porrati:2000cp; @Dilkes:2001av; @Babichev:2009jt] which arises when taking the massless limit —this is known as the van Dam–Veltman–Zakharov (vDVZ) discontinuity [@vanDam:1970vg; @Zakharov:1970cc]. Even in cases where the vDVZ discontinuity seems absent at the classical level, it has been known to reappear at the quantum level, either through anomalies in the case of spin 3/2, [@Duff:2002sm] or directly in the loops for massive spin-2 fields on AdS, [@Dilkes:2001av]. This paper extends the findings of Ref. [@deRham:2012ew] which discussed the quantum corrections in a class of Galileon theories corresponding to massive gravity in the *decoupling limit*.[^1]
In this paper we study how and at what scale the specific structure of the graviton gets detuned, and compare our conclusions to previous results.[^2] We emphasize we are *not* addressing the old cosmological constant problem, since $\sqrt{-g}\,\Lambda_{\rm CC}$ is one of the allowed ghost-free potential terms for the graviton, and it is therefore harmless for our discussion.
For the quantum stability analysis, we focus on quantum corrections arising at one-loop only, and leave the extension to higher order loops to a further study, where the mixing between graviton and scalar propagators within loops can occur [@2loops]. In particular this implies that either gravitons or matter field are running in the loops, but not *both* simultaneously. Furthermore, by working in dimensional regularization, we discard any measure issues in the path integral related to field redefinitions which show up in power law divergences. We thus focus on logarithmic running results which are independent of this measure factor—in the language of field theory we concentrate on the runnings of the couplings.
Moreover, our aim is to study the stability of the graviton [*potential*]{} against quantum corrections, rather than the whole gravity action. As a result it is sufficient to address the diagrams for which the external graviton legs have zero momenta (, we focus on the IR limit of the runnings). This approach is complementary to the work developed by Buchbinder in [@Buchbinder:2012wb] who used the Schwinger–DeWitt expansion of the one-loop effective action. This method allows one to obtain the Seeley–DeWitt coefficients associated with the curvature invariants generated by quantum corrections (see also Ref. [@Buchbinder:2007xq]). Our approach differs in two ways. First, we introduce a covariant coupling to the matter sector and obtain the quantum corrections generated by matter loops. And second, we go beyond the minimal model investigated by Ref. [@Buchbinder:2012wb] and study the quantum corrections to the full potential. As a by product, we do not focus on the radiatively generated curvature terms since these would also arise in GR and would therefore not be exclusive of theories of massive gravity.
For convenience, when computing the one-loop effective action, we consider a background configuration for the metric which is spacetime independent. Furthermore, we work in Euclidean space, and use the mostly $+$ signature convention. Massive gravity on a Minkowski reference metric is thus mapped to massive gravity on a flat Euclidean reference metric $\delta_{ab}$ in Euclidean space. Finally, we use units for which $\hbar=1$.
The remainder of this paper is organized as follows. In section \[sec:dRGT\] we review the tree level ghost-free covariant non-linear theory of massive gravity and briefly discuss the quantum corrections in the decoupling limit. We then move on to the full theory and concentrate in section \[sec:MatterLoops\] on the one-loop contributions arising from the coupling to external matter fields. We show they only imply a running of the cosmological constant and of no other potential terms for the graviton.
In section \[sec:GravitonLoops\] we discuss the one-loop contributions from the gravitons themselves, and show that whilst these destabilize the special structure of the potential, this detuning is irrelevant below the Planck scale. We then push the analysis further in section \[sec:1PI\] and show that even if the background configuration is large, as should be the case for the Vainshtein mechanism to work [@Vainshtein:1972sx], this will redress the one-loop effective action in such a way that the detuning remains irrelevant below the Planck scale. We end by summarizing our results and presenting some open questions in section \[sec:conclusion\]. The appendices collect extra material: appendix \[sec:appendixDimReg\] states our conventions for the dimensional regularization scheme, whilst appendix \[app:matter\] presents all the details of the calculation involving matter loops only.
Ghost-free massive gravity {#sec:dRGT}
==========================
In this section we review the ghost-free interactions in the theory of massive gravity. We also summarize the findings of Ref. [@deRham:2012ew] where the quantum stability of the classical theory in the decoupling limit was studied. This will be our starting point to investigate the way quantum corrections affect the general structure of the potential.
The presence of a square-root in the ghost-free realization of massive gravity makes its expression much more natural in the vielbein language [@Nibbelink:2006sz; @Chamseddine:2011mu; @Hinterbichler:2012cn; @Deffayet:2012zc] (see also Refs. [@Gabadadze:2013ria; @Ondo:2013wka]). In the vielbein formalism, the ghost-free potential is polynomial and at most quartic in the vielbein fields. To make use of this natural formulation, we will work throughout this paper in a ‘symmetric-vielbein inspired language’ where the metric is given by \[vielbeinForm\] g\_[ab]{}=(|\_[ab]{}+)\^2 (|\_[ac]{}+)(|\_[db]{}+)\^[cd]{}, where $\bar g_{ab}=\bar\gamma\ab^2=\bar\gamma_{ac}\bar\gamma_{bd}\delta^{cd}$ is the background metric, and $h\ab$ plays the role of the fluctuations. We stress that the background metric $\bar g_{ab}$ need not be flat, even though the reference metric $f\ab$ will be taken to be flat throughout this study, $f\ab=\delta\ab$.[^3]
In this language, when working around a flat background metric, $\bar \gamma_{ab}=\delta_{ab}$, the normal fluctuations about flat space are expressed in terms of $h$ as g\_[ab]{}-\_[ab]{}=h\_[ab]{}+h\_[ac]{}h\_[bd]{}\^[cd]{}. \[eq:defmetric\] The conversion from $g_{ab}$ to $h_{ab}$ is a field redefinition that will contribute a measure term in the path integral. This generates power law divergent corrections to the action which, since we will work in dimensional regularization, can be ignored. This reflects the fact that the physics is independent of such field redefinitions, and only the logarithmic runnings are physically meaningful for the purposes of our study.
Ghost-free potential {#potentialdRGT}
--------------------
A two-parameter family of potentials for the graviton has been proposed in Refs. [@deRham:2010ik; @deRham:2010kj]. These potential terms were built in such a way so as to remove higher derivative terms in the fields which would otherwise induce a propagating ghost degree of freedom [@Boulware:1973my; @Deffayet:2005ys]. Consider a graviton of mass $m$ \[eq:full\] L\_[mGR]{}=-(R-U(g,H)) , where the overall minus sign arises after Wick rotation to Euclidean space.[^4] In this action $\mathcal U$ is the potential and the tensor $H\ab$ is constructed out of the metric $g\ab$ and the four fields $\Phi^a$ [@Siegel:1993sk; @ArkaniHamed:2002sp], H=g-\_[cd]{}\_a \^c \_b \^d. \[eq:defHmunu\] For the purpose of this section, we can work with a flat background metric $\bar g\ab=\delta\ab$.
One can split the $\Phi^a$’s into a helicity-0 and-1 modes. For the sake of this argument, it is sufficient to focus on the helicity-0 mode and set the helicity-1 to zero, so that $\Phi^a=(x^a-\delta^{ab}\partial_b\hat{\pi})$, where $\pi=\Lambda_3^3 \,\hat \pi$ is the helicity-0 mode, and we recall that $\Lambda_3^3=m^2 \mpl$. Then, in terms of the helicity-0 and -2 modes, the tensor $H\ab$ becomes H=2 + \_a\^[ c]{} \_[bc]{}+2-\^[cd]{}\_[ac]{}\_[bd]{} . \[eq:defH\] In this notation $\hat{\Pi}\ab \equiv \partial_a \partial_b \pi/\Lambda_3^3$, and $h\ab \equiv \mpl \hat h\ab $ is the canonically normalized helicity-2 mode. Indices are lowered and raised with respect to the flat background metric, $\delta\ab$.
The fields have been introduced so as to restore diffeomorphism invariance. We are now free to set a gauge for the metric or the fields, and can choose in particular the unitary gauge, where $\Phi^a(x)=x^a$. In that gauge, the tensor $H\mn$ is simply given by $H\ab=2 \hat{h}\ab +\hat{h}_a^{\ c} \hat{h}_{cb}$. In terms of these ‘vielbein-inspired’ perturbations, the ghost-free potential becomes polynomial in $h\ab$, =- R - \^2 m\^2 \_[n=2]{}\^4[\_n \_n\[H\]]{} . \[eq:Lvielbein\] In unitary gauge the potential above is fully defined by \[eq:U2h\] \_2\[h\] &=& \^[abcd]{}\^[a’b’]{}\_[ cd]{} h\_[aa’]{}h\_[bb’]{}= -2 $h\ab h_{cd}-h_{ac}h_{bd}$\^[ac]{}\^[bd]{}\
\[eq:U3h\] \_3\[h\] &=& \^[abcd]{}\^[a’b’c’]{}\_[ d]{} h\_[aa’]{}h\_[bb’]{}h\_[cc’]{}\
\[eq:U4h\] \_4\[h\] &=& \^[abcd]{}\^[a’b’c’d’]{} h\_[aa’]{}h\_[bb’]{}h\_[cc’]{}h\_[dd’]{}, where $\E^{abcd}$ represents the fully antisymmetric Levi-Cevita [*symbol*]{} (and not tensor, so in this language $\E^{abc}_{\ \ \ d}=\delta_{dd'}\E^{abcd'}$, for example, carries no information about the metric).
The first coefficient is fixed, $\tilde{\alpha}_2=1$, whereas the two others are free. They relate to the two free coefficients of Ref. [@deRham:2010kj], as $\tilde{\alpha}_3=-2(1+\alpha_3)$ and $\tilde{\alpha}_4=-2(\alpha_3+\alpha_4)-1$ (where $\alpha_3$ and $\alpha_4$ are respectively the coefficients of the potential $\mathcal{U}_3$ and $\mathcal{U}_4$ in that language).[^5] The absence of ghost-like pathologies is tied to the fact that, when expressed in terms of $\pi$ uniquely, – are total derivatives.
Quantum corrections in the decoupling limit
-------------------------------------------
Before moving on to computing the quantum corrections to the potential arising from matter loops, we first review the radiative corrections within the decoupling limit as derived in Ref. [@deRham:2012ew]. In this case the helicities-2 and -0 of the graviton decouple from each other and become accessible separately, thereby creating a framework benefiting the visibility of the most important physical properties of the theory. In this limit, the usual helicity-2 mode of gravity can be treated linearly while the helicity-0 mode still contains non-linear interactions. The decoupling limit can be written in the compact form [@deRham:2010ik; @deRham:2010kj] $$\mathcal{L}=h^{\mu\nu}\Ein^{\alpha\beta}_{\ \ \mu\nu} h_{\alpha\beta}-
h^{\mu\nu}\sum_{n=1}^3 \frac{a_{n}}{\Lambda^{3(n-1)}_3} X^{(n)}_{\mu\nu}\!\left(\Pi\right),
\label{lagr1}$$ where the Lichnerowicz operator, $\Ein^{\alpha\beta}_{\ \ \mu\nu}$, acts on the metric perturbations as follows \^\_[ ]{} h\_=-12 $\Box h\mn-2\p_\alpha \p_{(\mu}h^\alpha_{\nu)}+\p_\mu\p_\nu h-\delta\mn (\Box h-\p_\alpha\p_\beta h^{\alpha\beta})$. \[eq:Lichnerowicz\] Here, $h_{(ab)}\equiv\frac{1}{2}(h_{ab}+h_{ba})$, and one can always set $a_1=1$ while the two other dimensionless coefficients $a_{2,3}$ are related to the two free parameters $\tilde \alpha_{2,3}$. The three tensors $X_{\mu\nu}^{(n)}$ represent the interactions of order $n$ in $\pi$ through $\Pi_{\mu\nu}$ between the helicities-2 and -0 modes $$\begin{aligned}
X^{(1)}_{\mu\nu}\left(\Pi\right)&=&{\mathcal{E}_{\mu}}^{\alpha\rho\sigma}
{{\mathcal{E}_\nu}^{\beta}}_{\rho\sigma}\Pi_{\alpha\beta}\ , \quad \nonumber \\
X^{(2)}_{\mu\nu}\left(\Pi\right)&=&{\mathcal{E}_{\mu}}^{\alpha\rho\gamma}
{{\mathcal{E}_\nu}^{\beta\sigma}}_{\gamma}\Pi_{\alpha\beta}
\Pi_{\rho\sigma} \ ,\ \ \textrm{and} \nonumber \\
X^{(3)}_{\mu\nu}\left(\Pi\right)&=&{\mathcal{E}_{\mu}}^{\alpha\rho\gamma}
{{\mathcal{E}_\nu}^{\beta\sigma\delta}}\Pi_{\alpha\beta}
\Pi_{\rho\sigma}\Pi_{\gamma\delta}\ .
\label{Xs}\end{aligned}$$ The natural question that arises is whether the parameter $a_n$ and the energy scale $\Lambda_3$ are radiatively stable. Using the antisymmetric structure of the interactions , one can roughly follow the same non-renormalization argument of Galileon theories [@Nicolis:2004qq; @Luty:2003vm; @Nicolis:2008in] to show that $a_{2}$ and $a_3$ do not get renormalized within the decoupling limit of the theory.[^6] The key point is that any external particle attached to a diagram has at least two derivatives acting on it. This in turn implies that the operators generated are all of the form $(\p^2 \pi)^{n_1}(\p^2 h)^{n_2}$, $n_1, n_2 \in \mathbb{N}$, and so are not of the same class as the original operators. This means that $a_2$ and $a_3$ are not renormalized. Furthermore, the new operators that appear in the 1PI are suppressed by higher powers of derivatives.
To be more precise, any external particle contracted with a field with two derivatives in a vertex contributes to a two-derivatives operator acting on this external particle—this is the trivial case. On the other hand, if we contract the external particles with fields without derivatives we could in principle generate operators with fewer derivatives. But now the antisymmetric structure of the interactions plays a crucial role. Take for instance the interaction $V \supseteq h^{\mu\nu} \mathcal{E}_{\mu}^{\;\;\alpha\rho\gamma} \mathcal{E}_{\nu\;\;\;\;\gamma}^{\;\;\beta\sigma}\Pi_{\alpha\beta}\Pi_{\rho\sigma}$, and contract an external helicity-2 particle with momentum $p_\mu$ with the helicity-2 field coming without derivatives in this vertex; the other two $\pi$-particles run in the loop with momenta $k_\mu$ and $(p+k)_\mu$. The contribution of this vertex gives $$\begin{aligned}
\mathcal A
\propto \int \frac{\mathrm{d}^4 k}{(2\pi)^4} G_k \,
G_{k+p}\ f^{\mu\nu}\, \mathcal{E}_{\mu}^{\;\;\alpha\rho\gamma} \,
\mathcal{E}_{\nu\;\;\;\;\gamma}^{\;\;\beta\sigma}
\, \, k_\alpha \, k_\beta \, (p+k)_\rho \, (p+k)_\sigma \cdots \,,\end{aligned}$$ where $G_k= k^{-2}$ is the Feynman massless propagator, and $f^{\mu\nu}$ is the spin-2 polarization tensor. The ellipses denote the remaining terms of the diagram, which are irrelevant for our argument. The only non-vanishing contribution to the scattering amplitude will come in with at least two powers of the external helicity-2 momentum $p_\rho p_\sigma$ $$\begin{aligned}
\mathcal A
\propto f^{\mu\nu} \mathcal{E}_{\mu}^{\;\;\alpha\rho\gamma}
\mathcal{E}_{\nu\;\;\;\;\gamma}^{\;\;\beta\sigma}\, \, p_\rho p_\sigma \int\frac{\mathrm{d}^4k}{(2\pi)^4} G_k \, G_{k+p} \,k_\alpha \, k_\beta \cdots \,,\end{aligned}$$ which in coordinate space corresponds to two derivatives, as argued above. The same argument applies to any other vertex, such as $h^{\mu\nu}X_{\mu\nu}^{(3)}(\Pi)$.
This is the essence of the ‘non-renormalization theorem’ in the decoupling limit of massive gravity: there are no quantum corrections to the two parameters $a_{2}$ and $a_3$, nor to the scale $\Lambda_3$. Moreover, the kinetic term of the helicity-2 mode is radiatively stable. We refer the reader to Ref. [@deRham:2012ew] for more details.
Propagator in unitary gauge
---------------------------
In this paper our goal is to go beyond the non-renormalization argument in the decoupling limit reviewed above and investigate the quantum corrections in the full non-linear theory. We choose to work in the unitary gauge in which the fields vanish and $\Phi^a=x^\alpha\delta^a_{\ \alpha}$. $h\mn$ encodes all the five physical degrees of freedom if it is massive (the two helicity-$\pm2$, the two helicity-$\pm1$ and the helicity-$0$ modes), and only the two helicity-$\pm2$ modes if it is massless. The Feynman propagator for the massless graviton is given by \[GR\_propagator\] G\^[(massless)]{}\_[abcd]{}&=&h\_[ab]{}(x\_1) h\_[cd]{}(x\_2)=f\^[(0)]{}\_[abcd]{} , in which $x_{1,2}$ are the Euclidean space coordinates, and where the polarization structure is given by \[f\] f\^[(0)]{}\_[abcd]{}=\_[a(c]{}\_[bd)]{}-12 \_[cd]{}. Here $\delta_{a(c}\delta_{bd)}\equiv\frac{1}{2} \delta_{ac} \delta_{bd}+\frac{1}{2}\delta_{ad}\delta_{bc}.$ For the massive graviton, on the other hand, the corresponding Feynman propagator is given by \[MG\_propagator\] G\^[(massive)]{}\_[abcd]{}&=&h(x\_1) h\_[cd]{}(x\_2)=f\^[(m)]{}\_[abcd]{} , with the polarization structure f\^[(m)]{}\_[abcd]{}=$\tilde \delta_{a(c}\tilde \delta_{bd)}-\frac 13 \tilde \delta\ab \tilde \delta_{cd}$ =+. \[eq:fmassive\] Notice that the polarization of the massive graviton is no longer proportional to $f^{(0)}_{abcd}$. Consequently when we take the massless limit, $m\to0$, we do not recover the GR limit, which is at the origin of the vDVZ-discontinuity [@vanDam:1970vg; @Zakharov:1970cc], \_[m0]{}f\^[(m)]{}\_[abcd]{} f\^[(0)]{}\_[abcd]{} . At first sight this might be worrisome since on solar system and galactic scales gravity is in very good agreement with GR. Nevertheless, on these small scales, the effects of massive gravity can be cloaked by the Vainshtein mechanism [@Vainshtein:1972sx], where the crucial idea is to decouple the additional modes from the gravitational dynamics via nonlinear interactions of the helicity-0 graviton.
The success of the Vainshtein mechanism relies on derivative interactions, which cause the helicity-0 mode to decouple from matter on short distances, whilst having observational signatures on larger scales. In this paper we will study how the Vainshtein mechanism acts explicitly at the quantum level, and how the quantum corrections do not diverge in the limit when $m\rightarrow 0$, even though the propagator does.
In the next section we focus on the IR behaviour of the loop corrections. Starting with loops of matter, we will see that the peculiar structure in has no effect on the computation of the quantum corrected effective potential at one-loop. This is because at one-loop the matter field and the graviton cannot both be simultaneously propagating in the loops if we consider only the contributions to the graviton potential. As a result, the quantum corrections are equivalent to those in GR, as is shown explicitly in section \[sec:MatterLoops\]. Only once we start considering loops containing virtual gravitons will the different polarization and the appearance of the mass in the propagator have an impact on the results, as we shall see in section \[sec:GravitonLoops\]. Furthermore, the graviton potential induces new vertices which also ought to be considered.
Matter loops {#sec:MatterLoops}
============
In the previous section, we have reviewed how the ‘non-renormalization theorem’ prevents large quantum corrections from arising in the decoupling limit of massive gravity. Since the coupling to external matter fields is suppressed by the Planck scale these decouple completely when we take $\mpl\to \infty$.
In this section, we keep the Planck scale, $\mpl$, finite. We look at the contributions from matter loops and investigate their effect on the structure of the graviton potential. For definiteness, we consider gravity coupled to a scalar field $\chi$ of mass $M$ and study one-loop effects. When focusing on the one-loop 1PI for the graviton potential, there can be no mixing between the graviton and the scalar field inside the loop (this mixing only arises at higher loops, which we discuss in further work [@2loops]). Furthermore, since we are interested in the corrections to the graviton potential, we only assume graviton zero momentum for the external legs. We use dimensional regularization so as to focus on the running of the couplings, which are encoded by the logarithmic terms.
Framework
---------
Our starting point is the Lagrangian for massive gravity to which we add a real scalar field $\chi$ of mass $M$, S=\^4x $\L_{\rm mGR}+\L_{\rm matter}$, with \[Lmatter\] Ł\_[matter]{}=$\frac 12 g^{ab}\p_a\chi \p_b \chi+\frac 12 M^2 \chi^2$. Note the sign difference owing to the fact that this is the Euclidean action. The Feynman propagator for the scalar field reads \[scalar\_propagator\] G\_&=&(x\_1) (x\_2)=. The mixing between the scalar field and the graviton is encoded in and is highly non-linear. Before proceeding any further it is convenient to perform the following change of variables for the scalar field \[Redefinition\] (g)\^[-1/4]{} , where $g\equiv \det\{g\ab\}$, so that the matter Lagrangian is now expressed as Ł\_[matter]{}=12g\^[cd]{}$\p_c \psi -\frac 14 \psi g^{ab}\p_c g\ab$$\p_d \psi-\frac 14 \psi g^{pq}\p_d g_{pq}$+12 M\^2 \^2 . Since we will only be considering zero momenta for the external graviton legs, we may neglect the terms of the form $\p g$.[^7] As a result, the relevant action for computing the matter loops in given by S\_[matter]{}=\^4x $\frac 12g^{cd}\p_c \psi \p_d \psi +\frac 12 M^2 \psi^2 $. In what follows we will compute the one-loop effective action (restricting ourselves to a scalar field in the loops only) and show explicitly that the interactions between the graviton and the scalar field lead to the running of the cosmological constant, but not of the graviton potential. This comes as no surprise since inside the loops the virtual scalar field has no knowledge of the graviton mass and thus behaves in precisely the same way as in GR, leading to a covariant one-loop effective action. When it comes to the potential, the only operator it can give rise to which is covariant is the cosmological constant. We show this result explicitly in the one-loop effective action, and then present it in a perturbative way, which will be more appropriate when dealing with the graviton loops.
One-loop effective action {#subsec:oneloopea}
-------------------------
The one-loop effective action $S_{1,\textrm{eff}}(g\ab, \psi)$ is given by e\^[- S\_[1,]{}(|g, |)]{}=e\^[- \^i $S_{ij}(\bar g\ab, \bar \psi)$ \^j]{}, where $\Psi_i$ is a placeholder for all the fields, $\Psi_i=\{g\ab, \psi\}$, and $S_{ij}$ is the second derivative of the action with respect to those fields, S\_[ij]{}(|g, |) |\_[g=|g, =|]{}. Here $\bar g\ab$ and $\bar{\psi}$ correspond to the background quantities around which the action for fluctuations is expanded. Since we are interested in the graviton potential part of the one-loop effective action, we may simply integrate over the scalar field and obtain e\^[- S\^[([matter-loops]{})]{}\_[1,]{}(|g)]{}=e\^[-$\frac{\delta^2 S}{\delta^2 \psi}|_{g\ab=\bar g\ab}$]{}. We therefore recover the well-known Coleman–Weinberg effective action, S\^[([matter-loops]{})]{}\_[1,]{}(|g)= 12 $\frac{\delta^2 S}{\delta^2 \psi}\Big|_{g\ab=\bar g\ab}$ =12 $\frac{\delta^2 S}{\delta^2 \psi}\Big|_{g\ab=\bar g\ab}$. Going into Fourier space this leads to Ł\^[([matter-loops]{})]{}\_[1,]{}(|g)&=& 12 $\frac 12 \bar g^{ab}k_a k_b+\frac 12 M^2$\
&=& 12 $\frac 12 \delta^{ab}\tilde k_a \tilde k_b+\frac 12 M^2$\
&& (\^2 ), where $\mu$ is the regularization scale and we restrict our result to the running piece. From the first to the second equality, we have performed the change of momentum $k_a \rightarrow \tilde k_a$ such that $g^{ab}k_a k_b = \delta^{ab}\tilde k_a \tilde k_b$. From this analysis, we see directly that the effect of external matter at one-loop is harmless on the graviton potential. This is no different from GR, since the scalar field running in the loops is unaware of the graviton mass, and the result is covariant by construction. This conclusion is easily understandable in the one-loop effective action (however, when it comes to graviton loops it will be harder to compute the one-loop effective action non-perturbatively and we will perform a perturbative analysis instead). For consistency, we apply a perturbative treatment to the matter fields as well in section \[subsec:renormLambda\].
*[**Higher Loops**]{}*.—
Before moving on to the perturbative argument, we briefly comment on the extension of this result to higher loops. Focusing on matter loops only then additional self-interactions in the matter sector ought to be included. Let us consider, for instance, a $\lambda \chi^3$ coupling. The matter Lagrangian will then include a new operator of the form $\mathcal{L}\supseteq \lambda\sqrt{g}\chi^3=\lambda g^{-1/4}\psi^3$, where $g=\det\{g\ab\}$. At $n$-loops, we have $n$ integrals over momentum, and $2(n-1)$ vertices $\lambda g^{-1/4}\psi^3$, so the $n$-loop effective action reads symbolically S\_n\^[([matter-loops]{})]{}(|g)= \_n$k_1^2,\cdots,k_n^2, M^2$, where $\mathcal{F}_n$ is a scalar function of the different momenta $k_j^2=\bar g^{ab}k_{j a}k_{j b}$.[^8] As a result one can perform the same change of variables as used previously, $k_j \to \tilde k_j$, with $k_j^2=\delta^{ab}\tilde k_{j a}\tilde k_{j b}\equiv \tilde k_j^2$. This brings $n$ powers of the measure $\sqrt{g}$ down so that the $n$-loop effective action is again precisely proportional to $\sqrt{g}$ S\_n\^[([matter-loops]{})]{}(|g)&=&g\^[n/2]{}\
&& \^[2(n-1)]{}M\^[6-2n]{} . The integral in square brackets is now completely independent of the metric $\bar g\ab$ and the $n$-loop effective action behaves as a cosmological constant. The same result holds for any other matter self-interactions. Once again this result is not surprising as this corresponds to the only covariant potential term it can be.
Perturbative approach {#subsec:renormLambda}
---------------------
In the previous subsection we have shown how at one-loop external matter fields only affect the cosmological constant and no other terms in the graviton potential. For consistency we show how this can be seen perturbatively. As mentioned previously, we use the ‘vielbein-inspired’ metric perturbation about flat space, as defined in Eq. . Including all the interactions between the graviton and the matter field, but ignoring the graviton self-interactions for now, the relevant action is then S &= \^4 x { h\^[ab]{}h\_[cd]{}\
& + (-1)\^n (n+1) (\^[ab]{})\^n \_a \_b +12M\^2 \^2}, \[eq:Lhxi\] where raising and lowering of the indices is now performed with respect to the flat Euclidean space metric, $\delta\ab$, since we are working perturbatively. Note that we are using the notation $(\hat{h}^{a b})^2 \equiv \hat{h}^{ac}\, \hat{h}_{c}^{\ b }$.
We now calculate the one-loop matter contribution to the $n$-point graviton scattering amplitudes. For simplicity of notation, we define the scattering amplitudes \^[ (n [pt]{})]{}\^[ (n [pt]{})]{}\_[a\_1 b\_1 a\_n b\_n]{}h\^[a\_1 b\_1]{}h\^[a\_n b\_n]{} . We start with the tadpole correction, using the dimensional regularization technique, which enables us to capture the running of the parameters of the theory.
*[**Tadpole**]{}*.—
At one-loop, the scalar field contributes to the graviton tadpole through the 3-vertex $\hat h^{ab}\p_a \psi \p_b \psi$ represented in Fig. \[Feynman\_diagrams\_1\]. Explicit calculation of this vertex gives \^[(1pt)]{}&=&=14 M\^4 \[\] J\_[M,1]{}, \[eq:1ptmatter\] where \[J1\] J\_[M,1]{}= , as explained in appendix \[sec:appendixDimReg\].
$\mathcal{A}^{\rm (1pt)}\ \ =$
[Scattering2]{}
*[**Beyond the tadpole**]{}*.—
We quote the result for the corrected $n$-point function here, and refer to appendix \[app:matter\] for all the details. Up to the $4$-point function and using Eqs. , , and , the counterterms which ought to added to the original action organize themselves into \[LCT\_2\] Ł\_[CT]{}&=&-$\mathcal{A}^{\rm (1pt)}+\frac{1}{2!} \mathcal{A}^{\rm (2pt)}
+\frac 1{3!} \mathcal{A}^{\rm (3pt)} +\frac 1{4!} \mathcal{A}^{\rm (4pt)}$\
&=&- ( \[\] + ( \[\]\^2 -\[\^2\] ) + ( \[\]\^3 +2\[\^3\] -3 \[\] \[\^2\])\
& & + ( \[\]\^4 -6\[\^4\] -6 \[\]\^2 \[\^2\] +3\[\^2\]\^2 +8 \[\] \[\^3\] ) ) J\_[M,1]{} \[eq:expansiondet\] .\
&=&- J\_[M,1]{} . \[eq:CC\_contribution\] Note that the last line is only technically correct if we include the zero-point function, which we can do (it is a vacuum bubble). We conclude that the matter loops renormalize the cosmological constant, which is the only potential term one can obtain from integrating out matter loops, in agreement with the findings of Refs. [@tHooft:1974bx; @Park:2010rp]. Importantly, matter loops do not affect the structure of the graviton potential.
*[**Higher $n$-point functions**]{}*.—
From the one-loop effective action argument, we know that all the $n$-point functions will receive contributions which will eventually repackage into the normalization of the cosmological constant. Seeing this explicitly at the perturbative level is nevertheless far less trivial, but we give a heuristic argument here. Taking the metric defined in Eq. , the expansion of the determinant of the metric to quartic order in $\hat{h}$ as given in is, in fact, *exact*. The finite nature of the running of the cosmological constant in is therefore no accident.
To show we would arrive at the same conclusion by explicit computation, consider the one-loop correction to the 5-point function. In this case, there are five Feynman diagrams which contribute at the same order for the quantum corrections, depicted in Fig. \[Feynman\_diagrams\_5pt\].
$\mathcal{A}^{\rm (5pt)}\ \ =$
[Scattering5pt]{} + + + +\
+ +\
From the interactions in the Euclidean action , we find \^[(5pt)]{} &=& ( \[\]\^5 +6 \[\^5\] - \[\] \[\^4\] +5 \[\^3\] \[\]\^2 -5 \[\^3\] \[\^2\] + \[\] \[\^2\]\^2 - \[\^2\] \[\]\^3 ) J\_[M,1]{}\
&&0 ,which vanishes identically in four dimensions, as noted in Ref. [@deRham:2010ik]. We can proceed in a similar manner to show that the same will be true for all the $n$-point functions, with $n>5$. This supports the consistency of the formalism introduced in and explicitly agrees with the findings for GR as well as with the direct computation of the one-loop effective action.
Having shown the quantum stability of the massive gravity potential at one-loop, one can see that the same remains true for any number of loops provided there are no virtual gravitons running in the internal lines.
Graviton loops {#sec:GravitonLoops}
==============
In the previous section we have studied in detail the quantum corrections to the potential for massive gravity arising from matter running in loops. We concluded that these quantum corrections could be resummed and interpreted as the renormalization of the cosmological constant. Therefore, we have shown that such loops are completely harmless to the special structure of the ghost-free interaction potential.
Now we push this analysis forward by studying quantum corrections from graviton loops. We start by considering one-loop diagrams, and since we are interested in the IR limit of the theory, we set the external momenta to zero, as before. We will again focus on the running of the interaction couplings, and thus apply dimensional regularization.
Based on studies within the decoupling limit [@deRham:2012ew], we expect the quantum corrections to the graviton mass to scale as $\delta m^2 \sim m^4/\mpl^2\sim 10^{-120} m^2$. Even though such corrections are parametrically small, a potential problem arises if they detune the structure of the interaction potential. If this happens, ghosts arising at a scale much smaller than the Planck mass could in general plague the theory, rendering it unstable against quantum corrections. To show how such corrections could arise, we organise the loop diagrams in powers of the free parameters $\tilde{\alpha}_3$ and $\tilde{\alpha}_4$ of Eq. .
Renormalization of the interactions
-----------------------------------
We start by studying the quantum corrections arising at the linear order in the potential parameters $\tilde{\alpha}_{2,3,4}$. Since $\mathcal{\tilde{U}}_2$ is precisely quadratic in $h$, it leads to no corrections. Next we focus on $\mathcal{\tilde{U}}_3$ in Eq. , which is cubic in $h$ and therefore can in principle renormalize the tadpole at one-loop. The tadpole contribution yields \[Tadpole\_alpha3\] \^[(1pt,3vt)]{} = -58\_3 \[h\] J\_[m,1]{} , which on its own is harmless (this would correspond to the potential $\tilde{\mathcal{U}}_1$ which we have not included in , but which is also ghost-free [@Hassan:2011hr]). The last potential term $\mathcal{\tilde{U}}_4$ is quartic in $h$, as shown in . This interaction vertex leads to quantum corrections to the 2-point function as shown in Fig. \[1loopalpha4\], \^[(2pt,4vt)]{} = \^[(2,4)]{}\_[abcd]{} h\^[ab]{}h\^[cd]{} = 5 \_4 $[h^2]-[h]^2$ J\_[m,1]{} \_2(h), where we have applied dimensional regularization with $J_{m,1}$ given in Eq. . This is nothing else but the Fierz–Pauli structure, which is ghost-free by construction [@Fierz:1939ix].[^9]
$\mathcal{A}^{\rm (2pt)}\ \ =$
[Scattering2grav]{}
Quadratic and other higher order corrections in $\tilde{\alpha}_3$ and $\tilde{\alpha}_4$ on the other hand are less trivial.
We will see however, that this optimistic result will not prevail for other corrections, which will induce the detuning of the interaction potential structure in Eqs. –. To see this we turn to the interactions coming from the Einstein–Hilbert term. Given we can write the Einstein–Hilbert term as \[EH\_term\] - R=h\^ \^\_[ ]{} h+ 1 h (h)\^2+ 1[\^2]{} h\^2 (h)\^2 + , \[eq:EHterm\] where $ \Ein^{\mu\nu}_{\ \ \alpha\beta} $ is the usual Lichnerowicz operator written explicitly in Eq. .
Contrary to the potential in –, Eq. contains an infinite numbers of interactions in $h$. Taking, for example, the third and quartic order interactions from the Einstein–Hilbert term, we find they do not generate any radiative correction to the tadpole, $\mathcal{A}^{\rm(1pt,3vt)}_{\rm EH}=0$, but they do contribute to the 2-point function as follows $$\label{EH3rd}
\mathcal{A}^{\rm (2pt)}_{\rm EH}=\frac{35}{12}\frac{m^4}{\mpl^2}(4[h^2]-[h]^2)J_{m,1}.$$ The result above also does not preserve the Fierz–Pauli structure and is thus potentially dangerous.
Detuning of the potential structure {#sec:detuning}
-----------------------------------
From the above we conclude that quantum corrections from graviton loops can spoil the structure of the ghost-free potential of massive gravity required at the classical level to avoid propagating ghosts. Interestingly, this detuning does not arise from the potential interactions at leading order in the parameters $\tilde{\alpha}_{3,4}$ but does arise from the kinetic Einstein–Hilbert term. Symbolically, the detuning of the potential occurs at the scale \[QC\] Ł\_[qc]{}\~ h\^n, where $m$ is the graviton mass. When working around a given background for $h=\bar h$ (which can include the helicity-0 mode, $\pi$), this leads to a contribution at quadratic order which does *not* satisfy the Fierz–Pauli structure, Ł\_[qc, |h]{}\~ h\^2 . Reintroducing the canonically normalized helicity-0 mode as $h_{\mu\nu}=\p_\mu \p_\nu \pi/m^2$, this implies a ghost for the helicity-0 mode Ł\_[qc, |h]{}\~ (\^2 )\^2\~ (\^2 )\^2 m\_=$\frac{\mpl}{\bar h}$\^[n/2]{}|h . For interactions with $n\ge 3$, the mass of the ghost, $m_{\textrm{ghost}}$, can be made arbitrarily *small* by switching on an arbitrarily *large* background configuration for $\bar h$. This is clearly a problem since *large* backgrounds ($\bar h \gtrsim \mpl$, or alternatively $\p^2 \bar \pi \gtrsim \Lambda_3^3$) are important for the Vainshtein mechanism [@Vainshtein:1972sx] to work and yet they can spoil the stability of the theory.
One-loop effective action {#sec:1PI}
=========================
We have shown that quantum corrections originated both from the potential as well as from the Einstein–Hilbert term in general destabilize the ghost-free interactions of massive gravity. This happens in a way which cannot be accounted for by either a renormalization of the coefficients of the ghost-free mass terms, or by a cosmological constant. This detuning leads to a ghost whose mass can be made arbitrarily small if there is a sufficiently large background source. From the decoupling limit analysis, we know that it is always possible to make the background source large enough without going outside of the regime of the effective field theory since we have |h \~|= | .
As an effective field theory we are allowed to make $\partial \partial \pi \gg \Lambda_3^3$ provided $\partial^3 \pi /(1+\partial^2 \pi/\Lambda_3^3 ) \ll \Lambda_3^4$. In other words, as long as derivatives of the background $\bar h$ are small $\partial \ll \Lambda$, the magnitude of the background may be large $\bar h \gg \mpl$.
The resolution of this problem in this case is that one also needs to take into account the redressing of the operators in the interaction potential. In this section we will investigate how the Vainshtein mechanism operates in protecting the effective action from the appearance of dangerous ghosts below the Planck scale.
Form of the potential at one-loop {#subsec:olea}
---------------------------------
Our previous approach involved explicit calculation of loop diagrams to evaluate the quantum corrections to the massive gravity potential. Here, we shall focus on the formalism of the one-loop effective action to confirm the destabilization result and provide a more generic argument. Since the quadratic potential $\tilde{\mathcal{U}}_2$ in has no non-linear interactions, we can take it as our sole potential term and consider all the graviton self-interactions arising from the Einstein–Hilbert term in Eq. . For simplicity, and without loss of generality, we therefore consider in what follows the specific theory of massive gravity Ł=- R - 14 \^2 m\^2 \_2\[h\]. \[eq:1lea\]
We start by splitting the field $h\mn$ into a constant background $\bar h\mn$ and a perturbation $\delta h\mn(x)$ which, in the language of the previous sections, will be the field running in the loops. We thus write $h\mn(x)=\bar h\mn+\delta h\mn(x)$. Up to quadratic order in the perturbation $\delta h$ Ł&=& 12 h\_$\Ein^{\mu\nu \alpha\beta} +m^2 \(\delta^{\mu\alpha} \delta^{\nu \beta}
- \delta^{\mu \nu} \delta^{\alpha \beta}$ h\_ +$ \frac{1}{\mpl}\bar h + \frac{1}{\mpl^2}\bar h^2 + \cdots$ (h)\^2\
&=& 12 h\_ $G^{-1 \ \mu\nu\alpha\beta}+M^{\mu\nu\alpha\beta}(\bar h)$h\^ h\_ \^ h\_, \[eq:Leff1loop\] where $M^{\mu\nu\alpha\beta}(\bar h)= \(\frac 1 \mpl \bar h + \frac 1{\mpl^2} \bar h^2+\cdots \)\p^2 $ symbolizes all the interactions in the Einstein–Hilbert term. $G^{-1}$ is the inverse of the massive graviton propagator. Following the same analysis of section \[subsec:oneloopea\], the one-loop effective action is then given by Ł\_[eff]{}&=&-12 $\frac{1}{\mu^2} \left\{G^{-1\ \mu\nu\alpha\beta}+M^{\mu\nu\alpha\beta}(\bar h)\right\}$\
& & - $\int\frac{\d^4k}{(2\pi)^4} \frac{f_{\mu\nu\alpha\beta}M^{\mu\nu\alpha\beta}}{\propm}
- \frac 12 \int \frac{\d^4k}{(2\pi)^4} \frac{f_{\mu\nu\alpha\beta}M^{\mu\nu a b } f_{abcd}M^{cd \alpha\beta}}{\propm^2} +\cdots $ , \[eq:olearedressed\] where $M^{\mu\nu\alpha\beta}(\bar{h})$ is expanded in Fourier space and depends explicitly on a derivative structure (and so, on the momentum $k$). Here $\mu$ denotes again a renormalization scale, which ought to be introduced as a consequence of the renormalization procedure, and to preserve the dimensional analysis. Eq. sources an effective potential which goes as $m^4 \mathcal{F}(\bar h / \mpl)J_{1,m} $ where $\mathcal{F}$ denotes an infinite series in powers of $\bar h$. This result implies a running of the effective potential.
To gain some insight on the form of this effective potential, we focus on the specific case of a conformally flat background where $\bar h\mn= \lambda \delta\mn$, for some real-valued $\lambda$. It follows[^10] M\^(|h=)=0, which means that in this case all interactions are lost and also vanishes. This implies that the effective potential for a generic $\bar h\mn$ has to be of the form Ł\_[eff]{}= c\_1 $[\bar h]^2-4[\bar h^2]$+$c_2[\bar h]^3+c_3[\bar h^2][\bar h]-(16 c_2+4c_3)[\bar h^3] $+, \[eq:Leffd\] for some coefficients $c_1$, $c_2$ and $c_3$. The explicit form of these coefficients can be read off by computing specific Feynman diagrams corresponding to the Einstein–Hilbert interactions, or by considering a more general background metric $\bar h\mn$. For instance, $c_1$ corresponds to the coefficient in Eq. , $c_1=\frac{35}{12}\frac{m^4}{\mpl^2}$. It is apparent that this structure is very different from that of the ghost-free potential of Eqs. –. This confirms the results obtained in the previous sections.
Mass of the ghost
-----------------
At what scale does this running arise? Let us first concentrate on the quadratic term in . Since the helicity-0 mode $\pi$ enters as $h\mn= \p_\mu\p_\nu \pi/m^2$, that term would lead to a correction of the form Ł\_[eff]{}\^[(2)]{}=$[\bar h]^2-4[\bar h^2]$ J\_[m,1]{} \~$\Box \pi$\^2 (m\^2/\^2) . This would excite a ghost at the Planck scale. Hence this contribution on its own is harmless. Next we consider the effect of the cubic interactions, Ł\_[eff]{}\^[(3)]{}= \[|h\]\^3 J\_1\~ $\Box \pi$\^3 (m\^2/\^2). \[eq:beforeV1\] We now elaborate on the general argument mentioned in section \[sec:detuning\]. We take a background configuration for $\pi$ wich is above the scale $\Lambda_3=(\mpl m^2)^{1/3}$ for the Vainshtein mechanism to work. This will induce a splitting of the helicity-0 mode $\p^2 \pi = \p^2 \bar \pi + \p^2 \delta \pi$, with $\p^2 \bar \pi \sim \mpl m^2/\kappa$ and $\kappa<1$. Then the operator in could lead to a ghost at a scale much lower than the Planck scale, Ł\_[eff]{}\^[(3)]{}\~$\Box \pi$\^2 (m\^2/\^2). \[eq:beforeV2\] Thus by turning on a *large* background, thereby making $\kappa$ *smaller*, the scale at which the ghost arises becomes smaller and smaller, and eventually comparable to $\Lambda_3$ itself. This renders the theory unstable, as argued in section \[sec:detuning\]. However, by assuming a large background we also need to understand its effect on the original operators via the Vainshtein mechanism [@Vainshtein:1972sx].
Vainshtein mechanism at the level of the one-loop 1PI
-----------------------------------------------------
The formalism of the one-loop effective action makes the Vainshtein mechanism particularly transparent as far as the redressing of the interaction potential is concerned. We further split the field $\bar h$ into a large background configuration $\bar{\bar{h}}\mn$ and a perturbation $\tilde h\mn\sim \p_\mu\p_\nu \pi/m^2$, such that $\bar{h}\mn=\bar{\bar{h}}\mn+\tilde h\mn$. Since $\bar{\bar{h}}\mn$ satisfies the equations of motion we have $\tilde{M}'(\bar{h})|_{\bar{\bar h}}=0$, with $\tilde{M}$ defined as in Eq. . We proceed as before and expand the one-loop effective action up to second order in $\tilde h\mn$, as follows Ł\_[eff]{}&=&-12 $ \frac{1}{\mu^2}\tilde{M}^{\mu\nu\alpha\beta}(\bar h)$\
&&-12 h \^2\
&\~& \[h\]\^2 \~ ()\^2 , where the last line is symbolic,[^11] and for simplicity of notation we have denoted the combination $\Xi \equiv \tilde{M}''(\bar{\bar{h}})/\tilde{M}(\bar{\bar{h}}) $. It follows that the mass of the ghost is $m^2_{\text{ghost}}=\Xi^{-1} \mpl^2 $. Provided we can show that $\Xi \lesssim 1$, the ghost will arise at least at the Planck scale and the theory will always be under control.
*[**Redressing the one-loop effective action**]{}*.—
The basis of the argument goes as follows. As explained previously, the Vainshtein mechanism relies on the fact that the background configuration can be large, and thus $\tilde{M}''$ can in principle be large, which in turn can make $\Xi$ large and lower the mass of the ghost. However, as we shall see in what follows, configurations with large $M''$ automatically lead to large $M$ as well. This implies that $\Xi$ is always bounded $\Xi \lesssim 1$, and the mass of the ghost induced by the detuning/destabilization of the potential from quantum corrections is always at the Planck scale or beyond.[^12]
*[**Computing the mass of the ghost**]{}*.—
To make contact with an explicit calculation we choose, for simplicity, the following background metric $\bar g\ab=\text{diag}\{\lambda_0^2, \lambda_1^2, \lambda_1^2, \lambda_1^2\}=\bar{\gamma}\ab^2$, and compute the possible combinations appearing in $\Xi$. We define $$\frac{\partial^2 \tilde{M}/\partial \bar{\bar{h}}_{\alpha\beta} \partial \bar{\bar{h}}_{\gamma\delta}}{\tilde{M}}\
\equiv \Xi_{(\alpha\beta,\gamma\delta)}\,, \ $$ and use units for which $\mpl^2=1$.
In general, the components of $\tilde{M}$ can be split into three categories. First, some components do not depend on the background, and are thus explicitly *independent* of $\lambda_0$ and $\lambda_1$. In this case $\tilde{M}''=\Xi=0$ trivially. Second, other components are of the form $$\tilde{M}\sim \frac{\lambda_0}{\lambda_1} \left( k_i k_j +m^2\right),
\label{eq:secondtildeM}$$ where $i$ and $j$ are spatial indices. In this case $\Xi_{(00,00)}=0$, whereas $\Xi_{(ii,ii)} \sim \lambda_1^{-2} \lesssim 1$.
Finally the remaining category contains terms which only depend on some power of the ratio of the components of the background metric, $\lambda_0/\lambda_1$ or $\lambda_1/\lambda_0$. The structure of the components of $\tilde{M}$ in this case is of the form$$\tilde{M}\sim \left(\frac{\lambda_1}{\lambda_0}\right)^p k_0^2
+\left(\frac{\lambda_0}{\lambda_1}\right)^q
\sum_i b_i k_i^2 \\ \ + \frac{\lambda_0}{\lambda_1} m^2,$$ in which $i$ denotes a spatial component, $b_i$, $p$ and $q$ are integer numbers such that $p, q \geq 1$. The last term represents the (non-vanishing) structure of the mass term.[^13] Let us work out the possible second derivatives in detail. On the one hand $$\Xi_{(00,00)}
\sim
\dfrac{\left(\frac{\lambda_1}{\lambda_0}\right)^p k_0^2 \, \lambda_0^{-2} +\left(\frac{\lambda_0}{\lambda_1}\right)^q \, \lambda_0^{-2}\, \sum_i b_i k_i^2 }{\left(\frac{\lambda_1}{\lambda_0}\right)^p k_0^2
+\left(\frac{\lambda_0}{\lambda_1}\right)^p \frac{\lambda_0}{\lambda_1} \sum_i b_i k_i^2 +\textrm{mass term}}\ \,,$$ where we have ignored factors of order unity to avoid clutter. At first sight this result seems troublesome as it appears to be dependent on the choice of background, and in particular on the hierarchy between $\lambda_0$ and $\lambda_1$. We will however show that this is *not* the case:
i. if $\lambda_0\sim\lambda_1\sim\lambda$, then $\Xi_{(00,00)}\sim \lambda^{-2}$. Since we are interested in incorporating the Vainshtein effect, we shall consider the case when $\lambda\gtrsim 1$, and thus $\Xi\lesssim 1$.
ii. if $\lambda_0 \gg \lambda_1 \gtrsim 1$, it follows $\Xi \lesssim \lambda_0^{-2} \lesssim 1$; the same holds true in the case $\lambda_1 \gg \lambda_0 \gtrsim 1$.
On the other hand, $$\Xi_{(jj,kk)}=
\Xi_{(jj,jj)}
\sim
\dfrac{\left(\frac{\lambda_1}{\lambda_0}\right)^p \, k_0^2 \,\lambda_1^{-2}+
\left(\frac{\lambda_0}{\lambda_1}\right)^q \, \lambda_1^{-2}\, \sum_i b_i k_i^2
+ \frac{\lambda_0}{\lambda_1^3} m^2}{\left(\frac{\lambda_1}{\lambda_0}\right)^p k_0^2
+\left(\frac{\lambda_0}{\lambda_1}\right)^q \sum_i b_i k_i^2+
\frac{\lambda_0}{\lambda_1}m^2}\ \ .$$ We repeat the previous analysis to show that, regardless of the possible hierarchy between $\lambda_0$ and $\lambda_1$, the quantum corrections will be parametrically small.
i. if $\lambda_0\sim\lambda_1\sim\lambda$, then $\Xi_{(jj,jj)}\sim \lambda^{-2}$, and hence $\Xi \lesssim 1$.
ii. if either $\lambda_0 \gg \lambda_1 \gtrsim 1$ or $\lambda_1 \gg \lambda_0 \gtrsim 1$, then $\Xi \lesssim \lambda_1^{-2} \ll 1$.
Similar conclusions can be drawn for the ‘mixed’ derivative $\Xi_{(00,ij)}$ or $\Xi_{(0i,0j)}$.
We have shown that, despite appearances, $\Xi\lesssim 1$ independently of the background and without loss of generality. Whenever the Vainshtein mechanism is relevant, that is, when $\lambda_0, \lambda_1 \gtrsim 1$, the redressing of the operators ensures that the mass of the ghost arises at least at the Planck scale. We therefore conclude that at the one-loop level the quantum corrections to the theory described by are under control.
Conclusions {#sec:conclusion}
===========
In a theory of gravity both the mass and the structure of the graviton potential are fixed by phenomenological and theoretical constraints. While in GR this tuning is protected by covariance, such a symmetry is not present in massive gravity. Nevertheless, the ‘non-renormalization theorem’ present in theories of massive gravity implies that these tuning are technically natural [@deRham:2012ew; @Nicolis:2004qq], and hence do not rely on the same fine-tuning as for instance setting the cosmological constant to zero. In this manuscript we have explored the stability of the graviton potential further by looking at loops of matter and graviton, assuming a covariant coupling to matter (see for instance Ref. [@deRham:2011qq] for a discussion of the natural coupling to matter and its stability).
When integrating out externally coupled matter fields, we have shown explicitly that the only potential contribution to the one-loop effective action is a cosmological constant, and the special structure of the potential is thus unaffected by the matter fields at one-loop.
For graviton loops, on the other hand, the situation is more involved—they *do* change the structure of the potential, but in a way which only becomes relevant at the Planck scale. Nevertheless, the Vainhstein mechanism that resolves the vDVZ discontinuity relies on a classical background configuration to exceed the Planck scale ($g\mn -\delta\mn\gtrsim 1$), without going beyond the regime of validity of the theory. A naïve perturbative estimate would suggest that on top of such large background configurations, the mass of the ghost could be lowered well below the Planck scale. However, this perturbative argument does not take into account the same Vainshtein mechanism that suppresses the vDVZ discontinuity in the first place. To be consistent we have therefore considered a non-perturbative background and have shown that the one-loop effective action is itself protected by a similar Vainshtein mechanism which prevents the mass of the ghost from falling below the Planck scale, even if the background configuration is large.
The simplicity of the results presented in this study rely on the fact that the coupling to matter is taken to be covariant and that at the one-loop level virtual gravitons and matter fields cannot mix. Thus at one-loop virtual matter fields remain unaware of the graviton mass. This feature is lost at higher loops where virtual graviton and matter fields start mixing.
Higher order loops are beyond the scope of this paper, but will be investigated in depth in Ref. [@2loops]. In this follow-up study, we will show how a naïve estimate would suggest that the two-loop graviton-matter mixing can lead to a detuning of the potential already at the scale $\mpl (m/M)^2$, where $m$ is the graviton mass and $M$ is the matter field mass. If this were the case, a matter field with $M\sim\Lambda_3$ would already bring the mass of the ghost down to $\Lambda_3$ which would mean that the theory could never be taken seriously beyond this energy scale (or its redressed counterpart, when working on a non-trivial background). However, this estimate does not take into account the very special structure of the ghost-free graviton potential which is already manifest in its decoupling limit. Indeed, in ghost-free massive gravity the special form of the potential leads to interesting features when mixing matter and gravitons in the loops.
To give an idea of how this mixing between gravitons and matter arises, let us consider the one-loop contribution to the scalar field two-point function depicted in Fig. \[Fig:ScalarField2pt\] if the scalar field does not have any self-interactions. We take the external leg of the scalar field to be on-shell, with momentum $q_a$ satisfying $\delta^{ab}q_a q_b+M^2=0$.[^14]
$\mathcal{A}_{\chi\chi}=$
[ScalarField2pt\_3]{} $+$
For the purpose of this discussion, it is more convenient to work in terms of the field $\chi$ directly rather than the redefined field $\psi$. First, diagram (a) gives rise to a contribution proportional to \^[(a)]{}\_\^4 k , where $k$ is the momentum of the virtual graviton running in the loop of diagram (a). By momentum conservation, the momentum $p$ of the virtual field $\chi$ in the loop is then $p_a=q_a-k_a$. Applying the on-shell condition for the external legs, $q^2+M^2=0$, we find \^[(a)]{}\_\^4 k $(k.q)^2+m^2 q^2$0 . So this potentially ‘problematic’ diagram that mixes virtual matter and gravitons (which could a priori scale as $m^{-4}$) leads to no running when the *external* scalar field is on-shell. In other words, at most this diagram can only lead to a running of the wave-function normalization and is thus harmless (in particular it does not affect the scalar field mass, nor does it change the ‘covariant’ structure of the scalar field Lagrangian).
Second, we can also consider the contribution from a pure graviton-loop in diagram (b). Since only the graviton runs in that loop, the running of the scalar field mass arising from that diagram is at most $\delta M^2= \frac{m^2}{\mpl^2}M^2 \ll M^2$ and is therefore also harmless. As a consequence we already see in this one-loop example how the mixing between the virtual graviton and scalar field in the loops keeps the structure of the matter action perfectly under control.
We would like to thank Michael Duff, Gregory Gabadadze, Kurt Hinterbichler, Rachel Rosen and Andrew J. Tolley for useful discussions, and Matteo Fasiello for comments on an early draft of this paper. CdR is supported by Department of Energy grant DE-SC0009946. LH is supported by the Swiss National Science Foundation. We acknowledge the use of the xAct package for Mathematica [@Brizuela:2008ra; @xAct].
Dimensional regularization {#sec:appendixDimReg}
==========================
For the one-loop diagrams we required the dimensional regularization technique to obtain the quantum corrections. A recurrent integral which appears in our calculations is of the form J\_[,n]{}= , \[eq:Jn\] where $\tilde{m}$ is a placeholder for whichever mass appears in the propagator. By symmetry we have = \_[\_1 \_[2j]{}]{} J\_[,n]{}, with the generalized Kronecker symbol, $$\delta_{\alpha_1 \cdots \alpha_{2j}}=
\delta_{\alpha_1 \alpha_2} \delta_{\alpha_3 \cdots \alpha_{2j}}+
\Big( \{\alpha_2\} \leftrightarrow \{\alpha_3, \cdots , \alpha_{2j}\}\Big)\ .$$ We also note that J\_[,n]{}= J\_[,1]{}. We do not need to express $J_{\tilde{m},1}$ explicitly in dimensional regularization, but can simply rely on these different relations to show how different diagrams repackage into a convenient form. It suffices to know that $J_{\tilde{m},1}$ contains the logarithmic divergence in $\tilde{m}$, which represents the running in renormalization techniques.
A closer look at the matter loops {#app:matter}
=================================
In section \[subsec:renormLambda\] we have calculated the quantum corrections arising from matter loops to the graviton tadpole. For completeness, we collect in this appendix the individual corrections from each Feynman diagram corresponding to a given $n$-point function.
*[**2-point function**]{}*.—
There are two Feynman diagrams which contribute to the corrected 2-point function, which arise respectively from the cubic and quartic interactions in the action .
$\mathcal{A}^{\rm (2pt)}\ \ =$
[Scattering2pt]{} +\
Evaluation of these one-loop diagrams shown in Fig. \[Feynman\_diagrams\_2pt\] gives \^[(2pt)]{}\_[(a)]{}&=& 2 \^[ab]{}\^[cd]{} =1[4]{} M\^4$2[\hat{h}^2]+[\hat{h}]^2$ J\_[M,1]{},\
\^[(2pt)]{}\_[(b)]{} &=&-3 =- \[\^2\] J\_[M,1]{} , so that the total contribution to the 2-point function is \^[(2pt)]{}&=&\^[(2pt)]{}\_[(a)]{}+\^[(2pt)]{}\_[(b)]{} =1[4]{} M\^4(\[\]\^2-\[\^2\]) J\_[M,1]{}. \[eq:2ptmatterapp\]
*[**3-point function**]{}*.—
The three-point scattering amplitude will receive corrections from the diagrams depicted in Fig. \[Feynman\_diagrams\_3pt\]
$\mathcal{A}^{\rm (3pt)}\ \ =$
[Scattering3pt]{} + +\
which give the following contributions \^[(3pt)]{}\_[(a)]{} &=& $[\hat{h}]^3+6[\hat{h}][\hat{h}^2]+8[\hat{h}^3]$ J\_[M,1]{},\
\^[(3pt)]{}\_[(b)]{} &=& - $[\hat{h}^2][\hat{h}]+2[\hat{h}^3]$ J\_[M,1]{} ,\
\^[(3pt)]{}\_[(c)]{} &=& 3 M\^4 \[\^3\] J\_[M,1]{} . We conclude the total 3-point function goes as \^[(3pt)]{} &=& $ 2[\hat{h}^3] +[\hat{h}]^3 -3 [\hat{h}] [\hat{h}^2] $ J\_[M,1]{} . \[eq:3pft\]
*[**4-point function**]{}*.—
The Feynman diagrams contributing to the corrected 4-point function are those in Fig. \[Feynman\_diagrams\_4pt\]
$\mathcal{A}^{\rm (4pt)}\ \ =$
[Scattering4pt]{} + + + +\
and they give the following contributions \^[(4pt)]{}\_[(a)]{} &=& M\^4 (12\[\^4\]+8\[\]\[\^3\]+3\[\^2\]\^2 +3 \[\^2\] \[\]\^2 + \[\]\^4) J\_[M,1]{},\
\^[(4pt)]{}\_[(b)]{} &=& $[\hat{h}^2]^2+2[\hat{h}^4]$ J\_[M,1]{} ,\
\^[(4pt)]{}\_[(c)]{} &=& 12 M\^4 $[\hat{h}^3] [\hat{h}] +2 [\hat{h}^4] $ J\_[M,1]{} ,\
\^[(4pt)]{}\_[(d)]{} &=& - $ [\hat{h}^2] [\hat{h}]^2
+2 [\hat{h}^2]^2 +2 [\hat{h}] [\hat{h}^3] +8 [\hat{h}^4] +2 [\hat{h}][\hat{h}^3]$ J\_[M,1]{} ,\
\^[(4pt)]{}\_[(e)]{} &=& -15 M\^4 \[\^4\] J\_[M,1]{} , so that the total 4-point function is given by \^[(4pt)]{} &=& $ [\hat{h}]^4 -6 [\hat{h}^4] -6 [\hat{h}^2] [\hat{h}]^2
+3 [\hat{h}^2]^2 +8 [\hat{h}] [\hat{h}^3]
$ J\_[M,1]{} . \[eq:4pft\] Eqs. , and have the precise coefficients to produce a running of the cosmological constant, as shown in Eq. .
[^1]: By decoupling limit we mean taking $m\rightarrow 0 $ and $\mpl\rightarrow \infty$ simultaneously, whilst keeping the energy scale $\Lambda_3\equiv
(m^2\mpl)^{1/3}$ fixed.
[^2]: Ref. [@Capper:1973bk] is an example of an early work which discussed the quantum corrections in GR in a similar way this paper explores the radiative corrections to massive gravity.
[^3]: In the Euclidean version of massive gravity both the dynamical metric $g\mn$ and the reference metric $f\mn$ have to be ‘Euclideanized’, $g\mn\to g\ab$ and $f\mn=\eta\mn \to \delta\ab$.
[^4]: As already mentioned, for the purposes of computing loop corrections, it is more convenient to work with the flat Euclidean reference metric, $\delta_{ab}$, after performing a Wick rotation $t \to -i \tau$, where $\tau$ is the Euclidean time.
[^5]: As in Ref. [@deRham:2010kj], this 2-parameter family of potential is the one for which there is no cosmological constant nor tadpole.
[^6]: The absence of the ghost in these theories is tightly related to the antisymmetric nature of their interactions, which in turn guarantees their non-renormalization. The same reasoning applies to the construction of the Lovelock invariants. For example, in linearized GR, linearized diffemorphism tells us that the kinetic term can be written using the antisymmetric Levi–Cevita symbols as $\mathcal{E}^{\mu\nu\alpha\beta}\mathcal{E}^{\mu'\nu'}_{\phantom{\mu'\nu'}\alpha\beta}R_{\mu\nu\mu'\nu'}$, which ensures the non-renormalization. Notice that gauge invariance alone would still allow for a renormalization of the overall factor of the linearized Einstein–Hilbert term, which does not occur in the decoupling limit.
[^7]: Such terms were kept in Ref. [@Park:2010rp] where the corrections involved higher order curvature invariants built out of the metric perturbations, but in this study we are only interested in the corrections to the graviton potential and not the higher curvature terms.
[^8]: Even if different momenta $k_j$ contract one can always reexpress them as functions of $k_j^2$, following a similar procedure to what is presented in appendix \[sec:appendixDimReg\].
[^9]: At the quadratic level, the Fierz–Pauli term is undistinguishable from the ghost-free potential term $\tilde{\mathcal{U}}_2$.
[^10]: This can be seen more explicitly, by writing the operator $M$ in terms of the background metric $\bar{\gamma}\ab$, recalling that $\bar g\ab=\bar{\gamma}_{ac}\bar{\gamma}_{bd}\delta^{cd}$ and the metric $g\ab$ is given in terms of $\bar{\gamma}\ab$ and the field fluctuation $h\ab$ as in . Then it follows that symbolically, M\^[abcd]{}\~(\^a\_|\_\^[b]{})(\^c\_|\_\^[d]{}) ( |[g]{}\^|[g]{}\^ |[g]{}\^[| ]{}\_\_), where the two first terms in bracket arise from the transition to the ‘vielbein-inspired’ metric fluctuation and the last term is what would have been otherwise the standard linearized Einstein–Hilbert term on a constant background metric $\bar{g}\ab$. Written in this form, $M$ is manifestly conformally invariant.
[^11]: In particular, by dimensional analysis, one should think of the schematic form for the effective Lagrangian as containing a factor of $1/\mu^2$, where $\mu$ carries units of $[\text{mass}]$.
[^12]: The only way to prevent $\Xi$ from being $\lesssim 1$ is to consider a region of space where some eigenvalues of the metric itself vanish, which would be for instance the case at the horizon of a black hole. However as explained in [@Deffayet:2011rh; @Koyama:2011xz; @Koyama:2011yg; @Berezhiani:2011mt], in massive gravity these are no longer coordinate singularities, but rather real singularities. In massive gravity, black hole solutions ought to be expressed in such a way that the eigenvalues of the metric never reach zero apart at the singularity itself. Thus we do not need to worry about such configurations here (which would correspond to $\lambda=1$ in what follows).
[^13]: The mass term can also arise in the form $(\lambda_1/\lambda_0) \, m^2$ for the components $\tilde{M}^{00ii}$, but the conclusions hereafter remain unchanged.
[^14]: Technically, in Euclidean space this means that the momentum is complex, or one could go back to the Lorentzian space-time for the purposes of this calculation, but these issues are irrelevant for the current discussion. Moreover, note that the on-shell condition is only being imposed for the external legs, and *not* for internal lines.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We analyse the renormalization of the of two-nucleon interaction with multiple subtractions in peripheral waves considering two chiral forces at N3LO. Phase shifts at low energies are then computed with several subtraction points below $\mu = 10~{\rm fm}^{-1}$. We show that for most peripheral waves the phase shifts have nearly no dependence on the renormalization scale. In two cases the phase shifts converge slowly as the renormalization scale approaches $\mu = 1 ~ {\rm fm}^{-1}$ and in one case the phase shifts presented oscillations with respect to the subtraction point $\mu$.'
author:
- |
E. F. Batista$^1$, S. Szpigel$^2$ and V. S. Timóteo$^3$\
[$^1$*Departamento de Ciências Exatas e Naturais, Universidade Estadual do Sudoeste da Bahia*]{}\
[$^2$ *Centro de Rádio-Astronomia e Astrofísica, Universidade Presbiteriana Mackenzie*]{}\
[$^3$ *Faculdade de Tecnologia - FT, Universidade Estadual de Campinas - UNICAMP*]{}
title: Renormalization of chiral nuclear forces with multiple subtractions in peripheral channels
---
Introduction
============
It is widely established in nuclear physics, that Quantum Chromodynamics (QCD) is in fact the underlying theory which describes the properties of strong nuclear forces. In this theory, the fundamental degrees of freedom are quarks which interact with each other via exchange of colored gluons. On the other hand, the strong nuclear force is also responsible for the binding of protons and neutrons in atomic nuclei. But according to QCD, nucleons are bound states of quarks and the nuclear force is considered as the residual part of the quark-quark interaction inside of the nucleon with exchange of gluons. Due to the property of asymptotic freedom, the running coupling constant is small enough at high energies to allow QCD be handled within a perturbative approach. However, at low energies, where mesons and nucleons are the relevant degrees of freedom, the running coupling constant becomes large and QCD is no longer perturbative. This strong non-perturbative nature of QCD at low-energies implies in several mathematical and computational difficulties in describing nucleon properties at this energy level.
Since QCD cannot be treated perturbatively at low energies, a new approach was developed to handle nuclear forces with degrees of freedom appropriate for low-energy systems. The idea was to use quantum field theory but replacing quarks and gluons degrees of freedom by pions and nucleons, keeping the fundamental properties of the underlying theory like chiral symmetry. This Effective Field Theory (EFT) scheme was already used in other systems to describe different types of interactions and Weinberg proposed an EFT approach to nuclear systems based on QCD. This idea generated a new branch in nuclear physics and allowed a deeper understanding of the nuclear force and few-nucleon systems. In particular, the two-nucleon system requires a non-perturbative extension of Chiral Perturbation Theory (ChPT) which works well in the case of pion-nucleon scattering.
Basically, an EFT is constructed by isolating the most relevant degrees of freedom and symmetries for the system under consideration and applying standard quantum field theory. In nuclear physics, replacing quarks and gluons by pions and nucleons means moving to a different (lower) energy scale and a connection between the symmetry properties of the underlying fundamental theory and the symmetries of their effective versions must be well-established. Thus, in a nuclear effective theories it is necessary to establish scale parameters which enable us to separate the high-energy components of the interaction from the low-energy part. In a series of papers [@wein], Weinberg proposed an effective field theory scheme for nuclear forces based on the chiral symmetry of QCD.
This approach was first applied by Ordóñez, Ray and van Kolck [@bira] and allowed the perturbative treatment of the $NN$ interaction. An expansion in powers of $(Q/\Lambda_\chi)^\nu$ is performed and $Q$ is a generic low momentum scale and $\Lambda$ is the chiral symmetry breaking scale which is approximately $1 GeV$. This expansion is controlled by a power counting scheme, called Weinberg Power Counting (WPC), which provides an hierarchical organization for the processes in few-nucleon systems. Following the WPC, the $NN$ interaction at leading order (LO) consists of one-pion-exchange (OPE) plus a contact term. At next-to-leading order (NLO), two-pion-exchange (TPE) and $\mathcal{O}(p^2)$ contact interactions are added, at next-to-next-to-leading order (N2LO) there is an additional set of TPE diagrams and, finally, at next-to-next-to-next-to-leading order (N3LO) corrections to both OPE and TPE are included along with $\mathcal{O}(p^4)$ contact interactions.
Regardless of how the chiral expansion is organized, the issue of how to renormalize the two-body interaction is of fundamental importance and has been subject of investigation for decades. Early works by Adhikari et al started by discussing the renormalization of two-body quantum hamiltonians [@adh]. Later, the problem was focused in the $NN$ interaction using renormalization group analysis and a new power counting was proposed by Kaplan, Savage and Wise [@ksw]. Discussions on the renormalization of singular and one-pion-exchange two-body interactions, power counting and renormalization of the three-body system are detailed by van Kolck et al [@vK; @NTvK]. Another renormalization group approach to two-body and nucleon-nucleon scattering was presented by Birse et al [@MCbirse] and a complete analysis of cutoff renormalization in configuration space was performed by the Granada group [@GRen]. Also, a comparison between renormalization in configuration and momentum spaces has been carried out in Ref. [@RxP].
Another renormalization approach for the $NN$ system consists of a hybrid scheme, where the LO contribution is treated non-perturbatively and the higher orders are handled perturbatively [@mnl11]. Results for $P$-waves and $D$-waves show that perturbative two-pion-exchange reproduces the experimental data up to $k_{\rm cm} \sim 300~{\rm MeV}$. Here we treat all terms non-perturbatively since in our renormalization scheme the pieces of the interaction are inserted as the subtractions are performed.
Apart from the divergences due to pion loops in the irreducible diagrams, the reducible diagrams also generate divergences in the scattering equation. To overcome this problem the most common employed method is introducing a cutoff regularization scale $\Lambda$ which limits the momentum integration, in the scattering equation, above this scale parameter resulting in a finite phase-shifts. The cutoff scheme handles the divergences by modifying the potential and keeping the scattering equation intact. The regularized interaction contains only low-momentum components and the cutoff scale is fixed at some scale, typically $\sim 2 - 3 ~{\rm fm}^{-1}$. A slight discomfort with this method comes from the fact that all physics above a certain momentum scale are excluded. Recently, a N3LO interaction has been optimized by an improved renormalization approach in configuration space which maintains the analytic structure of the scattering amplitude [@ekm15].
An alternative renormalization procedure referred as subtractive method or multiple subtractions, treat the divergences with a different perspective: instead of modifying the potential, as in cutoff method, and keeping the scattering equation untouched, here the interaction is kept intact and the scattering equation is modified by the introduction of subtractions in its kernel. The dependence of the phase shifts on the cutoff is replaced by a dependence on the subtraction point that can be later eliminated by using the renormalization group flow. With this procedure no components are neglected from the interaction and both low and high momentum components are included. Detailed descriptions of this approach can be found in Refs.[@NPA99; @PLB00; @PLB05; @npa07; @ijmp07; @PRC11; @JPG12].
In this work we employ the subtract kernel method to renormalize the Lippmann-Schwinger equation in N3LO and perform a detailed renormalization scale dependence analysis of the phase shifts in peripheral channels. The high angular momentum waves are interesting because the force in their channels contains no contact interactions and consists only of pion exchanges. The N3LO potential have contact contributions up to $D$-waves so that $F$ and higher waves have contributions purely from pion exchanges and no core given by contact interactions.
This paper is organized as follows: Section 2 describes the renormalization of N3LO interactions with five subtractions, Section 3 presents the numerical results for both uncoupled and coupled channels up to $J=6$ and our main conclusions are given in Section 4.
Renormalization of N3LO interactions
====================================
At any given order, the modern two-nucleon interactions can always be separated in two components: the pion exchange interactions and the contact terms, which parametrize the short range core of the interaction and are determined by fitting scattering data. The $NN$ potential is then written as $$V_{\rm NN} = V_{\rm pions} + V_{\rm cont} \; ,$$ where the first term contains one-pion-exchange and two-pion exchanges $$V_{\rm pions} = V_{\pi} + V_{2\pi} \; .$$
The power counting scheme organizes which set of Feynman diagrams must be included in each order in the chiral expansion: $$\begin{aligned}
V_{\rm pions} & = & V_{\rm pions}^{(0)} + V_{\rm pions}^{(2)} + V_{\rm pions}^{(3)} + V_{\rm pions}^{(4)} \cdots \; , \nonumber \\
V_{\rm cont} & = & V_{\rm cont}^{(0)} + V_{\rm cont}^{(2)} +V_{\rm cont}^{(4)} + \cdots \; ,
\label{pi+ct}\end{aligned}$$ where the superscript numbers in parentheses indicates the order in the chiral expansion. Note that in Eq. (\[pi+ct\]) for the contact contribution, all the odd powers cancel due to symmetry requirements. Hence, in the Weinberg power counting scheme, there is no contact interaction $V_{\rm cont}^{(3)}$ at the third order in the chiral expansion (N2LO). This actually breaks the order-by-order improvement of the chiral expansion when going from NLO to N2LO as shown very clearly in Ref. [@JPG12]. Nogga, Timmermans and van Kolck [@NTvK], looked at the interaction at leading order (LO) and found that additional counter terms that are not predicted by the Weinberg power counting are required in order to improve the description. Later, Valderrama showed that there are also problems in higher orders and they could be treated perturbatively [@mnl11].
For finite cutoff, however, Epelbaum et al. [@egm00]) showed that a better description of the phase-shifts is obtained when N2LO instead of NLO interactions are used (with the same contact terms). The same conclusion can be found by studying the $\chi^2 / {\rm datum}$ in E. Marji et al. [@marji13].
Now let us turn to the Renormalization of N3LO interactions with five subtractions. This approach introduces a renormalization scale $\mu$ (subtraction point), which denotes the momentum at which the subtractions are performed. For a given energy $E$ the Lippmann-Schwinger (LS) equation for the $T$-matrix in operator form is written as $$\begin{aligned}
T(E) = V + V~G_{0}^{+}(E)~T(E)\; ,
\label{LSE}\end{aligned}$$ where $V$ is the $NN$ potential at a given order in the chiral expansion and $G_{0}^{+}$ is the free Green’s function which, in terms of the free Hamiltonian $H_0$, is given by $$\begin{aligned}
G_{0}^{+}(E)= \frac{1}{E - H_0 + i \epsilon} \; .\end{aligned}$$
When bare potentials are introduced in the equation above, an ultraviolet divergence arises due to the implicit integral in the second term of the right-hand side of Eq. (\[LSE\]), which diverges when the momentum goes to infinity. In the standard cutoff procedure, the $NN$ potential $V$ is multiplied by a regularising function, $$V(p, p') \rightarrow V_{\Lambda}(p, p') \equiv \exp[-(p/\Lambda)^{2r}]~ V(p, p')~ \exp[-(p'/\Lambda)^{2r}] \; ,
\label{SCR}$$ where $\Lambda$ is the cutoff scale and $r \geq 1$. This function suppresses contributions from larger momenta, eliminating the ultraviolet divergences in the momentum integral. Non-relativistic nucleon-nucleon potentials based on chiral effective field theory with cutoff regularization provide a very accurate description of $NN$ scattering data below pion production threshold $E_{\rm lab} \sim 350~{\rm MeV}$.
The renormalization with multiple subtractions handles this problem in a different way since the $NN$ potential is not modified in favour of changing the Green’s function instead. The N3LO interactions require five subtractions to be renormalised with no cutoff and in this case the subtracted scattering equation is given by $$T^{(5)}_{\mu}(E) = V^{(5)}_{\mu}(E) + V^{(5)}_{\mu}(E)~G_{5}^{+}(E;-\mu^2)~T^{(5)}_{\mu}(E) \; ,
\label{LS5}$$ where $\mu$ is the subtraction scale, $V^{(n)}_{\mu}(E)$ is the driving term $$\begin{aligned}
V^{(n)}_{\mu}(E) = V^{(n-1)}_{\mu}(E) + V^{(n-1)}_{\mu}(E) \frac{(-\mu^2-E)^{n-1}}{(E-q^2)^{n}} V^{(n)}_{\mu}(E) \; ,\end{aligned}$$ which has to be calculated recursively, $G_{5}^{+}(E;-\mu^2)$ is the $5$-times subtracted Green’s function and $q$ is the the relative intermediate two nucleon momentum. $$\begin{aligned}
G_{5}^{+}(E) &=& {\cal F}_{5}(E;-\mu^2)~G_{0}^{+}(E) \; ,
\label{G5}\end{aligned}$$ where $${\cal F}_{5}(E;-\mu^2) = \left(\frac{\mu^2+E}{\mu^2+H_0} \right)^{5}$$ is a term that arises from the recursive nature of the Renormalization process and works as a form factor, being responsible for providing a regular $T$-matrix. Detailed expressions for the integral equations in the recursive calculation with partial-wave basis are given in Ref. [@JPG12] for the case of N2LO interactions with four subtractions.
The LS equation with five subtractions Eq. (\[LS5\]) has the same operator structure as the original equation Eq. (\[LSE\]), with the effective $NN$ potential $V$ replaced by the driving term $V^{(5)}_{\mu}(E)$ and the free Green’s function $G_{0}^{+}(E)$ replaced by the Green’s function with five subtractions $G_{5}^{+}(E;-\mu^2)$. The recursive driving term encodes the physical information apparently lost due to the removal of the propagation through intermediate states at the subtraction point $\mu$. Then, once the driving term is determined for a particular subtraction point, the subtracted Lippmann-Schwinger equation provides a renormalised solution for the $T$-matrix at any given energy $E$.
The driving terms $V^{(n)}_\mu$ are built recursively with the components of the nucleon-nucleon interaction. Here we use non-regulated N3LO interactions from Entem and Machleidt (EM) [@mach] and from Epelbaum, Glöckle and Meissner (EGM) [@epel]. The main difference between the two chiral forces is the Two-Pion Exchange part. The EGM potential uses Spectral Function Regularization (SFR) for the pion loop integrals resulting in a softer TPE component. Differences in the pionic part will then be compensated by changes in the Low-Energy Constants so that in the end the two forces give similar descriptions for the $NN$ system. Note that in the case of the EGM potential the SFR is still present for the loop integrals, but there is no smooth regulator function to suppress large momentum contributions. The smooth cutoff is also removed from the EM potential so that the interactions we are using are the original EM and EGM interactions with their cutoffs removed. Once the contact interactions are determined for both potentials, even with different off-shell behavior, the resulting on-shell scattering amplitudes are similar. Thus, we expect comparable results for cutoff regularization and subtractive renormalization at low energies. At high energies, $E_{\rm lab} > 200~{\rm MeV}$ the cutoff scheme is more efficient than the renormalization with multiple subtractions as far as describing the phase-shifts is concerned.
Numerical Results
=================
Here we work in a partial-wave relative momentum space basis and compute the phase shifts in each peripheral channel. However, a three-dimensional approach without any partial wave decomposition have also been employed [@hadi].
For numerical reasons, when implementing the Renormalization procedure, we solve the subtracted LS equation for the $K$-matrix using the principal value prescription. We then compute the neutron-proton phase-shifts for channels with angular momentum in the range $3 \leq J \leq 6$ with the N3LO potentials EM and EGM. Expressions for the phases-shits as functions of the on-shell $K$-matrix in coupled and uncoupled channels are given in Refs. [@batista].
For each wave, we consider several renormalization scales up to $\mu = 10~{\rm fm}^{-1}$, limiting the momentum integrations at $\Lambda = 30 ~{\rm fm}^{-1}$. In practice this means we have an infinite cutoff and the renormalization is completely imposed by the five subtractions, unlike in Refs. [@phill] where the cutoff still plays a role since only one subtraction was performed for the N2LO potential which requires four subtractions to allow an infinite cutoff.
In the case of the EGM potential, we have used an SFR cutoff of $\tilde\Lambda = 4~m_\pi (550~{\rm MeV})$, the most common choice. The only parameter of the EGM potential we changed was the cutoff $\Lambda\to\infty$. The EGM interaction depend on the SFR cutoff and the results are different if the SFR cutoff is modified, but this dependence is not related to the subtractions. We also believe that it is the SFR that drives the difference between the EM and EGM potentials as far as the renormalization scale dependence is concerned. Nevertheless, here we treated the SFR cutoff as an internal parameter of the EGM potential and looked only at the dependence of the phases on $\mu$.
The results for the N3LO-EM potential are displayed in Figures \[fig1\] to \[fig4\] for the uncoupled channels and in Figures \[fig5\] and \[fig6\] for the coupled channels. For the N3LO-EGM potential, the results are displayed in Figures \[fig7\] to \[fig10\] for the uncoupled channels and in Figures \[fig11\] and \[fig12\] for the coupled channels. The phase-shifts are compared to the Nijmegen partial wave analysis [@Nij]. For un updated high quality partial wave and error analysis, see the works from the Granada group. [@GRpwa].
Note that in the end we used different ranges of renormalization scales for EM and EGM and the reason is that the results for EGM converges faster as far as $\mu$ is concerned and there is no need to go above $2~{\rm fm}^{-1}$. In the case of EM, since there are still some variations in few waves for $\mu > 2~{\rm fm}^{-1}$ we extended the range of the subtraction point to $10~{\rm fm}^{-1}$. We believe this is related to the differences between EM and EGM that makes EGM softer than EM.
With few exceptions, in most of the channels we observe very small variation of the phases as we change the subtraction point, indicating that peripheral waves are nearly renormalization group invariant or fixed-points of the subtractive renormalization group. The exceptions are the $^3H_4$ and $^3I_5$ waves, which show slower convergence and the $^3G_5$ channel where the agreement with the partial wave analysis is only up to $\sim 50~{\rm MeV}$. Also, in the case of the EM potential, the phases for the $^3F_3$ wave present some oscillations when the renormalization scale approaches $\mu \sim 10 ~ {\rm fm}^{-1}$.
In this case of the coupled channels, the renormalization scheme is not failing but just requiring a slightly larger $\mu$ in the coupled channels, where we have the very singular tensor force. And this is what is different in the these channels compared to other F waves. The case of the uncoupled triplet 3F3 is somewhat different: the oscillation observed when going from $\mu = 6~ {\rm to}~ 9~{\rm fm}^{-1}$ is related to the TPE without SFR of the EM potential. This oscillation is not seen in the 3F3 wave when the EGM potential is used. An additional subtraction doesn’t modify the results and the $\mu$ dependence in the these channels, so only the minimum number of subtractions need to be performed (five at N3LO).
It is important to mention that the $\mu$-dependence of the phase-shifts is encoded in the recursive driving terms $V^{(n)}_\mu$ since they all depend on the renormalization scale. In the case of cutoff regularization in configuration space [@mnl06], the cutoff radius $R_S$ dependence has been traced to the most singular part of the interaction which can be attraction or repulsion, depending on the channel. Here we observe more dependence on the coupled channels due to the very singular tensor force and the $F$-wave issue is due to the differences between the TPE components of the EM and EGM potentials.
Conclusions and Outlook
=======================
So far we have renormalized the N3LO interactions with multiple subtractions in peripheral channels considering an infinite cutoff $\Lambda = 30 ~ {\rm fm}{-1}$. Only pions contribute to the nuclear force in these waves and the results are parameter free.
The five subtractions performed in the kernel of the LS equation provide finite $K$-matrix in peripheral waves and the resulting phase shifts are rather independent of the subtraction point with the exception of the $^3H_4$ and $^3I_5$ waves where the fixed point is reached at approximately $\mu \sim 1 ~ {\rm fm}^{-1}$ after the slow convergence shown in Figs. \[fig5\], \[fig6\], \[fig11\] and \[fig12\].
The oscillation in the $^3F_3$ channel suggests a closer look at this wave with the renormalization group flow equation that governs the driving terms (interactions) as the subtraction point is changed with the constraint of an invariant scattering amplitude.
The advances given by our approach when compared to cutoff regularization are: the NN force doesn’t have to be modified prior its insertion in the scattering equation; only the scattering equation is modified; the method is renormalization group invariant by construction and provides a non-relativistic flow equation for the driving terms that tell us how they change when the renormalization scale is modified in order to keep the amplitude invariant. Finally, we would like to point out that the cutoff looses any physical significance since our results were obtained with an extremely large value $\Lambda = 30 ~ {\rm fm}^{-1}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank financial support from FAPESP (grant 2016/07061-3), CNPQ (grant 306195/2015-1) and FAEPEX (grant 3284/16).
[99]{} S. Weinberg, [*Phys. Lett.*]{} B [**251**]{} (1990) 288; S. Weinberg, [*Nucl. Phys.*]{} B [**363**]{} (1991) 3; S. Weinberg, [*Phys. Lett.*]{} B [**295**]{} (1992) 114. C. Ordóñez, L. Ray and U. van Kolck, [*Phys. Rev. Lett.*]{} [**72**]{} (1994) 1982; [*Phys. Rev.*]{} C [**53**]{} (1996) 2086. S. K. Adhikari, T. Frederico and I. D. Goldman, [*Phys. Rev. Lett.*]{} [**74**]{} (1995) 487; S. K. Adhikari and T. Frederico, [*Phys. Rev. Lett.*]{} [**74**]{} (1995) 4572; S. K. Adhikari and A. Ghosh, [*J. Phys*]{} A [**30**]{} (1997) 6553. D. B. Kaplan, M. J. Savage and M. B. Wise, [*Phys. Lett.*]{} B [**424**]{} (1998) 390; D. B. Kaplan, M. J. Savage and M. B. Wise, [*Nucl. Phys.*]{} B [**534**]{} (1998) 329; D. B. Kaplan, M. J. Savage and M. B. Wise, [*Nucl. Phys.*]{} B [**478**]{} (1996) 629. P. F. Bedaque, H. W. Hammer and U. van Kolck, [*Phys. Rev. Lett.*]{} [**82**]{} (1999) 463; B. Long and U. van Kolck, [*Annals Phys.*]{} [**323**]{} (2008) 1304. A. Nogga, R. G. E. Timmermans and U. van Kolck, [*Phys. Rev.*]{} C [**72**]{} (2005) 054006. M. C. Birse, J. A. McGovern and K. G. Richardson, [*Phys. Lett.*]{} B [**464**]{} (1999) 169; T. Barford and M. C. Birse, [*Phys. Rev.*]{} C [**67**]{} (2003) 064006; K. G. Richardson, M. C. Birse and J. A. McGovern, hep-ph/9708435; M. C. Birse [*Phys. Rev.*]{} C [**77**]{} (2008) 047001. M. Pavon Valderrama and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**74**]{} (2006) 054001; M. Pavon Valderrama and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**74**]{} (2006) 064004; M. Pavon Valderrama and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**72**]{} (2005) 054002; M. Pavon Valderrama and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**70**]{} (2004) 044006. D.R. Entem, E. Ruiz Arriola, M. Pavon Valderrama and R. Machleidt, [*Phys. Rev.*]{} C [**77**]{} (2008) 044006; M. P. Valderrama, [*Phys. Rev.*]{} C [**84**]{}, 064002 (2011). E. Epelbaum, H. Krebs and U. -G. Mei[ß]{}ner, [*Eur. Phys. J.*]{} A [**51**]{} (2015) 53. T. Frederico, V. S. Timóteo and L. Tomio, [*Nucl. Phys.*]{} A [**653**]{} (1999) 209. T. Frederico, A. Delfino and L. Tomio, [*Phys. Lett.*]{} B [**481**]{} (2000) 143. V. S. Timóteo, T. Frederico, L. Tomio and A. Delfino, [*Phys. Lett.*]{} B [**621**]{} (2005) 109. V. S. Timóteo, T. Frederico, L. Tomio and A. Delfino, [*Nucl. Phys.*]{} A [**790**]{} (2007) 406c. V. S. Timóteo, T. Frederico, L. Tomio and A. Delfino, [*Int. J. Mod. Phys.*]{} E [**16 (9)**]{} (2007) 2822. V. S. Timóteo, T. Frederico, A. Delfino and L. Tomio, [*Phys. Rev.*]{} C [**83**]{} (2011) 064005. S. Szpigel and V. S. Timóteo, [*J. Phys.*]{} G [**39**]{} (2012) 105102. E. Epelbaum, W. Glöckle, and U. -G. Mei[ß]{}ner, [*Nucl. Phys.*]{} A [**671**]{} (2000) 295; [*Nucl. Phys.*]{} A [**637**]{} (1998) 107. E. Marji, A. Canul, Q. MacPherson, R. Winzer, Ch. Zeoli, D.R. Entem, R. Machleidt, [*Phys. Rev.*]{} C [**88**]{} (2013) 0540002. S. Bayegan, M. A. Shalchi and M. R. Hadizadeh, [*Phys. Rev.*]{} C [**79**]{} (2009) 304008; S. Bayegan, M. Harzchi and M. R. Hadizadeh, [*Nucl. Phys.*]{} A [**814**]{} (2008) 21; M. R. Hadizadeh and L. Tomio, [*Few. Body Syst.*]{} [**54**]{} (2013) 2227. D. R. Entem and R, Machleidt, [*Phys. Rev.*]{} C [**68**]{} (2003) 041001; R. Machleidt and D. R. Entem, [*Phys. Rept.*]{} [**503**]{} (2011) 1. E. Epelbaum, W. Glöckle, and U. -G. Mei[ß]{}ner, [*Nucl. Phys.*]{} A [**747**]{} (2005) 362; E. Epelbaum, [*Prog. Part. Nucl. Phys.*]{} [**57**]{} (2006) 654. E. Epelbaum, H. W. Hammer and U. -G. Mei[ß]{}ner, [*Rev. Mod. Phys.*]{} [**81**]{} (2009) 1773. E. F. Batista, S. Szpigel and V. S. Timóteo, AIP Conf. Proc. [**1625**]{} (2014) 205; E. F. Batista, S. Szpigel and V. S. Timóteo, Few-Body Syst. [**55**]{} (2014) 1049; E. F. Batista, S. Szpigel and V. S. Timóteo, J. Phys. Conf. Ser. [**630**]{} (2015) 012056. C. J. Yang, Ch. Elster and D. R. Phillips, [*Phys. Rev.*]{} C [**77**]{} (2008) 014002; C. J. Yang, Ch. Elster and D. R. Phillips, [*Phys. Rev.*]{} C [**80**]{} (2009) 034002; C. J. Yang, Ch. Elster and D. R. Phillips, [*Phys. Rev.*]{} C [**80**]{} (2009) 044002. V. G. J. Stoks, R. A. M. Klomp, C. P. F. Terheggen and J. J. de Swart, [*Phys. Rev.*]{} C [**49**]{} (1994) 2950. R. Navarro Perez, J. E. Amaro and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**88**]{} (2013) 024002; [*Phys. Rev.*]{} C [**88**]{} (2013) 069902; [*Phys. Rev.*]{} C [**88**]{} (2013) 064002; [*Phys. Rev.*]{} C [**91**]{} (2015) 029901; [*Phys. Rev.*]{} C [**89**]{} (2014) 064006. M. Pavon Valderrama and E. Ruiz Arriola, [*Phys. Rev.*]{} C [**74**]{}, 064004 (2006).
\
\
\
\
\
\
\
\
\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_1I6_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_3I6_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_1I6_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_3I6_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_1I6_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_3I6_rs1-5 "fig:")\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_1I6_rs6-9 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ uncoupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig4"}](EM_3I6_rs6-9 "fig:")
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3F4_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3H4_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_E4_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3F4_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3H4_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_E4_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3F4_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3H4_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_E4_rs1-5 "fig:")\
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3F4_rs6-9 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_3H4_rs6-9 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig5"}](EM_E4_rs6-9 "fig:")
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3G5_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3I5_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_E5_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3G5_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3I5_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_E5_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3G5_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3I5_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_E5_rs1-5 "fig:")\
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3G5_rs6-9 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_3I5_rs6-9 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig6"}](EM_E5_rs6-9 "fig:")
\
\
\
\
\
\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_1I6_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_3I6_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_1I6_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_3I6_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_1I6_rs1-2 "fig:")![(Color on-line) Phase-shifts in the $^1I_6$ and $^3I_6$ channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig10"}](EGM_3I6_rs1-2 "fig:")
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3F4_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3H4_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_E4_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3F4_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3H4_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_E4_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3F4_rs1-5 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_3H4_rs1-2 "fig:")![(Color on-line) Phase-shifts in the $^3F_4 - ^3H_4$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig11"}](EGM_E4_rs1-2 "fig:")
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3G5_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3I5_rs01-05 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_E5_rs01-05 "fig:")\
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3G5_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3I5_rs06-09 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_E5_rs06-09 "fig:")\
![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3G5_rs1 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_3I5_rs1 "fig:")![(Color on-line) Phase-shifts in the $^3G_5 - ^3I_5$ coupled channels calculated from the solution of the subtracted LS equation for the $K$-matrix with five subtractions for the N3LO-EGM potential for several values of the renormalization scale compared to the Nijmegen partial wave analysis.[]{data-label="fig12"}](EGM_E5_rs1 "fig:")
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Offline handwritten mathematical expression recognition is often considered much harder than its online counterpart due to the absence of temporal information. In order to take advantage of the more mature methods for online recognition and save resources, an oversegmentation approach is proposed to recover strokes from textual bitmap images automatically. The proposed algorithm first breaks down the skeleton of a binarized image into junctions and segments, then segments are merged to form strokes, finally stroke order is normalized by using recursive projection and topological sort. Good offline accuracy was obtained in combination with ordinary online recognizers, which are not specially designed for extracted strokes. Given a ready-made state-of-the-art online handwritten mathematical expression recognizer, the proposed procedure correctly recognized 58.22%, 65.65%, and 65.22% of the offline formulas rendered from the datasets of the Competitions on Recognition of Online Handwritten Mathematical Expressions(CROHME) in 2014, 2016, and 2019 respectively. Furthermore, given a trainable online recognition system, retraining it with extracted strokes resulted in an offline recognizer with the same level of accuracy. On the other hand, the speed of the entire pipeline was fast enough to facilitate on-device recognition on mobile phones with limited resources. To conclude, stroke extraction provides an attractive way to build optical character recognition software.'
author:
- 'Chungkwong Chan[^1] [^2]'
bibliography:
- 'main.bib'
title: 'Stroke extraction for offline handwritten mathematical expression recognition [^3]'
---
Keywords
: Character recognition; feature extraction; offline handwritten mathematical expression recognition; optical character recognition software; stroke extraction
2010 MSC
: 68T10; 68T45; 68U10
Introduction {#intro}
============
Mathematical expressions appear frequently in engineering and scientific documents. Since they contain valuable information, digitizing them would maximize their usability by enabling retrieval[@Zanibbi2012Recognition], integration to semantic web[@Marchiori2003], and other automated tasks. Compared with ordinary text, mathematical expressions can present some concepts more concisely because of their two-dimensional structure. However, such compact representations are difficult to be recognized mechanically.
People are used to writing mathematical expressions on paper or blackboard, so it is inconvenience to input them by using another way. Traditional input devices like keyboards are designed for sequence of characters, although spatial relations between symbols can be represented by markups such as TeX or MathML, entering mathematical expressions by typing in a computer language is not user-friendly at all, as new users are asked to learn a new language, remember a lot of commands and deal with miscellaneous errors. On the other hand, entering mathematical expressions with a graphical equation editor by choosing structural elements and symbols from toolboxes is inefficient for frequent users.
A recognition system turns handwritten mathematical expressions into machine manipulable syntax trees. Online recognition enables people to take notes or solve equations by writing on a touch-based device or dragging a mouse. On the other hand, offline recognition enables people to digitize existing manuscripts by scanning or record lecture notes on blackboards by taking photos. The main difference between two kinds of recognition is that trajectories are available to an online recognizer, whereas only bitmap images are given to an offline recognizer.
It is not surprising that online handwriting recognition systems often achieve a much higher accuracy. In fact, an online recognition problem can be converted to the corresponding offline problem by simply rendering the strokes. In the opposite direction, if the sequence of strokes can be recovered from a bitmap image, an online recognizer can also be applied to do offline recognition[@NISHIDA19951213]. The objective of this paper is to explore the feasibility of offline handwriting recognition via stroke extraction without designing specialized online recognizers.
The main concern is that recovery of strokes is unlikely to be perfect, so it may become a single point of failure. Fortunately, good online recognition systems can tolerate some input errors, since diversity of writing habits already affected them. Furthermore, if the underlying online recognition system has been retrained with extracted strokes, it should be able to adapt to them. In this case, accuracy of stroke extraction needs not be critical to that of offline recognition.
The main contributions of this paper include:
1. We propose a stroke extraction algorithm for handwritten mathematical expression, which enables reduction from offline recognition to online recognition.
2. We verify that the accuracy of the proposed system is comparable to that of other recent offline handwritten mathematical expression recognition systems, given a ready-made state-of-the-art online recognizer.
3. We demonstrate that the gap in accuracy between the proposed offline system and the underlying online system can be narrowed significantly by retraining.
The remainder of this paper is divided into four sections. Section \[sec:related\] reviews methods to recognize online and offline mathematical expressions, as well as methods to recover trajectories. Section \[sec:algorithm\] describes the proposed stroke extraction algorithm for mathematical expression in detail. Section \[sec:evaluation\] presents experimental results on standard datasets. Section \[sec:conclusion\] concludes the paper.
Related works {#sec:related}
=============
Online handwritten mathematical expression recognition {#subsec:relatedOnline}
------------------------------------------------------
Online handwritten mathematical expression recognition is a long-standing problem, a lot of works have been done since Anderson[@Anderson1967Syntax]. In the past decade, the problem attracted more and more attention because of the Competitions on Recognition of Online Handwritten Mathematical Expressions(CROHME)[@Mouch2016Advancing], which were held in 2011, 2012, 2013, 2014, 2016, and 2019.
Traditionally, the problem is further divided into symbol recognition and structural analysis[@Zanibbi2012Recognition]. For example, Álvaro et al.[@ALVARO201458] applied hidden Markov model to recognize the symbols and parser for a predefined two-dimensional stochastic context free grammar to analyze the structure.
In order to disambiguate symbols with structural information and vice versa, several approaches were proposed to integrate symbol recognition and structural analysis closely. Yamamoto et al.[@yamamoto] suggested parsing handwritten mathematical expression directly from strokes by using the Cocke-Younger-Kasami algorithm, so that symbol segmentation, character recognition, and structural analysis could be optimized simultaneously. Awal et al.[@AWAL201468] introduced another global approach that applies a segmentation hypothesis generator to deal with delayed strokes.
Recently, with the advances in computational power, end-to-end trainable recurrent neural networks became popular because of its ability to learn complex relations. For example, Zhang et al.[@8373726] developed an encoder-decoder framework for online handwritten mathematical expression recognition named Track, Attend, and Parse(TAP), which employs a guided hybrid attention mechanism.
Offline handwritten mathematical expression recognition {#subsec:relatedOffline}
-------------------------------------------------------
On the contrary, dedicated work on offline handwritten mathematical expression recognition was almost blank in literature until very recently. An offline task was first added to CROHME in 2019[@CROHME2019].
In the past, the closest problem addressed was the more constrained problem of printed mathematical expression recognition[@Garain2007OCR]. Again, in a typical system, symbols are first segmented and recognized, then the structure of the expression is analyzed[@Chan2000]. For instance, in the state-of-the-art system developed by Suzuki et al. [@Suzuki2003INFTY], symbols are extracted by connected component analysis and then recognized by a nearest neighbor classifier, after that, structural analysis is performed by finding a minimum spanning tree in a directed graph representing spatial relations between symbols.
Recently, Deng et al. [@pmlr-v70-deng17a] and Zhang et al. [@ZHANG2017196; @8546031] developed end-to-end trainable neural encoder-decoder models to translate image of mathematical expression into TeX code directly. They are quite similar to models for online recognition, except that convolutional neural networks are prepended to extract features. It should be noted that this method is so general that it can be applied to any image-to-markup problem, grammar of neither mathematical expression nor TeX is required by those systems explicitly because they can be learned from data.
Stroke extraction {#subsec:relatedStroke}
-----------------
Stroke extraction was studied for offline signature verification[@156588] and East Asian character recognition[@LIU20012339]. A typical stroke extractor detects candidates of sub-strokes first and then reassembles them into strokes by resolving ambiguities. Sub-strokes can be detected by breaking down the skeleton or approximating the image with geometrical primitives such as polygonal chains.
Lee et al.[@156588] designed a set of heuristic rules to trace the skeleton. Boccignone et al.[@BOCCIGNONE1993409] tried to reconstruct strokes by joining the pair of adjoining sub-strokes having the smallest difference in direction, length, and width repeatedly. Doermann et al.[@Doermann1995] proposed a general framework to integrate various temporal clues.
Jäger[@546812] reconstructed strokes by minimizing total change in angle between successive segments within a stroke. Lau et al.[@Lau2002Stroke] selected another cost function taking distance between successive segments and directions of the segments into account. Unfortunately, this kind of formulations is essentially traveling salesman problem which is NP-complete, so computing the optimum efficiently may not be possible if there are more than a few sub-strokes.
In order to prevent explosion of combinations, Kato et al.[@877517] restricted themselves to single-stroke script subjecting to certain assumptions on junctions, so that strokes can be extracted by traversal of graph. Nagoya et al.[@6424482] extended the technique to multi-stroke script under assumptions on how strokes are intersected.
Nevertheless, the effectiveness of stroke extraction to offline recognition has remained largely untested. Evaluations of existing stroke extraction methods were often performed visually or quantitatively with their only metrics on small private datasets. Resulting accuracy of offline recognition was seldom reported and limited to single character recognition, where the challenge of symbol segmentation was not addressed. Moreover, they strongly rely on specially designed structural matching methods, which can tolerate a variety of deformations, so they may not work well with ordinary methods for online recognition.
Offline to online reduction {#sec:algorithm}
===========================
Overview {#subsec:overview}
--------
Given a bitmap image containing a mathematical expression, it must be converted to a sequence of strokes before being passed to an online handwritten mathematical expression recognition engine. In more detail, key steps of the proposed offline handwritten mathematical expression recognition system are:
1. Adaptive binarization. Convert the possibly colored input image into a black and white image.
2. Stroke width transformation. Estimate stroke width for each foreground pixel.
3. Thinning. Skeletonize the binary image.
4. Decomposition of the skeleton. Break it down into segments and junctions.
5. Construction of an attributed graph. Segments and junctions form edges and vertexes of the graph respectively.
6. Simplification of the attributed graph. Remove vertexes and edges which are likely noises from the graph.
7. Reconstruction of strokes. Merge segments into strokes by a bottom up clustering.
8. Fixing double-traced strokes. Reuse some segments to join separated strokes.
9. Determination of stroke direction. Ensure that the order of points in each stroke conforms to the usual practice.
10. Stroke order normalization. Sort the strokes by the time they are expected to be written.
11. Online recognition. Invoke an online handwritten mathematical expression recognition engine to recognize the sequence of strokes extracted.
Preprocessing {#subsec:preprocessing}
-------------
Since skeleton roughly preserves the shape of strokes but much simpler, it is easier to trace strokes from the skeleton instead of the full image. Before skeletonization, a colored image should be binarized. The input image is first converted to a grayscale image by averaging the color channels(possibly weighted). Among the large number of binarization methods available, Sauvola’s method[@Sauvola2000Adaptive] is chosen. Compared with global thresholding techniques such as Otsu’s method[@4310076], such a local adaptive approach addressed commonly seen degradations including uneven illumination and random background noises. However, pixels that do not belong to the mathematical expression may still be marked foreground, text next to the expression and grid lines on a notebook for instance. Mathematical expression localization and separation are out of the scope of this paper.
After binarization, skeleton of the image is obtained by using a thinning method by Wang and Zhang [@Wang1989A], which is a variant of the original method by Zhang and Suen[@osti5085064] but preserves the shape of diagonal strokes better. Figure \[fig:thinning\] compares an image with its skeleton.
![Thinning[]{data-label="fig:thinning"}](chan1.pdf "fig:"){width=".5\linewidth"} \[fig:beforeThin\]
![Thinning[]{data-label="fig:thinning"}](chan2.pdf "fig:"){width=".5\linewidth"} \[fig:afterThin\]
For printed document recognition, skew detection and correction are often performed. However, they should not be applied to a single mathematical expression because the number of symbols may not be enough to estimate the angle reliably. To make thing worse in the present situation, symbols from a handwritten formula need not stick to a single baseline, so expressions like “$x^{x^x}$” may fool skew estimators based on line detection like Hough transformation.
Decomposition of skeleton {#subsec:separation}
-------------------------
After skeletonization, the skeleton is decomposed into segments and junctions, thus the skeleton can be viewed as a graph.
A foreground pixel having exactly two other foreground pixels in its 8-neighborhood where the two pixels are not 4-neighbor of each other is called a segment pixel. Other foreground pixels are called junction pixels. Figure \[fig:separation\] illustrates the rule.
![Identification of segment pixels and junction pixels[]{data-label="fig:separation"}](chan3.pdf){width="\linewidth"}
\
![Identification of segment pixels and junction pixels[]{data-label="fig:separation"}](chan4.pdf){width="\linewidth"}
A connected component of the set of segment pixels is called a segment. On the other hand, a connected component of the set of junction pixels is called a junction. The set of segments and the set of junctions can be computed by using any standard algorithm for connected component analysis[@HE201725]. In Figure \[fig:components\], segments are solid but junctions are not.
![Decomposition of skeleton into segments and junctions[]{data-label="fig:components"}](chan5.pdf){width="0.25\linewidth"}
For each segment $S_i$, its pixels can be listed in a way such that successive pixels are 8-neighbor of each other, more formally, $$S_i=\{p_{i,1},\dots,p_{i,\ell_i}\}$$ where $p_{i,k}$ is in the 8-neighborhood of $p_{i,k-1}$ for $k=2,\dots,\ell_i$. If $p_{i,1}$ is in the 8-neighborhood of $p_{i,\ell_i}$, the segment is topologically a circle and does not touch any other junction or segment unless $\ell_i=1$; otherwise, the segment is topologically a line segment, $p_{i,1}$ touches exactly one junction and so do $p_{i,\ell_i}$, other pixels in the segment never touch any other segment or junction.
For the sake of consistency, a “junction” is imposed to each looped segment to ensure that every segment has a start pixel and an end pixel, in addition, each touches a junction. Therefore, a junction can be considered as a vertex in the sense of graph theory, whereas a segment can be considered as an edge connecting two (possibly the same) vertexes. Furthermore, a path in this undirected graph corresponds to a possible trace of ink in the input image, a connected component in this graph corresponds to a connected component of the skeleton. Figure \[fig:rawGraph\] shows the graph coming from the same example as in Figure \[fig:components\].
Noise reduction {#subsec:noise}
---------------
Subtle features such as salt and pepper noises in the binarized image can affect the skeleton, salt noises result in really short segments whereas pepper noises result in isolated junctions. In addition, thinning may introduce distortions. Since they can distract stroke extractor and recognition engine, they should be discarded from the graph. Absolute threshold values are not used because they do not work in all resolutions. Observed that stroke width is uniform in a piece of handwriting, it is chosen to be a reference length.
Stroke width transformation is an image operator that assigns an estimated stroke width to each foreground pixel. It was proposed for scene text detection[@5540041], where strokes are considered as contiguous pixels having approximately constant stroke width locally. Using a straightforward viewpoint, stroke width of a pixel can be estimated by the minimum length of the four directional runs[@FAN20001881] passing through it as shown in Figure \[fig:swt\], where squares represent foreground pixels and arrows represent directional runs of the solid pixel. Under the above definition, stroke width transformation can be computed in linear time with respect to the size of binary image by caching the numbers of successive foreground pixels found in certain directions.
![Stroke width is the minimum of four directional run lengths[]{data-label="fig:swt"}](chan6.pdf){width="\linewidth"}
For each set of pixels, its width is estimated by the maximum stroke width among its pixels. Furthermore, the tip size of the pen is estimated by the average stroke width over all the segments. Now, the rules to reduce noises can be stated:
1. For each edge with a length smaller than a multiple of the average stroke width, remove it from the graph and merge its end points.
2. For each vertex with degree 0 and a width less than a multiple of the average stroke width, remove it from the graph.
Figure \[fig:simplifiedGraph\] shows the simplified graph coming from the same example as in Figure \[fig:components\].
![Skeleton graph[]{data-label="fig:simplify"}](chan7.pdf "fig:"){width="0.5\linewidth"} \[fig:rawGraph\]
![Skeleton graph[]{data-label="fig:simplify"}](chan8.pdf "fig:"){width="0.5\linewidth"} \[fig:simplifiedGraph\]
Stroke tracing {#subsec:tracing}
--------------
Clearly, an isolated vertex in the graph represents a dot in the mathematical expression, possibly a decimal point or part of a character like “i”. Therefore, a stroke containing a single point is extracted for each vertex with degree 0. In addition, a path in the skeleton graph indicates a candidate of stroke. Although there may be multiple ways to combine the edges into paths, some combinations are more likely to form strokes of a mathematical expression written by human being. Here are some heuristic principles:
- The total number of strokes should be minimal. Since letting the pen to leave the paper requires additional time, an unicursal way is preferred.
- The difference in directions between two successive segments should be as small as possible. Since turning suddenly requires slowing down, a fluent stroke is better to be smooth.
Subjecting to these considerations, each edge is assigned to exactly one path by a bottom up clustering. Initially, each edge forms a path on its own. While there is a pair of paths having a common end point, choose a pair such that the angle between them is the minimum, then merge them into one path. Repeat the procedure until no path can be merged.
It should be noted that the two principles may not always agree. If the number of strokes is considered more important, its minimum can be obtained by merging each path with circuits that have a common vertex with it, just like the algorithm that search for an Eulerian path.
Fixing double-traced strokes {#subsec:double}
----------------------------
Sometimes, a segment should be shared by more than one strokes or appeared in a stroke multiple times due to reentry during writing as shown in Figure \[fig:strokeMulti\]. The tracing procedure above would handle such cases incorrectly by producing too many strokes as in Figure \[fig:strokeSingle\].
![Identification of double-traced stroke[]{data-label="fig:fix"}](chan9.pdf "fig:"){width="\linewidth"} \[fig:strokeMulti\]
![Identification of double-traced stroke[]{data-label="fig:fix"}](chan10.pdf "fig:"){width="\linewidth"} \[fig:strokeSingle\]
In order to fix the double-traced strokes, a search for shared segments is needed, so that they can be used to reconnect separated strokes. Candidates of shared segments should meet the following criteria:
- The segment has two different end points, which are vertexes in the graph with an odd degree. Otherwise, the number of vertexes in the graph having an odd degree does not decrease when it is doubled.
- Each end point of the segment is also an end point of a path given by subsection \[subsec:tracing\] and the angle between them is not close to $\pi/2$. This condition can prevent the two strokes of the symbol “T” from being merged.
Stroke order normalization {#subsec:ordering}
--------------------------
Firstly, stroke direction detection is performed. The stroke tracing procedure gives rise to an ordering of points within a stroke naturally, however, the opposite ordering may also make sense. Since people usually write from left to right and top to bottom, a simple rule is sufficient to determine the direction of each stroke in most cases. Let the coordinates of the first and the last point of a stroke be $(x_{start},y_{start})$ and $(x_{end},y_{end})$ respectively, then the list of points should be reversed if $$2x_{end}+3y_{end}<2x_{start}+3y_{start}.$$
Finally, the order that people write down the strokes should be recovered. Although some mathematical expression recognition systems are stroke order free[@AWAL201468; @ALVARO201458], others use stroke order to prune the search space, so they are sensitive to stroke order[@SIMISTIRA201585; @Le2016]. In fact, stroke order normalization is a way to turn a stroke order dependent recognizer into a stroke order free one[@Le2019].
There are possibly multiple orders to write down the same formula. For example, someone prefers to write down the square root sign first, but the others write down the radicand first. Therefore, it is not always possible to recover the original order, what can be done is to assign a reasonable order. In some cases, mistakes made by the stroke extractor can also be viewed as a kind of normalization and may in fact enhance performance by eliminating unusual stroke order.
A hierarchical approach is applied to sort the strokes. To begin with, strokes are grouped by recursive projection, then the groups are sorted in a left to right and top to bottom manner. However, recursive X-Y cut cannot determine the order of symbols inside a square root, so strokes inside each group are sorted by a topological sort afterward. A stroke $T_i$ precedes another stroke $T_j$ if one of the following conditions holds:
- $T_i$ is on the left of $T_j$, where their projection to the $y$-axis(but not the $x$-axis) intersect;
- $T_i$ is on top of $T_j$, where their projection to the $x$-axis(but not the $y$-axis) intersect.
Further ambiguities are resolved by using the coordinates of the top left corner of the bounding boxes.
Figure \[fig:ordering\] illustrates how strokes are sorted, in which groups are separated by dotted lines and precedence relations are represented by arrows.
![Stroke order normalization[]{data-label="fig:ordering"}](chan11.pdf){width="\linewidth"}
Evaluation {#sec:evaluation}
==========
Datasets and evalation protocol
-------------------------------
The proposed system was evaluated on CROHME datasets, which are standard for mathematical expression recognition. Table \[tab:crohme\] summarizes the datasets. For each mathematical expression in a dataset, list of points in each stroke is provided together with ground truth. In addition to MathML representation of the expression, correspondence between symbols and strokes is also provided. Since a bitmap image of mathematical expression can be obtained by rendering the strokes, the datasets can be used to evaluate offline recognition systems as well. Following the settings of task 2(offline handwritten formula recognition) in CROHME 2019, formulas were rendered at resolution of $1000\times 1000$ pixels by using the script provided by the organizers. However, it should be noted that rendered images are different from scanned or camera captured mathematical expressions in the level of background noises, although removing them is another research topic.
[Xrrr]{} Year&Training&Validation&Test\
2014&8836&671&986\
2016&8836&986&1147\
2019&9993&986&1199\
Aligning with the offline task in CROHME 2019[@CROHME2019], expression level metrics computed from the symbol level label graphs of formulas were used to evaluate the proposed system. Structure rate measures the percentage of recognized expressions matching the ground truth if all the labels of symbols are ignored. Expression rate measures the percentage of recognized expressions matching the ground truth up to a certain number of labeling mistakes on symbols or spatial relations. On the other hand, stroke classification rate, symbol segmentation rate, symbol recognition rate, and metrics based on stroke level label graph are inapplicable because offline recognition does not produce correspondence between symbols and strokes.
Integration with a ready-made online recognizer {#subsec:myscript}
-----------------------------------------------
In the first experiment, the proposed stroke extractor [^4] was combined with version 1.3 of MyScript Math recognizer[^5], which is a state-of-the-art online handwritten mathematical expression recognition system, to form an offline recognizer. The online recognizer was customized with a grammar to eliminate candidates of symbols and constructs that never appeared in CROHME datasets. Table \[tab:customizedgrammar\] shows the customized grammar, where “$<start>$” is the start symbol.
[rrl]{} $<start>$ &::=& $<term>$\
$<term>$ &::=& $<operand>$\
&$|$& $<symbol2>$\
&$|$& $<symbol4>$\
&$|$& $<term> <term>$\
&$|$& $\frac{<term>}{<term>}$\
&$|$& $<operand>_{<term>}$\
&$|$& $<operand>^{<term>}$\
&$|$& $<operand>_{<term>}^{<term>}$\
&$\Big|$& $\begin{matrix}<symbol3>\\<term>\end{matrix}$\
&$\Bigg|$& $\begin{matrix}<term>\\<symbol4>\\<term>\end{matrix}$\
$<operand>$ &::=& $<symbol1>$\
&$|$& $\sqrt{<term>}$\
&$|$& $\sqrt[<term>]{<term>}$\
&$|$& $<symbol5> <term> <symbol6>$\
$<symbol1>$ &::=& ‘$0$’ $|$ ‘$1$’ $|$ ‘$2$’ $|$ ‘$3$’ $|$ ‘$4$’ $|$ ‘$5$’ $|$ ‘$6$’ $|$ ‘$7$’ $|$ ‘$8$’ $|$ ‘$9$’\
&$|$& ‘$A$’ $|$ ‘$B$’ $|$ ‘$C$’ $|$ ‘$E$’ $|$ ‘$F$’ $|$ ‘$G$’ $|$ ‘$H$’ $|$ ‘$I$’ $|$ ‘$L$’\
&$|$& ‘$M$’ $|$ ‘$N$’ $|$ ‘$P$’ $|$ ‘$R$’ $|$ ‘$S$’ $|$ ‘$T$’ $|$ ‘$V$’ $|$ ‘$X$’ $|$ ‘$Y$’\
&$|$& ‘$a$’ $|$ ‘$b$’ $|$ ‘$c$’ $|$ ‘$d$’ $|$ ‘$e$’ $|$ ‘$f$’ $|$ ‘$g$’ $|$ ‘$h$’ $|$ ‘$i$’ $|$ ‘$j$’ $|$ ‘$k$’\
&$|$& ‘$l$’ $|$ ‘$m$’ $|$ ‘$n$’ $|$ ‘$o$’ $|$ ‘$p$’ $|$ ‘$q$’ $|$ ‘$r$’ $|$ ‘$s$’ $|$ ‘$t$’ $|$ ‘$u$’\
&$|$& ‘$v$’ $|$ ‘$w$’ $|$ ‘$x$’ $|$ ‘$y$’ $|$ ‘$z$’ $|$ ‘$\Delta$’ $|$ ‘$\alpha$’ $|$ ‘$\beta$’ $|$ ‘$\gamma$’\
&$|$& ‘$\theta$’ $|$ ‘$\lambda$’ $|$ ‘$\pi$’ $|$ ‘$\sigma$’ $|$ ‘$\varphi$’ $|$ ‘$\phi$’ $|$ ‘$\mu$’ $|$ ‘$<$’ $|$ ‘$>$’ $|$ ‘$!$’\
&$|$& ‘$/$’ $|$ ‘$\infty$’ $|$ ‘$\sin$’ $|$ ‘$\cos$’ $|$ ‘$\tan$’ $|$ ‘$\lim$’ $|$ ‘$\log$’\
$<symbol2>$ &::=& ‘$+$’ $|$ ‘$-$’ $|$ ‘$\pm$’ $|$ ‘$\times$’ $|$ ‘$\div$’ $|$ ‘$\cdot$’ $|$ ‘$=$’ $|$ ‘$\prime$’ $|$ ‘$,$’ $|$ ‘$.$’\
&$|$& ‘$\ldots$’ $|$ ‘$\rightarrow$’ $|$ ‘$\exists$’ $|$ ‘$\in$’ $|$ ‘$\neq$’ $|$ ‘$\leq$’ $|$ ‘$\geq$’ $|$ ‘$\forall$’\
$<symbol3>$ &::=& ‘$\sum$’ $|$ ‘$\int$’ $|$ ‘$\lim$’\
$<symbol4>$ &::=& ‘$\sum$’ $|$ ‘$\int$’\
$<symbol5>$ &::=& ‘$($’ $|$ ‘$[$’ $|$ ‘$\{$’ $|$ ‘$|$’\
$<symbol6>$ &::=& ‘$)$’ $|$ ‘$]$’ $|$ ‘$\}$’ $|$ ‘$|$’
Table \[tab:crohme2014\] shows experimental results on CROHME 2014 test set. The first seven are online recognition systems participated in CROHME 2014[@CROHME2014]. The proposed system outperformed all participated systems in CROHME 2014 except MyScript. Since MyScript Math recognizer itself has evolved over the past few years, MyScript Interactive Ink version 1.3, which is up-to-date as of this writing, was evaluated in the online setting too. On the other hand, im2markup [@pmlr-v70-deng17a], Watch, Attend, and Parse(WAP)[@ZHANG2017196], CNN-BLSTM-LSTM[@LE2019255], Paired Adversarial Learning(PAL)[@PAL], and Multi-Scale Attention with Dense Encoder(MSD)[@8546031] are offline recognition systems, the proposed procedure achieved a better performance than them. Noted that the best results in the paper introducing WAP used online information to achieve an additional 2% gain in expression rate.
[Xrrrr]{} System & & Structure\
& Exact & $\leq 1$ & $\leq 2$& rate\
& & label & label &\
& & error & errors &\
\
São Paulo&15.01&22.31&26.57& -\
RIT, CIS&18.97&26.37&30.83& -\
RIT, DRPL&18.97&28.19&32.35& -\
Tokyo&25.66&33.16&35.90& -\
Nantes&26.06&33.87&38.54& -\
Politècnica de València&37.22&44.22&47.26& -\
MyScript&62.68&72.31&75.15& -\
MyScript 1.3&69.47&78.30&81.03& 82.86\
\
Harvard, im2markup&39.96&-&-&-\
USTC, WAP&44.4&58.4&62.2&-\
CAS, PAL&47.06&63.49&72.31&-\
TDTU, CNN-BLSTM-LSTM&48.78&63.39&70.18&-\
USTC, MSD&52.8&68.1&72.0&-\
**Proposed+MyScript 1.3**&58.22&71.60&75.15&77.38\
Table \[tab:crohme2016\] shows experimental results on CROHME 2016 test set. The first five are online recognition systems participated in CROHME 2016[@Mouchere2017ICFHR2016]. As expected, MyScript had a higher accuracy than the proposed system in all the metrics because the later was based on the former. The proposed system significantly outperformed all the remaining participated systems in CROHME 2016 without access to original strokes. On the other hand, WAP[@ZHANG2017196], CNN-BLSTM-LSTM[@LE2019255], and MSD[@8546031] are also offline recognition systems, the proposed system outperformed them.
[Xrrrr]{} System & & Structure\
& Exact & $\leq 1$ & $\leq 2$& rate\
& & label & label &\
& & error & errors &\
\
Nantes&13.34&21.02&28.33& 21.45\
São Paolo&33.39&43.50&49.17& 57.02\
Tokyo&43.94&50.91&53.70& 61.55\
Wiris&49.61&60.42&64.69& 74.28\
MyScript&67.65&75.59&79.86& 88.14\
MyScript 1.3&73.06&82.30&87.10& 88.58\
\
USTC, WAP&42.0&55.1&59.3&-\
TDTU, CNN-BLSTM-LSTM&45.60&59.29&65.65&-\
USTC, MSD&50.1&63.8&67.4&-\
**Proposed+MyScript 1.3**&65.65&77.68&82.56& 85.00\
The proposed system participated in CROHME 2019[@CROHME2019]. Table \[tab:crohme2019\] shows the final results of the competition. The proposed system was ranked the third place in the offline task.
[Xrrrr]{} System & & Structure\
& Exact & $\leq 1$ & $\leq 2$& rate\
& & label & label &\
& & error & errors &\
\
TUAT&39.95&52.21&56.54& 58.22\
MathType&60.13&74.40&78.57& 79.15\
PAL-v2&62.55&74.98&78.40& 79.15\
Samsung R&D 2&65.97&77.81&81.73& 82.82\
MyScript 1.3&77.40&85.82&87.99& 88.82\
MyScript&79.15&86.82&89.82& 90.66\
Samsung R&D 1&79.82&87.82&89.15& 89.32\
USTC-iFLYTEK&80.73&88.99&90.74&91.49\
\
TUAT&24.10&35.53&43.12&43.70\
Univ. Linz&41.49&54.13&58.88&60.02\
SCST-USTC&62.14&75.06&78.23&78.32\
PAL-v2& 62.89&74.98&78.40&79.32\
**Proposed+MyScript 1.3**& 65.22&78.48&83.07&84.90\
PAL& 71.23&80.31&82.65&83.82\
USTC-iFLYTEK& 77.15&86.82&88.99&89.49\
Although the accuracy of the proposed offline recognizer was good compared with other offline recognizers, it was still much lower than that of the underlying online recognizer. If the stroke extractor worked perfectly, the two should be the same. The gap in expression rate was much larger than the gap in structure rate, the observation indicates that structural analysis is less sensitive to errors in stroke extraction than symbol recognition.
In order to narrow the gap, one should bring output of the stroke extractor and expected input of the online recognizer closer. There are two ways to achieve that. The first way is to improve the stroke extractor, so that it can recover written strokes exactly. The second way is to retrain the online recognition engine, so that it can adapt to the artificial strokes.
Improving the stroke extractor seems to be the most obvious choice, but it is hard. Adding more heuristic rules would lead to serious maintainability issues, whereas marginal benefit is diminishing. Therefore, many researchers had given up in this direction. Developing data driven stroke extractors is possible, but it would defeat the purpose. If heavy models such as convolutional and recurrent neural networks are applied, it is pointless to predict trajectories instead of formulas themselves.
Retraining the online recognition engine is a realistic choice. If the recognizer has seen artificial strokes given by the stroke extractor during training, it should be able to learn what they mean. By applying this technique, the sense of stroke extracted is decoupled from the way written by human beings, thus the stroke extractor needs not worry too much about corner cases. Unfortunately, MyScript Math recognizer was not trainable by user, so the next experiment was switched to another online recognizer.
Integration with a trainable online recognizer {#subsec:tap}
----------------------------------------------
In the second experiment, the proposed stroke extractor was combined with TAP[^6], which is a published online handwritten mathematical expression recognition system[@8373726], to form an offline recognizer. The paper introducing TAP combined three online models, three offline models, and three language models to get the best results, but ensemble modeling was not used in this experiment, since the objective is to check if a pure online model can adapt to the artificial strokes.
Evaluation was performed on the CROHME 2016 dataset, only the official training set and validation set were used to train and validate the models respectively. The encoder of TAP consumes an 8-dimensional feature vector for a point at each time step, the decoder of TAP predicts a TeX token at each time step, the loss function involves alignment between points and TeX tokens. When trained on written strokes, trace points and alignment were available from the dataset directly. When trained on artificial strokes, trace points were extracted from rendered images and alignment were estimated with Hausdorff distance. In both cases, TeX code for each formula was converted from annotation in MathML.
Table \[tab:crohme2016tap\] shows experimental results on CROHME 2016 test set. The model which had been trained on written strokes performed much better on written strokes than on extracted strokes, the same phenomenon was already observed with MyScript. However, the model which had been trained on extracted strokes leaded to a much better offline recognizer, its accuracy almost caught up the online recognizer. The gap in expression rate between them was only 0.61%, noted that the gap in expression rate between the best online system and the best offline system participated in CROHME 2019 was 3.58%. The results verify that an online recognition system can adapted to the artificial strokes well by retraining. Another implication is that the essential difficulty between online recognition and offline recognition may be smaller than it was thought.
[llrrrr]{} Training/validation& Test & & Structure\
&& Exact & $\leq 1$ & $\leq 2$& rate\
&& & label & label &\
&& & error & errors &\
Written&Written&43.68&55.88&61.29&62.60\
Written&Extracted&23.63&38.71&47.17&51.79\
Extracted&Extracted&43.07&56.67&62.95&64.95\
Although the offline recognizer built upon TAP was not as good as the one built upon MyScript, there is room for improvement. For example, language models, ensemble models, augmented datasets, and expanded datasets were not applied in this experiment, whereas they were commonly used by other state-of-the-art mathematical expression recognition systems[@CROHME2019] to boost the accuracy significantly[@8373726; @LE2019255].
Efficiency {#subsec:efficency}
----------
In the last experiment, efficiency of the proposed system was examined on devices ranging from low-end mobile phones to powerful GPU server. Table \[tab:device\] shows details of those devices.
[llX]{} Device&Component&Specification\
Phone 1&CPU&Cortex-A7 4 cores @ 1.3 GHz\
&RAM&512MB\
Phone 2&CPU&Cortex-A53 2 cores @ 2.0 GHz + 6 cores @ 1.45 GHz\
&RAM&3GB\
Ultrabook&CPU&Intel Core m3-6Y30 4 cores @ 0.90GHz\
&RAM&4GB\
Desktop&CPU&Intel Core i5-7500 4 cores @ 3.40GHz\
&RAM&8GB\
Server&CPU&Intel Xeon Gold 6271C 2 cores @ 2.60GHz\
&RAM&32GB\
&GPU&NVIDIA Tesla V100 16GB\
Table \[tab:time\] shows the elapsed time used to recognize 1147 rendered mathematical expressions in the test set of CROHME 2016. Although the stroke extraction algorithm was implemented on CPU and not optimized for speed, it was still much faster than online recognition on all the devices tested. A modern low-end mobile phone(Phone 2) took about one second to recognize an image on average, whereas a clearly outdated mobile phone(Phone 1) took about two seconds. Although it is noticeable, the speed should be acceptable for on-device offline handwritten mathematical expression recognition, if users feel that online recognition is already fast enough.
[llrrrr]{} Device&Recognizer&\
&&Stroke&Online&Total&Mean\
&&extraction&recognition&&\
Phone 1&MyScript&1025&1653&2678&2.33\
Phone 2&MyScript&254&1123&1377&1.20\
Ultrabook&TAP&22&6908&6930&6.04\
Desktop&TAP&10&535&544&0.47\
&WAP&-&-&7845&6.84\
Server&TAP&34&152&186&0.16\
&WAP&-&-&417&0.36\
WAP[^7], which is a native offline recognizer[@ZHANG2017196], was slower than the proposed procedure on all the device tested. Like other modern optical character recognition systems, WAP uses a convolutional neural network which is computationally intensive[@Samsung] to extract features, so the speed relies heavily on the availability of powerful GPU. Since online recognition engines are usually less resource-hogging than native offline recognition engines, on-device online recognition is already integrated into various note-taking applications and input methods nowadays, whereas optical character recognition functionality usually relies on cloud based services. Therefore, stroke extraction provides an attractive way to improve user experience by avoiding unpredictable network latency and privacy issues.
Conclusion {#sec:conclusion}
==========
In this paper, a stroke extraction algorithm is proposed for handwritten mathematical expression, so that an offline recognizer can be built upon an online one. A proof-of-concept implementation of the proposed stroke extractor is publicly available as a free software. Given a ready-made state-of-the-art online recognition engine, good offline accuracy was achieved on CROHME datasets. Given a trainable online recognition system, an offline recognizer with the same level of accuracy was constructed by retraining it with extracted strokes. Noted that the underlying online recognizers are not specially designed for extracted strokes.
The proposed approach is especially preferable for real-time use cases on devices with limited resources. Since online recognizers generally occupy less memory and run faster than native offline recognizers, stroke extraction provides an efficient way to implement on-device offline recognition on mobile phones and tablets.
Stroke extraction is a general methodology to offline handwriting recognition. Besides handwritten mathematical expression, the same approach can be applied to other types of handwriting such as chemical expression, musical notation, and diagram in principle. Examining if specialized stroke extractors are needed for different types of handwriting is a future work.
Developing independent recognition systems for online and offline handwriting may no longer be necessary, since stroke extraction allows advances on online recognition to be propagated immediately to the offline case. Instead, online recognition system makers can enter the offline market without abandoning existing investments. Therefore, the potential of reduction from offline recognition to online recognition is justified.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
[^1]: School of Mathematics, Sun Yat-Sen University, 135 Xingang Xi Road, Guangzhou, 510275, China
[^2]: Email address: [email protected] or [email protected]
[^3]: Declarations of interest: none
[^4]: The proposed stroke extractor is publicly available as a free software at <https://github.com/chungkwong/mathocr-myscript> and <https://github.com/chungkwong/mathocr-myscript-android>.
[^5]: Documents of the recognizer is available at <https://developer.myscript.com/docs/interactive-ink/1.3/overview/about/>.
[^6]: The original version of TAP is available at <https://github.com/JianshuZhang/TAP>. Our version of programs, datasets and pretrained models are available at <https://github.com/chungkwong/mathocr-tap>.
[^7]: WAP is avaliable at <https://github.com/JianshuZhang/WAP>.
|
{
"pile_set_name": "ArXiv"
}
|
---
---
[**** ]{}\
Wajdi Dhifli^1^, Abdoulaye Baniré Diallo^1,\*^\
**[1]{} Computer Science department, University of Quebec at Montreal, Downtown station, Montreal, Quebec, Canada, H3C 3P8\
**
\* [email protected]
Abstract {#abstract .unnumbered}
========
Graph theory and graph mining constitute rich fields of computational techniques to study the structures, topologies and properties of graphs. These techniques constitute a good asset in bioinformatics if there exist efficient methods for transforming biological data into graphs. In this paper, we present Protein Graph Repository (PGR), a novel database of protein 3D-structures transformed into graphs allowing the use of the large repertoire of graph theory techniques in protein mining. This repository contains graph representations of all currently known protein 3D-structures described in the Protein Data Bank (PDB). PGR also provides an efficient online converter of protein 3D-structures into graphs, biological and graph-based description, pre-computed protein graph attributes and statistics, visualization of each protein graph, as well as graph-based protein similarity search tool. Such repository presents an enrichment of existing online databases that will help bridging the gap between graph mining and protein structure analysis. PGR data and features are unique and not included in any other protein database. The repository is available at <http://wjdi.bioinfo.uqam.ca/>.
Introduction {#introduction .unnumbered}
============
The advances in computational and biological techniques of protein studies have yielded enormous online databases. However, the complexity of protein structure requires adequate bioinformatics methods to mine these databases. The principles of graph theory have been adopted to investigate organic molecules [@Borgelt_2002] and proteins [@Milik_2003; @Huan_2005; @Dhifli_2014]. The tertiary structure captures homology between proteins that are distantly related in evolution. With the availability of more protein 3D-structures due to techniques such as X-ray crystallography, increasing efforts have been devoted to directly deal with them. A crucial step in the computational study of protein structures is to look for a convenient representation of their spatial conformations. The PDB format [@Berman_2000] represents the standard computer analyzable format that is used in online databases for representing macromolecular structures. Extensions of the PDB format have been proposed in the literature mainly mmCIF and PDBML/XML file formats [@Westbrook_2005]. The PDB format and its extensions mainly consist on spatial coordinates of atoms composing the considered macromolecule besides its biological description and experimental details with which it was obtained. Such representation prevent a direct use of the large repertoire of available data mining and graph theory tools to study protein structures. A possible representation of protein 3D-structure can be a graph of interconnected amino acids. Figure \[fig:protein\_to\_graph\] shows a real world example of the human hemoglobin protein and its corresponding graph. The graph representation preserves the overall structure and its components. Such representation can be considered as an alternative to existent representations, such as the PDB format [@Berman_2000].
![[**The Human Hemoglobin protein 3D structure (PDB-ID: 1GZX) and its corresponding graph.**]{} Each node in the graph represents an amino acid and each edge represents a spatial link (interaction) between two amino acids in the structure. The blue edges represent links from the primary structure and gray edges represent spatial links between amino acids that are distant in the primary structure. The Human Hemoglobin protein is composed of four subunits at the corners across a cavity at the center of the molecule. The example shows that the graph representation preserves the overall structure of the protein.[]{data-label="fig:protein_to_graph"}](protein_to_graph){width="100.00000%"}
![[**An example of two subgraphs corresponding to two recurrent substructures extracted from a dataset of 38 proteins (including the HFE(human) hemochromatosis protein) from the immunoglobin C1-set domains family.**]{} All the 38 proteins were transformed into graphs using PG-converter, then a frequent subgraph discovery was performed to discover recurrent substructures with a minimum support threshold of 30%. This example shows the mapping of both subgraphs on the original 3D-structure of the HFE(human) hemochromatosis protein.[]{data-label="fig:motif"}](motifs_protein_DS3_F1_1A6Z8_labels_graphs_2){width="100.00000%"}
It allows to fully exploit the potential of data mining and graph theory algorithms to perform complex studies such as the discovery of important substructures in protein families which can be performed through frequent subgraph mining, pattern recognition, and functional motif discovery. Figure \[fig:motif\] shows a real example of two subgraphs corresponding to two recurrent substructures in a dataset of 38 proteins from the immunoglobin C1-set domains family, and their corresponding mapping on the original 3D-structure of the HFE(human) hemochromatosis protein. Such substructures are relevant for protein classification, protein function prediction, protein folding, $etc$. For instance, we have previously explored the potential of graph representation in the classification of four protein 3D-structure datasets, including G-proteins, immunoglobin C1-set domains, C-type lectin domains, and protein kinases catalytic subunit [@Dhifli_2014]. Frequent subgraphs were mined and used as features for the classification of each dataset. The experimental results showed that this graph-based approach outperformed the most competitive bioinformatics approaches including structural alignment-based classification (using Dali [@Holm_2010]) and Blast-based classification [@Altschul_1990]. In this paper, we present *Protein Graph Repository* (PGR), an online repository of graphs representing all protein 3D-structures of the *Protein Data Bank* (PDB)[@Berman_2000]. PGR provides bioinformatics tools that facilitate the integration of graph theory techniques in the core of protein 3D-structure studies [@Dhifli_2014; @Stout_2008; @Dognin_2014].
Materials and Methods {#materials-and-methods .unnumbered}
=====================
Graph transformation of protein 3D structure {#graph-transformation-of-protein-3d-structure .unnumbered}
--------------------------------------------
Chemical interactions are the electrostatic forces that hold atoms and residues together, stabilizing proteins and forming molecules that give them their 3 dimensional shape [@Huan_2005; @Stout_2008; @Lovell_2003; @Saidi_2009]. These interactions are mainly:
- Covalent bonds between two atoms sharing a pair of valence electrons,
- Ionic bonds of electrostatic attractions between oppositely charged components,
- Hydrogen bonds between two partially negatively charged atoms sharing a partially positively charged hydrogen,
- Hydrophobic interactions where hydrophobic amino acids in the protein closely associate their side chains together,
- Van der Waals forces which represent transient and weak electrical attraction of one atom for another when electrons are fluctuating.
These interactions are supposed to be, in one form or another, the chemical analogues of the graph edges. Existing transformation approaches of protein 3D-structure into graph, similarly consider amino acids as graph nodes, but they differ in considering the edges in attempt to reflect the truly existing interactions. In the following, we present the main approaches in the literature that are used for building protein graphs in PGR. Let $G$ be a graph, $u$ and $v$ two nodes of $G$ ($u$,$v$ $\in$ $G$), $\Delta$ a function that computes the distance between pairs of nodes $\Delta(u,v)$, and $\delta$ a distance threshold.
- **Main Atom** abstracts each amino acid in only one main atom, $M_A$ [@Huan_2005; @Lovell_2003]. Two nodes representing two amino acids $u$ and $v$ are linked by an edge $e(u, v)$ if the euclidean distance between their two main atoms $\Delta(M_A(u), M_A(v))$ is below a distance threshold $\delta$. The main atom used in the literature is *$C_\alpha$* with usually $\delta \geq 7 $Å ** on the argument that $C_\alpha$ atoms define the overall shape of the protein conformation [@Huan_2005].
- **All Atoms** considers the distances between all pairs of atoms $\Delta(A_A(u), A_A(v))$, where $A_A(u)$ represents all atoms of $u$ ($A_A(u)=\forall atom \in u$)[@Saidi_2009]. Two nodes representing two amino acids $u$ and $v$ are linked by an edge $e(u, v)$ if the euclidean distance between any pair of atoms from both amino acids $\Delta(A_A(u), A_A(v))$ is below a distance threshold $\delta$. Although this increases the complexity of graph building, it allows detecting connections that were omitted using Main Atom [@Saidi_2009].
Main features of PGR {#main-features-of-pgr .unnumbered}
--------------------
The main bioinformatics features of PGR are listed in Figure \[fig:pgr\_features\].
![[**PGR main bioinformatics features.**]{}[]{data-label="fig:pgr_features"}](pgr_features3){width="100.00000%"}
#### Graph repository
We transformed all protein 3D-structures (of the PDB) into protein graphs (in PGR) using the described methods. The repository is enriched by a selection tool allowing the filtering and targeting of a specific population of proteins. Each protein graph can be displayed solely in a light-weight and interactive visualization interface using the best available visualization libraries including D3.js [@Bostock_2011] and Cytoscape [@Smoot_2011]. A set of the most important attributes for protein graph mining have been pre-computed including density, diameter, link impurity, $etc$. These attributes are presented with their Z-score according to all protein graph attribute distributions.
#### PG-converter
PGR also provides an online converter that allows to upload and transform protein 3D-structures into protein-graphs. Available transformation methods include All Atoms, Main Atom based on C$\alpha$, C$\beta$, amino acid centroid, side chain centroid, amino acid ray, amino acid ray and side chain orientation, or side chain ray.
#### PG-similarity
Furthermore, we provide a search tool based on the pairwise similarity of structural protein attributes. Such tool could constitute an asset for several biological tasks such as protein classification and function prediction. The pairwise similarity between two protein graphs is measured by the distance between their corresponding vector representation based on the structural and topological attributes. We selected a set of attributes from the literature that are interesting and efficient in describing connected graphs [@Li_2012; @Leskovec_2005]. In the following, we list and define the used structural and topological attributes:
1. **Number of nodes**: The total number of nodes in the protein graph, also called the graph order $|V|$.
2. **Number of edges**: The total number of edges in the protein graph, also called the graph size $|E|$.
3. **Average degree**: The degree of a node $u$, denoted $deg(u)$, represents the number of nodes adjacent to $u$. The average degree of a graph $G$ is the average value of the degrees of all nodes in $G$. Formally: $ deg(G) = \frac{1}{n} \sum^n_{i=1} deg(u_i)$ where $deg(u_i )$ is the degree of the node $u_i$ and $n$ is the number of nodes in $G$.
4. **Density**: The density of a graph $G=(V, E)$ measures how many edges are in $E$ compared to the maximum possible number of edges between the nodes in $V$. Formally: $ den(G) = \frac{2 \mid E\mid}{(\mid V\mid\ast (\mid V\mid -1))}$.
5. **Average clustering coefficient**: The clustering coefficient of a node $u$, denoted by $c(u)$, measures how complete the neighborhood of $u$ is, $i.e.$, $c(u)= \frac{2 e_u}{k_u (k_u - 1)}$ where $k_u$ is the number of neighbors of $u$ and $e_u$ is the number of connected pairs of neighbors. If all the neighbor nodes of u are connected, then the neighborhood of $u$ is complete and we have a clustering coefficient of 1. If no nodes in the neighborhood of $u$ are connected, then the clustering coefficient is 0. The average clustering coefficient of an entire graph $G$ having $n$ nodes, is given as the average value over all the nodes in $G$. Formally: $C(G)= \dfrac{1}{n} \sum_{i=1}^n c(u_i)$.
6. **Average effective eccentricity**: For a node $u$, the effective eccentricity represents the maximum length of the shortest paths between $u$ and every other node $v$ in $G$, $i.e.$, $e(u) = max\{d(u,v) : v\in V\}$. If $u$ is isolated then $e(u) = 0$. The average effective eccentricity is defined as $Ae(G)= \frac{1}{n}\sum_{i=1}^n e(u_i)$, where $n$ is the number of nodes of $G$.
7. **Effective diameter**: The effective diameter represents the maximum value of effective eccentricity over all nodes in the graph $G$, $i.e.$, $diam(G) = max\lbrace e(u)\mid u\in V\rbrace$ where $e(u)$ represents the effective eccentricity of $u$ as defined above.
8. **Effective radius**: The effective radius represents the minimum value of effective eccentricity over all nodes in the graph $G$, $i.e.$, $rad(G) = min\lbrace e(u)\mid u\in V\rbrace$ where $e(u)$ represents the effective eccentricity of $u$.
9. **Closeness centrality**: The closeness centrality measures how fast information spreads from a given node to other reachable nodes in the graph. For a node $u$, it represents the reciprocal of the average shortest path length between $u$ and every other reachable node in the graph, $i.e.$, $C_c(u) = \frac{n-1}{\sum_{v\in \lbrace V\setminus u\rbrace} d(u,v)}$ where $d(u,v)$ is the length of the shortest path between the nodes $u$ and $v$. For a graph $G$, we consider the average value of closeness centrality of all the nodes, $i.e.$, $C_c(G) = \frac{1}{n} \sum_{i=1}^n u_i$.
10. **Percentage of central nodes**: Here, we compute the ratio of the number of central nodes from the number of nodes in the graph. A node $u$ is considered as central point if the value of its eccentricity is equal to the effective radius of the graph, $i.e.$, $e(u) = rad(G)$.
11. **Percentage of end points**: It represents the ratio of the number of end points from the total number of nodes of the graph. A node $u$ is considered as end point if $deg(u) = 1$.
12. **Neighborhood impurity**: The impurity degree of a node $u$ belonging to a graph $G$, having a label $L(u)$ and a neighborhood (adjacent nodes) $N(u)$, is defined as $ImpurityDeg(u) = \mid L(v): v \in N(u), L(u)\neq L(v)\mid$. The neighborhood impurity of a graph $G$ represents the average impurity degree over all nodes with positive impurity.
13. **Link impurity**: An edge $\{u,v\}$ is considered to be impure if $L(u)\neq L(v)$. The link impurity of a graph $G$ with $k$ edges is defined as: $\frac{\mid\{u,v\}\in E: L(u)\neq L(v)\mid}{k}$.
14. **Label entropy**: It measures the uncertainty of labels. The label entropy of a graph $G$ having $k$ labels is measured as $E(G) = -\sum_{i=1 }^k p(l_i)\textit{ log }p(l_i)$, where $l_i$ is the $i^{th }$ label.
Illustrative Example {#illustrative-example .unnumbered}
--------------------
This section shows an illustrative example of a graph similarity search for the *Human Hemoglobin protein* (Figure \[fig:protein\_to\_graph\]) using PG-similarity. Figure \[fig:pg\_similarity\_example\] shows the different components of PG-similarity web interface. The graph similarity search based on the set of structural and topological attributes is performed to detect the top similar proteins for the considered tertiary structure. The graph representation used for the query is based on C$\alpha$ and the vector distance measure is the standardized euclidean distance.
Table \[tab:pg\_similarity\_example\] shows the obtained results for the top 10 most similar proteins. Similarly to the query 3D structure, all the obtained proteins are also Hemoglobin molecules and they are part of the same organism of the query protein, namely Homo sapiens. The similarity search was only based on the previously described structural and topological attributes with no additional knowledge about nor the query, neither the target structures. This demonstrates that PG-similarity allows an accurate detection of top similar proteins that is biologically meaningful. The selected similar could constitute an asset for several biological tasks such as protein classification and function prediction.
![[**Example of PG-Similarity search of structural neighbors of the Human Hemoglobin protein (PDB-ID: 1GZX).**]{}[]{data-label="fig:pg_similarity_example"}](pg_similarity_example){width="100.00000%"}
[**Rank**]{} [**PDB ID**]{} [**Distance**]{} [**Classification**]{} [**Molecule**]{} [**Taxonomy**]{}
-------------- ---------------- ------------------ ------------------------ ------------------ ------------------ -- --
1 1XYE 0.1072 Transport protein Hemoglobin Homo sapiens
2 1HGB 0.1092 Oxygen transport Hemoglobin Homo sapiens
3 1YE2 0.1299 Transport protein Hemoglobin Homo sapiens
4 1Y8W 0.1319 Transport protein Hemoglobin Homo sapiens
5 1Y7D 0.1353 Transport protein Hemoglobin Homo sapiens
6 1Y7C 0.1357 Transport protein Hemoglobin Homo sapiens
7 1DXV 0.1395 Oxygen transport Hemoglobin Homo sapiens
8 1XZ7 0.1475 Transport protein Hemoglobin Homo sapiens
9 1YGD 0.1493 Transport protein Hemoglobin Homo sapiens
10 1XXT 0.152 Transport protein Hemoglobin Homo sapiens
: [**Results of PG-similarity search for the top 10 similar structures to the Human Hemoglobin protein (PDB ID: 1GZX).**]{}
The used parameters for the query are: query PDB ID: 1GZX, graph building method: based on C$\alpha$, distance measure: standardized euclidean distance, number of most similars: 10.
\[tab:pg\_similarity\_example\]
Results and Discussion {#results-and-discussion .unnumbered}
======================
Regarding the graph transformation of protein 3D structures, both Main Atom and All Atoms suffer drawbacks. Since, the Main Atom technique abstracts amino acids into one main atom, it may omit possible edges between other atoms in the amino acids that are more close than their main atoms. Moreover, in the case of considering centroids of the amino acids as the main atoms, it may also suffer from two problems. In the case of two big amino acids, if their centroids are farther than the given distance threshold, they will be considered with no links while a real connection could be established between other close atoms (other than the centroids). In the case of small amino acids, if the distance between their centroids is smaller than the given distance threshold then they will be considered as connected while they can be disconnected in reality. The all Atoms technique overcomes both limitations by theoretically considering the distance between all the atoms in the amino acids, this highly increases the runtime and complexity of the technique. However, the authors proposed some heuristics to alleviate the complexity. For instance, they consider only the distance between the side chains’ centroids to decide whether their amino acids are connected or not, without regards to their chemical properties. This reduces the runtime but it may engender false edges.
Compared to the conventional distance matrix representation [@Holm_2010] *PG-similarity* measures the global structural and topological similarity between protein structures on a macro side, whereas distance matrix based similarity operates on a micro side and looks into every single detail in compared structures. Even though both similarity methods should be highly correlated and not diverge (as similar structures have similar topological descriptions), each method has its positive and negative sides. Distance matrix based methods has the advantage of detecting exact superposition and local matching sites, however, they are combinatorial and thus computationally costly. PG-similarity method is based on a vector embedding of graphs of protein structures based on a set of structural and topological attributes. This makes it unable to return local matches, however, such strategy makes it able to capture structural similarity in a very fast way. Moreover, some attributes, like clustering coefficient and neighborhood impurity, makes PG-similarity able to reveal hidden similarities that are undetected using existing methods.
With the growth of protein 3D-structures in online databases, the transformation of protein 3D-structures into graphs of interconnected amino acids and the application of graph mining concepts constitute a relevant feature for the development of rapid and efficient computational techniques. So far, PGR contains 188 252 graphs corresponding to 94 126 protein 3D-structures from the PDB. The 188 252 protein graphs are composed of 94 126 graphs created using Main Atom method with $C_\alpha$ and 94 126 graphs created using All Atoms method. PGR data will be regularly updated according to the PDB. PGR is an independent repository, but it can also act as a complementary resource to existing ones such as the PDB. PGR is apt for extension and additional services and functionalities will be added in the next coming versions.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by a grant of the Fonds de recherche du Québec - Nature et technologies to A.B. Diallo.
[10]{} Borgelt C, Berthold MR. Mining Molecular Fragments: Finding Relevant Substructures of Molecules. IEEE International Conference on Data Mining (ICDM). 2002;p. 51–58.
Milik M, Szalma S. Common Structural Cliques: A Tool for Protein Structure and Function Analysis. Protein Engineering. 2003;16:2003.
Huan J, Bandyopadhyay D, Wang W, Snoeyink J, Prins J, Tropsha A. Comparing Graph Representations of Protein Structure for Mining Family-Specific Residue-Based Packing Motifs. Journal of Computational Biology. 2005;12(6):657–671.
Dhifli W, Saidi R, Mephu Nguifo E. Smoothing 3[D]{} protein structure motifs through graph mining and amino-acids similarities. Journal of Computational Biology. 2014;21(2):162–172.
Berman HM, Westbrook JD, Feng Z, Gilliland G, Bhat TN, Weissig H, et al. The Protein Data Bank. Nucleic Acids Research. 2000;28(1):235–242.
Westbrook JD, Ito N, Nakamura H, Henrick K, Berman HM. the representation of archival macromolecular structure data in [XML]{}. Bioinformatics. 2005;21(7):988–992.
Holm L, Rosenstr[ö]{}m P. Dali server: conservation mapping in 3D. Nucleic Acids Research. 2010;38(Web-Server-Issue):545–549.
Altschul S, Gish W, Miller W, Myers E, Lipman D. Basic local alignment search tool. Journal of Molecular Biology. 1990;215:403–410.
Stout M, Bacardit J, Hirst JD, Smith RE, Krasnogor N. Prediction of topological contacts in proteins using learning classifier systems. Soft Computing. 2008;13:245–258.
Malod-Dognin N, Przulj N. GR-Align: fast and flexible alignment of protein 3D structures using graphlet degree similarity. Bioinformatics. 2014;30(9):1259–1265.
Lovell SC, Davis IW, Arendall WB, de Bakker PIW, Word JM, Prisant MG, et al. Structure validation by [C]{}$\alpha$ geometry: $\Phi$, $\Psi$ and [C]{}$\beta$ deviation. Proteins: Structure, Function, and Genetics. 2003;50(3):437–450.
Saidi R, Maddouri M, Mephu Nguifo E. Comparing graph-based representations of protein for mining purposes. In: Proceedings of the ACM KDD Workshop on Statistical and Relational Learning in Bioinformatics; 2009. p. 35–38.
Bostock M, Ogievetsky V, Heer J. D3: Data-Driven Documents. IEEE Trans Visualization & Comp Graphics (Proc InfoVis). 2011;.
Smoot M, Ono K, Ruscheinski J, Wang PL, Ideker T. Cytoscape 2.8: new features for data integration and network visualization. Bioinformatics. 2011;27(3):431–432.
Li G, Semerci M, Yener B, Zaki MJ. Effective graph classification based on topological and label attributes. Statistical Analysis and Data Mining. 2012;5(4):265–283.
Leskovec J, Kleinberg J, Faloutsos C. Graphs over time: densification laws, shrinking diameters and possible explanations. In: Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. KDD ’05. ACM; 2005. p. 177–187.
|
{
"pile_set_name": "ArXiv"
}
|
=1
*Dedicated to Percy Deift and Craig Tracy\
on the occasion of their 70th birthdays*
Introduction and statement of results {#section1}
=====================================
The Muttalib–Borodin ensemble {#section1.1}
-----------------------------
The Muttalib–Borodin biorthogonal ensemble is the following probability density function for $n$ particles on the half line $[0, \infty)$ $$\begin{gathered}
\label{eq:Bor1}
\frac{1}{Z_n} \prod_{j < k} (x_k-x_j) \prod_{j < k} \big(x_k^{\theta} - x_j^{\theta}\big) \prod_{j=1}^n e^{-nV(x_j)},
\qquad x_j \geq 0, \end{gathered}$$ with $\theta > 0$ and with an $n$-dependent weight function $w(x) = e^{-nV(x)}$ having enough decay at infinity. The model is named after Muttalib [@Mut95] who introduced it as a simplified model for disordered conductors in the metallic regime, and Borodin [@Bor99] who obtained profound mathematical results, in particular for Laguerre and Jacobi weights.
The model has attracted considerable attention in recent years. Random matrix models whose eigenvalues (or singular values) have the distribution were recently given in [@Che14; @ForWan15]. The model is also related to products of random matrices [@ForWan15; @KuiSti14].
In the large $n$ limit, the particles have an almost sure limiting measure $\mu^*$ which is the minimizer of $$\begin{gathered}
\label{eq:CRfunctional}
\frac{1}{2} \iint \log \frac{1}{|x-y|} d\mu(x) d\mu(y) +
\frac{1}{2} \iint \log \frac{1}{|x^{\theta} - y^{\theta}|} d\mu(x) d\mu(y)
+ \int V(x) d\mu(x) \end{gathered}$$ among all probability measures $\mu$ on $[0,\infty)$. This follows from large deviation results for and related models that were studied in [@BlLeToWi15; @But16; @EiSoSt11]. For $\theta = 1$, the functional reduces to the usual energy in the presence of an external field [@SaTo97].
The minimizer for was studied in detail by Claeys and Romano [@ClaRom13]. They found sufficient conditions for the minimizer to be supported on an interval $[0,a]$ for some $a > 0$. Forrester and co-authors [@ForLiu14; @FoLiZJ15] analyzed the equilibrium problem for with special potentials, and found expressions for the minimizers as Fuss–Catalan and Raney distributions, see also [@NeuVAs16].
It is the aim of this paper to show that for rational values of $\theta$, say $\theta = q/r$ with $q,r \in \mathbb N$, there is an equivalent minimization problem for a vector of $q+r-1$ measures. We expect that the vector equilibrium problem will be useful for subsequent asymptotic analysis. The special role of rational $\theta$ also appeared in the already mentioned work [@ClaRom13]. This paper gives finite term recurrence relations for the biorthogonal polynomials associated with , as well as a Christoffel–Darboux formula for the correlation kernel are given, but only for rational $\theta$.
In order to state our results we introduce the logarithmic energy $$\begin{gathered}
\label{eq:logenergy}
I(\mu) = \iint \log \frac{1}{|x-y|} d \mu(x) d\mu(y)
\end{gathered}$$ of a measure $\mu$ and the mutual energy $$\begin{gathered}
\label{eq:mutualenergy}
I(\mu,\nu) = \iint \log \frac{1}{|x-y|} d \mu(x) d\nu(y)
\end{gathered}$$ of two measures $\mu$ and $\nu$. Throughout we use for $j \in \mathbb Z$, $$\begin{gathered}
\label{eq:Deltaj}
\Delta_j = \begin{cases} [0,\infty), & \text{if $j$ is even}, \\
(-\infty,0], & \text{if $j$ is odd}.
\end{cases}\end{gathered}$$
Result for the case $\boldsymbol{\theta = 1/r}$ {#section1.2}
-----------------------------------------------
We first state the result for the case $\theta = 1/r$ with $r \in \mathbb N$.
\[thm11\] Let $V\colon [0,\infty) \to \mathbb R$ be continuous and $$\begin{gathered}
\lim_{x \to \infty} \big(V(x) - \log\big(1+x^2\big) \big) = + \infty.
\end{gathered}$$ Then there is a unique vector $(\mu_0^*, \mu_1^*, \ldots, \mu_{r-1}^*)$ of $r$ measures that minimizes the energy functional $$\begin{gathered}
\label{eq:VEproblem}
\sum_{j=0}^{r-1} I(\mu_j) - \sum_{j=0}^{r-2} I(\mu_j, \mu_{j+1})
+ \int V d\mu_0
\end{gathered}$$ among all vectors satisfying for every $j=0,\ldots, r-1$,
1. $\operatorname{supp}(\mu_j) \subset \Delta_j$,
2. $\mu_j(\Delta_j) = 1 - \frac{j}{r}$, and
3. $I(\mu_j) < +\infty$.
The measure $\mu_0^*$ has compact support, and it is the unique minimizer of the functional with $\theta = 1/r$ among probability measures on $[0,\infty)$.
The minimization problem for the energy functional is an example of a weakly admissible vector equilibrium problem in the sense of Hardy and Kuijlaars, see also below. The other measures $\mu_1^*, \ldots, \mu_{r-1}^*$ have full unbounded support, $\operatorname{supp}(\mu_j^*) = \Delta_j$ for $j=1, \ldots, r-1$. In fact $ \mu_j^* = \frac{1}{2} \operatorname{Bal}\big(\mu_{j-1}^* + \mu_{j+1}^*, \Delta_j \big)$, where $\operatorname{Bal}$ denotes the balayage onto $\Delta_j$, see below as well.
Result for rational $\boldsymbol{\theta}$ {#section1.3}
-----------------------------------------
For general rational $\theta = q/r$ with $q,r \in \mathbb N$, we first make the change of variables $x \mapsto x^{1/q}$, $y \mapsto y^{1/q}$ in the energy functional to obtain $$\begin{gathered}
\label{eq:CRfunctional2}
\frac{1}{2} \iint\! \log \frac{1}{|x^{1/q}-y^{1/q}|} d\nu(x) d\nu(y) +
\frac{1}{2} \iint\! \log \frac{1}{|x^{1/r} - y^{1/r}|} d\nu(x) d\nu(y)
+ \int \! \widehat{V}(x) d\nu(x),\!\!\!\! \end{gathered}$$ where $d\nu(x) = d\mu(x^{1/q})$, and $\widehat{V}(x) = V(x^{1/q})$. Note that $q$ and $r$ play a symmetric role in the energy functional .
\[thm12\] Let $\widehat{V}\colon [0,\infty) \to \mathbb R$ be continuous and $$\begin{gathered}
\label{eq:tildeVgrowth}
\lim_{x \to \infty} \big(\widehat{V}(x) - \log\big(1+x^2\big) \big) = + \infty.
\end{gathered}$$ Then there is a unique vector $(\nu_{-q+1}^*, \ldots, \nu_{-1}^*, \nu_0^*, \nu_1^*, \ldots, \nu_{r-1}^*)$ of $q+r-1$ measures that minimizes the energy functional $$\begin{gathered}
\label{eq:VEproblem2}
\sum_{j=-q+1}^{r-1} I(\nu_j) - \sum_{j=-q+1}^{r-2} I(\nu_j, \nu_{j+1})
+ \int \tilde{V} d\nu_0
\end{gathered}$$ among all vectors satisfying for every $j=-q+1,\ldots, r-1$,
1. $\operatorname{supp}(\nu_j) \subset \Delta_j$,
2. $\nu_j(\Delta_j) = m_j :=
\begin{cases} 1 - \frac{j}{r} & \text{if }j \geq 0, \vspace{1mm}\\ 1 - \frac{|j|}{q} & \text{if } j \leq 0, \end{cases}$
3. $I(\mu_j) < +\infty$.
The measure $\nu_0^*$ has compact support and it is the unique minimizer of the functional among probability measures on $[0,\infty)$, and $d\mu_0^*(x) = d\nu_0^*(x^{q})$ is the unique minimizer of with $\theta = q/r$.
As in Theorem \[thm11\] the other measures $\nu_j^*$ from Theorem \[thm12\] have full unbounded support: $\operatorname{supp}(\nu_j^*) = \Delta_j$ if $j \neq 0$.
The energy functional with $\hat{V} = 0$ and the normalizations $m_j$ as in condition ii) appeared in [@DuKu08 Theorem 2.3] where it describes the limiting eigenvalue distribution of banded Toeplitz matrices. The supports $\Delta_j$ of the measures, however, are more general curves in that case.
Theorem \[thm11\] is the special case $q=1$ of Theorem \[thm12\] and it is enough to prove the latter theorem. However, for sake of exposition we chose to state Theorem \[thm11\] separately as well.
Preliminaries {#section2}
=============
Vector equilibrium problems {#section2.1}
---------------------------
The unique existence of a minimizing vector of measures follows from the result of Hardy and Kuijlaars, which we recall here. The general setup of [@HaKu12] involves the following ingredients.
1. A sequence of closed subsets $\Delta_1, \ldots, \Delta_d$ of the complex plane, each with positive logarithmic capacity.
2. For each $i = 1,\ldots, d$, a lower semi-continuous function $V_i \colon \Delta_i \to \mathbb R \cup \{+\infty\}$ that is finite on a set of positive capacity.
3. A symmetric positive definite interaction matrix $C = (c_{ij})$ of size $d \times d$.
4. A sequence of positive numbers $m_1, \ldots, m_d$, such that $$\begin{gathered}
\label{eq:liminfVi}
\liminf_{|x| \to \infty, \, x \in \Delta_i} \left( V_i(x) -
\left(\sum_{j=1}^d c_{ij} m_j \right) \log \big(1 + |x|^2\big) \right) > -\infty
\end{gathered}$$ for every $i=1, \ldots, d$ for which $\Delta_i$ is unbounded.
Associated with the above data is the energy functional $$\begin{gathered}
\label{eq:JV}
J(\mu_1, \ldots, \mu_d) = \sum_{i,j=1}^d c_{ij} I(\mu_i,\mu_j) + \sum_{i=1}^d \int V_i(x) d\mu_i(x).
\end{gathered}$$ The problem is to minimize $J$ over $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$, or some subset of it. Here we use $\mathcal M_m(\Delta)$ to denote the set of positive Borel measures on $\Delta$ of total mass $m > 0$. Such a minimization problem is called a *weakly admissible vector equilibrium problem*.
The functional is not well-defined on all of $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$, since there is a problem with measures having overlapping supports, and with measures having unbounded supports. The problem with overlapping supports was solved by Beckermann et al. [@BeKaMaWi13]. To handle the situation with unbounded supports, a regularization of $J$ is introduced in [@HaKu12 Theorem 2.6] that comes from mapping the measures to the Riemann sphere and redefining $J$ accordingly. This procedure involves a modification of the mutual energy and the corresponding logarithmic energy to $$\begin{gathered}
\tilde{I}(\mu,\nu) = \iint \log \left( \frac{\sqrt{1+|x|^2} \sqrt{1+|y|^2}}{|x-y|} \right) d\mu(x) d\nu(y),\end{gathered}$$ and $\tilde{I}(\mu) = \tilde{I}(\mu,\mu)$, respectively.
We may call $\tilde{I}(\mu)$ and $\tilde{I}(\mu,\nu)$ the spherical (mutual) energy. Then $$\begin{gathered}
\label{eq:JV2}
\tilde{J}(\mu_1, \ldots, \mu_d)
= \begin{cases} \displaystyle \sum\limits_{i,j=1}^d c_{ij} \tilde{I}(\mu_i,\mu_j)
+ \sum\limits_{i=1}^d \int \tilde{V}_i(x) d\mu_i(x),
& \text{if all } \tilde{I}(\mu_i) < +\infty, \\
+\infty, & \text{otherwise},
\end{cases}
\end{gathered}$$ where $$\begin{gathered}
\label{eq:tildeVi}
\tilde{V}_i(x) = V_i(x) -
\left(\sum_{j=1}^d c_{ij} m_j \right) \log\big(1+|x|^2\big)
\end{gathered}$$ is an extension of since for vectors of measures with $I(\mu_i) < +\infty$ for all $i$, we have that $\tilde{J}(\mu_1, \ldots, \mu_d)
= J(\mu_1, \ldots, \mu_d)$. Note that $\tilde{V}_i$ is bounded away from $-\infty$ on $\Delta_i$ because of the lower-semicontinuity of $V_i$ and the growth condition at infinity.
The functional is thus defined on $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$ with values in $\mathbb R \cup \{+\infty\}$ only. In addition, $\tilde{J}$ has compact sublevel sets $\tilde{J}^{-1}((-\infty, \alpha])$ for any real $\alpha$ (which implies that it is lower semi-continuous), and it is strictly convex on the set where it is finite. As a consequence, there is a unique minimizer of $\tilde{J}$ on $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$, see [@HaKu12 Corollary 2.7].
Variational conditions {#section2.2}
----------------------
We use $U^{\nu}(z) = \int \log \frac{1}{|z-y|} d\nu(y)$, $z \in \mathbb C$, to denote the logarithmic potential of a measure $\nu$ and $$\begin{gathered}
\tilde{U}^{\nu}(z) = \int \log \left( \frac{\sqrt{1+|z|^2} \sqrt{1+|y|^2}}{|z-y|} \right) d\nu(y) \end{gathered}$$ to denote the spherical potential.
The variational conditions for the vector equilibrium problem were not discussed in [@HaKu12]. The following result is standard for the case $d=1$, and its extension to $d \geq 2$ is not difficult.
\[lem21\] Let $\vec{\mu}^*= (\mu_1^*, \ldots, \mu_d^*)$ be a vector of measures in $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$ with $\tilde{I}(\mu_i^*) < +\infty$ for all $i$. Suppose that for some constants $\ell_i$, $$\begin{gathered}
\label{eq:varcond}
2 \sum_{j=1}^d c_{ij} \tilde{U}^{\mu_j^*} + \tilde{V}_i
\ \begin{cases} = \ell_i & \text{on } \operatorname{supp}(\mu_i^*), \\
\geq \ell_i & \text{on } \Delta_i,
\end{cases}
\end{gathered}$$ for every $i=1, \ldots, d$. Then $\vec{\mu}$ is the minimizer of $\tilde{J}$ in $\prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$.
The functional $\tilde{J}$ is strictly convex on the set where it is finite. The strict convexity is due to the quadratic part $\tilde{J}_0(\vec{\mu}) = \sum\limits_{i,j=1}^d c_{ij} \tilde{I}(\mu_i,\mu_j)$ in the functional $\tilde{J}$ and it comes down to $$\begin{gathered}
\label{eq:varcond1}
\tilde{J}_0(\vec{\mu}) + \tilde{J}_0(\vec{\nu}) \geq 2 \sum_{i,j=1}^d c_{ij} I(\mu_i,\nu_j),
\end{gathered}$$ whenever $\vec{\mu}, \vec{\nu} \in \prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$, with strict inequality if $\vec{\mu} \neq \vec{\nu}$ and both $\tilde{J}(\vec{\mu}) < +\infty$, $\tilde{J}(\vec{\nu}) < +\infty$.
Let $\vec{\mu}^*$ be as in the lemma. Then for any $\vec{\mu} \in \prod\limits_{j=1}^d \mathcal M_{m_j}(\Delta_j)$, we find by integrating with respect to $\mu_i$, $$\begin{gathered}
\int \left(2 \sum_{j=1}^d c_{ij} \tilde{U}^{\mu_j^*} + \tilde{V}_i \right) d\mu_i
\geq \ell_i m_i, \end{gathered}$$ and so by summing over $i$, $$\begin{gathered}
\label{eq:varcond2}
2 \sum_{i,j=1}^d c_{ij} \tilde{I}(\mu_i, \mu_j^*) + \sum_{i=1}^d \int \tilde{V}_i d\mu_i
\geq \sum_{i=1}^d \ell_i m_i.
\end{gathered}$$
If we integrate with respect to $\mu_i^*$ and sum over $i$, we find an equality $$\begin{gathered}
2 \sum_{i,j=1}^d c_{ij} \tilde{I}(\mu_i^*, \mu_j^*) + \sum_{i=1}^d \int \tilde{V}_i d\mu_i^*
= \sum_{i=1}^d \ell_i m_i, \end{gathered}$$ which means[ $$\begin{gathered}
\label{eq:varcond3}
\sum_{i=1}^d \ell_i m_i = \tilde{J}(\vec{\mu}^*) + \tilde{J}_0(\vec{\mu}^*),
\end{gathered}$$ and in particular $\tilde{J}(\vec{\mu}^*) < +\infty$.]{}
Now we use with $\vec{\nu} = \vec{\mu}^*$, and combine it with and to find $$\begin{aligned}
\tilde{J}(\vec{\mu}) & = \tilde{J}_0(\vec{\mu}) + \sum_{i=1}^d \int \tilde{V}_i d\mu_i \geq
2 \int \sum_{i,j=1}^d c_{ij} I(\mu_i,\mu_j^*) - \tilde{J}_0(\vec{\mu}^*)
+ \sum_{i=1}^d \int \tilde{V}_i d\mu_i \\
& \geq \sum_{i=1}^d \ell_i m_i - \tilde{J}_0(\vec{\mu}^*) = \tilde{J}(\vec{\mu}^*).
\end{aligned}$$ Thus $\vec{\mu}^*$ is indeed the minimizer of $\tilde{J}$.
Nikishin interaction and balayage {#section2.3}
---------------------------------
In the present paper we are dealing with the interaction matrix $C = (c_{ij})$ where $$\begin{gathered}
\label{eq:Nikishin}
c_{ij} = \begin{cases} 1, & \text{if } i = j, \\
-\frac{1}{2}, & \text{if } |i-j| = 1, \\
0, & \text{otherwise}, \end{cases}
\end{gathered}$$ which is indeed a positive definite matrix, and sets $\Delta_j$ that alternate between the positive and negative real axis as in . The interaction matrix is characteristic for Nikishin systems in the theory of Hermite–Padé approximation [@NikSor91]. See also [@AptKui11; @Kui10] for surveys on the connections with Hermite-Padé approximation and random matrix theory.
In Theorem \[thm12\] we have $d=q+r-1$ and we use indices $i=-q+1, \ldots, r-1$, instead of $i=1,\ldots, d$. The total masses $m_j$ are as in condition ii) of Theorem \[thm12\]. The external fields are $V_0 = \widehat{V}$ and $V_i \equiv 0$ for $i \neq 0$. Then it is easy to see that the conditions (a), (b), (c), and (d) stated in Section \[section2.1\] are satisfied. In fact we have , since $$\begin{gathered}
\label{eq:masssums}
\sum_{j=-q+1}^{r-1} c_{ij} m_j = -\frac{1}{2} m_{i-1} + m_i - \frac{1}{2} m_{i+1}
= \begin{cases} 0, & \text{for } i \neq 0, \\
\frac{q^{-1} + r^{-1}}{2}, & \text{for } i = 0,
\end{cases} \end{gathered}$$ and then is satisfied because of .
Hence there exists a unique minimizing vector of measures $(\nu_{-q+1}^*, \ldots, \nu_{r-1}^*)$ for the energy functional $\tilde{J}$ with $$\begin{gathered}
\tilde{V}(x) = \widehat{V}(x) - \frac{q^{-1} + r^{-1}}{2} \log\big(1+|x|^2\big), \end{gathered}$$ see . We have to show that in addition $I(\nu_i^*) < +\infty$ for all $i$, and then we can conclude that it is also a minimizing vector for $J$.
The variational conditions from Lemma \[lem21\] are in this case $$\begin{gathered}
\label{eq:varcond4}
2 \tilde{U}^{\nu_0^*} - \tilde{U}^{\nu_{-1}^*} - \tilde{U}^{\nu_{1}^*}
+ \tilde{V} \ \begin{cases} = \ell_0 & \text{on } \operatorname{supp}(\nu_0^*), \\
\geq \ell_0 & \text{on } [0,\infty), \end{cases}
\end{gathered}$$ and for $j \in \{-q+1, \ldots, r-1\} \setminus \{0\}$,[$$\begin{gathered}
\label{eq:varcond5}
2 \tilde{U}^{\nu_j^*} - \tilde{U}^{\nu_{j-1}^*} - \tilde{U}^{\nu_{j+1}^*}
\ \begin{cases} = \ell_j & \text{on } \operatorname{supp}(\nu_j^*), \\
\geq \ell_j & \text{on } \Delta_j. \end{cases}
\end{gathered}$$ Here we set $\nu_{-q}^* = \nu_{r}^*=0$ so that , also hold for $j=-q+1$ and $j=r-1$.]{}
Note that $2 \nu_j^*$ and $\nu_{j-1}^* + \nu_{j+1}^*$ have the same total masses if $j \neq 0$. If holds then $2\nu_j^*$ is the *balayage measure* of $\nu_{j-1}^* + \nu_{j+1}^*$ onto $\Delta_j$. The balayage measure has full support $\Delta_j$, and equality holds in everywhere on $\Delta_j$. In addition, the constant $\ell_j$ is zero. So for $j \neq 0$ the variational condition is $$\begin{gathered}
\label{eq:varcond6}
2 \tilde{U}^{\nu_j^*} - \tilde{U}^{\nu_{j-1}^*} - \tilde{U}^{\nu_{j+1}^*} = 0 \qquad \text{on} \ \ \Delta_j.
\end{gathered}$$
If $x \mapsto \log(1+|x|^2)$ is integrable with respect to all measures, then reduces to $$\begin{gathered}
\label{eq:varcond8}
2 U^{\nu_0^*} - U^{\nu_{-1}^*} - U^{\nu_{1}^*}
+ \widehat{V} \ \begin{cases} = \ell & \text{on } \operatorname{supp}(\nu_0^*), \\
\geq \ell & \text{on } [0,\infty), \end{cases}\end{gathered}$$ for some constant $\ell$, while reduces to $$\begin{gathered}
\label{eq:varcond7} 2 U^{\nu_j^*} = U^{\nu_{j-1}^*} + U^{\nu_{j+1}^*}
\qquad \text{on} \ \ \Delta_j,
\end{gathered}$$ which is the more common form for the balayage in logarithmic potential theory, see [@SaTo97] where the discussion however is restricted to measures with compact support in $\mathbb C$.
Our strategy to prove Theorem \[thm12\] will be to establish the existence of a vector of measures $\vec{\nu}^* = (\nu_{-q+1}^*, \ldots, \nu_{r-1}^*)$ with $\operatorname{supp}(\nu_j^*) \subset \Delta_j$, $\nu_j^*(\Delta_j) = m_j$, such that the conditions and are satisfied. The measure $\nu_0^*$ will have compact support, and all other measures have full support. The density of $\nu_j^*$ will behave like $$\begin{gathered}
\frac{d\nu_j^*(x)}{dx} =
\begin{cases}
O\big(|x|^{-1-1/q}\big) & \text{for } j \geq 1, \\
O\big(|x|^{-1-1/r}\big) & \text{for } j \leq -1,
\end{cases} \qquad \text{as} \ \ |x| \to \infty. \end{gathered}$$
Proof of Theorem \[thm12\] {#section3}
==========================
An auxiliary result {#section3.1}
-------------------
The following is our main auxiliary result.
\[prop21\] Let $r \geq 2$ be an integer and let $a > 0$. Then there is a unique vector $(\mu_1^*, \ldots, \mu_{r-1}^*)$ of measures that minimizes the energy functional $$\begin{gathered}
\label{eq:CRfunctional3}
\sum_{j=1}^{r-1} I(\mu_j) - \sum_{j=1}^{r-2} I(\mu_j, \mu_{j+1}) + \int \log |x-a| d \mu_1(x)
\end{gathered}$$ among all vectors of measures satisfying for every $j=1, \ldots, r-1$,
1. $\operatorname{supp}(\mu_j) \subset \Delta_j$,
2. $\mu_j(\Delta_j) = 1 - \frac{j}{r}$, and
3. $I(\mu_j) < +\infty$.
Moreover, $$\begin{gathered}
\label{eq:CRfunctional3minimum}
\sum_{j=1}^{r-1} I(\mu_j^*) - \sum_{j=1}^{r-2} I(\mu_j^*, \mu_{j+1}^*) = - \frac{1}{2} \int \log|x-a| d\mu_1^*(x)
\end{gathered}$$ and $$\begin{gathered}
\label{eq:Umu1}
U^{\mu_1^*}(z) = \log \left| \frac{z^{1/r} - a^{1/r}}{z-a} \right|,
\qquad z \in \mathbb C. \end{gathered}$$
The minimization of under the conditions i), ii), iii) is a weakly admissible vector equilibrium problem for $r-1$ measures with total masses $m_j = 1 - \frac{j}{r}$, $j=1, \ldots, r-1$. We set $\mu_0^* = \delta_a$, the Dirac point mass at $a > 0$ and $m_0 = 1$. Also $\mu_r^* = 0$ and $m_r = 0$. Then by the discussion in Section \[section2\], there is a unique minimizer $(\mu_1^*, \ldots, \mu_{r-1}^*)$ for the extended functional. We are going to construct the measures $\mu_j^*$ explicitly. We show that these measures have densities with respect to Lebesgue measures that decay as $|x|^{-1 - 1/r}$ as $|x| \to \infty$, and that for $j=1, \ldots, r-1$, $$\begin{gathered}
2 U^{\mu_j^*} = U^{\mu_{j-1}^*} + U^{\mu_{j+1}^*}
\qquad \text{on} \ \ \Delta_j. \end{gathered}$$ This implies that also also $$\begin{gathered}
2 \tilde{U}^{\mu_j^*} = \tilde{U}^{\mu_{j-1}^*} + \tilde{U}^{\mu_{j+1}^*}
\qquad \text{on} \ \ \Delta_j. \end{gathered}$$ Then by Lemma \[lem21\] it follows that $(\mu_1^*, \ldots, \mu_{r-1}^*)$ is the minimizer for the extended functional, and since the function $x \mapsto \log(1+|x|^2)$ is integrable with respect to each of the measures, it then also follows that it is the minimizer of . We will see at the end of the proof that and hold as well.
We use a geometric construction based on the Riemann surface for the mapping $w = z^{1/r}$. The Riemann surface has $r$ sheets $$\begin{gathered}
\mathcal R_j =
\begin{cases} \mathbb C \setminus \Delta_1 = \mathbb C \setminus (-\infty,0],
& \text{for } j = 1, \\
\mathbb C \setminus (\Delta_{j-1} \cup \Delta_j) = \mathbb C \setminus \mathbb R, & \text{for } j=2,\ldots, r-1, \\
\mathbb C \setminus \Delta_{r-1}, & \text{for } j= r,
\end{cases}
\end{gathered}$$ where $\mathcal R_j$ is connected to $\mathcal R_{j+1}$ along $\Delta_j$ for $j=1, \ldots, r-1$ in the usual crosswise manner. There is one point at infinity that connects all $r$ sheets. Note that $\mathcal R_r = \mathbb C \setminus (-\infty,0]$ if $r$ is even, and $\mathcal R_r = \mathbb C \setminus [0,\infty)$ if $r$ is odd.
The Riemann surface has genus zero and $z = w^r$ is a rational parametrization of it. The rational function $$\begin{gathered}
\label{eq:Psi}
\Psi(w) = \frac{1}{r w^{r-1}(w-a^{1/r})}, \qquad z = w^r,
\end{gathered}$$ is meromorphic on the Riemann surface with a simple pole at $z = a$, a pole of order $r-1$ at $z=0$, and a zero of order $r$ at $z=\infty$. We use $\Psi_j$ to denote its restriction to $\mathcal R_j$. Explicitly, we then have $$\begin{gathered}
\label{eq:Psi1}
\Psi_1(z) = \frac{1}{r z^{1-1/r}(z^{1/r}-a^{1/r})},
\qquad z \in \mathcal R_1 = \mathbb C \setminus (-\infty,0]
\end{gathered}$$ with the principal branch of the $r$th roots. On the other sheets we have the same formula but with different choices of $r$th roots in $z^{1-1/r}$ and $z^{1/r}$. We always use $a^{1/r} > 0$.
The cuts $\Delta_j$ are oriented from left to right, and we use $\Psi_{j,\pm}(s)$, $s \in \Delta_j$, to denote the limit of $\Psi_j(z)$ as $z \to s$ with $ \pm \operatorname{Im} z > 0$. Then we define for $j=1, \ldots, r-1$, $$\begin{gathered}
\label{eq:mujdensity}
d\mu_j^*(s) = \frac{1}{2\pi i} (\Psi_{j,+}(s) - \Psi_{j,-}(s)) ds, \qquad s \in \Delta_j.\end{gathered}$$ This defines a real measure on $\Delta_j$ since $\Psi_{j,-}(s) = \overline{\Psi_{j,+}(s)}$ for $s \in \Delta_j$, but a priori it could be a signed measure. Suppose the density vanishes at an interior point $s \in \Delta_j$. Then $\Psi_{j,\pm}(s)$ is real, which implies that $\Psi(w)$ is real for some $w \in \mathbb C$ with $w^r = s$. From the formula for $\Psi$ it then easily follows that $w$ is real. However, if $w$ would be real and positive then $w^r$ would be on $[0,\infty)$ on the first sheet, and if $w$ would be real and negative then $w^r$ would be on $\mathbb R \setminus \Delta_{r-1}$ on the $r$th sheet. Since $w^r = s$ is on one of the cuts, we have a contradiction and we see that the density does not vanish at an interior, and therefore has a constant sign.
We compute the total masses by contour integration as in [@DuKu08]. We consider $j=1$ first. Then by $$\begin{gathered}
\int d\mu_j^*(s) = \frac{1}{2 \pi i}\int_{C} \Psi_1(s) ds, \end{gathered}$$ where $C$ is a contour that starts at $-\infty$ and follows the upper side of the cut $\Delta_1 = (-\infty,0]$, and goes back to $-\infty$ on the lower side of the cut. We deform the contour to a big circle $|s| = R$, and we pick up a residue condition from the pole at $s=a$. From we calculate the residue as $$\begin{gathered}
\lim_{z \to a} (z-a) \Psi_1(z) = \frac{1}{r a^{1-1/r}} \lim_{z \to a} \frac{z-a}{z^{1/r} - a^{1/r}} = 1.\end{gathered}$$ Hence $$\begin{gathered}
\int d\mu_1^*(s) = 1 - \lim_{R\to \infty} \frac{1}{2\pi i} \int_{|s|=R}
\Psi_1(s) ds \end{gathered}$$ with the circle $|s|=R$ oriented counterclockwise. Since for every $j=1, \ldots, r-1$, $$\begin{gathered}
\label{eq:Psiasymp}
\Psi_j(z) = \frac{1}{rz} + O\big(z^{-1-1/r}\big) \qquad \text{as} \ \ z \to \infty,
\end{gathered}$$ which easily follows from , we find $$\begin{gathered}
\label{eq:mu1mass} \int d\mu_1^*(s) = 1 - \frac{1}{r}.
\end{gathered}$$
Now consider $2 \leq j \leq r-1$. Then by and the fact that $\Psi_{j-1,\pm} = \Psi_{j,\mp}$ on $\Delta_{j-1}$, $$\begin{gathered}
\int d\mu_{j-1}^*(s) - \int d\mu_{j}^*(s)
= \frac{1}{2\pi i} \int_{\Delta_{j-1}} \left(-\Psi_{j,-}(s) - \Psi_{j,+}(s)\right) ds\\
\hphantom{\int d\mu_{j-1}^*(s) - \int d\mu_{j}^*(s) =}{}
- \frac{1}{2\pi i} \int_{\Delta_j} \left(\Psi_{j,+}(s)- \Psi_{j,-}(s) \right) ds \\
\hphantom{\int d\mu_{j-1}^*(s) - \int d\mu_{j}^*(s)}{} =
\frac{1}{2\pi i} \int_{\mathbb R} \left(\Psi_{j,-}(s) - \Psi_{j,+}(s)\right) ds,
\end{gathered}$$ since $\Delta_{j-1} \cup \Delta_j = \mathbb R$. Again by contour deformation this is $$\begin{gathered}
\int d\mu_{j-1}^*(s) - \int d\mu_{j}^*(s) = \lim_{R\to \infty} \frac{1}{2\pi i} \int_{|s|=R} \Psi_j(s) ds = \frac{1}{r}, \end{gathered}$$ where we used . Together with we conclude that $$\begin{gathered}
\int d\mu_j^*(s) = 1 - \frac{j}{r}, \qquad j =1, \ldots, r-1.
\end{gathered}$$ Since the total masses are positive, and the densities of the measures do not change sign, it now also follows that the measures are positive.
We introduce the Cauchy transforms of the measures $$\begin{gathered}
F_j(z) = \int \frac{d\mu_j^*(s)}{z-s}, \qquad z \in \mathbb C \setminus \Delta_j.\end{gathered}$$ Then by a similar contour integration argument, where now we pick up a residue contribution at $s=z$, while there is no contribution from infinity, we get $$\begin{gathered}
F_1(z) = \frac{1}{2\pi i} \int_C \frac{\Psi_1(s)}{z-s} ds = \frac{1}{z-a} - \Psi_1(z), \qquad z \in \mathbb C \setminus (-\infty,0] \label{eq:F1z}\end{gathered}$$ and for $j=2, \ldots, r-1$, $$\begin{gathered}
\label{eq:Fjz}
F_{j-1}(z) - F_j(z) = \frac{1}{2\pi i} \int_{\mathbb R} \frac{\Psi_{j,-}(s) - \Psi_{j,+}(s)}{z-s} ds = \Psi_j(z),\qquad z \in \mathbb C \setminus \mathbb R,\end{gathered}$$ where $F_r(z) = 0$. The identity $\Psi_{j,+} = \Psi_{j+1,-}$ on $\Delta_j$ then leads to $$\begin{gathered}
\label{eq:Fjbalayage}
F_{j,+}(x) + F_{j,-}(x) = F_{j-1}(x) + F_{j+1}(x), \qquad x \in \Delta_j,
\end{gathered}$$ for $j=2, \ldots, r-1$. By and , the identity also holds for $j=1$, if we agree that $$\begin{gathered}
F_0(z) = \frac{1}{z-a}. \end{gathered}$$
The measures have a density that decays like $|s|^{-1-1/r}$ as $|s| \to \infty$. This easily follows from the definitions and . Then $s \mapsto \log(1+s^2)$ is integrable for these measures, and the usual logarithmic potentials exist. By Sokhotskii–Plemelj formulas we have $$\begin{gathered}
2 \frac{d}{dx} U^{\mu_j^*}(x) = F_{j,+}(x) + F_{j,-}(x), \qquad x \in \Delta_j. \end{gathered}$$ Clearly also $$\begin{gathered}
\frac{d}{dx} U^{\mu_{j\pm 1}^*}(x) = F_{j\pm 1}(x), \qquad x \in \Delta_j. \end{gathered}$$ Then by integrating we obtain $$\begin{gathered}
\label{eq:Ujbalayage}
2 U^{\mu_j^*}(x) = U^{\mu_{j-1}^*}(x) + U^{\mu_{j+1}^*}(x), \qquad x \in \Delta_j.
\end{gathered}$$ There is no constant of integration in since $$\begin{gathered}
U^{\mu_i^*}(x) = (1-\tfrac{i}{r}) \log |x| + o(1) \qquad \text{as} \ \ |x| \to \infty, \end{gathered}$$ for each $i \in \{j-1, j,j+1\}$ and $\Delta_j$ is unbounded.
Thus we have reached the identity that we aimed for, as discussed in the beginning of the proof. It remains to verify and .
By we obtain for $j=1, \ldots, r-1$, $$\begin{gathered}
I(\mu_j^*) = \int U^{\mu_j^*} d\mu_j^* =
\frac{1}{2} \int \big( U^{\mu_{j-1}^*} + U^{\mu_{j+1}^*} \big) d\mu_j^*
= \frac{1}{2} \big(I(\mu_{j-1}^*, \mu_j^*) + I(\mu_j^*, \mu_{j+1}^*) \big),
\end{gathered}$$ which implies $$\begin{gathered}
\sum_{j=1}^{r-1} I(\mu_j^*) = \sum_{j=1}^{r-2} I(\mu_j^*, \mu_{j+1}^*) + \frac{1}{2}( I(\mu_0^*, \mu_1^*) + I(\mu_{r-1}^*, \mu_r^*) ).
\end{gathered}$$ We recall that $\mu_0^* = \delta_a$ and $\mu_r^* = 0$ and we obtain[ $$\begin{gathered}
\sum_{j=1}^{r-1} I(\mu_j^*) - \sum_{j=1}^{r-2} I(\mu_j^*, \mu_{j+1}^*) = \frac{1}{2} I(\delta_a, \mu_1^*)
= \frac{1}{2} \int \log \frac{1}{|x-a|} d\mu_1^*(x),
\end{gathered}$$ which is the identity in .]{}
Finally, we recall that by and $$\begin{gathered}
F_1(z) = \frac{1}{z-a} - \frac{1}{rz^{1-1/r}(z^{1/r} - a^{1/r})}, \qquad
z \in \mathbb C \setminus (-\infty,0], \end{gathered}$$ which after integration leads to $$\begin{gathered}
\int \log(z-s) d\mu_1^*(s) = \log(z-a) - \log\big(z^{1/r} - a^{1/r}\big). \end{gathered}$$ The constant of integration vanishes since both sides behave like $(1-1/r) \log z + o(1)$ as $z \to \infty$. Taking real parts we find .
We next extend Proposition \[prop21\] from point masses $\delta_a$ with $a > 0$ to general measures with compact support on $(0,\infty)$.
\[prop22\] Let $r \geq 2$ be an integer. Let $\mu$ be a probability measure on $(0,\infty)$ with compact support. Then there is a unique vector $(\mu_1^*, \ldots, \mu_{r-1}^*)$ of measures that minimizes the energy functional $$\begin{gathered}
\sum_{j=1}^{r-1} I(\mu_j) - \sum_{j=1}^{r-2} I(\mu_j, \mu_{j+1}) - I(\mu,\mu_1)\end{gathered}$$ among all vectors of measures satisfying for every $j=1, \ldots, r-1$,
1. $\operatorname{supp}(\mu_j) \subset \Delta_j$,
2. $\mu_j(\Delta_j) = 1 - \frac{j}{r}$, and
3. $I(\mu_j) < +\infty$.
Moreover, $$\begin{gathered}
\sum_{j=1}^{r-1} I(\mu_j^*) - \sum_{j=1}^{r-2} I(\mu_j^*, \mu_{j+1}^*) - I(\mu,\mu_1^*)\nonumber\\
\qquad{} = - \frac{1}{2} I(\mu,\mu_1^*) = - \frac{1}{2} \iint \log \left| \frac{x^{1/r} - y^{1/r}}{x-y} \right| d \mu(x) d\mu(y).\label{eq:CRfunctional4minimum}\end{gathered}$$
For $\mu = \delta_a$ this was done in Proposition \[prop21\].
Let $(\mu_1^*(a), \ldots, \mu_{r-1}^*(a))$ be the vector of measures that we obtain from $\delta_a$ as in Proposition \[prop21\]. Then for a general probability measure $\mu$ on $(0,\infty)$ with compact support, we put $$\begin{gathered}
\mu_j^* = \int \mu_j^*(a) d\mu(a), \qquad j = 1, \ldots, r-1. \end{gathered}$$ These are well-defined positive measures satisfying i), ii) and iii) of the proposition. The measures $\mu_j^*(a)$ have a density that decays as $|x|^{-1-1/r}$ as $|x| \to \infty$, and the same will be true for the measures $\mu_j^*$ since $\mu$ is compactly supported. Thus the logarithmic potentials exist, and $$\begin{gathered}
U^{\mu_j^*} = \int U^{\mu_j^*(a)} d\mu(a). \end{gathered}$$ The identity $$\begin{gathered}
2 U^{\mu_j^*(a)} = U^{\mu_{j-1}^*(a)} + U^{\mu_{j+1}^*(a)} \qquad \text{on} \ \ \Delta_j \end{gathered}$$ holds for every $a > 0$ by Proposition \[prop21\]. Integrating this with respect to $a$ and using Fubini’s theorem, we obtain $$\begin{gathered}
2 U^{\mu_j^*} = U^{\mu_{j-1}^*} + U^{\mu_{j+1}^*} \qquad \text{on} \ \ \Delta_j. \end{gathered}$$ As in the proof of Proposition \[prop21\] this leads to the first identity of .
For $j=1$ we get $$\begin{gathered}
U^{\mu_1^*}(z) = \int U^{\mu_1^*(a)}(z) d\mu(a)
= \int \log \left| \frac{z^{1/r} - a^{1/r}}{z-a} \right| d\mu(a),
\end{gathered}$$ see . Changing $z$ and $a$ to $x$ and $y$, and integrating over $d\mu(x)$, we obtain the second identity of .
Proofs of Theorems \[thm11\] and \[thm12\] {#section4}
==========================================
Theorem \[thm11\] is the special case $q=1$ of Theorem \[thm12\] and so it suffices to prove Theorem \[thm12\].
Let $\nu_0$ be a probability measure on $(0,\infty)$ with compact support, and let $(\nu_{-q+1}, \ldots, \nu_{-1}$, $\nu_0, \nu_1, \ldots, \nu_{r-1})$ be the minimizing vector of measures for under the assumptions i), ii) and iii) of Theorem \[thm12\], with $\nu_0$ fixed.
Then it is easy to see that $(\nu_1, \ldots, \nu_{r-1})$ is the minimizer for the vector energy problem of Proposition \[prop22\] with $\mu = \nu_0$, and hence by , $$\begin{gathered}
\label{eq:Inuplus}
\sum_{j=1}^{r-1} I(\nu_j) - \sum_{j=1}^{r-2} I(\nu_j, \nu_{j+1}) - I(\nu,\nu_1)
= - \frac{1}{2} \iint \log \left| \frac{x^{1/r} - y^{1/r}}{x-y} \right| d\nu_0(x) d\nu_0(y). \end{gathered}$$
Similarly, $(\nu_{-1}, \ldots, \nu_{-q+1})$ is the minimizer for the vector energy problem of Proposition \[prop22\] with $\mu= \nu_0$ and $q$ instead of $r$. Thus by again, $$\begin{gathered}
\label{eq:Inuminus}
\sum_{j=-q+1}^{-1}\! I(\nu_j) - \sum_{j=-q+1}^{-2} \! I(\nu_j, \nu_{j+1}) - I(\nu_{-1},\nu_0)
= - \frac{1}{2} \iint \!\log \left| \frac{x^{1/q} - y^{1/q}}{x-y} \right| d\nu_0(x) d\nu_0(y).\!\!\! \end{gathered}$$ The identity also holds in case $q=1$, since then $\nu_{-1}^*=0$ and both sides of are $0$.
From and we find that for a fixed $\nu_0$, the minimum of the energy functional taken over all $\nu_j$, for $j=-q+1, -1$, $j=1, \ldots, r-1$ satsifying items i), ii) of Theorem \[thm12\], is equal to $$\begin{gathered}
J(\nu_{-q+1}, \ldots, \nu_0, \ldots, \nu_{r-1}) =
I(\nu_0) - \frac{1}{2} \iint \log \left| \frac{x^{1/q} - y^{1/q}}{x-y} \right| d\nu_0(x) d\nu_0(y) \\
\hphantom{J(\nu_{-q+1}, \ldots, \nu_0, \ldots, \nu_{r-1}) =}{}
- \frac{1}{2} \iint \log \left| \frac{x^{1/r} - y^{1/r}}{x-y} \right| d\nu_0(x) d\nu_0(y)
+ \int \tilde{V}(x) d\nu_0(x), \end{gathered}$$ which reduces to with $\nu_0$ instead of $\nu$. Thus the component $\nu_0^*$ of the minimizer for the vector energy is also the minimizer of over probability measures on $[0,\infty)$.
We finally prove that $\nu_0^*$ has compact support. Let $\rho = \nu_{-1}^* + \nu_1^*$. Then $\nu_0^*$ is the minimizer of[ $$\begin{gathered}
\label{eq:nu0compact1}
\tilde{I}(\nu) - \tilde{I}(\nu,\rho) + \int \tilde{V} d \nu
\end{gathered}$$ among probability measures $\nu$ on $[0,\infty)$.]{}
=-1 If $x \mapsto \log(1+|x|^2)$ would be integrable with respect to $\nu_0^*$ then it would also be the minimizer of $$\begin{gathered}
I(\nu) - \tilde{I}(\nu,\rho) + \int \big(\tilde{V}(x) + \log \big(1+|x|^2\big)\big) d \nu(x) \nonumber\\
\qquad{} = I(\nu) + \int \left( \widehat{V}(x) - \int \log \frac{\sqrt{1+s^2}}{x-s} d\rho(s) \right) d\nu,\label{eq:nu0compact2}
\end{gathered}$$ which is a usual minimization problem for one measure with an external field $$\begin{gathered}
\widehat{V}(x) + \int \log \frac{x-s}{\sqrt{1+s^2}} d\rho(s) \end{gathered}$$ that is continuous on $(0,\infty)$ (since $\widehat{V}$ is continuous and $\rho$ is a measure on $(-\infty,0]$).
It is easy to see that $x-s > \sqrt{1+s^2}$ for $x > 1$ and $s < 0$. Thus $\int \log \frac{x-s}{\sqrt{1+s^2}} d\rho(s) > 0$ for $x > 1$, and it follows from that $$\begin{gathered}
\lim_{x \to \infty} \left( \widehat{V}(x) + \int \log \frac{x-s}{\sqrt{1+s^2}} d\rho(s) - \log\big(1+x^2\big) \right) = +\infty, \end{gathered}$$ which guarantees that has a minimizer with compact support. This minimizer also minimizes and thus coincides with $\nu_0^*$ which thus has compact support.
Theorem \[thm12\] is now fully proved.
A final remark {#section5}
==============
We consider the minimization problem for with $\theta = 1/r$. From Theorem \[thm11\] we obtain the following result that gives conditions that guarantee that the Cauchy transform of the minimizing measure is an algebraic function.
\[prop51\] Let $ \theta = 1/r$ be rational and suppose the external field $V\colon [0,\infty) \to \mathbb R$ is such that $V'$ is a rational function. Let $\mu^*$ be the probability measure that minimizes among all probability measures on $[0,\infty)$ and suppose that $\mu^*$ is supported on a finite union of intervals Then its Cauchy transform $$\begin{gathered}
F(z) = \int \frac{d\mu^*(s)}{z-s} \end{gathered}$$ is the solution of an algebraic equation of degree $r+1$. That is, there exist rational functions $c_j(z)$ for $j=0, \ldots, r$, such that $$\begin{gathered}
\sum_{j=0}^{r} c_j(z) F(z)^j + F(z)^{r+1}= 0, \qquad z \in \mathbb C. \end{gathered}$$
We turn to the vector equilibrium problem from Theorem \[thm11\] and denote the minimizing vector by $(\mu_0^*, \ldots, \mu_{r-1}^*)$ with $\mu_0^* = \mu^*$. The variational conditions are $$\begin{gathered}
2 U^{\mu^*}(x) - U^{\mu_1^*}(x) + V(x)
\ \begin{cases} = \ell, & \text{for } x \in
\operatorname{supp}(\mu^*), \\
\geq \ell, & \text{for } x \in [0,\infty), \end{cases}\end{gathered}$$ and for $j=1, \ldots, r-1$, $$\begin{gathered}
2 U^{\mu_j^*}(x) - U^{\mu_{j-1}^*}(x) - U^{\mu_{j+1}^*}(x) = 0 \qquad
\text{for} \ \ x \in \Delta_j,
\end{gathered}$$ where $\mu_r^* = 0$. We write $$\begin{gathered}
F_j(z) = \int \frac{d\mu_j^*(s)}{z-s}, \qquad j = 1, \ldots, r, \end{gathered}$$ and we find by differentiating the variational conditions $$\begin{aligned}
{3}
& F_{0,+}'(x) + F_{0,-}'(x) = F_1(x) + V'(x), \qquad && x \in \operatorname{supp}(\mu^*), & \nonumber\\
& F_{j,+}'(x) + F_{j,-}'(x) = F_{j-1}(x) + F_{j+1}(x), \qquad && x \in \Delta_j,& \label{eq:Fvarcond}\end{aligned}$$ for $j=1,\ldots, r-1$.
We construct a Riemann surface $\mathcal R$ with $r+1$ sheets $\mathcal R_j$, $j=0, \ldots, r$ given by $$\begin{gathered}
\mathcal R_0 = \mathbb C \setminus \operatorname{supp}(\mu^*), \\
\mathcal R_1 = \mathbb C \setminus ((-\infty,0] \cup \operatorname{supp}(\mu^*)), \\
\mathcal R_j = \mathbb C \setminus (\Delta_{j-1} \cup \Delta_j), \qquad \text{for} \ \ j = 2, \ldots, r-1, \\
\mathcal R_r = \mathbb C \setminus \Delta_{r-1},\end{gathered}$$ where $\mathcal R_0$ is connected to $\mathcal R_1$ along $\operatorname{supp}(\mu^*)$ and $\mathcal R_j$ is connected to $\mathcal R_{j+1}$ along $\Delta_j$ for $j=1, \ldots, r-1$ in the usual crosswise manner. After adding points at infinity we obtain a compact Riemann surface, since $\operatorname{supp}(\mu^*)$ consists of a finite union of intervals.
We define a function $\Psi$ on $\mathcal R$ by specifying on each of the sheets $$\begin{gathered}
\Psi (z) = \begin{cases}
V'(z) - F_0(z), & \text{for } z \in \mathcal R_0, \\
F_{j-1}(z) - F_{j}(z), & \text{for } z \in \mathcal R_j, \quad j = 1, \ldots, r.
\end{cases}\end{gathered}$$ Then $\Psi$ is meromorphic on each of sheets (since $V'$ is a rational function). Moreover, the variational conditions tell us that $\Psi$ extends to a meromorphic function on the full Riemann surface $\mathcal R$. Then also $V' - \Psi$ is a meromorphic function on $\mathcal R$ which agrees with $F_0$ on the zero sheet. Therefore $F = F_0$ satisfies an algebraic equation of degree $r+1$.
Note that $\mu^*$ minimizes $I(\mu) + \int (V - U^{\mu_1^*}) d\mu$ among all probability measures $\mu$ on $[0,\infty)$, and the external field $V - U^{\mu_1^*}$ is real analytic on $(0,\infty)$. If it were also real analytic at $0$, then it would follow from results in [@DeKrMc99] that $\mu^*$ is supported on a finite union of intervals. Maybe the methods of [@DeKrMc99] can be adapted to the present situation, and then the assumption in Proposition \[prop51\] about the finite number of intervals would be unnecessary.
The Riemann surface in the proof of Proposition \[prop51\] has genus $0$ if and only if $\operatorname{supp}(\mu^*) = [0,a]$ for some $a > 0$. This is the case if $V(x)= x$, and for more general conditions see [@ClaRom13 Theorem 1.8]. We note that this Riemann surface also appears in the paper of Forrester, Liu and Zinn-Justin [@FoLiZJ15], see Fig. 1 in that paper.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The author is supported by long term structural funding–Methusalem grant of the Flemish Government, by KU Leuven Research Grant OT/12/073, by the Belgian Interuniversity Attraction Pole P07/18, and by FWO Flanders projects G.0934.13 and G.0864.16.
[99]{}
Aptekarev A.I., Kuijlaars A.B.J., Hermite–[P]{}adé approximations and ensembles of multiple orthogonal polynomials, [*Russ. Math. Surv.*](http://dx.doi.org/10.1070/RM2011v066n06ABEH004771) **66** (2011), 1133–1199.
Beckermann B., Kalyagin V., Matos A.C., Wielonsky F., Equilibrium problems for vector potentials with semidefinite interaction matrices and constrained masses, [*Constr. Approx.*](http://dx.doi.org/10.1007/s00365-012-9165-z) **37** (2013), 101–134, [arXiv:1105.3088](http://arxiv.org/abs/1105.3088).
Bloom T., Levenberg N., Totik V., Wielonsky F., Modified logarithmic potential theory and applications, [*Int. Math. Res. Not.*](http://dx.doi.org/10.1093/imrn/rnw059), [t]{}o appear, [arXiv:1502.06925](http://arxiv.org/abs/1502.06925).
Borodin A., Biorthogonal ensembles, [*Nuclear Phys. B*](http://dx.doi.org/10.1016/S0550-3213(98)00642-7) **536** (1999), 704–732, [math.CA/9804027](http://arxiv.org/abs/math.CA/9804027).
Butez R., Large deviations principle for biorthogonal ensembles and variational formulation for the [D]{}ykema–[H]{}aagerup distribution, [arXiv:1602.07201](http://arxiv.org/abs/1602.07201).
Cheliotis D., Triangular random matrices and biorthogonal ensembles, [arXiv:1404.4730](http://arxiv.org/abs/1404.4730).
Claeys T., Romano S., Biorthogonal ensembles with two-particle interactions, [*Nonlinearity*](http://dx.doi.org/10.1088/0951-7715/27/10/2419) **27** (2014), 2419–2444, [arXiv:1312.2892](http://arxiv.org/abs/1312.2892).
Deift P., Kriecherbauer T., McLaughlin K.T.-R., New results on the equilibrium measure for logarithmic potentials in the presence of an external field, [*J. Approx. Theory*](http://dx.doi.org/10.1006/jath.1997.3229) **95** (1998), 388–475.
Duits M., Kuijlaars A.B.J., An equilibrium problem for the limiting eigenvalue distribution of banded [T]{}oeplitz matrices, [*SIAM J. Matrix Anal. Appl.*](http://dx.doi.org/10.1137/070687141) **30** (2008), 173–196, [arXiv:0704.0378](http://arxiv.org/abs/0704.0378).
Eichelsbacher P., Sommerauer J., Stolz M., Large deviations for disordered bosons and multiple orthogonal polynomial ensembles, [*J. Math. Phys.*](http://dx.doi.org/10.1063/1.3603994) **52** (2011), 073510, 16 pages, [arXiv:1102.0792](http://arxiv.org/abs/1102.0792).
Forrester P.J., Liu D.-Z., Raney distributions and random matrix theory, [*J. Stat. Phys.*](http://dx.doi.org/10.1007/s10955-014-1150-4) **158** (2015), 1051–1082, [arXiv:1404.5759](http://arxiv.org/abs/1404.5759).
Forrester P.J., Liu D.-Z., Zinn-Justin P., Equilibrium problems for [R]{}aney densities, [*Nonlinearity*](http://dx.doi.org/10.1088/0951-7715/28/7/2265) **28** (2015), 2265–2277, [arXiv:1411.4091](http://arxiv.org/abs/1411.4091).
Forrester P.J., Wang D., Muttalib–[B]{}orodin ensembles in random matrix theory – realisations and correlation functions, [arXiv:1502.07147](http://arxiv.org/abs/1502.07147).
Hardy A., Kuijlaars A.B.J., Weakly admissible vector equilibrium problems, [*J. Approx. Theory*](http://dx.doi.org/10.1016/j.jat.2012.03.009) **164** (2012), 854–868, [arXiv:1110.6800](http://arxiv.org/abs/1110.6800).
Kuijlaars A.B.J., Multiple orthogonal polynomial ensembles, in Recent Trends in Orthogonal Polynomials and Approximation Theory, [*Contemp. Math.*](http://dx.doi.org/10.1090/conm/507/09958), Vol. 507, Amer. Math. Soc., Providence, RI, 2010, 155–176, [arXiv:0902.1058](http://arxiv.org/abs/0902.1058).
Kuijlaars A.B.J., Stivigny D., Singular values of products of random matrices and polynomial ensembles, [*Random Matrices Theory Appl.*](http://dx.doi.org/10.1142/S2010326314500117) **3** (2014), 1450011, 22 pages, [arXiv:1404.5802](http://arxiv.org/abs/1404.5802).
Muttalib K.A., Random matrix models with additional interactions, [*J. Phys. A: Math. Gen.*](http://dx.doi.org/10.1088/0305-4470/28/5/003) **28** (1995), L159–L164, [cond-mat/9405084](http://arxiv.org/abs/cond-mat/9405084).
Neuschel T., Van Assche W., Asymptotic zero distribution of [J]{}acobi–[P]{}iñeiro and multiple [L]{}aguerre polynomials, [*J. Approx. Theory*](http://dx.doi.org/10.1016/j.jat.2016.02.002) **205** (2016), 114–132, [arXiv:1509.04542](http://arxiv.org/abs/1509.04542).
Nikishin E.M., Sorokin V.N., Rational approximations and orthogonality, *Translations of Mathematical Monographs*, Vol. 92, Amer. Math. Soc., Providence, RI, 1991.
Saff E.B., Totik V., Logarithmic potentials with external fields, [*Grundlehren der Mathematischen Wissenschaften*](http://dx.doi.org/10.1007/978-3-662-03329-6), Vol. 316, Springer-Verlag, Berlin, 1997.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Thermal dark matter that couples more strongly to electrons and photons than to neutrinos will heat the electron-photon plasma relative to the neutrino background if it becomes nonrelativistic after the neutrinos decouple from the thermal background. This results in a reduction in $N_{eff}$ below the standard-model value, a result strongly disfavored by current CMB observations. Taking conservative lower bounds on $N_{eff}$ and on the decoupling temperature of the neutrinos, we derive a bound on the dark matter particle mass of $m_\chi > 3-9$ MeV, depending on the spin and statistics of the particle. For $p$-wave annihilation, our limit on the dark matter particle mass is stronger than the limit derived from distortions to the CMB fluctuation spectrum produced by annihilations near the epoch of recombination.'
author:
- 'Chiu Man Ho and Robert J. Scherrer'
title: Limits on MeV Dark Matter from the Effective Number of Neutrinos
---
Roughly $20-25$% of the total energy content of the universe is in the form of non-baryonic dark matter. While a dark matter particle mass in the GeV range is often assumed, there has also been interest in masses in the MeV range. Dark matter with a mass in this range was invoked to explain the 511 keV $\gamma$-rays observed by INTEGRAL [@Boehm], and to explain the cosmic $\gamma$-ray background at $1-20$ MeV [@Ahn]. Supersymmetric models with MeV dark matter have been proposed [@Hooper], and MeV dark matter can arise in the context of the WIMPless dark matter model [@Feng]. MeV dark matter can have interesting effects on large-scale structure [@Kaplinghat].
We note here that a thermal MeV dark matter particle that couples more strongly to electrons and photons than to neutrinos will heat the electron-photon plasma when it becomes nonrelativistic before its abundance freezes out. If this occurs after the neutrinos decouple from the thermal background, then the ratio of the neutrino temperature to the photon temperature will be reduced, a process similar to the heating that occurs when the electron-positron pairs become nonrelativistic. The final result is a decrease in the effective number of neutrino degrees of freedom. This effect was first explored by Kolb et al. [@KTW] and more recently by Serpico and Raffelt [@SR] in the context of primordial nucleosynthesis. Recent CMB observations [@WMAP7B; @Dunkley; @Keisler] place severe lower bounds on $N_{eff}$, allowing us to constrain this process. (See also the earlier work of Ref. [@Dorosh], which examined heating of the photons relative to the neutrinos from decaying particles).
At recombination, the energy density in relativistic particles includes photons, whose temperature, $T_\gamma$, and therefore energy density is extremely well-measured, and a neutrino background with temperature $T_\nu = (4/11)^{1/3} T_\gamma$. The theoretical prediction for the effective number of neutrinos (assuming slight reheating of the neutrinos from early $e^+e^-$ annihilation) is $N_{eff} = 3.046$ [@Dolgov; @mangano1]. The neutrino density cannot be measured directly, but it can be inferred from measurements of the CMB. (For a discussion of the effect of $N_{eff}$ on the CMB fluctuations, see Refs. [@Bashinsky; @Hou]). The values of $N_{eff}$ from recent CMB observations, in combination with other cosmological data, are $N_{eff} = 4.34^{+0.86}_{-0.88}$ (68% CL) from WMAP [@WMAP7B], $N_{eff} = 4.56 \pm 0.75$ (68% CL) from the Atacama Cosmology Telescope [@Dunkley], and $N_{eff} = 3.86 \pm 0.42$ (68% CL) from the South Pole Telescope [@Keisler]. Archidiacono et al. [@Arch] used combined datasets to derive $N_{eff} = 4.08^{+0.71}_{-0.68}$ (95% CL). Clearly, the data favor values of $N_{eff}$ larger than the standard-model theoretical prediction, rather than smaller.
The extent of the heating from dark matter annihilation in the early universe can be derived from entropy conservation (see Refs. [@Weinberg; @KT], from which our discussion is derived). Our paper assumes a dark matter particle that couples much more strongly to electrons and photons than to neutrinos. The most natural example of such a particle is one that interacts with ordinary matter through an electromagnetic form factor, such as an electric or magnetic dipole [@Pospelov; @Sigurdson; @Gardner; @Masso; @Fitzpatrick; @Cho; @Heo1; @Heo2; @Banks; @Barger1; @Fortin; @Nobile; @Barger2; @Heo3], or an anapole moment [@HoScherrer2]. Dark matter particles in this category annihilate into Standard Model particles through the mediation of photons, while the models considered by Refs. [@Boehm; @Ahn; @Hooper; @Feng; @Kaplinghat] require the mediation of a new fermion or vector boson. In fact, the dark matter particles considered in Refs. [@Boehm; @Ahn; @Hooper; @Feng] could be relevant if their coupling with neutrinos is postulated to be suppressed. However, the model considered by Ref. [@Kaplinghat] requires that the dark matter particle couples to electrons and neutrinos equally, and so it is not relevant.
Let $\chi \bar \chi$ denote the pair of dark matter particles. To make our study general, we will allow a range of possibilities for the dark matter, including a self-conjugate scalar, a non-self-conjugate scalar, a spin-1/2 Majorana fermion or a spin-1/2 Dirac fermion. Thus, for the cases with self-conjugate and non-self-conjugate scalars, the notation $\chi \bar \chi$ really means $\chi \chi$ and $\chi \chi^\ast$ respectively. But for simplicity, we will keep the notation $\chi \bar \chi$ throughout the paper.
Consider first the case where the dark matter annihilates entirely after the neutrinos decouple, which occurs at a temperature of $T_{d} \approx 2-3$ MeV [@Dolgov; @Enqvist]. The total entropy prior to $\chi \bar \chi$ annihilation is proportional to $$S = \frac{R^3}{T}(\rho_{e^+e^-} + \rho_{\gamma} + \rho_{\chi \bar \chi} + p_{e^+e^-} +
p_{\gamma} + p_{\chi \bar \chi}),$$ while after $\chi \bar \chi$ annihilation it is $$S = \frac{R^3}{T}(\rho_{e^+e^-} + \rho_{\gamma} + p_{e^+e^-} +
p_{\gamma}).$$ For a relativistic particle, $p = \rho/3$, so following Ref. [@KT], we can write the total entropy density as $$s=\frac{\rho_{\textrm{tot}}+p_{\textrm{tot}}}{T} = \frac{2 \pi^2}{45} g_{*S} T^3,$$ where $g_{*S}$ is the total number of spin degrees of freedom for bosons, and 7/8 times the total number of spin degrees of freedom for fermions. Then the total entropy is $$S = \frac{2 \pi^2}{45} g_{*S} (RT)^3,$$ which is conserved through the process of any particle becoming nonrelativistic and annihilating. So the ratio of the final value of $RT$ after annihilation to the initial value of $RT$ prior to annihilation is $$\label{heating}
\frac{(RT)_f}{(RT)_i} = \left(\frac{g_{*Si}}{g_{*Sf}}\right)^{1/3},$$ where $g_{*Si}$ and $g_{*Sf}$ are the values of $g_{*S}$ for the relativistic particles in thermal equilibrium before and after annihilation, respectively. When the $\chi \bar \chi$ pairs annihilate after neutrino decoupling, the neutrinos do not share in the heating, so that $RT_\nu$ is constant and $T_\nu \propto R^{-1}$, while the photons and electron-positron pairs are heated as in Eq. (\[heating\]). Therefore, for the $\chi \bar \chi$ pairs with $g$ internal degrees of freedom, the ratio of $T_\nu$ to $T_\gamma$ after $\chi \bar \chi$ annihilation is: $$\label{ratf}
T_\nu/T_\gamma = \left[\frac{(7/8)4 + 2}{(7/8)4 + 2 + (7/8)g}\right]^{1/3},$$ if $\chi$ is a fermion, and $$\label{ratb}
T_\nu/T_\gamma = \left[\frac{(7/8)4 + 2}{(7/8)4 + 2 + g}\right]^{1/3},$$ if it is a boson. Taking, for example, the $\chi$ particle to be a spin-1/2 Majorana fermion gives $g = 2$, so that $T_{\nu}/T_\gamma = (22/29)^{1/3}$. Subsequent $e^+e^-$ annihilation further heats the photon temperature relative to the neutrino temperature by a factor of $(11/4)^{1/3}$, so that the final ratio of the neutrino temperature to the photon temperature would be $(88/319)^{1/3}$.
In terms of $N_{eff}$, the energy density for neutrinos is given by $$\rho_\nu = N_{eff}\left(\frac{7}{8}\right)\left(2\right)\left(\frac{\pi^2}{30}\right)
\left(\frac{T_\nu}{T_\gamma}\right)^4 T_\gamma^4.$$ Since $\rho_\nu$ at fixed $T_\gamma$ is the quantity that is inferred from CMB observations, a change in $T_\nu/T_\gamma$ will be interpreted as a change in $N_{eff}$, with $N_{eff} \propto
(T_\nu/T_\gamma)^4$. In this case, $\chi \bar \chi$ annihilation reduces the value of $T_\nu/T_\gamma$ relative to its value in the standard model by a factor of $(22/29)^{1/3}$, which corresponds to $N_{eff} = 3(22/29)^{4/3} = 2.1$, a value clearly excluded by the CMB observations.
This value of $N_{eff}$ corresponds to a dark matter particle with a mass well below the neutrino decoupling temperature. However, to derive a useful limit, we must consider what happens when $\chi$ annihilates during neutrino decoupling. Neutrino decoupling is not a sudden process, but for the purposes of our simplified calculation, we will take it to occur abruptly at a fixed temperature $T_d$, and we will assume that dark matter annihilations before $T_d$ fully heat the neutrinos, while those after $T_d$ heat only the photons and $e^+ e^-$ pairs. Let $I(T_\gamma)$ be given by (see, e.g., Ref. [@Weinberg] for a similar calculation):
$$\begin{aligned}
I(T_\gamma)&\equiv& \frac{S}{(RT_\gamma)^3} = \frac{1}{T_\gamma^4}\,(\rho_{e^+e^-} + \rho_{\gamma} + \rho_{\chi \bar \chi} + p_{e^+e^-} +
p_{\gamma} + p_{\chi \bar \chi}), \nonumber\\
\label{formula}
&=& \frac{11}{45} \pi^2 + \frac {g}{2 \pi^2} \int_{x=0}^\infty x^2 dx \left
(\sqrt{x^2 + (m_\chi/T_\gamma)^2} + \frac{x^2}{3\sqrt{x^2 + (m_\chi/T_\gamma)^2}}\right)
\left[\exp(\sqrt{x^2 + (m_\chi/T_\gamma)^2} \pm 1)\right]^{-1},\end{aligned}$$
where the plus (minus) sign is for a fermionic (bosonic) dark matter particle, and the variable of integration is $x = p_\chi/T_\gamma$. In the limit where all particles are fully relativistic, $I$ reduces to $(2 \pi^2/45)g_{*S}$; the integral in Eq. (\[formula\]) just quantifies the contribution to $I$ from $\chi \bar \chi$ as they become nonrelativistic.
As mentioned above, the $\chi \bar \chi$ annihilation will heat up photons relative to neutrinos only after neutrino decoupling. But this heating ends when the $\chi \bar \chi$ particles drop out of thermal equilibrium. Thus, the ratio of the neutrino temperature to the photon temperature due to $\chi \bar \chi$ annihilation alone is $$T_\nu/T_\gamma = \left[\frac{I(T_f)}{I(T_d)}\right]^{1/3},$$ where $T_f$ is the temperature at which the $\chi \bar \chi$ particles freeze out. Since $m_\chi/T_f \sim 20$ [@KT], it is obvious from Eq. (\[formula\]) that we can simply set $T_f = 0$ with negligible error: $$\label{Tratio}
T_\nu/T_\gamma = \left[\frac{I(0)}{I(T_d)}\right]^{1/3}.$$ The physical reason for this is that the $\chi \bar \chi$ abundance freezes out at a temperature of $T_f \sim m_\chi/20$, while most of the entropy from the $\chi \bar \chi$ annihilations is transferred to the thermal background when $T \sim m_\chi/3$. Of course, the temperature ratio given by Eq. (\[Tratio\]) must then be multiplied by an additional factor of $(4/11)^{1/3}$ from $e^+ e^-$ annihilations to obtain the final ratio of the neutrino temperature to the photon temperature.
In this approximation, the effective number of neutrinos as measured by CMB experiments will be given by $$N_{eff} = 3.046\left[\frac{I(0)}{I(T_d)}\right]^{4/3}.$$ The value of $N_{eff}$ as a function of $m_\chi/T_d$ is shown in Fig. 1, for a self-conjugate scalar boson ($g=1$), a non-self-conjugate scalar boson ($g=2$), a spin-1/2 Majorana fermion ($g=2$) and a spin-1/2 Dirac fermion ($g=4$).
=3.8truein
In fact, from Eqs. (\[ratf\])-(\[ratb\]), we can derive the $m_\chi \ll T_d$ limit for $N_{eff}$, namely $$N_{eff} = 3.046\left[\frac{11}{11 + (7/4)g}\right]^{4/3},$$ for fermionic $\chi$, and $$N_{eff} = 3.046\left[\frac{11}{11 + 2g}\right]^{4/3},$$ for bosonic $\chi$.
As noted earlier, neutrino decoupling is not a sudden process, so $T_d$ is not completely well-defined. Ref. [@Enqvist] gives a widely cited value of $T_d = 2.3$ MeV for the electron neutrinos, with the $\mu$ and $\tau$ neutrinos decoupling at a higher temperature. However, neutrino oscillations will tend to equilibrate the decoupling of all three neutrinos, an effect discussed in Refs. [@mangano1; @Hannestad]. Here we will simply take $T_d {\ {\raise-.5ex\hbox{$\buildrel>\over\sim$}}\ }2$ MeV as a conservative lower bound. Note that the presence of the additional relativistic energy density from the $\chi \bar \chi$ particles themselves will increase $T_d$, but this turns out to be a miniscule effect [@HoScherrer3].
Now we must determine a reasonable lower bound on $N_{eff}$. The combined results from Refs. [@WMAP7B; @Dunkley; @Keisler] are barely consistent with the standard model value of $N_{eff} = 3.046$. However, we will err on the side of caution and choose a lower bound of $N_{eff} > 2.6$, which is excluded at $2\sigma$ by all three sets of CMB observations.
These limits on $N_{eff}$ and $T_d$ can be combined with the results displayed in Fig. 1 to derive a lower bound on $m_\chi$. These bounds are $m_\chi {\ {\raise-.5ex\hbox{$\buildrel>\over\sim$}}\ }3$ MeV for the self-conjugate scalar boson, $m_\chi {\ {\raise-.5ex\hbox{$\buildrel>\over\sim$}}\ }6$ MeV for a two-component boson or fermion, and $m_\chi {\ {\raise-.5ex\hbox{$\buildrel>\over\sim$}}\ }9$ MeV for a Dirac fermion.
These limits are relevant for several models in the literature. As noted by Beacom and Yuksel [@Beacom], the model proposed in Ref. [@Boehm] actually requires positron injection at very low energies (${\ {\raise-.5ex\hbox{$\buildrel<\over\sim$}}\ }3$ MeV) to produce the 511 keV $\gamma$-rays observed by INTEGRAL [@Boehm]. But dark matter masses low enough to produce such particles from annihilations are ruled out by our limit. Thermal dark matter with the correct relic abundance interacting through an electric or magnetic dipole moment must have a mass less than $1-10$ GeV to avoid conflict with direct detection experiments [@Fortin]; our results shrink the allowed window from the other direction.
Our limits are complementary to several others in the literature. As noted, dark matter particles with masses in this range also affect primordial nucleosynthesis, and bounds can be placed from the observed element abundances, particularly helium-4. However, the effect on $N_{eff}$ as measured by the CMB appears to provide a better limit. For example, in the $1-10$ MeV mass range, Serpico and Raffelt [@SR] found a maximum reduction of only 0.002 in the primordial helium mass fraction. Using the results of Ref. [@Steigman], this corresponds to $\Delta N_{eff} = -0.15$, much smaller than the typical values in Fig. 1. However, there is no contradiction between our results and those of Ref. [@SR]. When $T_\nu/T_\gamma$ is reduced prior to primordial nucleosynthesis, there are actually two effects on the helium-4 abundance. First, the reduction in the expansion rate at fixed $T_\gamma$ reduces the helium-4 abundance, and this is the dominant effect, as noted by Serpico and Raffelt. However, there is a second effect which partially cancels the first: the decrease in the electron neutrino temperature reduces the weak interaction rates, which tends to increase the helium-4 abundance. Thus, the effect on BBN is smaller than if one reduced the overall expansion rate alone.
Another lower bound on $m_\chi$ comes from distortions to the CMB fluctuation spectrum due to annihilations near the epoch of recombination [@Padmanabhan; @Mapelli; @Zhang; @Galli; @Hutsi; @Finkbeiner]. This effect excludes dark matter with masses ${\ {\raise-.5ex\hbox{$\buildrel<\over\sim$}}\ }1-10$ GeV, a much tighter bound than ours (note that such annihilations also distort the [*spectrum*]{} of the CMB [@MTW; @Chluba], but these bounds are weaker given present observations). However, the CMB fluctuation bound only applies to $s$-wave annihilations, for which $\langle \sigma v \rangle$ does not change between the dark matter particle freeze-out and the epoch of recombination. For $p$-wave annihilations, the annihilation rate at recombination is generally negligible, and the CMB cannot be used to constrain such models. Therefore, this CMB constraint is applicable to the model considered in Ref. [@Feng] and a dark matter particle with a magnetic dipole moment [@Pospelov; @Sigurdson; @Gardner; @Masso; @Fitzpatrick; @Cho; @Heo1; @Heo2; @Banks; @Barger1; @Fortin; @Nobile; @Barger2; @Heo3]. It is not applicable to the models considered in Refs. [@Boehm; @Ahn; @Hooper] and a dark matter particle with an electric dipole moment [@Pospelov; @Sigurdson; @Gardner; @Masso; @Fitzpatrick; @Cho; @Heo1; @Heo2; @Banks; @Barger1; @Fortin; @Nobile; @Barger2; @Heo3] or an anapole moment [@HoScherrer2], because all of these models can be $p$-wave dominated. In these cases our limit provides the better constraint.
In contrast to the CMB constraint, our bounds do not depend on the velocity dependence of the annihilation cross section and therefore provide a good constraint in the case of $p$-wave annihilations. Indeed, the values of $N_{eff}$ derived in Refs. [@WMAP7B; @Dunkley; @Keisler] assume a standard recombination history, undistorted by dark matter annihilation, so it is unclear how $s$-wave annihilation at the epoch of recombination would affect the estimated values of $N_{eff}$. Of course, the reverse is also true; the bounds derived in Refs. [@Padmanabhan; @Mapelli; @Zhang; @Galli; @Hutsi; @Finkbeiner] do not take into account the effect we have outlined in this paper.
The bounds presented here can be evaded if the dark matter is asymmetric (see, e.g., Ref. [@LYZ] and references therein). Also, our bounds will be weakened to the extent that the dark matter couples to both the electron-photon plasma and to neutrinos. In fact, in the extreme opposite limit (coupling to neutrinos only), the $\chi \bar \chi$ annihilation heats the neutrinos instead of the photons, increasing $N_{eff}$ and providing better agreement with current observations [@Boehm2].
There is one obvious caveat to the bounds we have derived here. As noted earlier, the CMB limits on $N_{eff}$ are only in marginal agreement even with the standard model value for $N_{eff}$. If future observations show conclusive evidence that the observed $N_{eff}$ disagrees with the standard model, some mechanism will be required to generate the additional relativistic degrees of freedom, and this mechanism could also be invoked to erase the effects of the annihilating dark matter particle. (See, e.g., Ref. [@HoScherrer3]). Future PLANCK observations should help to resolve this issue. More precise observational bounds on $N_{eff}$ would also justify a more exact treatment of the effect outlined here, going beyond our simplifying assumption of sudden neutrino decoupling to a full numerical integration of the equations governing neutrino evolution in the early universe.
We thank D. Hooper for helpful discussions. C.M.H. and R.J.S. were supported in part by the Department of Energy (DE-FG05-85ER40226).
[99]{}
C. Boehm, D. Hooper, J. Silk, and M. Casse, , 101301 (2004).
K. Ahn and E. Komatsu, , 061301 (2005).
D. Hooper and K.M. Zurek, , 087302 (2008).
J. L. Feng and J. Kumar, , 231301 (2008).
D. Hooper, M. Kaplinghat, L. E. Strigari, and K. M. Zurek, , 103515 (2007).
E. W. Kolb, M. S. Turner, and T. P. Walker, , 2197 (1986).
P. D. Serpico and G. G. Raffelt, , 043526 (2004).
E. Komatsu, et al., Astrophys. J. Suppl. [**192**]{}, 18 (2011).
J. Dunkley, et al., , 52 (2011).
R. Keisler, et al., , 28 (2011).
A. G. Doroshkevich and M. Yu. Khlopov, Sov. Astron. Lett. [**9**]{}, 171 (1983).
A. D. Dolgov, Phys. Rept. [**370**]{}, 333 (2002).
G. Mangano, et al., Nucl. Phys. B [**729**]{}, 221 (2005).
S. Bashinsky and U. Seljak, , 083002 (2004).
Z. Hou, et al., arXiv:1104.2333 \[astro-ph\].
M. Archidiacono, E. Calabrese, and A. Melchiorri, , 123008 (2011).
S. Weinberg, [*Gravitation and Cosmology*]{}, (New York: Wiley, 1972).
E. W. Kolb and M. S. Turner, [*The Early Universe*]{}, (New York: Addison-Wesley, 1990).
M. Pospelov and T. ter Veldhuis, Phys. Lett. B [**480**]{}, 181 (2000).
K. Sigurdson, M. Doran, A. Kurylov, R. R. Caldwell, and M. Kamionkowski, , 083501 (2004); erratum, , 089903 (2006).
S. Gardner, , 055007 (2009).
E. Masso, S. Mohanty, and S. Rao, , 036009 (2009).
A. L. Fitzpatrick and K. M. Zurek, , 075004 (2010).
W. S. Cho, et al., Phys. Lett. B [**687**]{}, 6 (2010); erratum, Phys. Lett. B [**694**]{}, 496 (2011).
J. H. Heo, Phys. Lett. B [**693**]{}, 255 (2010).
J. H. Heo, Phys. Lett. B [**702**]{}, 205 (2011).
T. Banks, J. -F. Fortin, and S. Thomas, arXiv:1007.5515 \[hep-ph\].
V. Barger, W. -Y. Keung, and D. Marfatia, Phys. Lett. B [**696**]{}, 74 (2011).
J. -F. Fortin and T. M. P. Tait, , 063506 (2012).
E. Del Nobile, et al., JCAP [**1208**]{}, 010 (2012).
V. Barger, W. -Y. Keung, D. Marfatia, and P. -Y. Tseng, Phys. Lett. B [**717**]{}, 219 (2012).
J. H. Heo and C. S. Kim, arXiv:1207.1341 \[astro-ph\].
C. M. Ho and R. J. Scherrer, arXiv:1211.0503 \[hep-ph\].
K. Enqvist, K. Kainulainen, and V. Semikoz, Nucl. Phys. B [**374**]{}, 392 (1992).
S. Hannestad, , 083006 (2002).
C. M. Ho and R. J. Scherrer, arXiv:1212.1689 \[hep-ph\].
J. F. Beacom and H. Yuksel, , 071102 (2006).
G. Steigman, Ann. Rev. Nucl. Part. Sci. [**57**]{}, 463 (2007).
N. Padmanabhan and D. P. Finkbeiner, , 023508 (2005).
M. Mapelli, A. Ferrara, and E. Pierpaoli, Mon. Not. R. Astr. Soc. [**369**]{}, 1719 (2006).
L. Zhang, X. Chen, Y. -A. Lei, and Z. -g. Si, , 103519 (2006).
S. Galli, F. Iocco, G. Bertone, and A. Melchiorri, , 027302 (2011).
G. Hutsi, J. Chluba, A. Hektor, and M. Raidal, Astron. Astrophys. [**535**]{}, A26 (2011).
D. P. Finkbeiner, S. Galli, T. Lin, and T. R. Slatyer, Phys. Rev. D [**85**]{}, 043522 (2012).
P. McDonald, R. J. Scherrer, and T. P. Walker, , 023001 (2001).
J. Chluba and R. A. Sunyaev, Mon. Not. R. Astr. Soc. [**419**]{}, 1294 (2012).
T. Lin, H. -B. Yu, and K. M. Zurek, 063503 (2012).
C. Boehm, M. J. Dolan, and C. McCabe, arXiv:1207.0497 \[astro-ph\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Within Resonance Chiral Theory and in the context of QCD current correlators at next-to-leading order in $1/N_C$, we have analyzed the two-body form factors which include resonances as a final state . The short-distance constraints have been studied. One of the main motivations is the estimation of the chiral low-energy constants at subleading order, that is, keeping full control of the renormalization scale dependence. As an application we show the resonance estimation of some coupling, $L_{10}^r(\mu_0)=(-4.4\pm 0.9)\cdot 10^{-3}$ and $C_{87}^r(\mu_0)=(3.1\pm 1.1)\cdot 10^{-5}$.'
address:
- 'Departamento de Ciencias Físicas, Matem'' aticas y de la Computaci'' on, Universidad CEU Cardenal Herrera, c/Sant Bartomeu 55, E-46115 Alfara del Patriarca, Val\` encia, Spain'
- 'IFIC, Universitat de Val\` encia - CSIC, Apt. Correus 22085, E-46071 Val\` encia, Spain'
author:
- 'I. Rosell'
title: 'Current correlators and form factors in the resonance region [^1] '
---
Motivation
==========
A possibility to deal with Quantum Chromodynamics (QCD) at low energies is the use of effective field theories [@EFT]. Chiral Perturbation Theory (ChPT) is the effective field theory of QCD at very low energies [@ChPT] and it is constructed by using a perturbative expansion in the momenta and masses of the pseudo-Goldstone bosons. Chiral Perturbation Theory has been used up to next-to-next-to-leading order and one of its major problems is the estimation of the increasing number of low-energy constants (LECs), once the desired precision is growing.
The construction of an effective lagrangian in the resonance region is much more involved. The existence of many resonances at intermediate energies ($M_\rho < E < 2$ GeV), the absence of a mass gap to integrate out the heavier degrees of freedom and the lack of a natural expansion parameter makes difficult a formal effective field theory approach. Resonance Chiral Theory (RChT) provides a correct framework to incorporate the resonance fields [@RChT]. One considers the most general possible lagrangian, including all terms consistent with assumed symmetry principles. The key ingredients that allows the phenomenology are the $1/N_C$ expansion and the use of the information coming from QCD.
Assuming confinement, the $N_C \rightarrow \infty$ limit guarantees that meson dynamics are described by tree-level interactions of an effective local lagrangian including only meson degrees of freedom, higher corrections in $1/N_C$ being obtained by loop corrections [@NC]. On the other hand, the use of the short-distance information from QCD is fundamental to reduce the number of unknown couplings.
It is important to stress that the only model dependence of our approach is the cut of the tower of resonances. In other words, although in the context of the $1/N_C$ expansion an infinite number of resonances is required to recover the usual QCD results, RChT only considers the lightest resonance multiplets. It seems a good approximation taking into account that heavier contributions are expected to be suppressed by their masses. Furthermore, this idea is supported by the phenomenology.
Besides the logic interest of resonance lagrangians to make physics in the resonance region, one of its main motivations is the estimation of the chiral LECs. That is, one can integrate out the resonance fields to estimate the chiral couplings in terms of resonance parameters. This matching is more interesting when the short-distance information is used to reduce the number of unknown parameters. Put differently, our phenomenological approach can be understood as a bridge between high and low energies: chiral LECs cannot be obtained directly from the high-energy lagrangian (ChPT is a non perturbative approach), but RChT is able to match with both theories. Somehow Resonance Chiral Theory allows to transport the information from high energy to ChPT.
This estimation has been made at leading order (LO) in $1/N_C$ for all the ${{\cal O}}(p^4)$ and some ${{\cal O}}(p^6)$ LECs [@RChT]. Actually, one of the main motivations to work with RChT at one-loop level is precisely this estimation at next-to-leading order (NLO), since the leading estimation is unable to control the renormalization-scale dependence of the couplings, which are unknown and could be sizable. This project is a step towards the understanding of quantum loops within RChT [@NLO; @L10], which is necessary also to improve the hadronic contributions to distinguish new physics effects from Standard Model results in some observables.
Correlators and form factors
============================
Let us consider the two-point correlation function of two currents in the chiral limit: $$\begin{aligned}
\Pi_{X}^{\mu\nu}\! &\equiv& i \!\int \mathrm{d}^4x \, \mathrm{e}^{iqx}\;
\langle 0|T\left(J_X^\mu(x)J_X^\nu(0)^\dagger\right)|0\rangle \nonumber \\
&=& \left( -g^{\mu\nu} q^2 + q^\mu q^\nu \right)\,\Pi_{X}(q^2)\, , \nonumber \\
\Pi_{Y}\!&\equiv& i\!\int \mathrm{d}^4x \, \mathrm{e}^{iqx}\;
\langle 0|T\left(J_Y(x)J_Y(0)^\dagger\right)|0\rangle \, , \label{def1}\end{aligned}$$ where $J_X^\mu(x)$ can denote the vector or axial-vector currents and $J_Y(x)$ the scalar or pseudo-scalar densities.
At large $q^2$, the vector and axial-vector spectral functions tend to a constant whereas the scalar and pseudoscalar ones grow like $q^2$. Therefore, in the first case and considering that the spectral function is a sum of positive contributions corresponding to the different intermediate states and there is an infinite number of possible states, the absorptive contribution of a given state should vanish at infinite momentum transfer. Following the same argument, in the scalar and pseudoscalar case one would require spectral functions growing as a constant. However, the $SS-PP$ sum-rules, the Brodsky-Lepage counting rules [@brodsky-lepage] and the $1/q^2$ behavior of each one-particle intermediate cut seem to indicate that the vanishing assumption is reasonable.
Thus, we consider two sources of short-distance constraints. First, and taking into account that the optical theorem relates the spectral cuts with the corresponding two-body form factors, one can consider vanishing form factors. On the other hand, the matching between the resonance results with the ones obtained in the Operator Product Expansion (OPE) gives more constraints.
Within RChT we have calculated all two-body form factors associated with the scalar, pseudoscalar, vector and axial-vector currents [@L10] and have analyzed the high-energy constraints coming from well-behaved spectral functions.
The V-A correlator in RChT
--------------------------
We consider the difference between the two-point correlation function of two vector and two axial-vector currents, $\Pi(t)=\Pi_V(t)-\Pi_A(t)$, with $t=q^2$. Within RChT and at leading order, $\Pi(t)$ reads $$\label{eq:LO}
\Pi(t) = \frac{2F^2}{t} +\sum_i \left[
\frac{ 2\, F_{V_i}^2}{ M_{V_i}^2-t }
- \frac{2\, F_{A_i}^2}{ M_{A_i}^2-t} \right] ,$$ which involves an infinite number of vector and axial-vector resonance exchanges. At the NLO in $1/N_C$, $\Pi(t)$ has moreover one-loop corrections, $$\label{eq:Pi_structure}
\Pi(t) = \!
\frac{2 F^2}{t} +\! \sum_i \!\left[ \frac{2\, F_{V_i}^{r\,\, 2}}{M_{V_i}^{r\,\,2} - t}
- \frac{2\, F_{A_i}^{r\,\,2}}{M_{A_i}^{r\,\, 2} - t} \right] \!+ \widetilde{\Pi}(t) ,$$ being $\widetilde{\Pi}(t)$ the contributions associated with two-meson absorptive cuts. We have considered only the lowest-mass two-particle exchanges: two pseudo-Goldstone bosons or one pseudo-Goldstone boson and one resonance (higher thresholds are kinematically suppressed [@L10]). $\Pi(t)$ can be obtained from the spectral functions through a dispersive relation, up to a term which at NLO has the same structure as tree-level resonance exchanges. Therefore, this term can be absorbed by a redefinition of the resonance couplings and masses.
We have adopted the Single Resonance Approximation (SRA) as a first approach, where just the lightest resonances with non-exotic quantum numbers are considered. Making use of the Weinberg sum-rules, the $S-P$ sum rules and well-behaved form factors, one is able to have all the relevant resonance parameters in terms of the pion decay constant $F$ and resonance masses [@L10]. The only remanent thing before going to the phenomenology is to fix $F_V^r$ and $F_A^r$, which can be done by studying the asymptotic behavior of the observable at hand at subleading order, $$\begin{aligned}
F_V^{r\,\,2}&=& \frac{F^2\,M_A^{r\,\,2}}{M_A^{r\,\,2}-M_V^{r\,\,2}}
\! \left(1+\delta_{_{\rm NLO}}^{(1)}-\frac{M_V^2}{M_A^2}\delta_{_{\rm NLO}}^{(2)} \right)
,\nonumber \\F_A^{r\,\,2}&=& \frac{F^2\,M_V^{r\,\,2}}{M_A^{r\,\,2}-M_V^{r\,\,2}}
\! \left(1+\delta_{_{\rm NLO}}^{(1)}-\delta_{_{\rm NLO}}^{(2)} \right)
,\label{dmr}\end{aligned}$$ where $\delta_{_{\rm NLO}}^{(i)}$ parameterize the asymptotic expressions of the one-loop contribution $$\begin{aligned}
\widetilde{\Pi}(t) \!\!\!
&=\!\!\!& \frac{2F^2}{t}
\left( \delta_{_{\rm NLO}}^{(1)} + \widetilde{\delta}_{_{\rm NLO}}^{(1)} \ln{\frac{-t}{M_V^2}} \right)
\nonumber \\
&&+\frac{2F^2 M_V^2}{t^2} \!\left( \delta_{_{\rm
NLO}}^{(2)} \!+ \widetilde{\delta}_{_{\rm NLO}}^{(2)}
\ln\frac{-t}{M_V^2}\right)\!+\! \dots ,\end{aligned}$$ where the dots indicates subleading terms in the high-energy expansion. Note that the constraints $\widetilde{\delta}_{_{\rm NLO}}^{(1)}=\widetilde{\delta}_{_{\rm NLO}}^{(2)}=0$ give relations between masses, $M_A=M_V$ and $M_P=\sqrt{2}M_S$.
In order to avoid some incompatibilities between different short-distance constraints appearing in the Single Resonance Approximation [@L10], we have included additional resonance multiplets ($V'$ and $A'$). Obviously then there is much more couplings. Notwithstanding, we can use the known constraints coming from the $<VAP>$ Green-function analysis of Ref. [@VAP] and group some new couplings into expected tiny correction $\epsilon_{i}$, given by $$\begin{aligned}
\epsilon_{1}&=&\frac{F_{A'}^2}{F^2} -\frac{F_{V'}^2}{F^2} \,, \nonumber \\
\epsilon_{2}&=&\frac{F_{A'}^2 M_{A'}^2-F_{V'}^2 M_{V'}^2 }{F^2 M_V^2}\,, \nonumber \\
\epsilon_{3}&=&\frac{F_{V'} G_{V'}}{F^2}\, . \label{FVprime}\end{aligned}$$ In our numerical calculations we only take the $\epsilon_i$ corrections into account when they appear at LO in $1/N_C$. All the procedure is equivalent to the case explained before, the only change is that now $F_V^r$ and $F_A^r$ also depends on $\epsilon_i$, so Eq. (\[dmr\]) reads now $$\begin{aligned}
F_V^{r\,\,2}\!\!\!\!\!&=\!\!\!\!\!&\! \frac{F^2\,M_A^{r\,\,2}}{M_A^{r\,\,2}\!-\!M_V^{r\,\,2}}
\!\!\left[1\!+\!\epsilon_{1}\!+\!\delta_{_{\rm NLO}}^{(1)}\!
\!-\!\frac{M_V^2}{M_A^2}\!\! \left(\!\epsilon_{2}\!+\! \delta_{_{\rm NLO}}^{(2)}\!\right) \!\right]\!\!
,\nonumber \\
F_A^{r\,\,2}\!\!\!\!\!&=\!\!\!\!\!& \! \frac{F^2 \,M_V^{r\,\,2}}{M_A^{r\,\,2}\!-\!M_V^{r\,\,2}}
\!\!\left[1\!+\!\epsilon_{1}\!+\!\delta_{_{\rm NLO}}^{(1)}\!\!-\!\epsilon_{2}\!-\!\delta_{_{\rm NLO}}^{(2)} \right]\!\!
,\label{dmrbis}\end{aligned}$$ where of course the expressions of $\delta_{_{\rm NLO}}^{(i)}$ have changed [@L10]. The constraints $\widetilde{\delta}_{_{\rm NLO}}^{(1)}=\widetilde{\delta}_{_{\rm NLO}}^{(2)}=0$ are now used to fix $M_{V'}$ and $M_{A'}$.
All the complete expressions of present and next section are shown in Ref.[@L10].
The chiral couplings $L_{10}^r(\mu)$ and $C_{87}^r(\mu)$
========================================================
At very low energies $\Pi(t)$ is determined by ChPT: $$\begin{aligned}
\Pi(t) \!\!&\!\!=\!\!&\!\! \frac{2 F^2}{t} - 8 L_{10}^r(\mu) - \frac{\Gamma_{10}}{4\pi^2} \left( \frac{5}{3}-\ln \frac{-t}{\mu^2} \right)\nonumber \\
&& \!\! + \frac{t}{F^2} \bigg[ 16C_{87}^r (\mu) -\frac{\Gamma_{87}^{(L)}}{2\pi^2} \left( \frac{5}{3}-\ln \frac{-t}{\mu^2} \right) \nonumber \\
&& \qquad \qquad +{{\cal O}}\left(N_C^{0}\right) \bigg]+ {{\cal O}}\left(t^2\right) ,\label{eq:Pi_chpt}\end{aligned}$$ with $\Gamma_{10} = -1/4$ and $\Gamma_{87}^{(L)} = - L_9/2$. The couplings $F^2$, $L_{10}$ and $C_{87}/F^2$ are of ${{\cal O}}(N_C)$, while $\Gamma_{10}$ and $\Gamma_{87}^{(L)}/F^2$ are of ${{\cal O}}(N_C^0)$ and represent a NLO effect.
The large-$N_C$ limit in RChT
-----------------------------
The low-energy expansion of Eq. (\[eq:LO\]) determines, at LO in $1/N_C$, the chiral LECs appearing in Eq. (\[eq:Pi\_chpt\]): $$\begin{aligned}
L_{10} &=& - \frac{F_V^2}{4 M_V^2} + \frac{F_A^2}{4 M_A^2}
\approx -5.3 \cdot 10^{-3} ,\nonumber \\
C_{87}&=& \frac{F^2 F_V^2}{8 M_V^4} - \frac{F^2 F_A^2}{8 M_A^4}
\approx 4.3 \cdot 10^{-5} \, ,
\label{eq:L10-LO}\end{aligned}$$ where we have considered the SRA and have used the relations and inputs of Ref. [@L10].
Next-to-leading order corrections
---------------------------------
The low-energy expansion of Eq. (\[eq:Pi\_structure\]) determines now the couplings $L_{10}^r$ and $C_{87}^r$ at NLO in the $1/N_C$ expansion, with a control of the renormalization scale dependence. As it has been pointed out before, we have done the calculation under the SRA and including additional $V'$ and $A'$ [@L10]: $$\begin{aligned}
L_{10}^r(\mu_0)|^{\mathrm{SRA}} &=& (-5.2 \pm 0.4 ) \, \cdot \, 10^{-3}\, , \nonumber \\
L_{10}^r(\mu_0)|^{\mathrm{V'A'}}&=& \left(-3.6 \pm 0.9 \right)\cdot 10^{-3} \, , \nonumber \\
C_{87}^r(\mu_0)|^{\mathrm{SRA}}&=& (3.9\pm 0.6 )\, \cdot\, 10^{-5} \,, \nonumber \\
C_{87}^r(\mu_0)|^{\mathrm{V'A'}}&=& \left(2.2 \pm 1.1\right)\cdot 10^{-5}\,, \label{numbers}\end{aligned}$$ being $\mu_0=770$ MeV.
Conclusions
===========
Resonance Chiral Theory is an effective approach that allows to handle QCD at intermediate energies and is ruled by the $1/N_C$ expansion and constrained by the high-energy information. We have deeply investigated the constraints coming from well-behaved form factors in the spirit of correlators at NLO [@L10].
$10^3 \cdot L_{10}^r (\mu_0)$ $ 10^5 \cdot C_{87}^r (\mu_0)$
-------------------------- ------------------------------- ---------------------------------
This work $-4.4 \pm 0.9$ $ 3.1 \pm 1.1 $
Ref. [@bijnens-talavera] $-5.5 \pm 0.7$
Ref. [@davier-girlanda] $-5.13 \pm 0.19$
Ref. [@martin] $ -4.10 \pm 0.29 $ $3.85 \pm 0.13$
Ref. [@peris] $4.5 \pm 0.4$
: Comparison with different estimations.[]{data-label="comparison"}
One of the main motivations to deal with RChT at one-loop level is the estimation of the LECs, since tree-level predictions are unable to pin down the scale dependence of the ChPT couplings, which often are sizable. Here we have followed a general procedure described in Refs. [@NLO; @L10] to estimate the ${{\cal O}}(p^4)$ and ${{\cal O}}(p^6)$ couplings appearing in $\Pi(t)=\Pi_V(t)-\Pi_A(t)$, in terms of only the pion decay constant $F$ and resonance masses.
Combining the results of Eq. (\[numbers\]), obtained under the Single Resonance Approximation and including extra multiplets, one gets finally $$\begin{aligned}
L_{10}^r(\mu_0)&=& \left(-4.4 \pm 0.9\right)\cdot 10^{-3} ,
\nonumber \\
C_{87}^r(\mu_0)&=& \left(3.1 \pm 1.1\right)\cdot 10^{-5} .\end{aligned}$$ The general agreement with previous estimations is shown in Table 1. Note that our estimation is the only theoretical prediction at NLO.
Other chiral LECs can be estimated at next-to-leading order in $1/N_C$ by following the method explained here and applied to other observables. Work in this direction is in preparation.
[**Acknowledgments**]{}\
I wish to thank S. Narison for the organization of the conference and A. Pich and J.J. Sanz-Cillero for their helpful comments. This work has been supported in part by the Generalitat under grant GVPRE/2008/413, by the Spanish Government, under grants FPA2007-60323 and CSD2007-00042 (CPAN), and by the EU MRTN-CT-2006-035482 (FLAVIAnet).
[99]{}
A. Pich, arXiv:hep-ph/9806303. S. Weinberg, Physica A [**96**]{} (1979) 327; J. Gasser and H. Leutwyler, Annals Phys. [**158**]{} (1984) 142; Nucl. Phys. B [**250**]{} (1985) 465. G. Ecker et al., Nucl. Phys. B [**321**]{} (1989) 311; G. Ecker et al., Phys. Lett. B [**223**]{} (1989) 425; V. Cirigliano et al., Nucl. Phys. B [**753**]{} (2006) 139 . G. ’t Hooft, Nucl. Phys. B [**72**]{} (1974) 461; E. Witten, Nucl. Phys. B [**160**]{} (1979) 57. O. Cata and S. Peris, Phys. Rev. D [**65**]{} (2002) 056014; I. Rosell, J. J. Sanz-Cillero and A. Pich, JHEP [**0408**]{}, 042 (2004) ; JHEP [**0701**]{}, 039 (2007) ; I. Rosell, P. Ruiz-Femenía and J. Portolés, JHEP [**0512**]{} (2005) 020 ; Phys. Rev. D [**75**]{} (2007) 114011 ; I. Rosell, arXiv:hep-ph/0701248. A. Pich, I. Rosell and J. J. Sanz-Cillero, JHEP [**0807**]{} (2008) 014 . G. P. Lepage and S. J. Brodsky, Phys. Lett. B [**87**]{} (1979) 359; Phys. Rev. D [**22**]{} (1980) 2157; Phys. Rev. D [**24**]{} (1981) 1808. V. Cirigliano et al., Phys. Lett. B [**596**]{} (2004) 96 .
J. Bijnens and P. Talavera, Nucl. Phys. B [**489**]{} (1997) 387 .
M. Davier, L. Girlanda, A. Höcker and J. Stern, Phys. Rev. D [**58**]{} (1998) 096014 .
M. Gonzalez-Alonso, A. Pich and J. Prades, arXiv:0810.0760 \[hep-ph\]; arXiv:0810.2459 \[hep-ph\]. P. Masjuan and S. Peris, Phys. Lett. B [**663**]{} (2008) 61 .
[^1]: Talk given at the 14th International Quantum Chromodynamics Conference, 7–12th July (2008), Montpellier (France). IFIC/08-54 report.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This article gives an introduction for mathematicians interested in numerical computations in algebraic geometry and number theory to some recent progress in algorithmic number theory, emphasising the key role of approximate computations with modular curves and their Jacobians. These approximations are done in polynomial time in the dimension and the required number of significant digits. We explain the main ideas of how the approximations are done, illustrating them with examples, and we sketch some applications in number theory.'
author:
- 'Jean-Marc Couveignes[^1] and Bas Edixhoven'
title: Approximate computations with modular curves
---
Primary 65-D-99; Secondary 11-Y-40, 14-Q, 11-F-80, 11-G-18, 14-G-35, 14-G-40.
Drinfeld modules, $L$-functions, Weil conjecture.
Introduction {#sec:intro}
============
The purpose of this article is to give an introduction to the main results of the book [@book] and their generalization in the PhD thesis [@Bruin1] and in [@Bruin2], as well as some applications, and most of all to explain the essential role played by *approximate computations*. The intended reader is a mathematician interested in *numerical computations* in algebraic geometry or number theory.
The results concern fast algorithms in number theory and more precisely, fast computation of Fourier coefficients of modular forms. These coefficients, with Ramanujan’s $\tau$-function as a typical example, have deep arithmetic significance and are important in various areas of mathematics, from number theory and algebraic geometry to combinatorics and lattices.
The fastest previously known algorithms for computing these Fourier coefficients took exponential time, except in some special cases. The case of elliptic curves (Schoof’s algorithm) was at the birth of elliptic curve cryptography around 1985. The results mentioned above give an algorithm for computing coefficients of modular forms in polynomial time. For example, Ramanujan’s $\tau(p)$ with $p$ a prime number can be computed in time bounded by a fixed power of $\log p$.
Such fast computation of Fourier coefficients is itself based on the main result of the book: the computation, in polynomial time, of Galois representations over finite fields attached to modular forms by the Langlands program.
The computation of the Galois representations uses their realisation, following Shimura and Deligne, in the torsion subgroup of Jacobian varieties of modular curves. The main challenge is then to perform the necessary computations in time polynomial in the dimension of these nonlinear algebraic varieties. Exact computations involving systems of polynomial equations in many variables take exponential time. This is avoided by numerical approximations with a precision that suffices to derive exact results from them. Bounds for the required precision – in other words, bounds for the height of the rational numbers that describe the Galois representation to be computed – are obtained from Arakelov theory.
This article is organised as follows. Sections \[sec:afcaj\] and \[sec:psi\] are concerned with numerical methods used in the context of complex algebraic curves and their Jacobian varieties. Sections \[sec:ctd1\] and \[sec:ctd2\] describe how to get exact results about torsion points on modular curves using these numerical methods. Section \[sec:ctd1\] focuses on the genus $1$ curve $X_{11}$ while Section \[sec:ctd2\] deals with the general modular curve $X_\ell$. As an application, Section \[sec:avoq\] gives two examples of fast computation of coefficients of modular forms: Ramanujan’s $\tau$-function, and the classical sums of squares problem.
Algorithms for curves and Jacobians {#sec:afcaj}
===================================
Let $X$ be a connected, smooth, projective algebraic curve over the field $\CC$ of complex numbers. The set $X(\CC)$ of complex points of $X$ is a Riemann surface. Let $g$ be the genus of $X$ and let $(\omega_k)_{1\le k\le g}$ be a basis for the space of holomorphic differentials on $X$. We fix a point $b\in X(\CC)$ and we denote by $Y_b$ the set of homotopy classes of paths on $X(\CC)$ starting at $b$. The [*universal cover*]{} $f_b : Y_b \rightarrow X(\CC)$ maps every path to its end point. The [*fundamental group*]{} $\pi_1(X(\CC),b)\subset Y_b$ is the subset of (homotopy classes of) closed paths. It acts on $Y_b$, with quotient $X(\CC)$. We have an integration map $\phi_b : Y_b\rightarrow \CC^g$ defined by $$\phi_b(\gamma)=(\int_\gamma\omega_1, \ldots, \int_\gamma
\omega_g).$$ The image of $\pi_1(X(\CC),b)$ by $\phi_b$ is a lattice $\Lambda$ in $\CC^g$. It is called the [*lattice of periods*]{}. It is a free $\ZZ$-module of rank $2g$. The quotient $\CC^g/\Lambda$ is a complex torus. It is the set of complex points $J(\CC)$ on the [*Jacobian variety*]{} $J$ of $X$. The integration map $\phi_b :
Y_b\rightarrow \CC^g$ induces a map between the quotients $X(\CC)\rightarrow J(\CC)$. This map is a morphism of varieties $X\rightarrow J$. We call this morphism $\phi_b$ also. For every positive integer $k$ we denote $\phi_b^k : X^k\rightarrow J$ the morphism that maps $(P_1, \ldots, P_k)$ onto $\phi_b(P_1)+\dots+\phi_b(P_k)$. Since the image in $J$ does not depend on the ordering on the points $P_j$, we write $X^{(k)}$ for the $k$-th symmetric power of $X$. We note that $X^{(k)}$ is the quotient of $X^k$ by the action of the symmetric group. It is a nonsingular variety. We define the morphism $\phi_b^{(k)} : X^{(k)}\rightarrow J$ that maps $\{P_1, \ldots, P_k\}$ onto $\phi_b(P_1)+\dots+\phi_b(P_k)$. For $k=g$ the map $\phi_b^{(g)}$ is birational and surjective. It is not an isomorphism unless $g\leq 1$. Its fibers are projective linear spaces, mostly (but not all) points. A degree $g$ effective divisor $P= P_1+\dots+P_g$ is said to be [ *non-special*]{} if the map $\phi^{(g)}_b$ is a local diffeomorphism at $P$. Otherwise we say that $P$ is [*special*]{}. This definition does not depend on the chosen origin $b$. The set of special effective degree $g$ divisors is the singular locus of $\phi_b^{(g)}$. All these maps $\phi^{(k)}_b$ are called Abel-Jacobi maps. In particular $$\phi_b^{(g)}(\{P_1, \ldots, P_g\})=
\sum_{1\le j\le g}(\int_b^{P_j}\omega_k)_{k} \, \bmod \Lambda,$$ where we can integrate $\int_b^{P_j}\omega_k$ along any path between $b$ and $P_j$, provided we keep the same path for all $k$. We can apply the Abel-Jacobi map to any divisor on $X$. We set $\phi_b(\sum_j e_jP_j)=\sum_je_j\phi_b^{(1)}(P_j)$. We note that for degree zero divisors, the image does not depend on the origin $b$. A divisor is said to be [*principal*]{} if it is the divisor of a non-zero meromorphic function on $X$. Two divisors are said to be [*linearly equivalent*]{} when their difference is principal. Any principal divisor has degree zero. A degree zero divisor is principal if and only its image by $\phi_b$ is zero. So the set $J(\CC)=\CC^g/\Lambda$ of complex points on the Jacobian is canonically identified with the group $\Pic^0(X)$ of linear equivalence classes of degree zero divisors on $X$.
We now list important algorithmic problems related to the Abel-Jacobi map. We illustrate them on the simple example of the projective curve $X$ with equation
\[eq:X11\] Y\^2Z-YZ\^2=X\^3-X\^2Z.
This curve has genus $1$. We write $x=X/Z$ and $y=Y/Z$. The unique (up to a multiplicative constant) holomorphic differential on $X$ is $$\omega= \frac{dx}{2y-1}=\frac{dy}{x(3x-2)}.$$ We choose the point $b=[0:1:0]$ as origin for the integration map. For every computational problem we shall consider, we will also explain what can be proven when $X$ is a modular curve $X_\ell$ and $\ell$ (therefore $g$) tends to infinity. The definition of the modular curve $X_\ell$ is given in Section \[sec:ctd2\]. See also textbooks [@Diamond-Shurman; @Stein] where $X_\ell$ is often denoted $X_1(\ell)$.
Computing the lattice of periods {#sec:lattice}
--------------------------------
We first need a basis for the singular homology group $H_1(X(\CC),\ZZ)$. If $X$ is the genus one curve given by equation (\[eq:X11\]), such a basis can be deduced from the study of the degree two map $x : X\rightarrow \PU$ that sends $(x,y)$ onto $x$ and $[0:1:0]$ to $\infty$. This map is ramified at $\infty$ and the three roots of $4x^3-4x^2+1$. We lift a simple loop around $\infty$ and one of these three roots. We then lift a simple loop around $\infty$ and another root. We thus obtain two elements in $H_1(X(\CC),\ZZ)$ that form a basis for it.
Integrating a differential along a path is easy. We express the differential in terms of local coordinates. We then reduce to integrating converging power series. We integrate term by term. In case $X$ is the curve given in equation (\[eq:X11\]), we obtain a basis $(\Omega_1, \Omega_2)$ for the lattice $\Lambda$ of periods where $$\begin{aligned}
\Omega_1&=& 6.346046521397767108443973084,\\ \Omega_2&=&-3.173023260698883554221986542 + 1.458816616938495229330889613i.\end{aligned}$$
These calculation are made e.g. using the [@pari] system.
>a1=0;a2=-1;a3=-1;a4=0;a6=0;
>X=[a1,a2,a3,a4,a6];X=ellinit(X);
>X.omega
[6.346046521397767108443973084,
-3.173023260698883554221986542 + 1.458816616938495229330889613*I]
When dealing with general modular curves, an explicit basis for both the singular homology and the de Rham cohomology is provided by the theory of Manin symbols [@manin; @merel; @cremona; @frey; @Stein]. Computing (good approximations of) periods is then achieved in time polynomial in the genus and the required accuracy [@Couveignes2]. The practical side is described in [@Bosman1 §6.3]. Textbooks [@Cohen], [@cremona Chapter 3] give even faster techniques for genus $1$ curves, but we shall not need them.
Computing with divisor classes {#sec:divcla}
------------------------------
A degree zero divisor class can be represented by a point in the torus $\CC^g/\Lambda=J(\CC)$. It can also be represented by a divisor of the form
\[eq:div\] P\_1+…+P\_g-gb
in this class. This latter representation is not always unique. It is however unique for most classes because $\phi_b^{(g)}$ is birational. The addition problem in this context is the following: given two degree $g$ effective divisors $P=P_1+\dots+P_g$ and $Q=Q_1+\dots+Q_g$, one would like to compute a degree $g$ effective divisor $R=R_1+\dots+R_g$ such that the divisor class of $R-gb$ is the sum of the divisor classes of $P-gb$ and $Q-gb$. So we look for $g$ complex points $R_1$, …, $R_g$ such that $P_1+\dots+P_g+Q_1+\dots+Q_g-2gb$ is linearly equivalent to $R_1+\dots+R_g-gb$. This is achieved using the Brill-Noether algorithm [@BN; @Vol]. This algorithm uses a complete linear space $\cL$ of forms or functions. This space should have dimension $\ge 2g+1$. For example, assuming $g\ge 4$, we may take for $\cL$ the space of all holomorphic quadratic differential forms. We compute once for all a basis for this space. Then the Brill-Noether algorithm alternates several steps of two different natures. Sometimes we are given a form (function) and we want to compute its divisor. Sometimes we are given an effective divisor $D$ and we want to compute a basis for the subspace $\cL(-D)$ consisting of forms (functions) vanishing at this divisor.
The first problem (finding zeros of a given form) can be reduced, using a convenient coordinate system, to the following problem: given a power series $f(z)=\sum_{k\ge 0} f_kz^k$ with radius of convergence $\ge 1$, find approximations of its zeros in the disk $D(0,1/2)$ with center $0$ and radius $1/2$. It is clear (see [@Couveignes1 §5.4]) that, for the purpose of finding zeros, one can replace $f(z)$ by its truncation $\sum_{0\le k\le K}f_kz^k$ at a not too large order $K$. We then reduce to the classical problem of computing zeros of polynomials. A survey of this problem is given in [@Couveignes1 §5.3].
The second problem (finding the subspace of functions vanishing at given points) boils down to finding the kernel of the matrix having entries the values of the functions in the chosen basis of $\cL$ at the given points.
The only difficulty then is to control the conditioning of these two problems. This is done in two steps. We first prove [@Couveignes1 §5.4] that the zeros of a holomorphic function on a closed disk are well conditioned unless this function is small everywhere on this disk. We then prove [@Couveignes2 §12.7] that the form we consider cannot be small everywhere on any of the charts we consider, unless it has very small coordinates in the chosen basis of $\cL$.
The resulting algorithm for computing in the group of divisor classes of modular curves is polynomial time in the genus and the required direct accuracy [@Couveignes2 Theorem 12.9.1]. By [*direct accuracy*]{} we mean that the error is measured in the target space of the integration map, namely the torus $\CC^g/\Lambda$. Saying that the direct accuracy is bounded from above by $\epsilon$ means that the returned divisor $R'=R_1'+\dots+R_g'$ is such that $$\phi_b (R'-R) = \phi^{(g)}_b(R')-\phi^{(g)}_b(R)$$ is bounded from above by $\epsilon$ for the maxnorm in $\CC^g$. This does not necessarily imply that the $R_j$ are close to the $R_j'$. Indeed, in case $R=R_1+\dots+R_g$ is special, there exists a non-trivial linear pencil of divisors $R'$ such that $\phi_b^{(g)}(R')=\phi_b^{(g)}(R)$. Controlling the distance between $R$ and $R'$ will only be possible in some cases.
In the special case when $X$ is the curve given by equation (\[eq:X11\]) the map $\phi^{(1)}_b : X\rightarrow J$ is an isomorphism because the genus is $1$. Computing with divisor classes is then very simple and the Brill-Noether algorithm takes a simple form. The space $\cL$ consists of all degree $1$ homogeneous forms, and a basis for it is made of the three projective coordinates $X$, $Y$ and $Z$. Given $P$ and $Q$, one considers the unique projective line $\Delta_1$ through $P$ and $Q$. In case $P=Q$ we take $\Delta_1$ to be the tangent to $X$ at $P$. The line $\Delta_1$ meets $X$ at three points: $P$, $Q$ and a third point that we call $S$. We consider the unique projective line $\Delta_2$ through $S$ and the origin $b$. The line $\Delta_2$ meets $X$ at three points: $b$, $S$ and a third point that we call $R$. On can easily check that $P+Q$ is linearly equivalent to $b+R$ or equivalently $P-b+Q-b$ is equivalent to $R-b$. The coordinates of $R$ can be computed using very simple formulae [@silverman Chapter III]. We illustrate this using the [@pari] system. We call $P$ the point $[0:0:1]$. We first compute $Q$ such that $Q-b$ is linearly equivalent to $2(P-b)$. We write $Q-b\equiv 2(P-b)$ using the $\equiv$ symbol for linear equivalence. We then compute $R$ such that $R-b\equiv P-b+Q-b
\equiv 3(P-b)$. We then compute $S$ such that $S-b\equiv Q-b+R-b\equiv
5(P-b)$.
>P=[0,0];
>Q=elladd(X,P,P)
[1, 1]
>R=elladd(X,P,Q)
[1, 0]
>S=elladd(X,Q,R)
[0]
The answer for $S$ means that $S$ is just the origin $b=[0:1:0]$. So the divisor $P-b$ has order $5$ in the Picard group $\Pic(X)$, the group of divisors modulo linear equivalence.
The direct Jacobi problem {#sec:jaco}
-------------------------
Given a divisor on $X$ we want to compute its image by $\phi_b$ in the complex torus $J(\CC)=\CC^g/\Lambda$. It suffices to explain what to do when the divisor consists of a single point $P$. For every $1\le k\le g$ we then have to compute $\int_b^P\omega_k$. So we integrate $\omega_k$ along any path from $b$ to $P$. We split the chosen path in several pieces according to the various charts in our atlas for the Riemann surface $X(\CC)$. On every chart, the differentials $\omega_k$ can be expressed in terms of the local coordinate. We then reduce to computing integrals of the form $\int_0^{\frac{1}{2}}f(z)dz$ where $f(z)$ is holomorphic on the unit disk. Such an integral can be computed term by term. When $X$ is a modular curve, we have a convenient system of charts and a basis for $\cL$ consisting of forms having small coefficients in their expansions at every chart. There is long standing tradition with stating and proving bounds for these coefficients. It culminates with the so-called Ramanujan conjecture. This conjecture was proved by Deligne as a consequence of [@Deligne1] and his proof of the analog of the Riemann hypothesis in the Weil conjectures in [@Deligne2]. In case $X$ is the elliptic curve given by equation (\[eq:X11\]) we take for $P$ the point $[0:0:1]$ and find that $$\phi_b^{(1)}(P)=\int_b^P\omega = 2.538418608559106843377589234\bmod \Lambda.$$ This integral is computed using the [@pari] system.
> ellpointtoz(X,[0,0])
2.538418608559106843377589234
We notice that $$\phi_b^{(1)}(P)=\frac{2\Omega_1}{5}\bmod \Lambda.$$ So $5(P-b)$ is a principal divisor as already observed at the end of section \[sec:jaco\].
The inverse Jacobi problem {#sec:jacinv}
--------------------------
At this point we have two different ways of representing a degree zero class of equivalence of divisors. We can be given a divisor in this class like the one in equation (\[eq:div\]). Such a divisor will be called a [*reduced divisor*]{}. We can also be given a vector in $\CC^g$ modulo the lattice of periods $\Lambda$. It is of course very easy to compute with such vectors. We also have seen in section \[sec:divcla\] how to compute with reduced divisors. So both representations are convenient for computational purposes. We also have seen in section \[sec:jaco\] how to pass from a reduced divisor to the corresponding point in the torus $\CC^g/\Lambda$ applying the Abel-Jacobi map. We now consider the inverse problem: given a point $\alpha\bmod \Lambda$ in the torus $\CC^g/\Lambda$, find some $P=P_1+\cdots+P_g$ such that the reduced divisor $P-gb$ is mapped onto $\alpha \bmod \Lambda$ by $\phi_b$.
#### Using an iterative method
We can try an iterative method like the secant’s method. We illustrate the secant’s method in case $X$ is the curve given by equation (\[eq:X11\]) and
\[eq:alpha\] =(\_1+\_2)/11=0.2884566600635348685656 + 0.1326196924489541117573i.
Starting from $P_0=(50-50i,-223.147+547.739i)$ and $P_1=(20-20i,-54.587+137.965i)$ we obtain an approximation up to $10^{-26}$ after eighteen iterations. We use the [@pari] system and declare a function for the secant method.
>secant(alpha,P0,P1,K)=
{
local(f0,f1,x0,x1,x2,P2,P3);
for(k=1,K,
f0=ellpointtoz(X,P0)-alpha;f1=ellpointtoz(X,P1)-alpha;
x0=P0[1];x1=P1[1];
x2=x1-f1*(x1-x0)/(f1-f0);
P2=[x2,ellordinate(X,x2)[1]];P3=[x2,ellordinate(X,x2)[2]];
if(abs(P2[2]-P0[2])> abs(P3[2]-P0[2]) ,P2=P3,);
P0=P1;P1=P2;
);
return(P2);
}
The four parameters of this function are the target point in $\CC/\Lambda$, the two initial approximate values of $P$, and the number of iterations. We then type
>alpha=(omega1+omega2)/11;
x0=50-50*I;x1=20-20*I;
P0=[x0,ellordinate(X,x0)[2]];P1=[x1,ellordinate(X,x1)[2]];
secant(alpha,P0,P1,18)
Below are the results of iterations $14$ to $18$. We only give the values taken by the $x$-coordinate.
6.796891402429021881380876803 - 7.525836023544396684018482041i
6.796539495414535904114103146 - 7.525907619429540863361002543i
6.796539142100022043003057330 - 7.525908029913269174706910680i
6.796539142094915910541452272 - 7.525908029899464322147329306i
6.796539142094915911068237206 - 7.525908029899464321854796862i
#### The continuation method
Iterative methods only work if the starting approximation is close enough to the actual solution. Such an initial approximation can be provided by the solution of a different though close inverse problem. Coming back to our example, we will start from any point on $X$. Say $P_0=(0,0)$. We compute the image $\alpha_0\bmod \Lambda$ of $P_0$ by the integration map. We then choose any $P_{-1}$ that is close enough to $P_0$.
>P0=[0,0];
alpha0=ellpointtoz(X,P0);
Pm1=[0.1,ellordinate(X,0.1)[2]];
We now move slowly from $\alpha_0$ to $\alpha$. We set $\alpha_1=\alpha_0+0.1(\alpha-\alpha_0)$ and we solve the inverse problem for $\alpha_1$ using the secant’s method with initial values $P_{-1}$ and $P_0$.
>P1=secant(alpha0+0.1*(alpha-alpha0),Pm1, P0,5)
[0.218773824415936734050679268 - 0.0122309960881052801981765895*I,
0.0388323642082357612959944279 - 0.00390018046133107189481433241*I]
We now set $\alpha_2=\alpha_0+0.2(\alpha-\alpha_0)$ and we solve the inverse problem for $\alpha_2$ using the secant’s method with initial values $P_0$ and $P_1$.
>P2=secant(alpha0+0.2*(alpha-alpha0),P0, P1,5)
[0.410237833586311839505201998 - 0.0205989424813431290064696558*I,
0.111775424533436210193603161 - 0.00838376796781394064004855129*I]
We continue until we reach $\alpha$
>P3=secant(alpha0+0.3*(alpha-alpha0),P1, P2,5);
...
P9=secant(alpha0+0.9*(alpha-alpha0),P7, P8,5);
P10=secant(alpha,P8, P9,10)
[6.796539142094915911068237205 - 7.525908029899464321854796861*I,
-8.056577776742775028742861296 + 30.05694612451787404370259256*I]
This continuation method is very likely to succeed provided the integration map has a nice local behaviour all along the path from $\alpha_0$ to $\alpha$. This is how practical computations have been realised in [@Bosman1] for modular curves. It is however difficult to prove that this method works because the integration map $\phi^{(g)}_b$ has a singular locus as soon as $g>1$, and we do not know how to provably and efficiently find a path from $\alpha_0\bmod \Lambda$ to $\alpha\bmod \Lambda$ that keeps away from the singular locus.
Provably solving the inverse Jacobi problem {#sec:psi}
===========================================
We have presented in section \[sec:jacinv\] a heuristic algorithm for the inverse Jacobi problem. This algorithm is based on continuation. It seems difficult to prove it however because that would require a good control on the singular locus of the Jacobi map. In this section we present the algorithm introduced in [@Couveignes2]. This algorithm only requires a good control of the Jacobi map locally at a chosen divisor in $X^{(g)}$. This is a much weaker condition and it is satisfied for modular curves. An important feature of this algorithm is the use of fast exponentiation rather than continuation. The principle of fast exponentiation is recalled in section \[sec:fe\]. The algorithm for the inverse Jacobi problem itself is given in section \[sec:jil\]. Section \[sec:mp\] sketches the proof of this algorithm. Proving in this context means proving the existence of a Turing machine that returns a correct answer in a given time. One has to prove both the correctness of the result and a bound for the running time. This bound here will be polynomial in the genus of the curve and the required accuracy of the result.
Fast exponentiation in groups {#sec:fe}
-----------------------------
Assume we are given a group $G$. The group law in $G$ will be denoted multiplicatively. We assume that $G$ is [*computational*]{}. This means that we know how to represent elements in $G$, how to compare two given elements, how to invert a given element, and how to multiply two given elements.
The exponentiation problem in $G$ is the following: we are given an element $g$ in $G$ and an integer $e\ge 2$, and we want to compute $g^e$ as an element in $G$. A first possibility would be to set $a_1=g$ and to compute $a_k=a_{k-1}\times g$ for $2\le k\le e$. This requires $e-1$ multiplications in $G$. It is well known, however, that we can do much better. We write the expansion of $e$ in base $2$, $$e=\sum_{0\le k\le K}\epsilon_k2^k,$$ and we set $b_0=g$ and $b_{k}=b_{k-1}^2$ for $1\le k\le K$. We then notice that $$g^e= \prod_{0\le k\le K}b_k^{\epsilon_k}.$$ So we can compute $g^e$ at the expense of a constant times $\log e$ operations in $G$. The algorithm above is called [*fast exponentiation*]{} and it admits many variants and improvements [@gor]. Its first known occurrence dates back to Pingala’s Chandah-s[û]{}tra (before -200). See [@hist I,13].
Solving the Jacobi inverse problem by linear algebra {#sec:jil}
----------------------------------------------------
Recall that we have two different ways of representing an equivalence class of divisors of degree zero: reduced divisors or classes in the torus $\CC^g/\Lambda$. We have seen that both models are computational. The Abel-Jacobi map $\phi^{(g)}_b : X^{(g)}\rightarrow \CC^g/\Lambda$ is computational also. We want to invert it (although we know it is not quite injective). More precisely we assume we are given some $\alpha$ in $\CC^g$ and we look for a degree $g$ effective divisor on $X$ such that $\phi^{(g)}_b(P)=\phi_b (P-gb)=\alpha\bmod \Lambda$. It seems difficult to prove the heuristic methods given in section \[sec:jacinv\] for this purpose. So we present here a variant for which we can give a proof, at least when $X$ is a modular curve $X_\ell$. We illustrate this method in the case where $X$ is the curve given in equation (\[eq:X11\]). We still aim at the $\alpha$ given in equation (\[eq:alpha\]).
We need a non-special effective divisor $P_0$ of degree $g$. Since $g=1$ we can take any point on $X$. For example $P_0=(0,0)$. We note that the affine coordinate $x$ is a local parameter at $P_0$. We choose a small real number $\epsilon$. The smaller $\epsilon$ the better the precision of the final result. Here we choose $\epsilon = 0.0001$. We consider two points $P_1$ and $P_2$ that are very close to $P_0$. The first point $P_1$ is obtained by adding $\epsilon$ to the $x$-coordinate of $P_0$. The second point $P_2$ is obtained by adding $\epsilon i$ to the $x$-coordinate of $P_0$.
P0=[0,0];
P1=[0.0001,ellordinate(X,0.0001)[2]];
P2=[0.0001*I,ellordinate(X,0.0001*I)[2]];
We now compute the image $\alpha_1\bmod \Lambda$ of $P_1-P_0$ by the Abel-Jacobi map. We also compute the image $\alpha_2\bmod \Lambda$ of $P_2-P_0$. We note that $\alpha_1\bmod \Lambda$ is very close to $0\in \CC/\Lambda$. This is because $P_0$ and $P_1$ are close. We assume that $\alpha_1$ is the smallest complex number in its class modulo $\Lambda$. We make the same assumption for $\alpha_2$. Then $\alpha_1$ and $\alpha_2$ are two small complex numbers, and they form an $\RR$-basis of $\CC$. This is because the integration map $\phi^{(g)}_b$ is a local diffeomorphism at $P_0$ (or equivalently $P_0$ is a non-special divisor) and $\epsilon$ has been chosen small enough.
alpha1=ellpointtoz(X,P1)-ellpointtoz(X,P0);
alpha2=ellpointtoz(X,P2)-ellpointtoz(X,P0)-omega1-omega2;
Recall that our target in the torus $\CC/\Lambda$ is $\alpha \bmod \Lambda$ where $\alpha$ is the complex number given in equation (\[eq:alpha\]). So we compute the two real coordinates of $\alpha$ in the basis $(\alpha_1,\alpha_2)$.
>M=[real(alpha1), real(alpha2); imag(alpha1), imag(alpha2)];
coord=M^(-1)*[real(alpha),imag(alpha)]~
[-2884.566581407009845250155464, -1326.196933330853847302268151]~
We deduce that $\alpha$ is very close to $\alpha' = -2884\alpha_1-1326\alpha_2$. And the class $\alpha'\bmod \Lambda$ is the image by $\phi_b$ of $-2884(P_1-P_0)-1326(P_2-P_0)$. The linear equivalence class of the latter divisor is therefore a good approximation for our problem. There remains to compute a [*reduced divisor*]{} $P-gb$ in this class using the methods presented in section \[sec:divcla\]. Since the integers $2884$ and $1326$ are rather big, we use the fast exponentiation algorithm presented in section \[sec:fe\].
>coord=truncate(coord)
[-2884, -1326]~
>D1=ellsub(X,P1,P0);D2=ellsub(X,P2,P0);
P=elladd(X,ellpow(X,D1,coord[1]),ellpow(X,D2,coord[2]))
[6.798693122986621316758396123 - 7.528977879167267357619566769*I,
-8.059779911380488392224788509 + 30.07437308400090422713306570*I]
We now check that the image of $P-P_0$ by $\phi_b$ is close to $\alpha$
>ellpointtoz(P)
0.2884000018811813146007079855 + 0.1325999988977252987328424662*I
>alpha
0.2884566600635348685656351402 + 0.1326196924489541117573536012*I
For a better approximation we should start with a smaller $\epsilon$.
Matter of proof {#sec:mp}
---------------
The main concern when proving the algorithm in section \[sec:jil\] is to prove that we can find an initial divisor $P_0$ that is non-special. In fact we must guarantee a quantified version of this non-speciality condition. The differential of $\phi^{(g)}_b$ at $P_0$ should be non singular and its norm should not be too small. We can prove that such a condition holds true for modular curves [@Couveignes2 §12.6.7] because we have a very sharp description of these curves in the neighbourhood of the points called [*cusps*]{}. As a consequence we prove [@Couveignes2 Theorem 12.10.5] that the inverse Jacobi problem for modular curves can be solved in deterministic polynomial time in the genus and the required [*direct accuracy*]{}. Recall that [*direct accuracy*]{} means that the error is measured in the target space $\CC^g/\Lambda$. The main difference between the algorithm in this section and the one in section \[sec:jacinv\] is that we only need here to control the local behaviour of $\phi^{(g)}_b$ at $P_0$ while the algorithm in section \[sec:jacinv\] requires that the map $\phi^{(g)}_b$ be non-singular above the whole path from $\alpha_0$ to $\alpha$.
In some cases it will be desirable to control the [*inverse error*]{} that is the error on the output divisor $P$ in $X^{(g)}$. This will be possible when we can prove that $\phi^{(g)}: X^{(g)}\rightarrow
J$ is a local diffeomorphism at $P$ (that is $P$ is non-special). We will also need a lower bound for the norm of the differential of $\phi^{(g)}$ at $P$. Such a lower bound can be provided by arithmetic.
Computing torsion points I {#sec:ctd1}
==========================
In this and the next section we will assume that $X$ is a modular curve and $\ell$ a prime number. We will be interested in $\ell$-torsion points in the torus $J(\CC)=\CC^g/\Lambda$. A point $$\balpha = \alpha\bmod \Lambda$$ is an $\ell$-torsion point if and only if $\alpha$ lies in $\frac{1}{\ell}\Lambda$. So the $\ell$-torsion subgroup of $J(\CC)$ is $\frac{1}{\ell}\Lambda / \Lambda$ and it has cardinality $\ell^{2g}$. This group is also denoted $J[\ell]$.
Some of these torsion points carry important arithmetic information. The values taken by algebraic functions at these points generate interesting number fields. We want to compute these fields. In this section we will focus on a special case. We will assume that $X$ is the genus $1$ curve given in equation (\[eq:X11\]) and $\ell=11$. A more general situation will be studied in the next section \[sec:ctd2\]. We notice that the curve in equation (\[eq:X11\]) is indeed the modular curve known as $X_{11}$. Since $X$ has genus $1$, the map $\phi_b : X\rightarrow J$ is an isomorphism mapping $b=[0:1:0]$ onto the origin. So the affine coordinate $x$ and $y$ induce algebraic functions $x\circ \phi_b^{-1}$ and $y\circ \phi_b^{-1}$ on $J$. There are $11^2=121$ points of $11$ torsion in $J$ and $0$ is one of them. We will be interested in the values taken by $x\circ \phi^{-1}_b$ at the remaining $120$ points of $11$-torsion. On can check that $x\circ \phi^{-1}_b$ takes the same value at two opposite points. So there only remain $60$ values of interest. These are algebraic numbers and they form a single orbit under the action of the Galois group $\Gal(\bQ/\QQ)$. So it is natural to consider their annihilating polynomial
\[eq:divpol\] H(T)=\_[0= aJ\[11\]/[1]{}]{}(T-x(\^[-1]{}(a))).
This is an irreducible polynomial in $\QQ[T]$. Computing such polynomials is a cornerstone in the algorithmic of modular forms and Galois representations.
An algebraic approach {#sec:aaa}
---------------------
The polynomial in equation (\[eq:divpol\]) is known as the $11$-th division polynomial $\psi_{11}$ of the genus one curve $X$. For every $k\ge 1$ one can define the $k$-th division polynomial $\psi_k(T)$ to be the annihilating polynomial of the $x$-coordinates of all non-zero $k$-torsion points on $X$. These polynomials can be computed using recursion formulae [@enge Section 3.6] [@silverman Exercise 3.7] that follow from the simple algebraic form of the addition law on $X$. Using these recursion formulae we find $$H(T)=T^{60}-20T^{59}+112T^{58}+1855T^{57}+\cdots + 1321T^4 - 181T^3
+ 22T^2 - 2T + 1/11.$$ So we have an efficient algebraic method to compute $H(T)$. We will explain in section \[sec:ctd2\] why it seems difficult to us to generalize this algebraic method to curves of higher genus.
Using complex approximations {#sec:uca}
----------------------------
In this section we compute complex approximations of the coefficients of $H(T)$. We also explain how one can deduce the exact value of these coefficients from a sharp enough complex approximation. We have seen in sections \[sec:jacinv\] and \[sec:psi\] how to invert the map $\phi_b$. Given a point $a$ in the torus $\CC^g/\Lambda$ we can compute a complex approximation of some reduced divisor $P_a-gb$ such that $\phi_b (P_a-bg)=a$. Since here the genus is one, $P_a$ consists of a single point on $X$, and it is uniquely defined. In case $a=(\Omega_1+\Omega_2)/11$ we already found that the $x$-coordinate $x(P_a)$ of $P_a$ is $$6.796539142094915911068237206 - 7.525908029899464321854796862i$$ up to an error of $10^{-27}$. We let $a$ run over the $60$ elements in $(J[11]-\{0\})/\pm 1 $ and compute the $60$ corresponding values of $x(P_a)$ with the same accuracy. We then compute their sum and find it is equal to $20$ up to an error of $10^{-25}$. This suggests that the coefficient of $T^{59}$ in $H(T)$ is $-20$. In order to turn this heuristic into a proof, we need some information about the coefficients of $H(T)$. We know that these coefficients are rational numbers. We need an upper bound on their [*height*]{}. The [ *height*]{} of a rational number is the maximum of the absolute values of its numerator and denominator. We explain in the next section \[sec:recover\] how a good approximation and a good bound on the height suffice to characterise and compute a rational number. In case $X$ is the curve given in equation (\[eq:X11\]) an upper bound on the height of the coefficients of $H(T)$ can be proved by elementary means. For example we know that the denominator of these coefficients is either $1$ or $11$. In case $X$ is a modular curve, similar bounds will be necessary. These bounds have been proved by the second author in collaboration with de Jong in [@Ed-Jo1] and [@Ed-Jo2], using Arakelov theory and arithmetic geometry together with a result of Merkl in [@Merkl] on upper bounds for Green functions.
All the coefficients of $H(T)$ are computed in the same way. They are symmetric functions of the $x(P_a)$, so we can compute sharp approximations for them. We deduce their exact values using an a priori bound on their height.
Recovering a rational number from a good approximation {#sec:recover}
------------------------------------------------------
In the previous section \[sec:uca\] we claimed that a rational number $x=a/b$ can be recovered from a sharp enough complex approximation, provided we have an a priori bound on the height of $x$. We recall that the height of a rational number $a/b$, with $a$ and $b$ integers that are relatively prime, is $\max\{|a|,|b|\}$. The rational number $x=a/b$ is known if we know an upper bound $h$ for its height and an approximation $y$ of it (in $\RR$, say), with $|x-y|<1/(2h^2)$. Indeed, if $x'=a'/b'$ also has height at most $h$, and $x'\neq x$, then $$|x-x'| = \left|\frac{a}{b}-\frac{a'}{b'}\right| =
\left|\frac{ab'-ba'}{bb'}\right| \geq \frac{1}{|bb'|} \geq 1/h^2.$$ We also note that there are good algorithms to deduce $x$ from such a pair of an approximation $y$ and a bound $h$, for example, by using continued fractions, as we will now explain.
In practice we will use rational approximations $y$ of $x$. Every rational number $y$ can be written uniquely as $$[a_0,a_1,\ldots,a_n] = a_0+\cfrac{1}{a_1+
\cfrac{1}{\ddots\genfrac{}{}{0pt}{0}{}{a_{n-1}+\cfrac{1}{a_n}}}}\;,$$ where $n\in\ZZ_{\geq0}$, $a_0\in\ZZ$, $a_i\in\ZZ_{>0}$ for all $i>0$, and $a_n>1$ if $n>0$. To find these $a_i$, one defines $a_0:=\lfloor
y\rfloor$ and puts $n=0$ if $y=a_0$; otherwise, one puts $y_1:=1/(y-a_0)$ and $a_1=\lfloor y_1\rfloor$ and $n=1$ if $y_1=a_1$, and so on. The rational numbers $[a_0,a_1,\ldots,a_i]$ with $0\leq i\leq n$ are called the *convergents* of the continued fraction of $y$. Then one has the following well-known result (see Theorem 184 from [@Hardy-Wright]).
\[prop\_cont\_frac\] Let $y$ be in $\QQ$, $a$ and $b$ in $\ZZ$ with $b\neq 0$, and $$\left|\frac{a}{b}-y\right|<\frac{1}{2b^2} .$$ Then $a/b$ is a convergent of the continued fraction of $y$.
Computing torsion points II {#sec:ctd2}
===========================
In this section we describe how we compute the fields of definition of certain torsion points in Jacobians of modular curves. We recommend [@Diamond-Shurman] to those who are interested in an introduction to the theory of modular forms.
Let $\SL_2(\ZZ)$ denote the group of $2$ by $2$ matrices with coefficients in $\ZZ$ and with determinant one. It acts on the complex upper half plane $\HH$ via fractional linear transformations
(
a & b\
c & d
)z = .
The standard fundamental domain $F$ for $\SL_2(\ZZ)$ acting on $\HH$ (see Figure \[figbas1\]) consists of the $z$ with $|z|\geq1$ and $|\Re(z)|\leq1/2$.
$$\begin{tikzpicture}[scale=2,thick]
\fill[black!20!white] (0.5,2) -- (0.5,0.866) arc (60:120:1cm)
-- (-0.5,2)-- cycle;
\draw (-2,0) -- (2,0);
\draw (-0.5,2) -- (-0.5,0) node[anchor=north]{$-1/2$};
\draw (0.5,2) -- (0.5,0) node[anchor=north]{$1/2$};
\draw (1,0) arc (0:180:1cm) node[anchor=north]{$-1$};
\draw (1,0) -- (1,0) node[anchor=north]{$1$};
\draw (0,1.2) -- (0,1.2) node[anchor=south]{$F$};
\end{tikzpicture}$$
It is not bounded, hence not compact. Viewing $\HH$ as the open northern hemisphere in ${\mathbb P}^1(\CC)=\CC\cup\{\infty\}$, with boundary the equator ${\mathbb P}^1(\RR)$, we see that the closure $\ol{F}$ of $F$ in ${\mathbb P}^1(\CC)$ is the union of $F$ and the point $\infty$.
For every prime number $\ell$ we let $\Gamma_\ell$ denote the subset of $\SL_2(\ZZ)$ consisting of the $(\begin{smallmatrix}a & b\\ c & d\end{smallmatrix})$ with $c$, $a-1$ and $d-1$ divisible by $\ell$. Then $\Gamma_\ell$ is a subgroup of $\SL_2(\ZZ)$, of index $\ell^2-1$. We assume that $\ell\geq 5$ from now on. Then the action of $\Gamma_\ell$ on $\HH$ is free. Each $z$ in $\HH$ has a neighbourhood $U$ such that all $\gamma U$ for $\gamma$ in $\Gamma_\ell$ are disjoint. The quotient $\Gamma_\ell\backslash\HH$ is therefore a Riemann surface that we denote by $Y_\ell$, and the quotient map $\HH\to Y_\ell$ is a covering map, that is, each point $y$ in $Y_\ell$ has an open neighbourhood $U$ such that the inverse image of $U$ in $\HH$ is the disjoint union of copies of $U$, indexed by the inverse image of $y$.
The Riemann surface $Y_\ell$ is not compact. A fundamental domain $F_\ell$ in $\HH$ for $\Gamma_\ell$ can be gotten as the union of the $\gamma F$, where $\gamma$ ranges over a set of representatives of $\Gamma_\ell\backslash\SL_2(\ZZ)/\{1,-1\}$. Such a set consists of $(\ell^2-1)/2$ elements and it can easily be found. We can compactify $Y_\ell$ to a compact Riemann surface $X_\ell$ by adding $\ell-1$ points, called *cusps*, the points of ${\mathbb P}^1(\RR)$ that lie in the closure of $F_\ell$ in ${\mathbb P}^1(\CC)$. These points lie in fact in ${\mathbb P}^1(\QQ)$ and can easily be written down. All this leads to an explicit topological and analytic description of $X_\ell$. It is covered by coordinate disks around the cusps. For example, the function
q, ze\^[2iz]{},
restricted to the set of $z$ with $\Im(z)>1/\ell$, induces a coordinate on a disk in $X_\ell$ around the cusp $\infty$. Indeed, the image under $q$ of this region is the punctured disk of radius $e^{-2\pi/\ell}$ around $0$, and the cusp $\infty$ fills the puncture. The genus $g_\ell$ of $X_\ell$ is equal to $(\ell-5)(\ell-7)/24$. For $\ell=11$ the genus is $1$, and indeed, $X_{11}$ is the elliptic curve $X_{11}$ given by equation (\[eq:X11\]).
It is of course a miracle that such an analytically defined Riemann surface as $X_{11}$ is defined over $\QQ$, that is, can be described as a curve in a projective space given by a equations with coefficients in $\QQ$. But this is true for all $\ell$, and it is explained as follows, for $\ell>13$. The theory of modular forms gives that the $\CC$-vector spaces $\Omega^1(X_\ell)$ of holomorphic differentials on $X_\ell$ have bases consisting of $1$-forms $\omega$ whose pullback to $\HH$ is of the form $(\sum_{n\geq1}a_nq^n){\cdot}(dq)/q$ with all $a_n$ in $\ZZ$. Quotients of such $\omega$ and $\omega'$ in $\Omega^1(X_\ell)$ then provide sufficiently many rational functions on $X_\ell$ to embed it into a projective space, such that the image is given by homogeneous polynomial equations with coefficients in $\QQ$.
We let $J_\ell$ denote the Jacobian variety of $X_\ell$. It is also defined over $\QQ$, as well as its group law. This means that the group law is described by quotients of polynomials with coefficients in $\QQ$. Therefore, for all $P$ and $Q$ in $J_\ell$ and for each $\sigma$ in $\Aut(\CC)$, the automorphism group of the field $\CC$, we have $\sigma(P+Q)=\sigma(P)+\sigma(Q)$. For each integer $m\geq1$ the kernel $J_\ell[m]$ of the multiplication by $m$ map is finite (it consists of $m^{2g_\ell}$ elements) and preserved by the action of $\Aut(\CC)$. This implies that all $P$ in $J_\ell[m]$ have coordinates in the algebraic closure $\bQ$ of $\QQ$ in $\CC$, that is, for each rational function $f$ on $J_\ell$ that is defined over $\QQ$ and has no pole at $P$, the value $f(P)$ of $f$ at $P$ is in $\bQ$. The analytic description above of $X_\ell$ gives us an analytic description of $J_\ell$.
We are interested in certain subgroups $V_\ell$ of the $\ell$-torsion subgroup $J_\ell[\ell]$ of $J_\ell$ that are invariant under the Galois group $\Gal(\bQ/\QQ)$ and consist of $\ell^2$ elements. These $V_\ell$ can be described explicitly and efficiently in terms of certain operators called Hecke operators on the first homology group of $X_\ell$. The whole point is to understand them algebraically, with their $\Gal(\bQ/\QQ)$-action.
The subgroup $V_\ell$ defines a commutative $\QQ$-algebra $A_\ell$ of dimension $\ell^2$ as $\QQ$-vector space, the coordinate ring of $V_\ell$ over $\QQ$. This algebra $A_\ell$ consists of the functions $f\colon V_\ell\to\bQ$ with the property that for all $\sigma$ in $\Gal(\bQ/\QQ)$ and all $P$ in $V_\ell$ we have $f(\sigma
(P))=\sigma(f(P))$. Addition and multiplication are pointwise. Each $f_\ell$ in $A_\ell$ with the property that the $f_\ell(P)$ are all distinct is a generator, and $A_\ell$ is then given as $\QQ[T]/(H_{f_\ell})$, with $$H_{f_\ell} = \prod_{P\in V_\ell}(T-f_\ell(P)) \quad\text{in $\QQ[T]$.}$$
A direct approach for computing $A_\ell$ or $H_{f_\ell}$ algebraically, as in Section \[sec:aaa\] in the case of the division polynomial $\psi_{11}$, is very unlikely to succeed in time polynomial in $\ell$, because in the case of $V_\ell$ one has to work with the algebraic variety $J_\ell$, whose dimension grows quadratically with $\ell$. Writing down polynomial equations with coefficients in $\QQ$ for $J_\ell$ and $V_\ell$ is probably still possible, in time polynomial in $\ell$. But computing a $\QQ$-basis of $A_\ell$ from the equations in a standard way uses Groebner basis methods, which, as far as we know, take time exponential or even worse in the number of variables, that is, exponential or worse in $\ell$.
For this reason we replace, in [@book], exact computations by approximations. There are then two problems to be dealt with. The first is to show that $f_\ell$ can be chosen so that the logarithm of the height of the coefficients of $H_{f_\ell}$, that is, the number of digits of their numerator and denominator, does not grow faster than a power of $\ell$. This problem is solved in [@Ed-Jo1], [@Merkl] and [@Ed-Jo2], using arithmetic algebraic geometry and analysis on Riemann surfaces. The second problem is to show that for the same choice of $f_\ell$, the values $f_\ell(P)$ at all $P$ in $V_\ell$ can be approximated in $\CC$ with a precision of $n$ digits in time polynomial in $n+\ell$. This is done in [@Couveignes2]. The chapters [@Bosman1] and [@Bosman2] contain real computations using the method of Section \[sec:jacinv\], for prime numbers $\ell\leq 23$.
Let us now explain how we choose $f_\ell$ (up to some technicalities; the precise setup is given in [@Ed3 §8.2]) and say some words about the approximation of the $f_\ell(P)$. Standard functions on Jacobian varieties such as $J_\ell$ are theta functions. But a problem is that these are usually given as power series in $g_\ell$ variables, and as $g_\ell$ grows this can make the number of terms that must be evaluated for a sufficiently good approximation grow exponentially in $\ell$. In other words, we know no method to approximate their values fast enough (of course, it is not excluded that such methods do exist). Our solution is to transfer the problem from $J_\ell$ to $X_\ell^{g_\ell}$, via the Abel-Jacobi map. We choose $h_\ell$ a suitable non-constant rational function on $X_\ell$, defined over $\QQ$, of small degree and with small coefficients. Then we take as origin a suitable divisor of degree $g_\ell$ on $X_\ell$, defined over $\QQ$. This divisor is carefully chosen in [@Ed3] to have the following property: for each $P$ in $V_\ell$ there is a [*unique*]{} effective divisor $Q_P = Q_{P,1}+\cdots+Q_{P,g_\ell}$ on $X_\ell$, such that its image under the Abel-Jacobi map is $P$. Then we define $f_\ell(P)=h_\ell(Q_{P,1})+\cdots+h_\ell(Q_{P,g_\ell})$. Rather magically, the problem of power series in many variables has disappeared. The function $h_\ell$ is locally given by a power series in [*one*]{} variable. We evaluate it at each $Q_{P,i}$ separately. The Abel-Jacobi map (see Section \[sec:afcaj\]) is given by a sum of $g_\ell$ integrals of $g_\ell$-tuples of holomorphic $1$-forms in one variable. The analytic description above of $X_\ell$ and $J_\ell$ should make it clear that the Abel-Jacobi map and the function $h_\ell$ can be well approximated with standard tools. That means that the only remaining problem is the inversion of the Abel-Jacobi map, that is, the approximation of the divisors $Q_P$, but that was discussed and solved in Sections \[sec:afcaj\] and \[sec:psi\]. The main result obtained in [@book] is the following theorem.
\[thmcompVl\] There is a deterministic algorithm that on input a prime number $\ell\geq 11$ computes the $\QQ$-algebra $A_\ell$ in time polynomial in $\ell$.
A [*probabilistic*]{} algorithm for computing $A_\ell$ is also given in [@book]. It relies on $p$-adic approximations rather than complex approximations. In [@Couveignes3] it is explained how such $p$-adic approximations can be computed efficiently. From a theoretical point of view, a probabilistic algorithm is not quite as satisfactory as a deterministic one. From a practical point of view, it is just as good. In our case the probabilistic algorithm has a simpler proof than the deterministic one. And Peter Bruin [@Bruin1; @Bruin2] has been able to generalize it to a much wider class of $V_\ell$-like modular spaces. Finding a similar generalization for the deterministic algorithm is an open problem at this time.
Applications and open questions {#sec:avoq}
===============================
The main motivation for all the work done in [@book] is the application in number theory to the fast computation of coefficients of modular forms. Instead of attempting to present this in the most general case we give two examples: Ramanujan’s $\tau$-function, and powers of Jacobi’s $\theta$-function.
Recall that $q\colon \HH\to\CC$ is the function $z\mapsto e^{2\pi iz}$. The *discriminant modular form* $\Delta$ is the holomorphic function on $\HH$ given by the converging infinite product
\[eqn5.1\] = q\_[n1]{}(1-q\^n)\^[24]{}.
The holomorphic function $\Delta$ has a power series expansion in $q$,
= \_[n1]{}(n)q\^n,
whose coefficients, which are integers, define *Ramanujan’s $\tau$-function*. It can be shown that for all $(\begin{smallmatrix}a & b\\ c & d\end{smallmatrix})$ in $\SL_2(\ZZ)$, and for all $z$ in $\HH$, we have
\[eqn5.3\] () = (cz+d)\^[12]{}(z).
Functions $f\colon\HH\to\CC$ that are given by a power series $\sum_{n\geq 1}a_n(f)q^n$ with this symmetry under the action of $\SL_2(\ZZ)$ on $\HH$ with the exponent $12$ replaced by an integer $k$ are called cuspidal modular forms of weight $k$ on $\SL_2(\ZZ)$. The complex vector spaces $S(\SL_2(\ZZ),k)$ of cuspidal modular forms of weight $k$ are finite dimensional. The dimension grows roughly as $k/12$. More precisely, for $k<12$ the space $S(\SL_2(\ZZ),k)$ is zero, and $S(\SL_2(\ZZ),12)$ is one-dimensional, generated by $\Delta$. The fact that each $g$ in $\GL_2(\QQ)$ with $\det(g)>0$ acts on $\HH$ and normalises $\SL_2(\ZZ)$ up to finite index leads to operators $T_{k,g}$ on the $S(\SL_2(\ZZ),k)$. These operators are named after Hecke. For each integer $n\geq1$ there is an operator $T_{k,n}$; for $n$ prime it is the one induced by the matrix $(\begin{smallmatrix}n&0\\0&1
\end{smallmatrix})$ and for general $n$ it is a bit more complicated. As the space $S(\SL_2(\ZZ),12)$ is one-dimensional, each $T_{12,n}$ acts on it as multiplication by a scalar. This scalar turns out to be the coefficient $\tau(n)$ of $q^n$ in the power series of $\Delta$. Well known relations between the Hecke operators imply relations between the $\tau(n)$ that are summarised in the identity of Dirichlet series, for $s$ in $\CC$ with real part $\Re(s)$ large enough:
\_[n1]{} (n)n\^[-s]{} = \_p(1-(p)p\^[-s]{}+p\^[11]{}p\^[-2s]{})\^[-1]{}.
Here the product is over all prime numbers, and both sides converge for $\Re(s)>13/2$. In fact, it is a famous theorem of Deligne ([@Deligne1] and [@Deligne2]) that for all primes $p$ one has
\[eq:ineq\] |(p)|2p\^[11/2]{},
as conjectured by Ramanujan.
The identity of Dirichlet series shows that for $n$ and $m$ with greatest common divisor $1$ we have $\tau(nm)=\tau(n)\tau(m)$, and that for $p$ prime and $r\geq2$ we have $\tau(p^r)=\tau(p)\tau(p^{r-1})-p^{11}\tau(p^{r-2})$. Therefore, the computation of $\tau(n)$ is reduced to that of the $\tau(p)$ for $p$ dividing $n$. We can now state one of the main theorems of [@book].
\[theorem:tau\] There is a deterministic algorithm that on input an integer $n\geq1$ together with its factorisation into prime factors computes $\tau(n)$ in time polynomial in $\log n$.
Before this result, the fastest known algorithms to compute $\tau(n)$ took time exponential in $\log n$. For example, if one computes the product in (\[eqn5.1\]) up to order $n$ by multiplying the necessary factors, then one spends time at least linear in $n$. To prove the theorem, it suffices to show that for $p$ prime $\tau(p)$ can be computed in time polynomial in $\log p$. This will be done using Theorem \[thmcompVl\].
The fact that modular forms have an enormous amount of symmetry as in (\[eqn5.3\]) is certainly powerful, but it does not suffice at this point. What is needed is Galois symmetry, which is also behind Deligne’s famous result mentioned above. A lot could be said on this, but this is not an appropriate place for that.
In a nutshell: modular forms give elements in de Rham cohomology of complex algebraic varieties defined over $\QQ$, and the singular homology with torsion coefficients $\ZZ/\ell\ZZ$ of those complex varieties can be defined algebraically (Grothendieck, Artin, Deligne) and therefore has an action by $\Gal(\bQ/\QQ)$.
For example, $\Delta$ gives rise, for every prime integer $\ell\ge 11$, to a certain subgroup $V_\ell$ of the $\ell$-torsion of the Jacobian $J_\ell$ of $X_\ell$. This subgroup $V_\ell$ has cardinality $\ell^2$. For $\ell\neq p$ the image of $\tau(p)$ in $\ZZ/\ell\ZZ$ is determined by the action of $\Gal(\bQ/\QQ)$ on this $V_\ell$. The addition map $V_\ell\times
V_\ell\to V_\ell$ induces a $\QQ$-algebra morphism called co-addition from $A_\ell$ to $A_\ell\otimes A_\ell$, that is, from $\QQ[T]/(H_{f_\ell})$ to $\QQ[T_1,T_2]/(H_{f_\ell}(T_1),H_{f_\ell}(T_2))$. Computing $\tau(p)$ modulo $\ell$ (for $p\neq\ell$) is then done by reducing $A_\ell$ with its co-addition modulo $p$ and computing on this reduction $A_{\ell,p}$ a certain relation between the co-addition and the Frobenius map that sends $a$ in $A_{\ell,p}$ to $a^p$, just as in Schoof’s algorithm for elliptic curves (see Section 1.2 of [@Ed1 §1.2]). For more detail the interested reader is referred to Section 2.4 of [@Ed2 §2.4] and the references therein. The point is that this advanced machinery can actually be used for computing $\tau(p)\bmod \ell$ in time polynomial in $\log p$ and $\ell$.
In order to recover the actual value of $\tau(p)$ as an integer, we compute $\tau(p)$ modulo several small primes $\ell$. If the product of these small primes is bigger than $4p^{5.5}$ then we deduce $\tau(p)$ using inequality (\[eq:ineq\]) and Chinese remainder theorem [@Cohen 1.3.3].
We now come to our second example: the classical question in how many ways a positive integer $n$ can be written as a sum of $d\geq 1$ squares of integers. Let us write $r_d(n)$ for this number, that is, $r_d(n)=\#\{x\in\ZZ^d : x_1^2+\cdots+x_d^2=n\}$. Then $r_d(n)$ is the coefficient of $q^n$ in the formal power series $\theta_d$, with
\_d = \_[n0]{}r\_d(n)q\^n = \_[x\^d]{}q\^[x\_1\^2++x\_d\^2]{}= (\_[x\_1]{}q\^[x\_1\^2]{}) (\_[x\_d]{}q\^[x\_d\^2]{}) = \_1\^d .
The formal power series $\theta_1$ defines a holomorphic complex function on the complex upper half plane $\theta\colon\HH\to\CC$ by viewing $q$ as the function $q\colon z\mapsto e^{2\pi iz}$. Poisson’s summation formula then shows that for all $z\in\HH$ we have
(-1/4z) = (-2iz)\^[1/2]{}(z),
where the square root is continuous and positive for $z$ in $i{\cdot}\RR_{>0}$. This functional equation for $\theta$, together with the obvious one $\theta(z+1)=\theta(z)$, imply that $\theta$ is a modular form of weight $1/2$, and therefore that $\theta_d$ (interpreted as a function on $\HH$) is a modular form of weight $d/2$.
This fact is the origin of many results concerning the numbers $r_d(n)$. The famous explicit formulas for the $r_d(n)$ for even $d$ up to $10$ due to Jacobi, Eisenstein and Liouville (see [@Milne2] and Chapter 20 of [@Hardy-Wright]) owe their existence to it. In order to state these formulas, let $\sum_{d|m}$ denote summation over the positive divisors $d$ of $m$, with the convention that there are no such $d$ if $m$ is not an integer, and let $\chi\colon\ZZ\to\CC$ be the map that sends $n$ to $0$ if $n$ is even, to $1$ if $n$ is of the form $4m+1$ and to $-1$ if $n$ is of the form $4m-1$. Then we have: $$\begin{aligned}
r_{2}(n) & = 4\sum_{d|n}\chi(d), \\
r_{4}(n) & =8\sum_{2\nmid d|n}d + 16\sum_{2\nmid d|(n/2)}d, \\
r_{6}(n) & = 16\sum_{d|n}\chi\left(\frac{n}{d}\right)d^2
-4\sum_{d|n}\chi(d)d^2, \\
r_{8}(n) & = 16\sum_{d|n}d^3 - 32\sum_{d|(n/2)}d^3 +
256\sum_{d|(n/4)}d^3, \\
r_{{10}}(n) & = \frac{4}{5}\sum_{d\mid n}\chi(d)d^4 +
\frac{64}{5}\sum_{d\mid n}\chi\left(\frac{n}{d}\right)d^4
+\frac{8}{5}\sum_{d\in\ZZ[i],\, |d|^2=n}d^4.
\end{aligned}$$ In the last formula, $\ZZ[i]$ is the set of Gaussian integers $a+bi$ in $\CC$ with $a$ and $b$ in $\ZZ$.
Using these formulas, the numbers $r_d(n)$ for $d$ in $\{2,4,6,8,10\}$ can be computed in time polynomial in $\log n$, if $n$ is given with its factorisation in prime numbers. This is not the case for the formulas that were found a bit later by Glaisher for $r_d(n)$ for some even $d\geq12$. We give the formula that he found for $d=12$, as interpreted by Ramanujan:
r\_[12]{}(n) = 8\_[d|n]{}d\^5 -512\_[d|(n/4)]{}d\^5 + 16a\_n, \_[n1]{}a\_nq\^n = q\_[m1]{}(1-q\^[2m]{})\^[12]{}.
Computing $a_n$ by multiplying out the factors $1-q^{2m}$ up to order $n$ takes time at least linear in $n$, hence exponential in $\log
n$. We know of no direct way to compute the $a_n$ in time polynomial in $\log n$, even if $n$ is given with its factorisation. However, $\sum_{n\geq1}a_nq^n$ is a modular form, and therefore we *can* compute $a_n$ in time polynomial in $\log n$, if $n$ is given with its factorisation, *via* the computation of Galois representations. The same is true for the $r_d(n)$ for all even $d$. The explicit formulas for $d\leq 10$ correspond precisely to the cases where the Galois representations that occur are of dimension one, whereas for $d\geq 12$ genuine two-dimensional Galois representations always occur, as proved by Ila Varma in her master’s thesis [@Varma].
We conclude that from an algorithmic perspective the classical problem of computing the $r_d(n)$ for even $d$ and $n$ given with its factorisation into primes is solved for *all* even $d$. The question as to the existence of *formulas* has a negative answer, but for *computations* this does not matter.
#### Open questions
Finally, we should point out that the algorithms in theorems \[thmcompVl\] and \[theorem:tau\], despite their polynomial time complexity, are not so practical at present. However, Bosman’s computation of the $V_\ell$ associated with $\Delta$ for $\ell$ in $\{13, 17, 19\}$ enabled him to further study Lehmer’s conjecture on the values of $\tau(n)$ modulo $n$. See Lygeros and Rozier [@Ly-Ro] for a more classical experimental approach. A challenge for the near future is to design and implement a practical variant of these algorithms.
[10]{}
J. Bosman, Computations with modular forms and Galois representations. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 129–157.
J. Bosman, Polynomials for projective representations of level one forms. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 159–172.
B. Edixhoven and J.-M. Couveignes editors, with contributions by J. Bosman, J.-M. Couveignes, B. Edixhoven, R. de Jong and F. Merkl, *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011.
A. Brill and M. Noether, Über die algebraischen Functionen und ihre Anwendung in der Geometrie. *Mathematische Annalen.* **7** (1874), 269–310.
P. Bruin, Modular curves, Arakelov theory, algorithmic applications. PhD-thesis, Leiden, 2010. Available on-line at: `http://hdl.handle.net/1887/15915`
Peter Bruin, Computing coefficients of modular forms. *Publications math[é]{}matiques de Besan[ç]{}on.* (2011), 19–36. Available on-line at: `http://pmb.univ-fcomte.fr/2011.html`
H. Cohen, *A course in computational algebraic number theory.* Graduate Texts in Mathematics 138. Springer, Berlin, 1993.
J.-M. Couveignes, Computing complex zeros of polynomials and power series. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 95–128.
J.-M. Couveignes, Approximating $V_f$ over the complex numbers. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 257–336.
J.-M. Couveignes, Computing $V_f$ modulo $p$. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 337–370.
J.E. Cremona, *Algorithms for modular elliptic curves.* Cambridge University Press, London, 1997.
B. Datta and A.N. Singh, *History of Hindu Mathematics.* Motilal Banarsi Das, Lahore, 1935.
P. Deligne, Formes modulaires et repr[é]{}sentations $l$-adiques. S[é]{}minaire Bourbaki **355** (1969).
P. Deligne, La conjecture de Weil. I. Inst. Hautes Études Sci. Publ. Math. **43** (1974), 273–307.
F. Diamond and J. Shurman, *A first course in modular forms.* GTM 228, Springer, Berlin, 2005.
B. Edixhoven, Introduction, main results, context. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 1–27.
B. Edixhoven, Modular curves, modular forms, lattices, Galois representations. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 29–68.
B. Edixhoven, Description of $X_1(5l)$. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 173–185.
B. Edixhoven and R. de Jong, Applying Arakelov theory. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 187–201.
B. Edixhoven and R. de Jong, Bounds for Arakelov invariants of modular curves. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 217–256.
A. Enge, *Elliptic curves and their applications to cryptography, an introduction.* Kluwer Academic, New York, 1999.
G. Frey and M. Müller, Arithmetic of modular curves and applications. In *On Artin’s conjecture for odd $2$-dimensional representations.* Lecture Notes in Math. 1585, Springer, Berlin, 1994.
D. M. Gordon, A Survey of Fast Exponentiation Methods. J. Algorithms **27**(1) (1998), 129–146.
G.H. Hardy and E.M. Wright, *An introduction to the theory of numbers.* Fifth edition. Clarendon Press, New York, 1979.
N. Lygeros and O. Rozier, A new solution to the equation $\tau(p)\equiv 0\pmod p$. J. Integer Seq. **13** (2010), no. 7, Article 10.7.4, 11 pp.
Y. Manin, Parabolic points and zeta function of modular curves. Math. USSR Izvestija **6** (1972), no. 1, 19–64.
L. Merel, Universal Fourier expansions of modular forms. In *On Artin’s conjecture for odd $2$-dimensional representations.* Lecture Notes in Math. 1585, Springer, Berlin, 1994, 59–94.
F. Merkl, An upper bound for Green functions on Riemann surfaces. In *Computational Aspects of Modular Forms and Galois Representations.* Annals of Mathematics Studies 176, Princeton University Press, Princeton, NJ, 2011, 203–215.
S.C. Milne, Infinite families of exact sums of squares formulas, Jacobi elliptic functions, continued fractions, and Schur functions. Ramanujan J. **6** (2002), no. 1, 7–149.
C. Batut, K. Belabas, D. Bernardi, H. Cohen, and M. Olivier, [*User’s guide to PARI/GP (version 2.3.1)*]{}. `http://pari.math.u-bordeaux.fr.`
J. Silverman, *The arithmetic of elliptic curves.* Lecture Notes in Math. 106, Springer, Berlin, 1986.
W.A. Stein, *Modular forms, a computational approach.* With an appendix by Paul E. Gunnells. Graduate Studies in Mathematics, 79. American Mathematical Society, Providence, RI, 2007.
I. Varma, Finding elementary formulas for theta functions associated to even sums of squares. Indag. Math. (N.S.) **22** (2011), 12-26.
E. Volcheck, Computing in the [J]{}acobian of a plane algebraic curve. In *Algorithmic Number Theory Conference.* Lecture Notes in Computer Sciences, volume 877, Springer 1994, 221-233,
[^1]: Research supported by ANR (project ALGOL ANR-07-BLAN-0248) and by DGA ma[î]{}trise de l’information.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show how QKD on a multi-user, multi-path, network can be used to establish a key between any two end users in an asynchronous fashion using the technique of bit-transport. By a suitable adaptation of our previous secret-sharing scheme we show that an attacker has to compromise all of the intermediate relays on the network in order to obtain the key. Thus, two end users can establish a secret key provided they trust at least one of the network relays.'
author:
- |
*Stephen M Barnett*$^{1}$* *&* Simon JD Phoenix*$^{2}\bigskip$\
$^{1}$Dept. of Physics, University of Strathclyde, Glasgow G4 0NG, UK\
$^{2}$Khalifa University, PO Box 127788, Abu Dhabi, UAE
title: Asynchronous QKD on a Relay Network
---
Introduction
============
The elegant and startingly original theoretical idea of Quantum Key Distribution (QKD) \[1\] has developed into a mature technology \[2\] with commercial systems readily available. Nevertheless, for a variety of reasons, it still remains something of a curiosity amongst security professionals. There is a sense in which the technology, however beautiful, addresses a non-existent problem since the threat models used by security professionals rarely put key distribution at the top of the list, with good reason. Existing key distribution mechanisms are considered to be more than adequate to address the perceived risk. Furthermore, with suitable key-expansion algorithms, there is little in practice that QKD can achieve that a conventional classical system cannot.
The security of QKD is based on different principles, however, and conventional security techniques largely rely on unproven (but reasonable) assumptions \[3\]. So, for example, the security proof for a block cipher used in a suitable mode for key expansion, rests on the assumption that the cipher is a pseudorandom permutation. Whether we choose a QKD system or a conventional classical system for our key distribution, we still have to rely to some extent on our confidence in the underlying principles on which the security is based.
Another difficulty with QKD is the limitation imposed by the nature of the technology. The technique relies on the transmission of single quanta, or at least a reasonable approximation to them. Any network element which is too lossy, or actively processes the signal in some way, will destroy the capability of the quantum channel to transmit keys. Thus, installing the technology on realistic networks poses something of a technical challenge. Whilst stable and tested solutions exist for point-to-point links, extending this to a network application is not straightforward and relies on the introduction of additional trusted network elements to enable the system to span reasonable distances and to route the signal between the required end points of the network. Good progress, however, has been made in developing the basic technique to work on more realistic communication networks \[4,5\].
The above comments notwithstanding it is likely that QKD will find application as part of an overall security solution for some situations and networks. Furthermore, the current threat model to key distribution will significantly alter as more progress is made towards the development of a working quantum computer that can process strings of qubits of sufficient size to pose a threat to existing public-key mechanisms \[6,7\]. Whilst classical key distribution techniques based on symmetric cryptography can address the threat posed by quantum computation, it is by no means certain that these will be an obvious natural choice over a QKD solution should the need arise for a widespread overhaul of the existing key distribution techniques based on public-key cryptography.
In this paper we look at how the bit-transport technique for QKD \[8\] can be used on a network in an *asynchronous* fashion to establish keys between any end-users of the network. The technique requires that the network relays act as intermediaries to correlate various QKD transmissions together. We show that with a suitable arrangement of relays an attacker has to compromise all of the relays on any particular channel in order to obtain the key. Thus, instead of having to trust *all* of the relays on a channel, the end users only have to trust *at least one*. We achieve this by a suitable adaptation of our ‘drop-out’ technique \[9\] for single QKD channels.
A Single-Relay QKD Channel
==========================
A relay on a QKD channel is used, primarily, to increase the distance. The conventional way of achieving this is for Alice and Bob to each establish separate quantum[^1] keys with the intermediate relay. In an obvious notation Alice ($A$) establishes a key with the relay ($R$) which we label $QK_{AR}$. Bob establishes a different key $QK_{RB}$ where we have used the order of the indices here to denote the ‘direction’ of key establishment which we take to be the direction in which the quanta are transmitted. The final key, $K$, between Alice and Bob can be established in a linkwise fashion.
If one relay is not enough to span the distance between $A$ and $B$ with a quantum key transmission then, clearly, we can use any number of intermediate relays which each establish a separate quantum key $QK_{R_{j}R_{k}}$. Once all of the keys between the various network entities have been established, the final key between Alice and Bob can again be established in a linkwise fashion.
In our previous work \[8\] we showed how we can use intercept/resend relays \[10\] to establish an end-to-end key over an extended distance with no loss of effective final key size. The primary concern with any relay system is that the relays have to be trusted intermediaries. In conventional, or intercept/resend, operation the compromise of a single relay compromises the entire channel. In \[9\] we showed how it was possible to modify the transmission protocol by adding a single relay (at least) so that an attacker needs to compromise *all* of the relays on a channel in order to obtain the key. This technique employs the notion of quantum secret sharing developed for multi-path networks in \[11\] in which we create distinct logical paths on a single channel by randomly dropping out the relays from the channel.
Asynchronous Bit-Transport on the Channel
-----------------------------------------
Let us consider a channel over which Alice and Bob desire to establish a quantum key. Further, let us suppose that a single relay is required to span the distance so that the channel is of the form Alice $\longrightarrow$ Relay $\longrightarrow$ Bob. There are two ways in which the relay can be operated; in link-by-link mode, or in intercept/resend mode to establish an end-to-end key. We have shown how bit transport can be used to establish an end-to-end key with intercept/resend relays \[8\], we’ll now consider how bit transport can be used to establish a key between Alice and Bob in an asynchronous fashion with the help of the relay. The relay establishes an independent QKD channel with Alice and Bob, respectively. On each channel the bits are sifted with public announcement of the coding basis and the ‘bad’ channels discarded. At the end of this process Alice and the relay, and Bob and the relay, possess a set of data that should, in an ideal world and in the absence of an eavesdropper, be in perfect respective agreement. The 4 sets of data can be checked for errors. If the error rate is not too high then the data sets can be saved and labelled.
After many such transmissions Alice and the relay have $n$ sets of data $S_{AR}^{(k)}$ and $\bar{S}_{AR}^{(k)}~~(1\leq k\leq n)$, where the bar denotes the relay’s data set which could differ slightly from the corresponding set of Alice if there are errors on the channel. The relay and Bob have similar sets of data, $S_{RB}^{(k)}$ and $\bar{S}_{RB}^{(k)}$ and we assume that they have $m$ such sets. The elements of each set consist of a tuple $(t,b)$ where $t$ is the timeslot and $b$ is the bit value. So for a given pair of sets, $S_{AR}^{(k)}$ and $\bar{S}_{AR}^{(k)}$, we would have elements $(t,b)$ for Alice and elements $(t^{\prime},b^{\prime})$ for the relay. We would have $t=t^{\prime}$ but in the presence of errors we would have $b=b^{\prime}$ for most, but not all, of the timeslots. Let us suppose that Alice and the relay now conduct a secure error-correcting process so that at the end of this, and with a suitable re-labelling of the timeslot values, we have that $t=t^{\prime}$ and $b=b^{\prime}$ for all elements. Let us label the sets after error-correction by $\Sigma$. We further suppose that Bob and the relay perform the same process on their sets so that they also have identical sets with elements $\left( \tau,\beta\right) $.
Now if Alice and Bob wish to establish a key then the relay can choose one of the error-corrected sets $\Sigma_{RB}^{(k^{\prime})}$ and one of the error-corrected sets $\Sigma_{AR}^{(k)}$ essentially at random (or select from those which initially had a lower error rate, for example). The relay can then choose the timeslots, at random, from the sets such that $b=\beta$ and simply announce the respective timeslots to both Alice and Bob. Alice and Bob will then share the same set of bits which can subsequently be used as a key. An eavesdropper has to have collected all of the data between Alice, Bob and the relay in order to have any chance of getting any information about the key, because she cannot know in advance which sets of data the relay will choose. Of course, once the timeslots have been announced then Alice and Bob need to perform a privacy amplification \[2,7\] on their data in order to eliminate the possible information the eavesdropper could have gleaned to a negligible level[^2].
There are variations on this protocol. For example, the error-correction need not be done before the linking of the timeslots by the relay (although this will increase the effective error rate on the final data). The participants on the channel could, with a collection of error-corrected sets, decide to XOR these sets together to reduce the potential information of the eavesdropper. With the necessity to establish only a short key and with a potentially large number of sets to choose from this technique could reduce the eavesdropper’s information to negligible levels in a similar fashion to the standard privacy amplification procedure. Alternatively, the privacy amplification could be done on the error-corrected sets $\Sigma$, individually, before the bit-transport by linkage of the timeslots. Furthermore, the linkage on this channel need not be initiated by the relay. Alice, for example, could begin the linkage by announcing the timeslots she wishes to use which then get correlated by the relay to suitable timeslots of Bob’s. The linkage need not be restricted to the selection of a single respective set of the participants. Indeed, elements from different sets could be chosen at random provided that they have the same bit value. The main limitation of this asynchronous key-establishment, whichever protocol variation is adopted, is that the relay has to be entirely trusted by Alice and Bob.
QKD Channels with Multiple Relays
=================================
It is clear that this process can be extended to channels which require multiple relays. Let us consider a channel between Alice and Bob that requires 2 relays to span the distance. Thus we have a channel of the form : $$A\longrightarrow R_{1}\longrightarrow R_{2}\longrightarrow B$$
Independent QKD transmissions are run on the channels $A\longrightarrow
R_{1},R_{1}\longrightarrow R_{2}$ and $R_{2}\longrightarrow B$ so that at the end of many such runs, and after sifting and error-correction, Alice and $R_{1}$ share $n$ sets $\Sigma_{AR_{1}}^{(i)}$, the two relays share $m$ sets $\Sigma_{R_{1}R_{2}}^{(j)}$, and $R_{2}$ and Bob share $l$ sets $\Sigma
_{R_{2}B}^{(k)}$. For convenience we shall assume that all of these sets are of the same size with cardinality $N$. The timeslot index is just an integer identifying a transmission instance, thus each of these sets consists of elements of the form $(t,b)$ with $1\leq t\leq N$ and $b\in\{0,1\}$. Each set is therefore an ordered list of $N$ bit values. An example for $N=10$ is given below, in which Alice and the first relay have selected[^3] the 4$^{th}$ set from their list of $n$ sets, the relays have selected the 2$^{nd}$ set from their list of $m$ sets, and the second relay and Bob have selected the 7$^{th}$ set from their list of $l$ sets. In practice $N$ will be orders of magnitude greater than 10.
\[c\][|c|c|c|c|]{}$N$ & $\Sigma_{AR_{1}}^{(4)}$ & $\Sigma_{R_{1}R_{2}}^{(2)}$ & $\Sigma_{R_{2}B}^{(7)}$\
1 & 0 & 1 & 0\
2 & 1 & 1 & 0\
3 & 1 & 1 & 1\
4 & 0 & 0 & 1\
5 & 1 & 1 & 0\
6 & 1 & 1 & 1\
7 & 0 & 1 & 0\
8 & 0 & 0 & 1\
9 & 1 & 0 & 1\
10 & 0 & 1 & 1\
Let us suppose that Alice and Bob wish to establish a key of length 4 bits. We’ll consider the case where the relays select the key to be used. $R_{1}$ and $R_{2}$ communicate and agree on 4 elements chosen at random from their set $\Sigma_{R_{1}R_{2}}^{(2)}$. For example, we suppose that they agree on the following list (6,1,9,3) giving the key 1101. Now $R_{1}$ chooses, at random, the indices from the set $\Sigma_{AR_{1}}^{(4)}$ that will give this key and communicates the list of chosen indices to Alice. So, for example, in order to communicate the key 1101 to Alice, $R_{1}$ might transmit the list (3,2,8,9). It is important that once an index value has been selected it is eliminated from any subsequent choice. The relay $R_{2}$ performs the same process with Bob and, for example, might transmit the index list (4,9,2,10). At the end of this process both Alice and Bob will share the key 1101.
As we have noted above, the selection of the bit values need not be restricted to a single set. These values could be chosen randomly from all available sets. In this case each bit value index must be accompanied by another integer which indexes the set from which it is taken. Thus a list of tuples must be transmitted. So, for example, the first relay could send Alice the list $[(12,4),(2,1),...(34,10)]$ which would indicate that the first bit of the key is the 4$^{th}$ element of their set $\Sigma_{AR_{1}}^{(12)}$ and so on.
Of course, more sophisticated schemes for key establishment can be envisaged, rather than just the straight linkage of the timeslots. For example, Alice and the first relay could partition the data in their error-corrected sets into 10 bit blocks (say), where the elements of each block are selected at random by publicly agreeing on a random sequence. In effect, a common random permutation is applied to the set. The bit values of the key can then be established by announcing a block and determining the parity. This procedure combines an element of privacy amplification into the key establishment.
Again, the main limitation on this technique, from a security perspective, is that each network element knows the final key and so each network element (that is, the relays) needs to be trusted. If any one relay is compromised then the key between Alice and Bob can be determined by the attacker. We can adapt our previous secret sharing technique \[9\] to alleviate this problem so that an attacker has to compromise *all* of the relays on the channel. Let’s look at an example of how this works.
Securing a Multiple-Relay Channel
---------------------------------
Let us consider the following channel $$A\longrightarrow R_{1}\longrightarrow R_{2}\longrightarrow R_{3}\longrightarrow B$$
Now let us suppose that a successful QKD transmission can be performed between network elements that are, *at most*, 2 steps away. The following quantum keys between Alice and Bob can therefore be established by utilizing the bit-transport technique outlined above;
\[c\][cc]{} &\
$QK_{AR_{1}R_{2}R_{3}B}$ & $AR_{1}R_{2}R_{3}B$\
$QK_{AR_{2}R_{3}B}$ & $A-R_{2}R_{3}B$\
$QK_{AR_{1}R_{2}B}$ & $AR_{1}R_{2}-B$\
$QK_{AR_{1}R_{3}B}$ & $AR_{1}-R_{3}B$\
$QK_{AR_{2}B}$ & $A-R_{2}-B$
We can see that there is at least one channel in which a given relay does not participate. So Alice and Bob establish these separate quantum keys and simply XOR them together to obtain their final key. None of the relays now possess this final key, and in order to obtain the key an attacker must compromise all of the relays in the channel. If only one is not compromised, and therefore trusted, the attacker cannot obtain the final key. For this particular example Alice and Bob could establish either of the following final quantum keys[^4]: $$\begin{aligned}
QK_{AB} & =QK_{AR_{1}R_{2}R_{3}B}\oplus QK_{AR_{2}R_{3}B}\oplus
QK_{AR_{1}R_{2}B}\oplus QK_{AR_{1}R_{3}B}\oplus QK_{AR_{2}B}\\
& \\
QK_{AB}^{\prime} & =QK_{AR_{2}R_{3}B}\oplus QK_{AR_{1}R_{2}B}\oplus
QK_{AR_{1}R_{3}B}\oplus QK_{AR_{2}B}$$
The second key here would seem intuitively more preferable since relays 1 and 3 only participate in 2 out of the 4 channels. It is obvious how to extend this to any general network configuration. Indeed, on multi-path networks we could employ a combination of this single-channel secret sharing and the multi-path secret sharing developed in \[11\]. Furthermore, on multi-path networks we could choose any path, at random, for each key bit we wish to establish using the bit-transport technique. The attacker would then be in the position of having to compromise *all* of the relays on the multi-path network, or collect *all* the data exchanged on all possible paths.
Conclusions
===========
We have shown how, with a suitable adaptation of our previous bit-transport and secret-sharing techniques \[8,9\], an asynchronous quantum key can be established between any two users on a network in such a way that an attacker has to compromise all of the intermediate network elements to obtain the final key. Indeed, in order to obtain even a limited amount of information about the key the attacker must collect the data between all network elements, even on different paths. Of course, the standard operating assumptions of a normal single link QKD channel must be observed. So, for example, the public communications between the various elements must be authenticated and any side-channel information must be protected.
**Acknowledgement**
SMB thanks the Royal Society and the Wolfson Foundation for financial support
**References**
1. Bennett, C.H.; Brassard, G. Quantum Cryptography: Public Key Distribution and Coin Tossing. *Proceedings of the IEEE International Conference on Computers, Systems, and Signal Processing*, Bangalore, India, p.175, 1984
2. Gisin, N.; Ribordy, G.; Tittel, W.; Zbinden, H. Quantum Cryptography, *Rev. Mod. Phys*., **2002**, *74*, 145–195
3. See, for example, Katz, J.; Lindell, Y. *Introduction to Modern Cryptography*; Chapman & Hall: Florida, 2008
4. Zbinden, H.; Gisin, N.; Huttner, B.; Muller, A.; Tittel, W. Practical aspects of quantum key distribution. *J. Cryptology,* **1998**, *11***,** 1-14
5. Sasaki, M.; *et.al.* Field test of quantum key distribution in the Tokyo QKD Network. *Optics Express*, **2011,** *19* (11), 10387-10409
6. Shor, P.W. Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer. *SIAM J. Comput*., **1997**, *26* (5), 1484–1509
7. Barnett, S.M. *Quantum Information*; Oxford University Press: Oxford, 2009
8. Barnett, S.M.; Phoenix, S.J.D. Extending the Reach of QKD Using Relays. *Proceedings of the IEEE GCC Conference*, Dubai, UAE, pp.140-142, 2011
9. Barnett, S.M.; Phoenix, S.J.D. Securing a Quantum Key Distribution Relay Network Using Secret Sharing, *Proceedings of the IEEE GCC Conference*, Dubai, UAE, pp.143-145, 2011
10. Bechmann-Pasquinucci, H.; Pasquinucci, A. Quantum Key Distribution with Trusted Quantum Relay. **2005,** http://arxiv.org/abs/quant-ph/0505089 (accessed Mar 19, 2012)
11. Beals, T.R.; Sanders, B.C. Distributed relay protocol for probabilistic information theoretic security in a randomly-compromised network. *Third International Conference on Information Theoretic Security* (ICITS), pp.29–39, 2008
[^1]: We use the term ‘quantum key’ here merely to describe the process by which the key has been established, that is, by quantum key distribution. There is, of course, nothing quantum about the key itself!
[^2]: Although privacy amplification can be performed after the bit-transport (as we have discussed here) this is not optimal because the linkage of the timeslots gives the attacker a greater number of bits of the key that she knows with certainty.
[^3]: We assume that such a selection has been made for the purposes of explanation of the asynchronous bit-transport technique.
[^4]: Of course there are other final quantum keys that can be established. It is important, however, that any relay has not participated in the establishment of all of the keys used in the XOR. Thus, for example, if we tried to establish a key $QK_{AR_{2}R_{3}B}\oplus QK_{AR_{1}R_{2}B}$ then relay 2 knows the final key and the attacker would only have to compromise this relay in order to obtain the final key.
|
{
"pile_set_name": "ArXiv"
}
|
\#1\#2 \#1
------------------------------------------------------------------------
\#1
\#1[[**[\#1]{}**]{}]{}
Introduction
============
In a system of correlated electrons the external magnetic field can induce several interesting effects. Theoretically the least understood are those phenomena, where the magnetic field couples to the motion of charge carriers. In recent years it has been realized that the understanding of the anomalous temperature- and doping-dependence of the Hall effect [@ong] is one of the most challenging questions in connection with the normal state of cuprates, being representatives of two-dimensional (2D) doped magnetic insulators. Here even the sign of the effect is theoretically controversal [@brin1; @assa]. The diamagnetic contribution to the d.c. susceptibility has been much less investigated [@wals], although it is closely related to the Hall conductivity [@rojo]. It emerges from the orbital motion of mobile carriers. For noninteracting electrons the contribution corresponds to the Landau diamagnetism, which is largely temperature independent. In analogy to the Hall effect and other anomalous properties of the normal state in cuprates, one could expect anomalies also in the diamagnetic contribution. So far, however, both experimental and theoretical answers are lacking.
Magnetic field dependence of the eigenstates of tightbinding electrons is nontrivial even in the absence of any electron correlations [@hofs], in particular when the dependence of the ground state on the field strength $B$ and electron density is investigated [@hase; @nori]. There have been only few analogous studies of correlated systems. Recently, the ground state of a single hole in the 2D $t$-$J$ model in the presence of magnetic field [@bera] has been studied. The main message is that for finite (but not very small) $B$ the energy is reduced by an amount proportional to parameter $t$, and the result in such a doped insulator was interpreted in terms of the composite nature of quasiparticles (QP) [@bera; @bera1]. Another evident observation is, however, the difficulty to extract a reasonable result from studies of small systems.
The aim of this paper is to elaborate on the problem of a single hole in a magnetic insulator in the presence of a homogeneous magnetic field. We study the planar $t$-$J$ model [@rice] as a prototype model for strongly correlated electrons and electronic properties of cuprates, $$H=-t\sum_{\langle ij\rangle s}({\rm e}^{i \theta_{ij}}
\tilde{c}^\dagger_{js}\tilde{c}^{\phantom{\dagger}}_{is}+
\text{H.c.})+J\sum_{\langle ij\rangle}\vc{S}_i\!\cdot\!\vc{S}_j ,
\label{model}$$ where $\tilde{c}^\dagger_{is},\tilde{c}_{is}$ are fermionic operators, projecting out states with the double occupancy. We consider the system in a homogeneous field $B$, perpendicular to the plane, and use for convenience the Landau gauge, where $$\theta_{ij}=\f{e}{\hbar}\vc{d}_{ij}\!\cdot\!\vc{A}(\vc{r}_i),\qquad
\vc{A}=B(0,x,0),
\label{theta}$$ with $\vc{d}_{ij}=\vc{r}_j-\vc{r}_i$. The relevant parameter for the strength of $B$ is the dimensionless flux per plaquette $\alpha=2\pi
Ba_0^2 /\phi_0$, where $\phi_0=h/e$ is the unit quantum flux, and the relevant regime is $-\pi <\alpha <\pi$. Further on we set the lattice spacing $a_0=1$, as well as $h=k_B=1$.
In the following we restrict our study of the model Eq. (\[model\]) to the case of a single hole doped into a magnetic insulator. The idea is that results for a single hole (spin polaron) remain relevant for the regime of finite, but low, hole concentration $c_h \ll 1$. Here a semiconductor-like picture implies, so that most measurable quantities, assuming the independence of spin polarons, should simply scale with $c_h$. E.g., the diamagnetic susceptibility should behave as $\chi\propto c_h$.
The ground state of the spin polaron at $B=0$ has been studied extensively both by analytical and numerical approaches, and can be considered as one of few rather settled problems within the theory of correlated systems. Still, here at least two substantially different regimes have to be distinguished.
At finite $J>0$ (as relevant for cuprates with $J/t\sim0.3$) the ground state of a hole in an antiferromagnetic (AFM) spin background has the property of a quasiparticle (QP) with $S=1/2$ and a well defined dispersion $\varepsilon_0(\vc{k})$. Consistent results have been obtained for $\varepsilon_0(\vc{k})$ using the self-consistent Born approximation (SCBA) [@schm; @mart], perturbation expansion [@prel], numerical approaches including both the exact diagonalization of small systems and the quantum Monte Carlo method [@dago]. Calculations reproduce a minimum at $\vc{k}^*=(\pm\pi/2,\pm\pi/2)$, which is very anisotropic, i.e.$\mu=m_{\perp}/m_{\parallel}\sim5$ for $J/t\sim0.3$. This indicates a very weak dispersion along the AFM zone boundary, connecting $\vc{k}=\vc{k}^*$ with $\vc{k}=\vc{k}^{**}=(\pi,0)$, $(0,\pi)$. Studying small systems [@dago] the latter dispersion is not easy to reproduce correctly. E.g., on a frequently studied system of $4\times4$ sites, states with $\vc{k}^*$ and $\vc{k}^{**}$ are degenerate, so pronounced finite size effects are expected. Since a small $B$ just probes the effective mass of the QP, it is not surprising that results obtained on small lattices are not reliable or can be even misleading [@bera]. On the other hand, ARPES measurements on undoped cuprates [@well] show a more isotropic minimum around $\vc{k}^*$. The explanation seems to be beyond the simple $t$-$J$ model, and the additional effect is attributed to the next-nearest-neighbor hopping ($t'$) term [@hybe].
The behavior at $J=0$ is quite different. As shown by Nagaoka [@naga], the ground state is ferromagnetic (FM) with $S=S_{\rm max}$ and momentum $\vc{k}=0$, where the QP is a simple hole in the filled band of polarized electrons with an unrenormalized band mass. Nevertheless, close to this simple QP branch there is a large density of complicated excited states, where the hole motion is predominantly incoherent [@brin]. Therefore it is expected that even moderate temperature $T>0$ should have a considerable effect.
The paper is organized as follows. Section II is devoted to the study of a single hole at $J=0$ and arbitrary $B$. Results are obtained via the high-$T$ expansion and the Lanczos diagonalization technique for small systems at $T=0$ and for $T>0$ as well. In Sec. III we consider the AFM case with $J>0$. Here the analysis of small systems at finite $B$ is employed together with the study of the ground state using the cumulant expansion in $t/J$, which has proven to be very informative for $B=0$ [@prel]. In the last section, Sec. IV, our results are summarized and a brief discussion on the magnitude of the susceptibility and its relation to the Hall constant is presented.
$J=0$ case
==========
High-$T$ expansion
------------------
To study a single hole, as described by the model Eq.(\[model\]) with $J=0$ and $B>0$, we first use the standard high-$T$ expansion (HTE). Its application is in this case simple, since the only expansion parameter is $t/T$, while $B$ remains arbitrary. The free energy $$F=-T\,{\rm ln}\, Z=-T\,{\rm ln}\,{\rm Tr}\,
e^{-\tilde{\beta}\tilde{H}},\label{free}$$ is within the high temperature expansion expressed in terms of moments $\mu_n$ and cumulants $\lambda_n$ $$\begin{aligned}
{\rm ln}\, Z&=&{\rm ln}\,{\rm Tr}\,1+{\rm
ln}\,\left[1+\sum_{n=1}^\infty{\frac{\tilde{\beta}^n}{n!}\mu_n}\right]
\nonumber \\
&=& {\rm ln}\,{\rm Tr}\,1+
\sum_{n=1}^\infty{\frac{\tilde{\beta}^n}{n!}\lambda_n},\label{lnz}\end{aligned}$$ where $\tilde \beta=t/T$, $\tilde H=H/t$ and $$\mu_n=(-1)^n {\rm Tr}\, \tilde H^n/{\rm Tr}\,1.\label{mu}$$
Moments $\mu_n$ can be expressed as a sum over ${n\choose n/2}^2$ closed graphs (paths). Counting different spin configurations for $B=0$ which remain unchanged [@brin] after the performed path, each graph contributes a weight $2^{f-r+1}$. Here $f$ is the number of cycles in the spin permutation resulting from a hole traversing the graph, and $r$ is the number of different sites in the graph. For $B>0$ the only change comes from the contribution of the enclosed magnetic flux, so that the weight becomes $$w_n=2^{f-r+1}\,e^{im\alpha},$$ where $m$ is the area of the graph in units of $a_0^2$.
Here it is helpful to choose a $45^\circ$ rotated coordinate system so that a 2D graph decouples into a direct product of two 1D graphs. In this way it is straightforward to generate nonequivalent graphs numerically. We were able to evaluate $\mu_n$ and $\lambda_n$ up to the order $n=18$. In Table \[tab1\] lowest cumulants ($n\le 6$) $\lambda_n=\sum_m\lambda_{mn}\cos m\alpha$ are presented for illustration, while higher cumulants are available upon request.
$\lambda_{nm}$ 0 $1$ 2 3 4 5 6
---------------- --------- ------------------- ---------------- ---------------- ----------------- ------------ -------------
1
4
-20 2
472 -48 $\f{3}{2}$
-24518 5992 -198 $\f{3}{2}$ $\f{1}{4}$
2207234 $-\f{2703635}{4}$ $\f{65195}{2}$ $-\f{1065}{2}$ $-\f{1155}{32}$ $\f{5}{8}$ $\f{5}{32}$
: Cumulants $\lambda_n=\sum_m{\lambda_{nm}\cos m\alpha}$.
\[tab1\]
From Eqs.(\[free\],\[lnz\],\[mu\]) the series for the orbital susceptibility (per one hole) can be generated, $$\chi=\mu_0{\partial^2 F \over \partial
B^2}{\bigg\vert}_{B=0},\label{defchi}$$ $$\frac{\chi}{\chi_0}=-\frac{1}{12}\tilde{\beta}^{3}+
\frac{13}{120}\tilde{\beta}^{5}-
\frac{2087}{16128}\tilde{\beta}^{7}+\frac{8161}{53760}\tilde
{\beta}^{9}-\cdots,\label{chi}$$ where $\chi_0=\mu_0e^2a_0^4t/\hbar^2$.
There is no unique procedure for the extrapolation of the power series Eqs.(\[lnz\],\[chi\]) to low $T$. For the present problem the most reasonable approach seems to be via the density of states $\rho(\varepsilon)$ and their moments [@brin], $$\begin{aligned}
Z&=&\int{\rho(\varepsilon)\,e^{-\beta\varepsilon}{\rm
d}\varepsilon},\nonumber \\
\mu_n&=&\int{\varepsilon^n\rho(\varepsilon){\rm
d}\varepsilon}.\label{dens}\end{aligned}$$
The density of states can be expanded in terms of Legendre polynomials, $$\rho(\varepsilon)=\sum_{\ell=0}^\infty{B_\ell{\rm
P}_\ell(\varepsilon)},\qquad
\mu_n=\sum_{\ell=0}^n{C_{n\ell}B_\ell},\label{leg}$$ with coefficients $$C_{n\ell}={\Gamma(\f{n}{2}+\f{1}{2})\Gamma(\f{n}{2}+1)\over
2\Gamma(\f{n}{2}+\f{\ell}{2}+\f{3}{2})
\Gamma(\f{n}{2}-\f{\ell}{2}+1)},\label{coef}$$ for $n>\ell$ and even $\ell+m$, while $C_{n\ell}=0$ otherwise.
The density of states is used to extrapolate both $F$ and $\chi$ to low $T$. After solving the linear equations (\[leg\]) for $B_\ell$, we can calculate the susceptibility $$\frac{\chi}{\chi_0}=\frac{1}{Z\tilde\beta}
\left[\frac{\partial^2\!Z}{\partial\alpha^2}-
\frac{1}{Z}\left(\frac{\partial Z}{\partial\alpha}\right)^2\right],
\label{chit}$$ where the same set of equations as in (\[dens\]-\[coef\]) holds for derivatives with respect to $\alpha$ as well.
Small system diagonalization
----------------------------
In analysis of the $t$-$J$ model the Lanczos technique for the exact diagonalization of small systems has been already extensively employed [@dago], predominantly in the investigation of the ground state and their static and dynamic properties. Recently a method combining the Lanczos procedure and the random sampling has been introduced [@jakl1] which allows calculation of finite-temperature properties in small correlated systems. The method has been used in the study of various response function within the $t$-$J$ model at $T>0$ [@jakl1]. The application is particularly simple for static quantities, which can be expressed as expectation values of conserved quantities [@jakl2]. The calculational effort is comparable to the ground state evaluation. In particular the average energy $\langle
E\rangle$, $$\begin{aligned}
\langle E\rangle&\approx&\frac{N_{st}}{KZ}\sum_{n=1}^{K}
\sum_{m=0}^{M-1}|\langle n|\psi_m^n\rangle|^2E_m^n
{\rm e}^{-\beta E_m^n},\nonumber\\
Z&\approx&\frac{N_{st}}{K}\sum_{n=1}^{K}
\sum_{m=0}^{M-1}|\langle n|\psi_m^n\rangle|^2{\rm e}^{-\beta E_m^n},
\label{lan}\end{aligned}$$ should be evaluated in this way, where $|\psi_m^n\rangle, E_m^n$ are respectively approximate eigenfunctions and energies obtained by the diagonalization within the orthonormal set, generated from the initial functions $|n\rangle$ in $M$ Lanczos steps. The $K$ initial functions $|n\rangle$ are chosen at random, while $N_{st}$ is the dimension of the complete basis. Note that it is enough to choose $M,K
\ll N_{st}$. For more detailed explanations we refer to Refs.[@jakl1; @jakl2].
The introduction of finite $B>0$ in the model, Eq.(\[model\]), reduces the translational symmetry and thus for a given system size increases the required minimal basis set. We are at present able to consider the problem of a single mobile hole in canonical 2D systems with $N = 16,18,20$ sites [@oitm], and periodic boundary conditions (p.b.c.).
It is nontrivial to incorporate phases due to a homogeneous $B$, being at the same time compatible with p.b.c. It is well known [@frad] that this can be accomplished only for quantized magnetic fields $B=m
B_0$, where $B_0=\phi_0/N$ is the smallest field corresponding to the unit quantum flux per system. To incorporate such $B$ in small systems (tilted squares), the following procedure is used: a) phases $\theta_{ij}$ for all hops inside squares are left as given within the particular Landau gauge, b) phases attributed to hops across the square boundaries are subject to the condition that the magnetic flux in each plaquette remain the same, i.e. $B'\equiv B~\rm{mod}(\phi_0)$ (up to the addition of an unit flux per plaquette). These boundary requirements lead to a set of linear equations which have solutions only for $B=mB_0$. Only for an untilted square lattice with $N=L\times
L$ sites the phases can be expressed in a simple form as $$\begin{aligned}
\theta_{(i_x,i_y)(i_x,i_y+1)}&=& B i_x , \nonumber\\
\theta_{(i_x,i_y)(i_x+1,i_y)}&=& 0,\qquad\quad i_x<L,\nonumber\\
\theta_{(L,i_y)(1,i_y)}&=& B i_y L.
\label{bound}\end{aligned}$$
Results
-------
Let us first discuss the polaron internal energy $\langle\varepsilon(\alpha)\rangle$ as obtained from the high-$T$ expansion $$\langle\varepsilon\rangle=-{\partial({\rm ln}\,Z)\over\partial\beta}=
-t\sum _{n=0}^\infty{{\tilde\beta}^n \over n!}\lambda_{n+1}.
\label{energ}$$ From cumulants $\lambda_{nm}$ in Table I we see that the $\alpha$ dependence of $\langle\varepsilon\rangle$ first enters within the order $\tilde\beta^3$. Such a term originates from a hole hopping around a loop, contributing to ${\rm Tr}\,\tilde H^4$ when all spins are equally polarized, in analogy to the processes contributing to the Hall constant [@brin1]. As a result, $\alpha$ dependence of $\langle\varepsilon\rangle$ is vanishing fast ($\propto(t/T)^3$) for $T>t$.
In Fig. \[fig1\] we present $\langle\varepsilon(\alpha)\rangle$ for lower $T$, i.e. $T=t$ and $T=0.2~t$. The results of the high-$T$ extrapolation and finite size calculations performed for $N=16,18$ and 20 are compared. For most points the agreement between both methods is quite satisfactory. On the other hand, there are clearly visible anomalies at $\alpha = n \pi/2$ for $N=16$ and less pronounced at $\alpha=0$ for $N=18, 20$. It is straightforward to explain these discrepancies within the high-$T$ expansion. In small systems there are some additional processes due to p.b.c., which lead to changes of $\langle\varepsilon(\alpha)\rangle$ relative to an infinite system. E.g., in the $N=4\times4$ system contributions from graphs in lowest field-dependent order $\tilde{\beta}^3$, representing four consecutive hops in the $x$- or $y$-direction cancel, except for $\alpha= n
\pi/2$, changing the cumulant $$\Delta\lambda_4=1,$$ as seen in Fig. \[fig1\].
It should be observed that for $N=16$ this correction to $\langle
\varepsilon\rangle$ is within the leading order of $\tilde\beta^3$, while analogous corrections in larger systems, e.g. for $N=18$, emerge only in higher orders. This confirms that on small systems the calculation of $B$-induced diamagnetic currents is more delicate than the evaluation of most of the static polaron properties. Nevertheless, at $J=0$ finite size effects are rather well under control, at least in comparison to the AFM case $J>0$ presented in Section III.
Consistency of high-$T$ expansion and small system ($N>16$) results allows reliable extrapolation of the susceptibility $\chi$ to quite low $T\sim 0.1~t$, using the procedure via the density of states $\rho(\varepsilon)$, Eqs.(\[dens\]-\[chit\]). The result is presented in Fig. \[fig2\], and as expected $\chi$ is diamagnetic. While for $T\gg t$ the susceptibility $\chi$ is proportional to $\beta^3$, the variation is less steep for $0.1~t<T<t$, where the variation is closer to $\chi\propto\beta^{\eta}$ with $\eta<1$.
It is quite delicate to approach $T=0$ within the $J=0$ model. The ground state for a single hole is within the sector of maximal total spin $S=S_{\rm max}$ [@naga], however excited states are numerous and close in energy, so that the transition between the regime of an incoherent hole propagation and the regime of the large FM polaron appears to happen at surprisingly low $T^*/t \sim 0.1$. This is consistent with the well established fact that the FM-polarized ground state is very sensitive to any change of parameters. The behavior of a single hole is simple only strictly at $T=0$. It is expected that the QP at $B=0$ behaves according to the Nagaoka theorem, i.e. as a free hole with energy $\varepsilon=-4~t$ in a filled band of spinless fermions. Looking only in the sector $S=S_{\rm max}$, a finite field ($B>0$) should increase the ground state energy according to the cyclotron frequency, i.e. $\delta
\varepsilon\sim eB/m^*=Bt$.
In Fig. \[fig3\] we show both the absolute g.s. energy $\varepsilon$ and the lowest energy in the sector $S=S_{\rm max}$ as calculated within small systems with $N=16,18$ and 20. For the Nagaoka sector $S=S_{\rm max}$ the behavior is for $\alpha < \pi/4$ clearly of the cyclotron type, while for higher $\alpha$ there are some visible commensurability anomalies identical to the study of spinless fermions in a magnetic field [@kohm]. On the other hand, deviations of the absolute ground state from the naive result are much more pronounced. First, even the smallest $B=B_0$ leads to the instability of the $S=S_{\rm max}$ state, and the actual spin of the g.s. is $S<S_{\rm max}$. Nevertheless, $\delta\varepsilon$ remains quite close to the cyclotron value. For higher $B>B_0$ the ground state saturates quite abruptly to lowest spin $S=1/2$ and g.s. energy $\varepsilon$ is much lower than in the $S=S_{\rm max}$ case. Note, however, that even an approximate validity of the simple cyclotron-frequency argument, $\delta\varepsilon\propto|B|$, implies a divergent susceptibility $\chi(T$$\to$$0)\to\infty$, as found also from the high-$T$ expansion.
$J>0$: Antiferromagnetic spin background
========================================
$J>0$ introduces quite a different QP behavior. At $T=0$ the spin background corresponds to an AFM with long range order. The ground state of the spin polaron is well understood, and corresponds within the $t$-$J$ model to $S=1/2$ and $\vc{k}=\vc{k}^*$, with a weak dispersion along the AFM zone boundary.
At least for weak magnetic fields it is expected that the QP description of the hole is still valid, resulting in a cyclotron motion with the linear-in-field dependence of the QP energy. Such behavior, is questioned by recent small-system diagonalisation results [@bera]. It should be reminded that results obtained from the exact diagonalisation can be misleading due to finite size effects, if the cyclotron radius is comparable to the linear size of the system. With p.b.c. even more pronounced effects compared to those already discussed for $J=0$ in Sec. II can be expected. For the AFM case they arise from larger ground state degeneracy, and from the near-degeneracy of the QP dispersion along the AFM zone boundary.
Cumulant expansion method
-------------------------
At $T$=$0$ we can consider a QP moving in an ordered AFM by performing the expansion starting in the limit $t/J \ll 1$. Here the standard cumulant-expansion (CE) procedure for the ground-state energy, as first considered by C. Bloch [@bloc], is followed.
The implementation of the method in the zero-field case with a single hole in the $t$-$J$ model has been already given in some detail elsewhere [@prel]. There the diagonal part of the Heisenberg spin interaction in Eq. (1) has been taken as the unperturbed Hamiltonian $H_0$ and the ground-state energy $\varepsilon_0(\vc{k})$, expressed relative to the undoped AFM ground state, has been obtained as a double power series in $u=t/J$ and $\gamma = J_\perp /J$ $$\varepsilon_0(\vc{k})=J\sum_{n,m}a_{n,m}(\vc{k})u^n({\gamma\over
2})^m\,,$$ where the $\vc{k}$-dependence of the expansion coefficients merely reflects the underlying translational invariance of the problem.
A finite magnetic field breaks the translational invariance of the system, so the above result, obtained from the nondegenerate theory is not valid. Indeed, the unperturbed ground state becomes highly degenerate, allowing for any hole position in an otherwise perfectly Néel-ordered spin state, thus the degenerate perturbation theory must be used. The version of the secular determinant as given in [@mess] is used, which up to the fourth order in perturbation $H'$ reads $$\begin{aligned}
(E&-&E_0)\delta_{ij}\nonumber\\&=&<i|H'{Q_0\over \eta_0}H'|j>+
<i|H'{Q_0\over\eta_0}H'{Q_0\over\eta_0}H'{Q_0\over \eta_0}H'|j>\nonumber\\
&-&\sum_\ell<i|H'{Q_0\over\eta_0}H'|\ell><\ell |H'{Q_0\over
\eta_0^2}H'|j>\nonumber\\
&+&\cdots
\label{dpex}\end{aligned}$$ The (bra)kets denote the set of degenerate states – positions of the hole – with the unperturbed energy $E_0$, whereas $P_0$ and $Q_0$ are the projectors onto this subset of states and its complement, respectively. $H'$ is taken to be the sum of the hopping part and the transverse spin part of $H$ in (1), while $\eta_0$ stands for the energy denominator $E_0-H_0$. A contribution to the matrix element $M(i,j)=<i|\cdots |j>$ resulting from a particular process $\cal
O_\mu$ involving products of operators as in (\[dpex\]) is calculated from the associated graph [@prel; @bloc] in any order of perturbation expansion. However, as the hole moves from some initial position $i$ to a final but equivalent position $j$ under the action of a particular string of operators $\cal O_\mu$ along some path, it acquires a definite phase factor $\theta_{ij}$, depending on the path itself. Two such paths are depicted in Fig. \[fig4\].
In the chosen Landau gauge the phase associated with any link in the $y$-direction is given by the $x$-coordinate of that link. Thus, the total phase $\theta_{ij}$ acquired by the hole along the path $\cal
C$, is $$\theta_{ij}({\cal C})=\int_{\cal C}\vc{A}\cdot {\rm d}\vc{s}
= i_x(j_y-i_y)\alpha + \phi_{ij}({\cal C}).$$ Here $\phi_{ij}({\cal C})$ is the phase relative to the initial point at $i$. Care should be taken of the proper orientation in which the link is traversed. Thus, referring to Fig. \[fig4\], phase along ${\cal C}_\mu$ is $\phi= -2\alpha$, whereas along the path ${\cal
A}_\mu$ $\phi= 4\alpha$.
\#1[[\#1]{}]{}
(60,60)(0,-5) (10,5)(10,0)[5]{}[(0,1)[50]{}]{} (5,10)(0,10)[5]{}[(1,0)[50]{}]{} (40,20)[(0,1)[30]{}]{} (20,20)[(1,0)[20]{}]{} (20,20)[(0,1)[10]{}]{} (20,30)[(-1,0)[10]{}]{} (10,30)[(0,1)[10]{}]{} (10,40)[(1,0)[20]{}]{} (30,40)[(0,1)[10]{}]{} (30,50)[(1,0)[10]{}]{} (20,20) (40,40) (40,25)[(0,0)]{} (40,35)[(0,0)]{} (20,25)[(0,0)]{} (40,45)[(0,0)]{} (10,35)[(0,0)]{} (30,45)[(0,0)]{} (18,17)[(0,0)\[lb\]]{} (38,37)[(0,0)\[lb\]]{} (9,35)[(0,0)\[r\]]{} (18.5,25)[(0,0)\[r\]]{} (29,45)[(0,0)\[r\]]{} (42.5,25)[(0,0)\[r\]]{} (42.5,35)[(0,0)\[r\]]{} (44,45)[(0,0)\[r\]]{} (24.5,43.5)[(0,0)\[t\]]{} (34.5,23.5)[(0,0)\[t\]]{} (10,1.5)[(0,0)\[b\]]{} (20,1.5)[(0,0)\[b\]]{} (30,1.5)[(0,0)\[b\]]{} (40,1.5)[(0,0)\[b\]]{} (50,1.5)[(0,0)\[b\]]{} (30,-4)[(0,0)\[b\]]{}
We have generated all the paths in order $u^n\gamma^m$ with $(n,m)$ up to $(2,4)$, $(4,4)$, $(6,3)$ and $(8,2)$. Each contribution to the perturbation series in order $(n,m)$ is then given by the magnitude $\omega(i,j)$ of the matrix element $\langle i|{\cal O}_\mu^{(r)}|
j\rangle$ and the phase $\theta_\mu(i,j)$, where ${\cal O}_\mu^{(r)}$ refers to a definite product of operators of order $r=n+m$ along the path from $i\to j$ [@prel].
The secular equation (\[dpex\]) becomes a difference equation for the on-site amplitudes $f_i$ on a rectangular grid with $N=L_x\times L_y$ sites $$\begin{aligned}
\varepsilon f_i&=&\sum_{j} M(i,j) f_j,\nonumber\\
M(i,j)&=&\sum_{n,m}u^n\gamma^m \sum_{{\cal C}_\mu}\omega_\mu(i,j )
e^{\imath\theta_\mu(i,j)},\label{sr2d}\nonumber\\\end{aligned}$$ where $i$ and $j$ run only over one sublattice, $M(i,j)$ is a sum of contributions along different paths ${\cal C}_\mu$, and the energy $\varepsilon\equiv\varepsilon(\alpha)$ is again measured with respect to the g.s. of the undoped AFM state. Referring again to Fig. \[fig4\] the paths ${\cal C}_\mu$ and ${\cal A}_\mu$ would then first appear in order $u^8\gamma^4$ and $u^4\gamma^2$, respectively.
In the chosen gauge Eq. (\[theta\]), the system is translationally invariant along the $y$-direction. Thus, the ansatz $f_i=g_{i_x}\exp(\imath k_y i_y)$ reduces the above equation to a difference equation in one dimension, where $\ell=i_x$, $$\begin{aligned}
\varepsilon g_\ell=&&\sum_{\ell'} {\widetilde
M}(\ell,\ell')g_{\ell'},\qquad 1\le \ell,\ell'\le
L_x,\nonumber\\
{\widetilde M}(\ell,\ell')={\sum_\tau}'e^{\imath\tau
Q}&&\sum_{n,m}u^{n}\gamma^m
\sum_{{\cal
C}_\mu}\omega_\mu(\ell'-\ell,\tau)e^{\imath\phi_\mu(\ell'-\ell,\tau)}
\nonumber\\ Q=&&k_y+\alpha \ell,\label{eige}\end{aligned}$$ and the summation over $\tau$ is restricted to run over values for which $\ell'-\ell+\tau$ is even, whereas $k_y\in
\lbrack0,2\pi\rbrack$. Note also that $\omega$ and $\phi$ do not depend on the initial point $\ell$, but only on the path $\cal
C$. Taking $\alpha=2\pi p/L_x$, one can impose the p.b.c. also in the $x$ direction.
The eigenvalue problem of Eq.(\[eige\]) is solved numerically for $L_x\gg\xi$, where $\xi=13$ is the smallest linear size of the region visited by the hole, within the order of perturbation series here considered. In Fig. \[fig5\] we plot the g.s. energy $\varepsilon$ as a function of $\alpha$ at $J/t=2$ and the isotropic exchange $\gamma=1$, evaluated for $L_x=128$. A linear-in-field dependence of $\delta\varepsilon(\alpha)=\varepsilon(\alpha)-\varepsilon(0)$ in Fig. \[fig5\] is evident for small $\alpha$, implying that the hole may still be described as a QP exhibiting cyclotron motion. However, after the initial rise an almost monotonic decrease is observed and the minimum of $\varepsilon(\alpha)$ is achieved for $\alpha=\pi$. The figure includes also the respective data from exact diagonalization for $N=20$. Although both sets of data do not agree in detail, which can be partially attributed to the perturbational character of the CE results, the overall behavior is remarkably similar, including the initial rise in $\delta\varepsilon(\alpha)$ for $N=20$ and the cusp-like behavior close to $\alpha=\pi/2$. This holds to a lesser degree for other commensurability points, e.g. for $\alpha=\pi/4$ and $3\pi/4$.
The experimentally relevant value of $J/t$ is $\sim 0.3$. Technically, it is possible to perform a Pad[' e]{}-like extrapolation of $\varepsilon(\alpha)$ to larger values of $u$, in analogy to the $B=0$ case [@prel]. Still we do not attempt to perform such an extrapolation here. Relying on previous experience [@prel] that no crossover in the QP behavior sets in down to $J/t
\ll 1$ (where Nagaoka regime takes over) we believe that the qualitative conclusions on QP behavior for finite $B$ remain valid in the physically relevant regime $J<t$ as well.
Small-system results
--------------------
Using the numerical finite temperature technique described in Sec. IIB and applied to the $J=0$ case the results for $J>0$ are also obtained. In the following we choose $J/t=0.4$, which is close to the situation in cuprates, and consider the contribution of the single hole energy $\varepsilon=E(N_h$=$1)- E(N_h$=$0)$ to the internal energy at $T=0$ and $T>0$ ($\langle\varepsilon\rangle$). First note that for $T\gg {\rm max}(t,J)$, the leading order of the high-$T$ expansion is independent of $J$ and in this limit results for $J=0$ and $J>0$ match.
In Fig. \[fig7\] the intermediate temperature hole energies $\langle\varepsilon(\alpha)\rangle$ are evaluated at two different $T$ within different system sizes. Comparing results with those for $J=0$ in Fig. \[fig1\], several conclusions can be reached. At higher $T/t=1$ energies $\langle\varepsilon(\alpha)\rangle$ for $J=0$ and $J/t=0.4$ are qualitatively similar. As a function of $\alpha$ both cases correspond approximately to the simple $\cos
\alpha$ variation. It is however evident that finite $J$ considerably reduces (by a factor $\sim 4$) the total energy span.
Relative to the $J=0$ case it is also clear that finite size effects become more pronounced when approaching the low-$T$ regime. This is related to the near-degeneracy of the quasiparticle dispersion $\varepsilon_0({\vc k})$ at the bottom of the QP band, which is poorly reproduced in small systems. Nevertheless, numerical results establish quite consistently the crossover at $T\sim T^*\propto J$.
For $T<T^*$ energy $\langle\varepsilon(\alpha)\rangle$ shows an overall opposite, i.e. a paramagnetic-like variation with $\alpha$, as found in [@bera] and reproduced above within the CE with respect to $t/J$. The results in this temperature region are more size dependent. At $T/t=0.2$ in Fig. \[fig7\] we can find a maximum for $\alpha>0$ only in systems with $N=18,20$ sites, while the $N=16$ case shows a different behavior. The deviation within the latter system is due to additional symmetry of the $4\times4$ system (hypercube).
Quite similar results are obtained for the ground state energy $\varepsilon$, shown in Fig. \[fig8\]. Again, $N=18$ and $N=20$ systems yield very similar $\varepsilon(\alpha)$, while $N=16$ results deviate especially for ‘commensurate’ values $\alpha=0,\pi/2, \pi$. The qualitative trend of $\varepsilon(\alpha)$ agrees well with the CE results on Fig. 5. The similarity is in the existence of (rather shallow) maximum at low $\alpha$ for largest $N$, indicating a QP behavior with a cyclotron motion in weak $B$ and also a pronounced reduction of $\varepsilon$ for $\alpha >\pi/2$. The only difference is in the absence of commensurability dips in Fig. 8, which we attribute to the very particular system shapes used in the calculation, and to the almost degenerate dispersion along the magnetic Brillouin zone boundary.
Finally, let us present in Fig. \[fig9\] results for the orbital susceptibility $\chi(T)$, Eq.(\[defchi\]). For $J=0$ the calculation is performed via high-$T$ expansion as already discussed in Sec. 2. Since we rely only on discrete values of $\alpha$, the analysis of finite-system data for $J/t=0.4$ is on the other hand less reliable.
Moreover, the variation of $F(\alpha)$ is quite subtle at smallest $\alpha_{\rm min}$, as evidenced also from $\varepsilon$ in Fig. \[fig8\]. In Fig. \[fig9\] the susceptibility $\chi$ is presented as obtained from Eq.(\[defchi\]), with parabolic fit for $F(\alpha)$ using only $\alpha=0$ and $\alpha=\alpha_{\rm min}$. This procedure could be questionable for $T<T^*$, but qualitative behavior is still quite instructive. Relative to the $J=0$ case, the diamagnetism is suppressed by $J>0$ at higher $T\gg J$. In an intermediate regime $T\sim J$, susceptibility $\chi$ appears even to change sign, i.e. becomes paramagnetic. Only at low $T<T^*$ a pronounced strongly $T$-dependent diamagnetic response is again observed, consistent with the QP cyclotron motion at $T=0$. It is an interesting observation that an exact solution of the problem on a single plaquette (at $T>0$) subjected to an effective staggered AFM field, reproduces qualitative features of Fig. \[fig9\].
Conclusions
===========
Our study shows that the calculation of the effects of finite magnetic field, coupled to orbital motion of electrons, becomes delicate in models of correlated electrons, e.g. within the $t$-$J$ model considered in this work. Results for magnetic observables such as the diamagnetic susceptibility appear to be strongly influenced by finite-size effects, which are hard to overcome in available system sizes. Both for $J=0$ and $J>0$ some of deviations appear at ‘commensurate’ values of $\alpha$ within the given system geometry, and are particularly large for $N=16$. We have shown that it is possible to understand such finite-size effects within the high-$T$ expansion or within the $t/J$ perturbation expansion as a contribution of additional graphs due to p.b.c. Still it is impossible to eliminate them systematically in most interesting physical regimes. These effects lead to a nonmonotonous variation of observables, e.g. $\langle\varepsilon(\alpha)\rangle$, which in turn leads to an enhanced uncertainty in $\chi(T)$.
The $J=0$ case seems both easier to study and to understand. High-$T$ expansion and small systems show a continuous transition from the high-$T$ regime of incoherent hopping to the Nagaoka state at $T=0$, with a monotonous increase of the diamagnetic $\chi$. At $T>0$ where the HTE is reliable the variation of the energy with field $\langle\varepsilon(\alpha)\rangle$ is quite close to a simple $\cos\alpha$ form. Nevertheless the asymptotic behavior at low $T$ is not simple to establish, since the nature of low lying states (above the Nagaoka state) is complicated.
The behavior of the AFM $J>0$ polaron is more involved. While at $T\gg
J$ the exchange scale $J$ is not important and results qualitatively follow those for $J=0$, new physics emerges for $T\lesssim J$. A nearly flat $\langle\varepsilon(\alpha)\rangle$ at intermediate regime $T\sim J$ is quite remarkable, and leads to a vanishing diamagnetic $\chi$ (or even change of its sign). It seems, that here $J>0$ diminishes and even destroys emerging coherence of QP. Only at lower $T<J$ the coherence is established and the known dynamical picture of a coherent AFM polaron is dominating the behavior in lowest fields $\alpha$. Reliable results are however difficult to obtain even for $T=0$, since in small systems their variation with $\alpha$ is very sensitive to the system shape, boundary conditions etc. due to the very anisotropic and degenerate QP dispersion.
The CE results are very instructive, but it is not straightforward to make them quantitative for $J<t$. For $J>t$, where the CE series converges quite rapidly, there is a clear evidence in weak fields that the QP is exhibiting cyclotron motion. The commensurability effects are quite pronounced and agree with finite-cluster data for large $J/t>1$, see Fig. \[fig5\]. On the other hand, these effects are not evident in small-system results for $J/t<1$. This non-agreement is attributed in part to specific lattice shapes and p.b.c. The flatness of the dispersion seems to be the main reason for the structureless character of small-system data in the region of small $\alpha$ for $J/t<1$.
Here we would like to point to the close similarity with the challenging theoretical problem of the Hall effect, which is difficult to approach even for very low doping, e.g. for a single hole [@prel1]. Similarly to the orbital susceptibility the Hall effect emerges due to coupling to orbital currents. The Hall constant is given by $R_H=\sigma_{xy}/B\sigma_{xx}^2$, where the off-diagonal conductivity $\sigma_{xy}$ can be related to the orbital susceptibility $\chi$ as [@rojo] $$\sigma_{xy}=B\frac{\partial\chi}{\partial c_h}
\frac{\partial c_h}{\partial \mu},$$ where $\mu$ is the chemical potential. In our case of very low hole doping, i.e. in the semiconductor-like regime, the susceptibility $\chi$ scales linearly with $c_h$; moreover $\partial
c_h/\partial\mu=-\beta c_h$, so that $\sigma_{xy}\propto\chi$. The high-$T$ expansion of the Hall constant $R_H^*(T)$ (the high-frequency value) is analogous to that of $\chi(T)$ [@brin1]. On the other hand, crossing the scale $T\sim J$ remains the challenge, whereby it seems that at this intermediate $T$ the hole-like $R^*_H$ is even reduced with respect to its high-$T$ value [@brin1; @assa]. Experiments [@ong] indicate, that $R_H$ recovers for $T<T^*$, varying strongly with $T$ and approaches the well known quasiclassical result for $T\to 0$ [@prel1]. Note that our results for $\chi(T)$, Fig. \[fig9\], indicate just on such behavior.
Let us finally comment on the magnitude of the diamagnetic susceptibility. Since we are evaluating the case of a single hole, at low doping $c_h\ll 1$ the observable diamagnetic contribution to the susceptibility (per unit cell) should be $\chi=\zeta c_h\chi_0$, where $\zeta$ is dimensionless value, presented in Figs. \[fig2\] and \[fig9\]. It is convenient to compare these values to the spin susceptibility (per unit cell) of the planar undoped AFM for $T<T^*$, where $\chi_s\sim4.0\mu_0\mu_B/J$ [@jakl2; @sing]. Setting $m_t=\hbar^2/2ta_0^2$ the ratio can be expressed as $$\frac{\chi}{\chi_s}=K\zeta c_h\frac{J}{t}\left(\frac{m_e}{m_t}\right)^2,
\label{rat1}$$ where $K\sim 2.8$ is a numerical constant. Taking the standard values for cuprates $t=0.4$ eV, $J/t=0.3$ and $a_0=0.38$ nm, the above relation reduces to $\chi/\chi_s\sim1.9\,c_h\,\zeta$.
Another relation can be obtained for the Pauli susceptibility of a half-filled band of free tight-binding electrons, where a constant (average) density of states is assumed for simplicity. This gives a similar relation $$\frac{\chi}{\chi_P}=K'\zeta c_h\left(\frac{m_e}{m_t}\right)^2
\sim9.3\,c_h\,\zeta\,,
\label{rat2}$$ with $K'\sim 4.0$. To estimate the actual value of $\chi/\chi_s$ in Eq. (27) we take $c_h\sim 0.15$, e.g. as in the ’optimal’ doping regime, and $\zeta\sim-0.1$ in the region $T<T^*$. This gives $\chi/\chi_s\sim-0.03$, which is of the same order of magnitude as the experimentally measured value [@wals]. Note, however, that below the crossover temperature $T^*$ $\zeta$ becomes strongly temperature dependent, as opposed to the usual $T$-independent Landau-type diamagnetism in Fermi liquids. Since it is difficult to distinguish different contributions to the actual susceptibility in experiments, it remains to be seen whether such $T$-dependent $\chi$ really appears in cuprates and analogous systems.
One of the authors (P.P.) wishes to thank X. Zotos for helpful suggestions concerning the introduction of a magnetic field in small systems.
This work was supported by Ministry of Science and Technology of Slovenia under Project No. J1-6166-0106/97.
\* E-mail: `[email protected]`
For a review see e.g. N. P. Ong, in [*Physical Properties of High Temperature Superconductors*]{}, ed. by D. M. Ginsberg (World Scientific, Singapore, 1990), Vol. 2, p. 459. W. Brinkman and T. M. Rice, Phys. Rev. B [**4**]{}, 1566 (1971); B. S. Shastry, B. I. Shraiman, and R. R. P. Singh, Phys. Rev. Lett. [**70**]{}, 2004 (1993). F. F. Assaad and M. Imada, Phys. Rev. Lett. [**74**]{}, 3868 (1995). R. E. Walstedt, R. F. Bell, L. F. Schneemeyer, J. V. Waszczak, and G. P. Espinosa, Phys. Rev. B [**45**]{}, 8074 (1992); M. Miljak, V. Zlati' c, I. Kos, J. D. Thompson, P. C. Canfield, and Z. Fisk, Sol. St. Commun., [**85**]{}, 519 (1993). A. G. Rojo, G. Kotliar, and G. S. Canright, Phys. Rev. B [**14**]{}, 9140 (1993). D. R. Hofstadter, Phys. Rev. B [**14**]{}, 2239 (1976). Y. Hasegawa, P. Lederer, T. M. Rice, and P. B. Wiegmann, Phys. Rev. Lett. [**63**]{}, 907 (1989). F. Nori and Y.-L. Lin, Phys. Rev. B [**49**]{}, 4131 (1994); Y.-L. Lin and F. Nori, Phys. Rev. B [**53**]{}, 13374 (1996). P. B' eran, Phys. Rev. B [**54**]{}, 1391 (1996). P. B' eran, D. Poilblanc, and R. B. Laughlin, Nucl. Phys. B [**B473**]{}, 707 (1996). T. M. Rice, in [*Proceedings of the Les Houches Summer School, Session LVI*]{}, ed. by B. Doucot and J. Zinn-Justin (Elsevier, Amsterdam, 1995), p. 19. S. Schmitt-Rink, C. Varma, and A. Ruckenstein, Phys. Rev. Lett. [**60**]{}, 2783 (1988); C. L. Kane, P. A. Lee, and N. Read, Phys. Rev. B [**39**]{}, 6880 (1989). G. Martinez and P. Horsch, Phys. Rev. B [**44**]{}, 317 (1991). P. Prelovšek, I. Sega, and J. Bonča, Phys. Rev. B [**42**]{}, 10706 (1990). For a review, see E. Dagotto, Rev. Mod. Phys. [**66**]{}, 763 (1994). B. O. Wells [*et al.*]{}, Phys. Rev. Lett. [**74**]{}, 964 (1995). M. S. Hybertsen [*et al.*]{}, Phys. Rev. B [**41**]{}, 11068 (1990); T. Tohyama and S. Maekawa, Phys. Rev. B [**49**]{},3596 (1994); A. Nazarenko, K. J. E. Vos, S. Haas, E. Dagotto, and R. J. Gooding, Phys. Rev. B [**51**]{}, 8676 (1995). Y. Nagaoka, Phys. Rev. [**147**]{}, 392 (1966). W. Brinkman and T. M. Rice, Phys. Rev. B [**2**]{}, 6880 (1970). J. Jaklič and P. Prelovšek, Phys. Rev. B [**49**]{}, 5065 (1994); Phys. Rev. Lett. [**74**]{}, 3411 (1995); [**75**]{}, 1340 (1995); Phys. Rev. B [**52**]{}, 6903 (1995). J. Jaklič and P. Prelovšek, Phys. Rev. Lett. [**77**]{}, 892 (1996). J. Oitmaa and D. D. Betts, Can. J. Phys. [**56**]{}, 897 (1978). E. Fradkin, [*Field Theories of Condensed Matter Systems*]{}, (Addison-Wesley, Redwood City, 1991), Frontiers in Physics Vol. 82, p. 252. M. Kohmoto, Phys. Rev. B [**39**]{}, 11943 (1989);\
Y. Hasegawa, Y. Hatsugai, M. Kohmoto, and G. Montambaux, Phys. Rev. B [**41**]{}, 9174 (1990). C. Bloch, Nucl. Phys. [**6**]{}, 329 (1958). A. Messiah, [*Quantum Mechanics*]{} (Nord-Holland, Amsterdam, 1962), Vol. 2, p. 717. Note that within the CE approach the g.s. is AFM-like to any finite order of the perturbation series. The hole can thus visit sites of one sublattice only. P. Prelovšek, Phys. Rev. B [**55**]{}, 9219 (1997). R. R. P. Singh and R. L. Glenister, Phys. Rev. B [**46**]{}, 11871 (1992).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider cross-layer design of delay optimal transmission strategies for energy harvesting transmitters where the data and energy arrival processes are stochastic. Using Markov decision theory, we show that the value function is weakly increasing in the queue state and weakly decreasing in the battery state. It is natural to expect that the delay optimal policy should be weakly increasing in the queue and battery states. We show via counterexamples that this is not the case. In fact, the delay optimal policy may perform 8–17% better than the best *monotone* policy.'
author:
- 'Borna Sayedana and Aditya Mahajan[^1]'
bibliography:
- 'IEEEabrv.bib'
- 'mybibfile.bib'
title: Counterexamples on the monotonicity of delay optimal strategies for energy harvesting transmitters
---
Energy harvesting transmitters, Markov decision processes, monotone policy, power-delay trade-off.
Introduction
============
Latency is an important consideration in many Internet of Things (IoT) applications which provide real-time and/or critical services. Often IoT devices are battery powered and harvest energy from the environment. In such situations, intelligent transmission strategies are needed to mitigate the unreliability of available energy and provide low-latency services.
In this paper, we investigate the cross-layer design of delay optimal transmission strategies for energy harvesting transmitters when both the data arrival and the energy arrival processes are stochastic. Our motivation is to characterize qualitative properties of optimal transmission policies for such model. For example, in queuing theory, it is often possible to establish that the optimal policy is monotone increasing in the queue length [@stidham1989monotonic; @gallisch1979monotone]. Such a property, in addition to being intuitively satisfying, simplifies the search and implementation of the optimal strategies. Such monotonicity properties are also known to hold for cross-layer design of communication systems when a constant source of energy is available at the transmitter [@berry2000power]. So it is natural to ask if such qualitative properties hold for energy harvesting transmitters.
Partial answers to this question for throughput optimal policies for energy harvesting transmitters are provided in [@zafer2008optimal; @ahmed2016optimal; @sinha2012optimal; @mao2014joint; @kashef2012optimal; @shaviv2018online]. Under the assumption of backlogged traffic or the assumption of a deterministic data arrival process or the assumption of deterministic energy arrival process, these papers show that the optimal policy is weakly increasing in the queue state and/or weakly increasing in the battery state. There are other papers that investigate the structure of delay or throughput optimal policies under the assumption of a deterministic energy arrival process [@yang2012optimal; @tutuncuoglu2012optimum; @ozel2011transmission].
There are some papers which investigate the problem of delay optimization for energy harvesting transmitters [@ozel2011transmission; @Fawaz2018; @Sharma2018], but they don’t characterize the structure of delay-optimal policies rather provide numerical solutions or propose low-complexity heuristic policies or only establish structural properties of value functions.
Our main result is to show that in contrast to the structure of the delay optimal policy when constant source of energy is available [@berry2000power] and the structure of throughput optimal policies for energy harvesting communication systems [@zafer2008optimal; @ahmed2016optimal; @sinha2012optimal; @mao2014joint; @kashef2012optimal; @shaviv2018online], the delay optimal policy for energy harvesting communication systems is not necessarily monotone in battery or queue state. We present counterexamples to show that the delay optimal policy need not be weakly increasing in queue state nor in the battery state. Furthermore, the performance of the optimal policy is about 8–17% better than the performance of the best monotone policy. These counterexamples continue to hold for i.i.d. fading channels as well.
### Notation {#notation .unnumbered}
Uppercase letters (e.g., $E$, $N$, etc.) represent random variables; the corresponding lowercase letters (e.g., $e$, $n$, etc.) represent their realizations. Cursive letters (e.g., $\mathcal L$, $\mathcal B$, etc.) represent sets. The sets of real, positive integers, and non-negative integers are denoted by ${\mathds R}$, ${\mathds Z}_{> 0}$, and ${\mathds Z}_{\ge 0}$ respectively. The notation $[a]_{L}$ is a short hand for $\min \{a,L\}$.
Model And Problem Formulation {#sec:model}
=============================
![Model of a transmitter with energy-harvester[]{data-label="fig:model"}](model){width="\linewidth"}
Consider a communication system shown in Fig \[fig:model\]. A source generates bursty data packets that have to be transmitted to a receiver by an energy-harvesting transmitter. The transmitter has finite buffer where the data packets are queued and a finite capacity battery where the harvested energy is stored. The system operates in discrete time slots. The data packets and the energy that arrive during a slot are available only at the beginning of the next slot.
At the beginning of a slot, the transmitter picks some data packets from the queue, encodes them, and transmits the encoded symbol. Transmitting a symbol requires energy that depends on the number of encoded packets in the symbol. At the end of the slot, the system incurs a delay penalty that depends on the number of packets remaining in the queue.
Time slots are indexed by $k \in {\mathds Z}_{\geq 0}$. The length of the buffer is denoted by $L$ and the size of the battery by $B$; $\mathcal{L}$ and $\mathcal B$ denote the sets $\{0,1,\ldots,L\}$ and $\{0,1,\ldots,
B\}$, respectively. Other variables are as follows:
- $N_{k} \in \mathcal{L}$: the number of packets in the queue at the beginning of slot $k$.
- $A_{k} \in \mathcal{L}$: the number of packets that arrive during slot $k$.
- $S_{k} \in \mathcal{B}$: the energy stored in the battery at the beginning of slot $k$.
- $E_{k} \in \mathcal{B}$: the energy that is harvested during slot $k$.
- $U_{k}$: the number of packets transmitted during slot $k$. The feasible choices of $U_{k}$ are denoted by $\mathcal{U}(N_k,
S_k)$ where $$\mathcal{U}(n,s) \coloneqq \{ u \in \mathcal{L} :
u \le n \text{ and } p(u) \le s \},$$ where $p(u)$ denotes the amount of power needed to transmit $u$ packets. In our examples, we model the channel as a band-limited AWGN channel with bandwidth $W$ and noise level $N_0$. The capacity of such a channel when transmitting at power level $P$ is $W \log_2( 1 + P/(N_0W))$. Therefore, for such channels we assume $p(u) = \lfloor N_0 W (2^{u/W} -
1) \rfloor$. In general, we assume that $p:\mathcal{L} \rightarrow \mathds{R}_{\geq 0}$ is a strictly convex and increasing function with $p(0) = 0$.
The dynamics of the data queue and the battery are $$N_{k+1} = [N_{k} - U_{k} + A_{k}]_{L}
\quad\text{and}\quad
S_{k+1} = [S_{k} - p(U_{k})+E_{k}]_{B}.$$
Packets that are not transmitted during slot $k$ incur a delay penalty $d(N_{k} - U_{k})$, where $d:\mathcal{L} \rightarrow \mathds{R}_{\geq 0}$ is a convex and increasing function with $d(0) = 0$.
The data arrival process $\{A_{k}\}_{k\geq 0}$ is assumed to be an independent and identically distributed process with pmf (probability mass function) $P_A$. The energy arrival process $\{E_{k}\}_{k \geq 0}$ is an independent process that is also independent of $\{A_{k}\}_{k \geq 0 }$ with pmf $P_E$.
The number $U_{k}$ of packets to transmit are chosen according to a scheduling policy $f \coloneqq \{f_{k}\}_{k \geq 0}$, where $$U_{k} = f_{k}(N_{k},S_{k}), \quad U_k \in \mathcal{U}(N_k, S_k).$$
The performance of a scheduling policy $f$ is given by $$\label{eq:cost}
J(f) \coloneqq {\mathds{E}}^{f} \Big[ \sum_{k = 0}^{\infty} \beta^{k} d(N_{k}-U_{k})
\Bigm| N_{0} = 0,S_{0} =0 \Big],$$ where $\beta \in (0,1)$ denotes the discount factor and the expectation is taken with respect to the joint measure on the system variables induced by the choice of $f$.
We are interested in the following optimization problem.
\[prob:main\] Given the buffer length $L$, battery size $B$, power cost $p(\cdot)$, delay cost $d(\cdot)$, pmf $P_A$ of the arrival process, pmf $P_E$ of the energy arrival process, and the discount factor $\beta$, choose a feasible scheduling policy $f$ to minimize the performance $J(f)$ given by .
Dynamic Programming Decomposition {#sec:DP}
=================================
The system described above can be modeled as an infinite horizon time homogeneous Markov decision process (MDP) [@puterman2014markov]. Since the state and action spaces are finite, standard results from Markov decision theory imply that there exists an optimal policy which is time homogeneous and is given by the solution of a dynamic program. To succinctly write the dynamic program, we define the following Bellman operator: Define the operator $\mathscr B : [\mathcal{L} \times \mathcal{B} \to
{\mathds R}] \to [\mathcal{L} \times \mathcal{B} \to {\mathds R}]$ that maps any $V: \mathcal{L} \times \mathcal{B} \rightarrow \mathds{R}$ to $$\begin{gathered}
\label{DyP}
\big[\mathscr{B} V\big](n,s) =
\min_{u \in \mathcal{U}(n,s)}\Big\{ d(n-u) \\+
\beta{\mathds{E}}\big[ V([n - u + A]_{L}, [s - p(u) + E]_{B}) \big]
\Big\},\end{gathered}$$ where $A$ and $E$ are independent random variables with pmfs $P_A$ and $P_E$. Then, an optimal policy for the infinite horizon MDP is given as follows [@puterman2014markov].
Let $V^*: \mathcal {L}\times \mathcal {B} \to {\mathds R}$ denote the unique fixed point of the following equation: $$\label{eq:bellman}
V(n,s) = [\mathscr{B} V](n,s),
\quad \forall (n,s) \in \mathcal L \times \mathcal B.$$ Furthermore, let $f^*$ be such that $f^*(n,s)$ attains the minimum in the right hand side of . Then, the time homogeneous policy $f^{*,\infty} = (f^*, f^*, \dots)$ is optimal for Problem \[prob:main\].
The dynamic program described in can be solved using standard algoirthms such as value iteration, policy iteration, or linear programming algorithms [@puterman2014markov].
Properties of the value function
--------------------------------
Let $\mathcal {M}$ denote the family of the functions $V:{\cal{L}} \times
{\cal{B}} \rightarrow \mathds{R}$ such that for any $s \in {\mathcal{B}}$, $V(n,s)$ is weakly increasing in $n$ and for any $n \in \mathcal{L}$, $V(n,s)$ is weakly decreasing in $s$. Furthermore, let $\mathcal F_{s}$ denote the family of functions $f \colon
\mathcal L \times \mathcal B \to \mathcal U$ such that for any $n \in \mathcal{L}$, $f(n,s)$ is weakly increasing in $s$. Similarly, let $\mathcal F_{n}$ be family of functions $f \colon
\mathcal L \times \mathcal B \to \mathcal U$, such that for any $s \in \mathcal B$, $f(n,s)$ is weakly increasing in $n$.
\[thm:monotone\] The optimal value function $V^* \in \mathcal{M}$.
The proof is presented in the Appendix.
Proposition \[thm:monotone\] says that if we follow an optimal policy, the optimal cost when starting from a smaller queue state is lower than that starting from a larger queue state. Similarly, the optimal cost when starting from a larger battery state is lower than that starting from a smaller battery state. Such a result appears to be intuitively obvious.
One might argue that it should be the case that the optimal policy should be weakly increasing in state of the queue, and weakly increasing in the available energy in the battery. In particular, if it is optimal to transmit $u$ packets when the queue state is $n$, then (for the same battery state) the optimal number of packets to transmit at any queue state larger than $n$ should be at least $u$. Similarly, if it is optimal to transmit $u$ packets when the battery state is $s$, then (for the same queue state) the optimal number of packets to transmit at any battery state larger than $s$ should be at least $u$. In the next section, we present counterexamples that show both of these properties do not hold. The code for all the results is available at [@code].
Counterexamples on the monotonicity of optimal policies {#sec:counterexample}
=======================================================
On the monotonicity in queue state {#sec:queue}
----------------------------------
to
Consider the communication system with a band-limited AWGN channel where $\mathcal{L} = 5 $, $\mathcal{B} = 5 $, $\beta = 0.99$, $N_0 = 2.0$, $W = 1.75$ (thus, $p(u) = \lfloor 3.5 \cdot (2^{(u/1.75)}-1) \rfloor$), $d(q) = q$, data arrival distribution $P_A = [0.59,0.41,0,0,0]$, and energy arrival distribution $P_E =[0.385,0.23,0.385,0,0]$.
The optimal policy for this system (obtained by policy iteration [@puterman2014markov]) is shown in Fig. \[fig:ex1a\], where the rows correspond to the current queue length and the columns correspond to the current energy level. Note that the policy is not weakly increasing in queue state (i.e, $f^{*} \notin \mathcal{F}_{n}$). In particular, $f(3,4) < f(2,4)$. Given that the optimal policy is not monotone, one might wonder how much do we lose if we follow a monotone policy instead of the optimal policy. To characterize this, we define the best queue-monotone policy as: $$f^{\circ}_{n} = \arg\min_{f \in \mathcal{F}_{n}}
\left\{ \max_{(n,s) \in \mathcal L \times \mathcal B}
\bigl| V(n,s) - V^*(n,s) \bigr| \right\}$$ and let $V^{\circ}_{n}$ denote the corresponding value function.
The best monotone policy cannot be obtained using dynamic programming and one has to resort to a brute force search over all monotone policies. For the model described above, there are $86400$ monotone policies.[^2] The best monotone policy obtained by searching over these is shown is Fig. \[fig:ex1b\]. The worst case difference between the two value functions is given by $$\alpha_{n} = \max_{(n,s) \in \mathcal L \times \mathcal B}
\left\{
\frac{ V^\circ(n,s) - V^*{n}(n,s)}
{ V^*(n,s) }
\right\}
= 0.1764.$$
Thus, for this counterexample, the best queue-monotone policy performs $17.46\%$ worse than the optimal policy.
On the monotonicity in battery state {#sec:battery}
------------------------------------
Consider the communication system described in Sec. \[sec:queue\] but with the data arrival distribution $P_A=[0.33,0.67,0,0,0]$ and the energy arrival distribution $P_E = [0.033,0.934,0.033,0,0]$.
The optimal policy (obtained using policy iteration [@puterman2014markov]) is shown in Fig. \[fig:ex2a\]. Note that the policy is not weakly increasing in the battery state (i.e $f^{*} \notin \mathcal{F}_{s}$). In particular, we have that $f^{*}(5,2) > f^{*}(5,3)$.
to
Given that optimal policy is not monotone, the previous question arises again that how much do we lose if we follow a monotone policy instead of the optimal policy. To characterize this, we define the best battery-monotone policy as: $$\begin{aligned}
f^{\circ}_{s} = \arg\min_{f \in \mathcal{F}_{s}}
\left\{ \max_{(n,s) \in \mathcal L \times \mathcal B}
\bigl| V(n,s)-V^*(n,s) \bigr| \right\}\end{aligned}$$ and let $V^{\circ}_{s}$ denote the corresponding value function.
As before, we find the best monotone policy by a a brute force search over all $303750$ monotone battery-policies. The resultant policy is shown in Fig. \[fig:ex2b\]. The worst case difference between the two value functions is given by $$\alpha_{s} = \max_{(n,s) \in \mathcal L \times \mathcal B}
\left\{
\frac{ V^\circ(n,s) - V^*{s}(n,s)}
{ V^*(n,s)}
\right\}
= 0.0860.$$
Thus, for this counterexample, the best battery-monotone policy performs $8.60\%$ worse than the optimal policy.
Counterexamples for fading channels {#sec:fading}
===================================
to
Channel model with i.i.d. fading
--------------------------------
Consider the model in Sec. \[sec:model\] where the channel has i.i.d. fading. In particular, let $H_k \in \mathcal H$ denote the channel state at time $k$ and $g(H_k)$, where $g : \mathcal H \to {\mathds R}_{> 0}$, denote the attenuation at state $H_k$. Thus, the power needed to transmit $u$ packets when the channel is in state $h$ is given by $p(u)/g(h)$. We assume that $\{H_k\}_{k \ge 0}$ is an i.i.d. process with pmf $P_H$ that is independent of the data and energy arrival processes $\{A_k\}_{k \ge 0}$ and $\{E_k\}_{k \ge 0}$.
On the monotonicity in queue state {#sec:fading-queue}
----------------------------------
Consider the model in Sec. \[sec:queue\] with $N_{0} = 1$, $W = 1.75$, and an i.i.d. fading channel where $\mathcal H = \{1,2,3\}$, $g(\cdot) = \{0.4, 0.7, 0.8\}$ and $P_H =
[0.15, 0.25, 0.6]$. The optimal policy for this model (obtained using policy iteration) is shown in Fig. \[fig:ex3a\]–\[fig:ex3c\]. Note that for all $h$, the optimal policy in not monotone in the queue length.
In this case, there are $(4320) \times (1296) \times (362) \approx 10^{8}$ monotone policies. Therefore, a brute force search to find the best monotone policy is not possible. We choose a heuristic monotone policy $f^\circ_n$ which differs from $f^*$ only at the following points: $f^\circ_n(1,2,1) = 0$, $f^\circ_n(4,3,1)=1$, $f^\circ_n(5,s,1)=1$, for $s \in \{3,4\}$, $f^\circ_n(2,3,2) = 1$, $f^\circ_n(5,1,2) = 1$, $f^\circ_n(3,5,2) = 2$, $f^\circ_n(3,5,2) = 2$, $f^\circ_n(5,4,2) = 2$, $f^\circ_n(5,1,3) = 1$ and $f^\circ_n(3,4,3) = 2$. The policy $f^\circ_n$ may be thought of as the queue-monotone policy that is closest to $f^*$. Let $V^\circ_n$ denote the corresponding value function. The worst case difference between the two value functions is given by $$\alpha_{n} = \max_{(n,s,h) \in \mathcal L \times \mathcal B \times \mathcal H}
\frac{\bigl| V^*(n,s,h) - V^\circ_{n}(n,s,h)\bigr|}
{\bigl| V^*(n,s,h) \bigr|}
= 0.2052.$$ Thus, the heuristically chosen queue-monotone policy performs $20.52\%$ worse than the optimal policy.
On the monotonicity in the battery state {#sec:fading-battery}
----------------------------------------
Consider the model in Sec. \[sec:battery\] with $N_{0} = 1.55$, $W = 1.75$, and an i.i.d. fading channel where $\mathcal H = \{1,2\}$, $g(\cdot) = \{0.75, 0.80\}$, and $P_H =
[0.3,0.7]$. The optimal policy for this model (obtained using policy iteration) is shown in Fig. \[fig:ex4a\]–\[fig:ex4b\]. Note that for $h
\in \{1,2\}$, the optimal policy is not monotone in the battery state.
In this case, there are $(629856) \times (30375019) \approx 10^{10}$ monotone policies. Therefore, a brute force search is not possible. As before, we choose a heuristic policy $f^\circ_s$ which is the battery-monotone policy that is closest to $f^*$. In particular, $f^\circ_s$ differs from $f^*$ only at two points: $f^\circ_s(5,3,1) = 1$ and $f^\circ_s(5,3,2) = 1$. Let $V^\circ_s$ denote the corresponding value function. The worst case difference between the two value functions is given by $$\alpha_{s} = \max_{(n,s,h) \in \mathcal L \times \mathcal B \times \mathcal H}
\frac{\bigl| V^*(n,s,h) - V^\circ_{s}(n,s,h)\bigr|}
{\bigl| V^*(n,s,h) \bigr|}
= 0.1114.$$ Thus, the heuristically chosen battery-monotone policy performs $11.14\%$ worse than the optimal policy.
Conclusion
==========
In this paper, we consider delay optimal strategies in cross layer design with energy harvesting transmitter. We show that the value function is weakly increasing in the queue state and weakly decreasing in the battery state. We show via counterexamples that the optimal policy is not monotone in queue length nor in the available energy in the battery.
Discussion about the counterexamples
------------------------------------
One might ask why the optimal policy is not monotone in the above model. The standard argument in MDPs to establish monotonicity of the optimal policies is to show that the value-action function is submodular in the state and action. The value-action function is given by $${H}(n,s,u) = d(n-u) +
\beta{\mathds{E}}\big[ V([n - u + A]_{L}, [s - p(u) + E]_{B}) \big]$$ A sufficient condition for the optimal policy to be weakly increasing in the queue length is:
1. for every $s \in \mathcal{B}$, $H(n,s,u)$ is submodular in $(n,u)$.
Note that since $d(\cdot)$ is convex, $d(n-u)$ is submodular in $(n,u)$. Thus, a sufficient condition for (S1) to hold is:
1. for all $s\in \mathcal{B}$, ${\mathds{E}}\big[ V([n - u + A]_{L}, [s -
p(u) + E]_{B}) \big]$ is submodular in $(n,u)$.
Since submodularity is preserved under addition, a sufficient condition for (S2) to hold is:
1. for all $s \in \mathcal{B}$, $V(n-u, s-p(u))$ is submodular in $(n,u)$.
By a similar argument, it can be shown that a sufficient condition for the optimal policy to be weakly increasing in battery state is:
1. for all $n \in \mathcal{L}$, $V(n-u, s-p(u))$ is submodular in $(s,u).$
We have not been able to identify sufficient conditions under which (S3) or (S4) hold. Note that if the data were backlogged, then we do not need to keep track of the queue state; thus, the value function is just a function of the battery state. In such a scenario, (S4) simplifies to $V(s-p(u))$ is submodular in $(s,u)$. Since $p(\cdot)$ is convex, it can be shown that convexity of $V(s)$ is sufficient to establish submodularity of $V(s-p(u))$. This is the essence of the argument given in [@mao2014joint; @sinha2012optimal].
Similarly, if the transmitter had a steady supply of energy, then we do not need to keep track of the battery state; thus, the value function is just a function of the queue state. In such a scenario, (S3) simplifies to $V(n-u)$ is submodular in $(n,u)$. It can be shown that convexity of the $V(n)$ is sufficient to establish submodularity of $V(n-u)$. This is the essence of the argument given in [@berry2000power].
In our model, data is not backlogged and energy is intermittent. As a result, we have two queues—the data queue and the energy queue—which have coupled dynamics. This coupling makes it difficult to identify conditions under which $V(n-u,s-p(u))$ will be submodular in $(n,u)$ or $(s,u)$.
Implication of the results
--------------------------
In general, there are two benefits if one can establish that the optimal policy is monotone. The first advantage is that monotone policies are easier to implement. In particular, one needs a $(L+1) \times (B+1)$-dimensional look-up table to implement a general transmission policy (similar to the matrices shown in Figs. \[fig:ex1\] and \[fig:ex2\]). In contrast, one only needs to store the thresholds boundaries of the decision regions (which can be stored in a sparse matrix) to implement a queue- or battery-monotone policy. Our counterexamples show that such a simpler implementation will result in a loss of optimality in energy-harvesting systems.
The second advantage is that if we know that the optimal policy is monotone, we can search for them efficiently using monotone value iteration and monotone policy iteration [@puterman2014markov]. Our counterexamples show that these more efficient algorithms cannot be used in energy-harvesting systems.
One might want to restrict to monotone policies for the sake of implementation simplicity. However, if the system does not satisfy properties (S3) and (S4) mentioned in the previous section, then dynamic programming cannot be used to find the best monotone policy. Thus, one has to resort to a brute force search, which suffers from the curse of dimensionality.
Monotonicity of Bellman operator
--------------------------------
\[lem:monotone\] Given $V \in \mathcal{M}$ and any $n \in \mathcal{L}$, $s \in \mathcal{B}$, and $u \in \mathcal{U}(n,s)$, let $$ {H}(n,s,u) = d(n-u) +
\beta{\mathds{E}}\big[ V([n - u + A]_{L}, [s - p(u) + E]_{B}) \big].$$ Define $W = \mathscr{B} V$, i.e., $$W(n,s) = \min_{u \in \mathcal{U}(n,s)} {H}(n,s,u),
\quad \forall n \in \mathcal L, s \in \mathcal B.
\label{eq:W}$$ Then, for all $n \in \mathcal{L}$, $s \in
\mathcal{B}$, and $u \in \mathcal{U}(n,s)$, we have:
1. ${H}(n,s,u) \leq {H}([n+1]_L,s,u)$.
2. ${H}(n,s,n) \leq {H}([n+1]_L,s,[n+1]_L)$.
3. ${H}(n,[s+1]_B,u) \leq {H}(n,s,u)$.
As a consequence of the above, $W \in \mathcal{M}$.
The properties of $H$ follow from the monotonicity of $d(\cdot)$ and the fact that monotonicity is preserved under expectations. The details are omitted due to lack of space. To prove that $W \in \mathcal M$, we consider any $n \in \mathcal L$ and $s
\in \mathcal B$ and let $f(n,s)$ denote an arg min of the right hand side of . Now there are two cases: $f(n+1,s) \neq n+1$ and $f(n+1,s) = n+1$.
1. Suppose $u^* = f(n+1,s) \neq n+1$. Then, it must be the case that $u^* \in \mathcal U(n,s)$. Thus, $$\begin{aligned}
W(n+1,s) &= H(n+1, s, u^*)
\stackrel{(a)}\ge H(n, s, u^*) \\
& \ge \min_{u \in \mathcal U(n,s)} H(n,s,u)
= W(n,s),
\end{aligned}$$ where $(a)$ follows from Property 1.
2. Suppose $u^* = f(n+1,s) = n+1$. Then, it must be the case that $p(n+1) \le s$ and, therefore, $p(n) \le s$. Hence $n \in \mathcal
U(n,s)$. Thus, $$\begin{aligned}
W(n+1,s) &= H(n+1,s,n+1)
\stackrel{(b)}\ge H(n,s,n) \\
& \ge \min_{u \in \mathcal U(n,s)} H(n,s,u)
= W(n,s),
\end{aligned}$$ where $(b)$ follows from Property 2.
As a result of both of these cases, we get that $$\begin{aligned}
\label{pf:1}
W(n,s) \leq W(n+1,s).
\end{aligned}$$
Now let $u^{*} = f(n,s)$, recall that $\mathcal{U} (n,s) \subseteq \mathcal{U} (n,s+1)$ then $u^{*} \in \mathcal{U}(n,s+1)$ thus $$\begin{aligned}
\label{pf:2}
W(n,s) &= H(n, s, u^*)
\stackrel{(c)}\ge H(n, s+1, u^*)\notag \\
& \stackrel{(d)}\ge \min_{u \in \mathcal U(n,s+1)} H(n,s+1,u)
= W(n,s+1),
\end{aligned}$$ Where $(c)$ follows from Property 3 and $(d)$ follows from the fact that $u^{*} \in \mathcal{U}(n,s+1)$.
From and we infer $W \in \mathcal{M}$.
Proof of Proposition \[thm:monotone\]
-------------------------------------
Arbitrarily initialize $V^{(0)} \in \mathcal M$ and for $n \in {\mathds Z}_{>0}$, recursively define $
V^{(n+1)} = \mathscr{B} V^{(n)}.
$ Since $V^{(0)} \in \mathcal M$, Lemma \[lem:monotone\] implies that $V^{(n)} \in \mathcal{M}$, for all $n \in
{\mathds Z}_{>0}$. Since monotonicity is preserved under the limit, we have that $\lim_{n \rightarrow \infty} V_{0}^{(n)} \in \mathcal{M}$. By [@puterman2014markov], $
\lim_{n \rightarrow \infty} V_{0}^{(n)} = V.
$ Hence, $V \in \mathcal{M}$.
[^1]: This work was supported in part by Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant RGPIN-2016-05165.
[^2]: \[fnt\]Due to the power constraint $\mathcal U(n,s)$, it is not possible to count the number of monotone functions using combinatorics. The number above is obtained by explicit enumeration.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Sidecoin is a mechanism that allows a snapshot to be taken of Bitcoin’s blockchain. We compile a list of Bitcoin’s unspent transaction outputs, then use these outputs and their corresponding balances to bootstrap a new blockchain. This allows the preservation of Bitcoin’s economic state in the context of a new blockchain, which may provide new features and technical innovations.'
author:
- |
Joseph Krug & Jack Peterson\
www.sidecoin.net
title: 'Sidecoin: a Snapshot Mechanism for Bootstrapping a Blockchain'
---
Snapshots and Their Uses
========================
A snapshot records all public addresses in the Bitcoin network with a substantial balance (over about 20 cents in USD) and sorts them from highest to lowest balance. This eliminates millions of addresses from places like Satoshi Dice that have small amounts of bitcoin in them, called ‘dust’. We then save these addresses, along with their balances, and an alternative representation of the address known as a hash160.[^1]
Snapshots allow us to do a few things. They can be used to examine addresses and balances to find high value holders, or to see the distribution of wealth among Bitcoiners: the economic state of the Bitcoin network. However, a much more fun use is to create altcoins – or, as coins created using this method are called, *spinoff* coins.[^2] To create a spinoff coin, we first take a snapshot of Bitcoin’s blockchain, then create an altcoin with an initial distribution based on Bitcoin’s. *Sidecoin* is the reference implementation for bootstrapping a new blockchain using Bitcoin’s present state. This technology allows bitcoin owners to trustlessly claim a proportional amount of a new cryptocurrency.
Imagine a coin development team like Gridcoin’s[^3]. Gridcoin uses a proof-of-work that, in principle, benefits society through computation, by directing it to to fold proteins for cancer research. Their team would probably be pleased if a tiny bit of the enormous mining power used to mine bitcoins could be redirected to their *useful* proof-of-work system. It provides a principled, transparent distribution method with which ‘pre-mined’ cryptocurrencies can disseminate their tokens. One example is Ripple, which does not use mining at all. Taking a snapshot, and allowing Bitcoiners to claim coins on the Gridcoin or Ripple networks, removes the profit gained by releasing new cryptocurrencies which do not contribute new technology. Coin developers can still profit by allowing mining after the initial claiming, but only by providing a true improvement to Bitcoin or other altcoins.
A snapshot-based distribution bootstraps an altcoin with a much larger user base than they otherwise would have had: any bitcoin owner is automatically ‘bought-in’ to the altcoin, provided they decide to claim their coins. Bitcoin holders who think an altcoin does not provide sufficient technological advancements over Bitcoin can simply claim their new coins, then sell them. Altcoin creators who believe that their coins hold value due to useful features are incentivized to buy more after the initial sharp selloff, because they will profit once the market realizes the value of their innovations.
Taking a Snapshot
=================
There were several steps involved in creating our snapshot file. First, we downloaded the Bitcoin blockchain using `bitcoind`, then we parsed all the balances in the Bitcoin blockchain, along with their corresponding public keys.[^4] Next, we converted these hash160s into the commonly used Bitcoin address: the base-58 strings seen everywhere, on `blockchain.info`, in your wallet, etc. We then used a shell script to manipulate the text to sort these addresses, hash160s, and their corresponding balances into a tab-delimited file. An altcoin developer could simply use this file, embed it in the first non-genesis block, then allow users to claim their coins through our coin verification process.
To create the snapshot file, we use Znort’s blockchain parser[^5], modified to use memory-lean hash tables. We then use a shell script[^6] to call the appropriate functions on the block parser. We specify the maximum block number as well as the number of addresses to include in the snapshot; a value of about 2,000,000 includes all addresses with balances[^7] above about $\$0.20$. Next, the hash160s are converted into addresses[^8], which is also written to the snapshot file. After parsing, a `snapshotToImport.txt` file is output containing the balance, hash160, and address. An example entry is:
``` {language="json"}
180893019187.00000000 8c1d15231afa4868330f8af694ba637b69fdc2d7 14CKu2rJN2f8fdPrtmRLWChhXESgN5qaA7
```
The snapshot file has as many rows as the number of addresses we specified to the blockparser.
Incorporating a Snapshot into a Blockchain’s Initial State
==========================================================
Once the snapshot file has been created, it must to be recorded on the new blockchain. We incorporate the snapshot data as a series of unspent transaction outputs into block one of the new blockchain. Then, we add a claim transaction function and make it available via RPC call. This allows bitcoin owners to trustlessly claim their coins in the spinoff blockchain. Finally, we discuss how to create your own snapshot and incorporate it in a spinoff.
First, we mine the genesis block. Next, we incorporate the snapshot data into the first block.[^9] We begin by creating coinbase transactions[^10] for each of the addresses in the snapshot:
``` {language="json"}
Input:
<pubKeyHash from snapshot>
Output:
Value: <Bitcoin address's balance>
scriptPubKey: OP_DUP OP_HASH160 <pubKeyHash> OP_EQUALVERIFY OP_CHECKSIG
```
where the Bitcoin address’s balance is the balance recorded in the snapshot. After these transactions are loaded into block one, we find the block’s hash, process the block (i.e., the same thing a miner does upon finding a new block), then submit it to the network.
Finally, we provide a trustless method of claiming these transactions. Claim transactions give our users who own bitcoins a way of claiming their portion of sidecoins. To do this, the user must submit a *claim transaction*:
``` {language="json"}
Input:
<previous transaction ID>
<index of the output being claimed> (should be 0)
scriptSig: <signature> <pubKey>
```
where `<pubKey>` is the public key of the user’s Bitcoin address. This input Script has successfully unlocked the new coins on the Sidecoin network! These coins are locked up in an address belonging to the user’s Sidecoin wallet:
``` {language="json"}
Output:
Value: <Bitcoin address's balance>
scriptPubKey: OP_DUP OP_HASH160 <Sidecoin pubKeyHash> OP_EQUALVERIFY OP_CHECKSIG
```
To facilitate submitting these claim transactions, we created a new RPC command, `claimtx`. Users call this by inputting `claimtx` followed by the user’s Bitcoin address:
``` {language="json"}
claimtx "1F1tAaz5x1HUXrCNLbtMDqcw6o5GNn4xqX"
```
The *claimtx* function (in `rpcserver.cpp`) searches through the snapshot file to find the address’s corresponding hash160 and unspent output in block one. Then, it automatically builds the claim transaction (described above), locking the coins in an output belonging to a new address in the user’s Sidecoin wallet:
``` {language="json"}
signrawtransaction
'0100000001ec70650bf05a75d62f3bcf83d183064e118d39070c2a7237d33ee9a4d4930da20000000000ffffffff01606b042a010000001976a914345cd34789f945f0cbd952ce254bf5246e63be0c88ac00000000' '[{"txid":"a20d93d4a4e93ed337722a0c07398d114e0683d183cf3b2fd6755af00b6570ec","vout":0,"scriptPubKey":"76a914f4dfe70369cc98dc8113bba69bda324e4f1386a088ac"}]'
```
Next, the user is given instructions to open the RPC console of their Bitcoin client containing the address which held the bitcoins the user is claiming at the time the snapshot was taken. The user pastes in the output supplied by the `claimtx`, which is the unsigned transaction appended to the RPC command `signrawtransaction`. The user’s bitcoin client will then sign this transaction.
Finally, the user takes the results, appends them to the command `sendrawtransaction`, and submits it using Sidecoin’s RPC console:
``` {language="json"}
sendrawtransaction "01000000017b1eabe0209b1fe794124575ef807057c77ada213\
8ae4fa8d6c4de0398a14f3f00000000494830450221008949f0\
cb400094ad2b5eb399d59d01c14d73d8fe6e96df1a7150deb38\
8ab8935022079656090d7f6bac4c9a94e0aad311a4268e082a7\
25f8aeae0573fb12ff866a5f01ffffffff01f0ca052a0100000\
01976a914cbc20a7664f2f69e5355aa427045bc15e7c6c77288\
ac00000000"
```
The user has now unlocked their sidecoins! This transaction is like any other transaction: it just spends unspent outputs. The key difference is that it involves data from two different blockchains! We do the claim transaction in this manner to trustlessly verify that the user owns the Bitcoin private key associated with the address to which the sidecoins have been assigned.
In order to create a snapshot for use on one’s own computer, one simply runs the snapshot script which asks the user a few questions about which block they’d like to take the snapshot in and how many addresses they’d like in the snapshot. After this, take the `snapshotToImport.txt` file and paste it in the *Appdata/Sidecoin/balances* folder on your respective OS. Place a `sidecoin.conf` file there, as well. Open the `snapshot.h` file in Sidecoin’s source folder and set `GENESIS_SWITCH` to true. Then, launch Sidecoin, and it will mine the mainnet, testnet, and regtest genesis blocks. (Upon completion, the block hashes and Merkle root fields in `chainparams.cpp` need to be updated.)
After the genesis block is mined, launch Sidecoin again and it will load the transactions into block one from the snapshot file, then it will add block one to the blockchain. If it outputs `accepted` everything has been done correctly! Next, change `GENESIS_SWITCH` to false. Finally, relaunch Sidecoin to see a new successful spinoff implementation.
For users who want to use an already created spinoff coin, one can simply wait to download block one from other peers. However, when the network is first starting off this is very slow and impractical due to the large size of block one. Instead, one can try to load block one in on their own by enabling the `GENESIS_SWITCH` and editing the `blockOne.nNonce` and `nTime` to the values that they were mined with if not already set. Then place the `snapshotToImport.txt` file in the *balances* directory, start Sidecoin, wait for it to load the block, then disable the `GENESIS_SWITCH`.[^11]
Proof-of-Concept
================
Sidecoin[^12] is our proof-of-concept implementation of a snapshot/spinoff mechanism. It is a fork of Bitcoin Core 0.9.1. It does not support pay-to-script-hash addresses in the snapshot, nor does it support simplified claim verification schemes, as proposed on Bitcointalk.[^13]
We made a few changes to our fork of Bitcoin Core to make Sidecoin. We had to change a couple things in the `CheckBlock` code segments: allowing multiple coinbase transactions, as well as modifying block size limits, signature operations limits, and transaction limits. We only made these modifications for block one, which is determined as the block with a previous block hash equivalent to the genesis block. We also added functions allowing Bitcoin to read in snapshot data from a flat file into block one, the `claimtx` function, and manually loading block one into the blockchain, similar to how the genesis block is built.
Finally, we recognize that most of the utility of snapshots and spinoffs will ultimately be replaced by the interoperability provided by Sidechains.[^14] However, for those looking to make an altcoin and dramatically boost community adoption, spinoffs could be a solution. Another possible application of the snapshot is using it to bootstrap a replacement for Bitcoin in the unlikely event something catastrophic occurs.
[^1]: ‘hash160’ is shorthand for the 160-bit hash created by SHA256 followed by RIPEMD160.
[^2]: `http://bitcointalk.org/index.php?topic=563972.0`
[^3]: `http://www.gridcoin.us`
[^4]: More precisely, the hash160, which is derived from the public key.
[^5]: `https://github.com/znort987/blockparser`
[^6]: `https://github.com/AugurProject/SideCoin/blob/master/snapshot`
[^7]: The balances are given in Satoshis.
[^8]: `http://bitcoin.stackexchange.com/questions/5021/how-do-you-get-the-op-hash160-value-from-a-bitcoin-address`
[^9]: The genesis block is loaded from memory every time `bitcoind` is started. However, aside from the genesis block, only the last one-hundred blocks are loaded into memory. Since the snapshot block is so large, we do not want it loaded into memory every time the coin daemon is called!
[^10]: `http://bitcoin.org/en/developer-guide#transaction-data`
[^11]: The text file is needed for anyone looking to claim a transaction anyway, so this a nice method. An alternative method is to paste the blocks and chainstate folders from a peer who has a blockchain with block one in it into one’s Sidecoin app data folder. Block one is stored as a checkpoint, so there is not much concern of someone falsifying the snapshot and giving a faulty block one.
[^12]: `https://github.com/AugurProject/SideCoin`
[^13]: `http://bitcointalk.org/index.php?topic=563972.0`
[^14]: `http://www.blockstream.com/sidechains.pdf`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '*Symbolic Computation* and *Satisfiability Checking* are two research areas, both having their individual scientific focus but sharing also common interests in the development, implementation and application of decision procedures for arithmetic theories. Despite their commonalities, the two communities are rather weakly connected. The aim of our newly accepted project (H2020-FETOPEN-CSA) is to strengthen the connection between these communities by creating common platforms, initiating interaction and exchange, identifying common challenges, and developing a common roadmap from theory along the way to tools and (industrial) applications. In this paper we report on the aims and on the first activities of this project, and formalise some relevant challenges for the unified community.'
author:
- Erika Ábrahám
- John Abbott
- Bernd Becker
- 'Anna M. Bigatti'
- Martin Brain
- Bruno Buchberger
- Alessandro Cimatti
- 'James H. Davenport'
- Matthew England
- Pascal Fontaine
- Stephen Forrest
- Alberto Griggio
- Daniel Kroening
- 'Werner M. Seiler'
- Thomas Sturm
bibliography:
- 'literature.bib'
title: |
: Satisfiability Checking meets\
Symbolic Computation\
(Project Paper)
---
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank the anonymous reviewers for their comments. We are grateful for support by the H2020-FETOPEN-2016-2017-CSA project (712689) and the ANR project ANR-13-IS02-0001-01 SMArT. Earlier work in this area was also supported by the EPSRC grant EP/J003247/1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Multipath propagation of radio waves in indoor/outdoor environments shows a highly irregular behavior as a function of time. Typical modeling of this phenomenon assumes the received signal is a stochastic process composed of the superposition of various altered replicas of the transmitted one, their amplitudes and phases being drawn from specific probability densities. We set out to explore the hypothesis of the presence of deterministic chaos in signals propagating inside various buildings at the University of Calgary. The correlation dimension versus embedding dimension saturates to a value between 3 and 4 for various antenna polarizations. The full Liapunov spectrum calculated contains two positive exponents and yields through the Kaplan-Yorke conjecture the same dimension obtained from the correlation sum. The presence of strange attractors in multipath propagation hints to better ways to predict the behaviour of the signal and better methods to counter the effects of interference. The use of Neural Networks in non linear prediction will be illustrated in an example and potential applications of same will be highlighted.'
author:
- 'C. Tannous'
- 'R. Davies'
- 'A. Angus'
date: 'March 16, 2001'
title: 'Strange Attractors in Multipath propagation: Detection and characterisation.'
---
Introduction
============
Multipath propagation of radio waves in indoor or outdoor environments shows a highly irregular behavior as a function of time \[1\]. The characterization of radio channels in mobile or in building propagation is important for addressing issues of design, coding, modulation and equalization techniques tailored specifically to combat time and frequency dispersion effects.\
Irregular behavior of the received signal has prompted researchers in the past to model the channel with stochastic processes. One of the earliest linear models in this vein is the Turin et al. model \[2\] in which the impulse response of the channel is written as a superposition of replicas of the transmitted signal delayed and having altered amplitudes and phases.\
A number of models exist differing in the numbers of replicas of the signal or in the type of probability distributions from which the amplitudes and phases are drawn. Also, different stochastic processes are used in the generation of delay times of received replicas. The popular choice for the amplitude probability density functions (PDF) are Rayleigh or Rice PDFs depending on whether a weak or strong line of sight propagation exists; nevertheless other PDFs have been used such the lognormal, Nakagami-m or uniform. The phases ought to be drawn from PDFs compatible with the ones selected for the corresponding amplitudes; nevertheless the most popular choice found in the literature is the uniform $[0 - 2 \pi]$ distribution. Delay times are usually extracted from either stationary of stationary Poisson processes although in some cases the Weibull PDF is used.
A HYPOTHESIS OF DETERMINISTIC CHAOS
===================================
Although an assumption of stochastic behavior in mobile or indoor propagation is ubiquitous, in the present work we set out to explore the hypothesis that the indoor communication channel displays deterministically chaotic behavior. This question is important in many respects and the tools to answer it readily exist. These tools are based on the determination of the correlation dimension of the strange attractor associated with the multipath profile considered as a real valued time-series x(t). Using delay coordinates \[3\] one forms the m-dimensional delay vector ${\bf X}(t) = [x(t),x(t-T) ...x(t-(m-1)T]$ with delay $T$ and computes the correlation sum C(r) which is the ratio of the number of pairs of delay vectors (the distance between which is less than r) to the total number of pairs. From this, the correlation dimension $\nu$ is defined as the logarithmic slope of C(r) versus r for small r. For a true stochastic process, $\nu$ increases with m without showing any saturation. In contrast, for deterministic chaos, $\nu$ saturates at a value, the next integer greater than which, represent the minimum number of non-linear recursion or differential equations from which it originates. If the profile turns out to be deterministic, the modulation, coding and detection/demodulation techniques ought to be adapted accordingly in order to account for this fact; otherwise one has to rely upon techniques capable of handling stochastic signals. Let us illustrate this by an analysis of multipath measurements we have made in an indoor environment.
EXPERIMENT AND CORRELATION DIMENSION ANALYSIS
=============================================
The propagation environment from which the data are collected are hallways in the Engineering Building at The University of Calgary \[4\]. The transmitted power was 10 dBm fed into a half wave dipole antenna with a matching balun. The receiving antenna was a cross-polarized dipole array. The co-polarized antennas (CPA) and crosspolarized antennas (XPA) profiles referred to in Figure 1 are from a single measurement run and points from both profiles were obtained in coincident pairs. The terms CPA and XPA simply refer to the relative state of polarisation between the transmit and receive antennas. The receiving hardware was specially developed to measure diversity characteristics and gives an accurate reference between distance from transmitting to receiving antennas and received signal strength. The same measurement procedure was employed as for the arbitrarily polarized data set. We have estimated the correlation dimension for the sets of data: Arbitrarily Polarized Antennas (APA, 6000 data points), co-polarized (CPA, 3800 points) and Cross-polarized antennas (XPA, 3800 points) by the box-counting method of Grassberger and Procaccia \[3\]. There are a number of limitations and potential pitfalls with correlation dimension estimation that have been discussed by various authors \[5\]. In addition, the number of operations it takes to estimate C(r) is $O(N^{2})$ where $N$ is the number of collected experimental points. Recently Theiler \[5\] devised a powerful box-assisted correlation sum algorithm based on a lexicographical ordering of the boxes covering the attractor, reducing the number of operations to $O(Nlog(N))$ and incorporating several test procedures aimed at avoiding the previous pitfalls. Before we used Theiler’s algorithm, we made some preliminary tests against well known cases. We generated uniform random numbers, Gaussian random numbers, and numbers z(n) according to the logistic one-dimensional map at the onset of Chaos: z(n+1) = a z(n)(1-z(n)) with a = 3.5699456 and in the fully developed Chaotic regime at a=4. In the first two cases we found $\nu$ approximately equal to m as expected in purely stochastic series whereas $\nu$ saturated respectively at 0.48 and 0.98 (we used 2000 points only) for the logistic map indicating the presence of a low dimensional attractor and deterministic Chaos (the exact correlation dimensions for the logistic map is 0.5 at a=3.5699456 and 1 for a=4.). We tested as well the Hénon two dimensional map, the Lorenz three dimensional system of non-linear differential equations, the Rössler three and four dimensional systems as well as an infinite dimensional system, the delay-differential Mackey-Glass equation whose attractor dimension is tunable with the delay time. The results we found for the various correlation dimensions agreed with all the results known in the literature to within a few percents. Then we went ahead and examined the $\nu$ vs. m curves for the three sets of experimental data along with a set of 6000 Rayleigh and band-limited Rayleigh distributed numbers which constitute prototypic received envelopes. Our results, in Figure 1 show that, while $\nu \sim m$ for the pure Rayleigh case (with a slope equal to one), and $\nu \propto m$ for the band limited Rayleigh case (with a slope smaller than one) the $\nu$ vs. m curves for the three examined experimental sets of data start linearly with m then show saturation indicating the presence of a low dimensional attractor (whose dimension is about 4 for CPA and XPA data whereas it is slightly above 4 for the APA situation). This finding is in line with the Ruelle criterion \[6\] that sets an upper bound on the possible correlation dimension one can get from any algorithm of the Grassberger-Procaccia type. This upper bound is set by the available number of data points $N$ in the time series. The dimension that can be detected should be much smaller than $2 \hspace{2mm} log_{10}(N)$. Since we used 3800 and 6000 points respectively, the upper bound for the detectable correlation dimension in our case is about 7.16 to 7.56. We respect this bound since our correlation dimensions are around 4. Nevertheless the presence of Chaos is going to be confirmed through another route, the spectrum of the Liapunov exponents that will be discusssed next.
SPECTRUM OF THE LIAPUNOV EXPONENTS
==================================
The spectrum of Liapunov exponents is very important in the study of dynamical systems. If the largest exponent is positive, this is a very strong indication for the presence of Chaos in the time series originating from the dynamical system. The reciprocal of this exponent is the average prediction time of the series and the sum of all the positive exponents (if more than one is detected like in hyper-chaotic systems such as the Rössler four dimensional system of non-linear differential equations or the large delay Mackey-Glass equation) is the Kolmogorov entropy rate of the system. The latter gives a quantitative idea about the information processes going on in the dynamical system. In addition, with the Kaplan-Yorke conjecture, the full spectrum gives the Hausdorff dimension of the strange attractor governing the long time evolution of the dynamical system. We have calculated the Liapunov exponents of the data with four different methods. Firstly, we determined the largest exponent $\lambda_{max}$ from the exponential separation of initially close points on the attractor and averaging over several thousand iterations. Second, we determined the largest Liapunov exponent from the correlation sum with the help of the relation $C(r) \sim r^{\nu} exp(-m T \lambda_{max})$ valid for large values of the embedding dimension $m$ and small values of r. Finally, we determined all exponents with two different methods: the Eckmann et al. method \[7\] and the Brown et al.’s \[8\]. Our results for the spectrum of exponents is shown in figures 2 and 3. We tried several embedding delay times T and several approximation degrees for the tangent mapping polynomial (as allowed in the Brown et al. \[8\] algorithm ) without observing major changes in the spectrum. Several time series (Logistic map, Hénon, Lorenz, Rössler and Mackey-Glass) were tested for the sake of comparison to results obtained with the experimental data. In addition, the Liapunov exponents saturate smoothly as they should for large embedding dimension. Then we applied the Kaplan-Yorke conjecture to get the dimension of the attractor: Using the following typical numbers we obtained for the exponents $\lambda_{max}=\lambda_{1}=18.06, \lambda_{2}=1.88, \lambda_{3}=-8.85, \lambda_{4} =-24.94,
\lambda_{5}=-68.80$ and using the formula:
$$D=j+\frac{\sum_{i}\lambda_{i}}{|\lambda_{j+1}|}$$
where the summation is over i=1,2...j. The $\lambda_{i}$’s are ordered in a way such that they decrease as i increases. We determine j from the conditions $\sum_{i} \lambda_{i} > 0 \hspace{2mm} \mbox{and:} \hspace{2mm} \lambda_{j+1} + \sum_{i}\lambda_{i} < 0$. We get j=3 and D=3.44 for the strange attractor dimension (called its Liapunov dimension). The total sum of the Liapunov exponents is negative ( equal to -82.65) as it should be for dissipative systems with a strange attractor. The value of the attractor dimension will be confirmed from the spectrum of singularities or the mutifractal spectrum in the next section.
MULTIFRACTAL SPECTRUM
=====================
The generalized dimension may be used to characterize non-uniform fractals for which there are different scaling exponents for different aspects of the fractal, so-called multifractals. For these, there are two scaling exponents, one generally called $\tau$, for the support of the fractal, and one called q, for the measure of bulk of the fractal. In general, $\tau(q)= (q-1)D_q$, where $D_q$ is the generalized dimension. Multifractals have been employed to characterize multiplicative random processes, turbulence, electrical discharge, diffusion-limited aggregation, and viscous fingering \[9\]. Multifractals have this in common: there is a non-uniform measure (growth rate, probability, mass) on a fractal support. Besides the exponents, $\tau$ and q, and the generalized dimension $D_q$, there is another method for characterizing multifractals. This depends upon the use of the mass exponent $\alpha$, and the multifractal spectrum, $f(\alpha)$ \[9\]. A graph of the multifractal spectrum explicitly shows the fractal dimension, f, of the points in the fractal with the mass exponent (or scaling index), $\alpha$. We have estimated $D_q$ by use of the generalised moments of the correlation sum with a window chosen carefully enough to avoid temporal correlation effects. We have developed a program that computes the generalized correlation sum using a box-assisted method. Our program is based on one written by Theiler \[5\]. Several modifications had to be made to the straightforward box-assisted correlation sum method. In addition, our program allows for logarithmic scaling with the distance parameter r. From a log-log graph of the generalized correlation sums, appropriate scaling regions can be identified, for each order, q. Least-squares fits to these scaling regions yields a sequence of generalized correlation dimensions, $D_q$, for values of q between $\pm \infty$. We have found computation of $D_q$ for integers in the interval \[-10,10\] and $D_{-\infty}$ and $D_{+\infty}$ to be sufficient. From the $D_q$, we calculate the $\tau(q)= (q-1)D_q$. We then perform a Legendre transform to obtain the $f(\alpha)$ curve. We do this by first fitting a smooth curve (a hyperbola was considered to be adequate) to the $\tau(q)$ curve. With an analytic expression for the $\tau(q)$ curve, we can compute the Legendre transform in closed form. The domain of $f(\alpha)$ may be found from $D_{-\infty}$ and $D_{+\infty}$; we assume that $\alpha$ is confined to this region, and that $f(\alpha)$ is 0 at these points. The values of $f(\alpha)$ for $D_{-\infty} \leq \alpha \leq D_{+\infty}$ are calculated as min\[${q\alpha-\tau(q)}$\], the minimum being taken over q. We found that this procedure, although complex, corrects for the known numerical sensitivities of the Legendre transform. We checked that our method for obtaining $f(\alpha)$ gave the same results as those found in the literature for the Logistic map and the strongly dissipative circle map \[9\]. The $f(\alpha)$ curve for the multipath data is shown in Figure 4. It may be seen that the peak value of $f(\alpha)$, corresponding to the box-counting dimension at q=0, is about 3.7. This is consistent with our above findings from the correlation dimension and the full spectrum of Liapunov exponents. Our further research in this area concerns the prediction of the received signal intensity, from our above hypothesis of the presence of deterministic chaos.
NON LINEAR PREDICTION
=====================
We applied the above findings to the non linear prediction of multipath profiles considering that each point on the envelope of the measured signal is a function of some number of past points in the series, deviating from the traditional wave superposition approach. More precisely, we write:
$$y(n+1)=F[y(n),y(n- 1),y(n-2),y(n-3) ...y(n-m+1)]$$
with F an $m$ dimensional map and y(n) the value of the signal x(t) sampled at timestep n. Expressing F as a sum of sigmoidal and linear functions \[10\] we determine the unknown weights through the Marquardt least squares minimisation method \[11\] in order to achieve the best fit to the data. Our results comparing the onestep prediction to the multipath data analysed above are displayed in figure 5. The goodness of fit between the measured and predicted envelopes for a map dimension m=5 is another indication of the soundness of the approach. This is confirmed in Figure 6 where we display the normalised prediction error versus the embedding dimension. A minimum is observed in the prediction error for an embedding equal to 5 or 6. In Figure 6, we started always from the same initial weights and let the system run through 150 iterations searching for the least squares minimum for 200 data points and later on making 300 one step ahead predictions. For embedding dimension larger or equal to 7 the minimisation procedure stopped because of the presence of zero-pivot in the least-squares matrices. The presence of a minimum in the prediction error around an embedding equal to 5 or 6 complies again with the value of embedding dimension used previously in the correlation dimension analysis, the Liapunov spectrum and the Kaplan-Yorke conjecture.\
DISCUSSION
==========
We stress that although the profiles we examined were found to be chaotic in all three experimental configurations with confirmations from the Liapunov spectrum and non linear prediction studies, indicating that we would be able to describe our data with a set of at most 5 non-linear differential or algebraic equations, investigation in other propagation situations is needed. Nevertheless, in our investigations we observed a significant amount of consistency between the various methods of detecting Chaos and characterising it using embedding dimensions $m$ beyond the minimum $m_{min}$ required by the Takens theorem ($m_{min} > 2d+1$, where d is the dimension of the strange attractor). Assuming the hypothesis of the presence of Chaos in a given multipath profile is firmly established, many avenues become possible. For instance, one might consider devising ways for controlling the signal propagation by altering slightly some accessible system parameter and improving the performance characteristics of the channel \[12\]. Shaw \[13\] has introduced the concept that a deterministically chaotic system can generate entropy. The consequences of this observation is important for the design of communication equipment when the channel is a chaotic system. For one, it implies that information at the receiver about the state of the channel is lost at a mean rate given by the Kolmogorov entropy. For another it implies that a channel estimator should be adapted to the mathematical nature of the set of non linear equations describing the channel as shown in the previous paragraph. Our studies up to this date have shown that this approach is valid in an indoor situation but not in an outdoor one. This might be due to the confined geometry one encounters inside buildings and the boundary conditions for the electromagnetic fields leading to a low dimensional system of non linear equations giving birth to the observed chaotic behaviour. Our studies in this direction are in progress.\
[**Acknowledgements**]{}
We thank James Theiler, Jean-Pierre Eckmann and Reggie Brown for sending us their computer programs and correspondance, as well as Halbert White for some unpublished material. C.T. thanks Sunit Lohtia and Bin Tan for their friendly help with the manuscript.\
A.A.M. Saleh and R. A.Valenzuela: “A Statistical Model For Indoor Multipath Propagation”, IEEE [**SAC-5**]{}, 138 (1987). G. L.Turin, F. D. Clapp, T. L. Johnston, S. B. Fine and D. Lavry: “A Statistical Model of Urban Multipath Propagation”, IEEE [**VT-21**]{}, 1 (1972). P. Grassberger and I. Procaccia: “Characterisation of Strange Attractors ”, Phys. Rev. Lett. [**50**]{}, 346 (1983). R. J. Davies: “In-Building UHF Propagation Studies”, MSc Thesis, University of Calgary (1989) Unpublished. J. Theiler: “Efficient algorithm for estimating the correlation dimension from a set of discrete points”, Phys. Rev. [**A36**]{}, 4456 (1987). D. Ruelle: “Deterministic Chaos: the science and the fiction”, Proc. R. Soc. London [**A427**]{}, 241 (1990). J-P Eckmann, S. Oliffson Kamphorst, D. Ruelle and S. Ciliberto: “Liapunov exponents from time series”, Phys. Rev. [**A34**]{}, 4971 (1986). R. Brown, P. Bryant and H.D.I. Abarbanel:“ Computing the Liapunov spectrum of a dynamical system from observed time series’ Phys. Rev. [**A43**]{}, 2787 (1991). G. Paladin and A. Vulpiani: ”Anomalous scaling laws in multifractal objects“, Phys. Rep. [**156**]{}, 141 (1987). H. White: ”Some asymptotic results for learning in single hidden-layer feedforward network models“ J. Am. Stat Association [**84**]{}, 1003 (1989). D. W. Marquardt: ” An algorithm for least squares estimation of non-linear parameters“, J. Soc. Ind. App. Math. [**11**]{}, 431 (1963). E. Ott, C. Grebogi and J. A. Yorke: ”Controlling Chaos“, Phys. Rev. Lett. [**64**]{}, 1196 (1990). R. S. Shaw: ”Strange attracctors, Chaotic behaviour and information flow", Z. Naturforsch [**36a**]{}, 80 (1981).
**Figure Captions**
- Correlation dimensions vs embedding dimension: Full squares are for Rayleigh distributed points; full triangles are for handlimited Rayleigh distributed points. Full diamonds are for experimental results in the XPA case whereas open diamonds correspond to the CPA case and open squares to the APA case.
- Liapunov exponent spectrum from Eckmann et al. \[7\] method versus embedding dimension for the APA data (since APA data consist of the largest number of points, 6000).
- Liapunov exponent spectrum from Brown et al. method \[8\] for the same data as those of Fig.2. A linear tangent mapping is used to fit the dynamics. A very similar spectrum is obtained for a second order polynomial.
- Spectrum of generalised dimensions $f(\alpha)$ versus a for the APA data used in fig.2. The value at the maximum of $f(\alpha)$ corresponding to the Hausdorff dimension of the strange attractor agrees with the minimum bound obtained from fig.1 and with the Kaplan-Yorke conjecture (see text). The spectrum is obtained through embedding in 10 dimensions. This happens to be enough, given the values obtained for the various generalised dimensions.
- Measured envelope (APA data used in Fig.2 continuous curve) and its one step prediction (dashed curve) from the Neural Network fit to the five dimensional map F (eq.2). The training is over the first 200 first points.
- Normalised prediction error versus embedding. Starting from the same initial weights, we trained the neural network, for a given embedding, over the first 200 points with a Marquardt minimisation standard deviation parameter equal to 0.01 and total number of 150 iterations. Once, the parameters at the minimum error are found we made a one-step ahead prediction over the next 300 points and calculated the resulting squared error divided by the total of points. One sees a minimum for an embedding dimension around 5 or 6. For an embedding equal to 7 or larger, a large error or no convergence (null pivot encountered in the least square error matrices) were observed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper we investigate the existence, decay and concentration behavior of solutions for the following class of fractional relativistic Schrödinger equations: $$\left\{
\begin{array}{ll}
(-\Delta+m^{2})^{s}u + V(\operatorname{\varepsilon}x) u= f(u) &\mbox{ in } {\mathbb{R}}^{N}, \\
u\in H^{s}({\mathbb{R}}^{N}), \quad u>0 &\mbox{ in } {\mathbb{R}}^{N},
\end{array}
\right.$$ where $\operatorname{\varepsilon}>0$ is a small parameter, $s\in (0, 1)$, $m>0$, $N> 2s$, $(-\Delta+m^{2})^{s}$ is the fractional relativistic Schrödinger operator, $V:{\mathbb{R}}^{N}{\rightarrow}{\mathbb{R}}$ is a continuous potential satisfying a local condition, and $f:{\mathbb{R}}{\rightarrow}{\mathbb{R}}$ is a continuous subcritical nonlinearity. The approach is based on a variant of the extension method and applying appropriate variational techniques.
address: 'Vincenzo AmbrosioDipartimento di Ingegneria Industriale e Scienze Matematiche Università Politecnica delle MarcheVia Brecce Bianche, 1260131 Ancona (Italy)'
author:
- Vincenzo Ambrosio
title: 'The nonlinear fractional relativistic Schrödinger equation: existence, decay and concentration results'
---
Introduction
============
In this paper we consider the following class of nonlinear fractional elliptic problems: $$\label{P}
\left\{
\begin{array}{ll}
(-\Delta+m^{2})^{s}u + V(\operatorname{\varepsilon}x) u= f(u) &\mbox{ in } {\mathbb{R}}^{N}, \\
u\in H^{s}({\mathbb{R}}^{N}), \quad u>0 &\mbox{ in } {\mathbb{R}}^{N},
\end{array}
\right.$$ where $\operatorname{\varepsilon}>0$ is a small parameter, $s\in (0, 1)$, $N> 2s$, $m>0$, $V:{\mathbb{R}}^{N}{\rightarrow}{\mathbb{R}}$ is a continuous potential and $f:{\mathbb{R}}{\rightarrow}{\mathbb{R}}$ is a continuous nonlinearity. The nonlocal operator $(-\Delta+m^{2})^{s}$ appearing in is defined via Fourier transform by $$(-\Delta+m^{2})^{s}u:=(2\pi)^{-\frac{N}{2}}\mathcal{F}^{-1}((|k|^{2}+m^{2})^{s}\mathcal{F}u)$$ for any $u:{\mathbb{R}}^{N}{\rightarrow}{\mathbb{R}}$ belonging to the Schwartz space $\mathcal{S}({\mathbb{R}}^{N})$ of rapidly decaying functions, or equivalently (see [@FF; @LL]) $$\begin{aligned}
\label{FFdef}
(-\Delta+m^{2})^{s}u(x):=m^{2s}u(x)+C(N,s) m^{\frac{N+2s}{2}} P.V. \int_{{\mathbb{R}}^{N}} \frac{u(x)-u(y)}{|x-y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|x-y|)\, dy,\end{aligned}$$ where $P. V.$ stands for the Cauchy principal value, $K_{\nu}$ is the modified Bessel function of the third kind (or Macdonald function) of index $\nu$ (see [@ArS; @Erd]) which satisfies the following well-known asymptotic formulas for $\nu\in {\mathbb{R}}$ and $r>0$: $$\begin{aligned}
&K_{\nu}(r)\sim \frac{\Gamma(\nu)}{2} \left(\frac{r}{2}\right)^{-\nu} \mbox{ as } r\rightarrow 0, \mbox{ for } \nu>0, \label{Watson1}\\
&K_{\nu}(r)\sim \sqrt{\frac{\pi}{2}}r^{-\frac{1}{2}} e^{-r} \mbox{ as } r \rightarrow \infty, \mbox{ for } \nu\in {\mathbb{R}}\label{Watson2},\end{aligned}$$ and $C(N, s)$ is a positive constant whose exact value is given by $$C(N, s):=2^{-\frac{N+2s}{2}+1} \pi^{-\frac{N}{2}} 2^{2s} \frac{s(1-s)}{\Gamma(2-s)}.$$ Equations involving $(-\Delta+m^{2})^{s}$ arise in the study of standing waves $\psi(x, t)$ for Schrödinger-Klein-Gordon equations of the form $$\begin{aligned}
\imath \frac{\partial \psi}{\partial t}=(-\Delta+m^{2})^{s}\psi-f(x, \psi), \quad \mbox{ in } {\mathbb{R}}^{N}\times {\mathbb{R}}, \end{aligned}$$ which describe the behavior of bosons. In particular, when $s=1/2$, the operator $\sqrt{-\Delta+m^{2}}-m$ plays an important role in relativistic quantum mechanics because it corresponds to the kinetic energy of a relativistic particle with mass $m>0$. If $p$ is the momentum of the particle then its relativistic kinetic energy is given by $E=\sqrt{p^{2}+m^{2}}$. In the process of quantization the momentum $p$ is replaced by the differential operator $-\imath \nabla$ and the quantum analog of the relativistic kinetic energy is the free relativistic Hamiltonian $\sqrt{-\Delta+m^{2}}-m$. Physical models related to this operator have been widely studied over the past 30 years and there exists a huge literature on the spectral properties of relativistic Hamiltonians, most of it has been strongly influenced by the works of Lieb on the stability of relativistic matter; see [@ES; @Herbst; @LY1; @LY2] for more physical background. On the other hand, there is also a deep connection between $(-\Delta+m^{2})^{s}$ and the theory of Lévy processes. Indeed, $m^{2s}-(-\Delta+m^{2})^{s}$ is the infinitesimal generator of a Lévy process $X^{2s, m}_{t}$ called $2s$-stable relativistic process having the following characteristic function $$E^{0} e^{\imath k\cdot X^{2s, m}_{t}}=e^{-t[(|k|^{2}+m^{2})^{s}-m^{2s}]}, \quad k\in {\mathbb{R}}^{N};$$ we refer to [@BMR; @CMS; @ryznar] for a more detailed discussion on relativistic stable processes. When $m=0$, the previous operator boils down to the fractional Laplacian operator $(-\Delta)^{s}$ which has been extensively studied in these last years due to its great applications in several fields of the research; see [@BuVa; @DPV; @MBRS] for an introduction on this topic. In particular, a great interest has been devoted to the existence and multiplicity of solutions for fractional Schrödinger equations [@Laskin1] like $$\label{ANNALISA}
\operatorname{\varepsilon}^{2s}(-\Delta)^{s}u + V(x) u= f(u) \mbox{ in } {\mathbb{R}}^{N},$$ and the asymptotic behavior as $\operatorname{\varepsilon}{\rightarrow}0$; see for instance [@AM; @Aampa; @Armi; @DDPW; @FFV; @FS] and the references therein.
When $m>0$ and $\operatorname{\varepsilon}=1$ in , some interesting existence, multiplicity, and qualitative results of solutions for can be found in [@Ajmp; @BMP; @CS1; @CZN1; @CZN2; @FV; @Ikoma; @SecchiJDDE], while only one result [@CS2] treats with the semiclassical analysis $\operatorname{\varepsilon}{\rightarrow}0$ of a fractional Hartree equation involving $\sqrt{-\operatorname{\varepsilon}^{2}\Delta+m^{2}}$.
Motivated by the above papers, in this work we focus our attention on the concentration phenomenon of solutions to as $\operatorname{\varepsilon}{\rightarrow}0$. Along this paper, we suppose that $V:{\mathbb{R}}^{N}{\rightarrow}{\mathbb{R}}$ is a continuous function which satisfies the following conditions due to del Pino and Felmer [@DF]:
there exists $V_{1}\in (0, m^{2s})$ such that $-V_{1}:=\inf_{x\in {\mathbb{R}}^{N}}V(x)$,
there exists a bounded open set $\Lambda\subset {\mathbb{R}}^{N}$ such that $$-V_{0}:=\inf_{x\in \Lambda} V(x)<\min_{x\in \partial \Lambda} V(x),$$ with $V_{0}>0$. We also set ${\mathcal{M}}:=\{x\in \Lambda: V(x)=-V_{0}\}$. Without loss of generality, we may assume that $0\in {\mathcal{M}}$.
Concerning the nonlinearity $f$, we assume that $f:{\mathbb{R}}{\rightarrow}{\mathbb{R}}$ is continuous, $f(t)=0$ for $t\leq 0$, and $f$ fulfills the following hypotheses:
$\lim_{t{\rightarrow}0} \frac{f(t)}{t}=0$,
$\limsup_{t{\rightarrow}\infty} \frac{f(t)}{t^{p}}<\infty$ for some $p\in (1, {2^{*}_{s}}-1)$, where ${2^{*}_{s}}:=\frac{2N}{N-2s}$ is the fractional critical exponent,
there exists $\theta\in (2, {2^{*}_{s}})$ such that $0<\theta F(t)\leq t f(t)$ for all $t>0$,
$\frac{f(t)}{t}$ is increasing for $t>0$.
The main result of this work can be stated as follows:
\[thm1\] Assume that $(V_1)$-$(V_2)$ and $(f_1)$-$(f_4)$ are satisfied. Then, for every small $\operatorname{\varepsilon}>0$, there exists a solution $u_{\operatorname{\varepsilon}}$ to such that $u_{\operatorname{\varepsilon}}$ has a maximum point satisfying $$\lim_{\operatorname{\varepsilon}{\rightarrow}0} {\rm dist}(\operatorname{\varepsilon}x_{\operatorname{\varepsilon}}, {\mathcal{M}})=0,$$ and for which $$0<u_{\operatorname{\varepsilon}}(x)\leq Ce^{-c|x-x_{\operatorname{\varepsilon}}|} \quad \forall x\in {\mathbb{R}}^{N},$$ for suitable constants $C, c>0$. Moreover, for any sequence $(\operatorname{\varepsilon}_{n})$ with $\operatorname{\varepsilon}_{n}{\rightarrow}0$, there exists a subsequence, still denoted by itself, such that there exist a point $x_{0}\in \mathcal{M}$ with $\operatorname{\varepsilon}_{n}y_{\operatorname{\varepsilon}_{n}}{\rightarrow}x_{0}$, and a positive least energy solution $u\in {H^{s}({\mathbb{R}}^{N})}$ of the limiting problem $$(-\Delta+m^{2})^{s}u-V_{0} u=f(u) \mbox{ in } {\mathbb{R}}^{N},$$ for which we have $$u_{\operatorname{\varepsilon}_{n}}(x)=u(x-y_{\operatorname{\varepsilon}_{n}})+\mathcal{R}_{n}(x)$$ where $\lim_{n{\rightarrow}\infty} \|\mathcal{R}_{n}\|_{{H^{s}({\mathbb{R}}^{N})}}=0$.
The proof of Theorem \[thm1\] is obtained through suitable variational techniques. Firstly, we start by observing that $(-\Delta+m^{2})^{s}$ is a nonlocal operator and that does not scale like the fractional Laplacian operator $(-\Delta)^{s}$. More precisely, the first operator is not compatible with the semigroup ${\mathbb{R}}_{+}$ acting on functions as $t*u\mapsto u(t^{-1}x)$ for $t>0$. This fact does not permit to adapt in a simple way the same arguments performed to deal with $(-\Delta)^{s}$. Nevertheless, we overcome these difficulties by using a variant of the extension method [@CS] for $(-\Delta+m^{2})^{s}$ (see [@CZN1; @FF; @StingaT]) which permits to study via local variational methods a degenerate elliptic equation in a half-space with a nonlinear Neumann boundary condition. Clearly, some additional difficulties arise in the investigation of this problem because we have to handle the trace terms of the involved functions and thus a more careful analysis will be needed.
Due to the lack of informations on the behavior of $V$ at infinity, we carry out a penalization argument [@DF] which consists in modifying appropriately the nonlinearity $f$ outside $\Lambda$, and thus consider a modified problem whose corresponding energy functional fulfills all the assumptions of the mountain pass theorem [@AR]. Then we need to check that, for $\operatorname{\varepsilon}>0$ small enough, the solutions of the auxiliary problem are indeed solutions of the original one. This goal will be achieved by combing an appropriate Moser iteration argument [@Moser] with some elliptic regularity estimates established in [@FF]. To our knowledge this is the first time that the penalization trick is used to study the concentration phenomena for the fractional relativistic Schrödinger operator $(-\Delta+m^{2})^{s}$ for all $s\in (0, 1)$ and $m>0$. When $m=0$, namely when reduces to after rescaling, we refer the interested reader to [@AM; @Aampa; @Armi] for similar approaches. Finally, we show that the solutions of have an exponential decay, contrary to the case $m=0$ for which the solutions of satisfy the power-type decay $|x|^{-(N+2s)}$ as $|x|{\rightarrow}\infty$; see [@Armi; @FFV; @FQT]. To investigate the decay of solutions to , we construct a suitable comparison function and we carry out some refined estimates which take care of an adequate estimate concerning $2s$-stable relativistic density with parameter $m$ found in [@GR], and that the modified Bessel function $K_{\nu}$ has an exponential decay at infinity. We stress that exponential type estimates for equations like , appear in [@CS2; @CZN1] where $s=\frac{1}{2}$, $V$ is bounded, $f$ is a Hartree type nonlinearity, and in [@FV] where $s\in (0, 1)$, $m=1$, $V\equiv 0$ and $|f(u)|\leq C u^{p}$ for some $p\in (1, {2^{*}_{s}}-1)$. Anyway, our approach to obtain the decay estimate is completely different from the above mentioned papers, it is more general and we believe that can be applied in other situations to deal with fractional problems driven by $(-\Delta+m^{2})^{s}$. We conclude this introduction by pointing out that in view of the techniques developed here and the recent result in [@Ana], we are preparing a work [@Agn] in which we obtain a multiplicity result for when $f$ is a Beresticky-Lions type nonlinearity [@BL]. The paper is organized as follows. In section $2$ we collect some notations and preliminary results which will be used along the paper. In section $3$ we introduce a penalty functional in order to apply suitable variational arguments. The section $4$ is devoted to the proof of Theorem \[thm1\]. Finally, we give some interesting results for $(-\Delta+m^{2})^{s}$ in the appendix.
preliminaries
=============
Notations and functional setting
--------------------------------
We denote the upper half-space in ${\mathbb{R}}^{N+1}$ by $${\mathbb{R}}^{N+1}_{+}:=\{(x, y)\in {\mathbb{R}}^{N+1}: y>0\},$$ and for $(x, y)\in {\mathbb{R}}^{N+1}_{+}$ we consider the Euclidean norm $|(x, y)|:=\sqrt{|x|^{2}+y^{2}}$.
Let $p\in [1, \infty]$ and $A\subset {\mathbb{R}}^{N}$ be a measurable set. We indicate by $L^{p}(A)$ the set of measurable functions $u: {\mathbb{R}}^{N}{\rightarrow}{\mathbb{R}}$ such that $$|u|_{L^{p}(A)}:=\left\{
\begin{array}{ll}
\left(\int_{A} |u|^{p}\, dx\right)^{1/p}<\infty &\mbox{ if } p<\infty, \\
{\rm esssup}_{x\in A} |u(x)| &\mbox{ if } p=\infty.
\end{array}
\right.$$ When $A={\mathbb{R}}^{N}$, we simply write $|u|_{p}$ instead of $|u|_{L^{p}({\mathbb{R}}^{N})}$. With $\|w\|_{L^{p}({\mathbb{R}}^{N+1}_{+})}$ we will always denote the norm of $w\in L^{p}({\mathbb{R}}^{N+1}_{+})$.
Let $D\subset {\mathbb{R}}^{N+1}$ be a bounded domain, that is a bounded connected open set, with boundary $\partial D$, we denote by $\partial' D$ the interior of $\overline{D}\cap \partial{\mathbb{R}}^{N+1}_{+}$ in ${\mathbb{R}}^{N}$, and we set $\partial ''D:=\partial D\setminus \partial'D$. For $R>0$, we put $$\begin{aligned}
&B^{+}_{R}:=\{(x,y)\in \mathbb{R}^{N+1}_{+}: |(x,y)|<R\}, \\
&\Gamma_{R}^{0}:=\partial'B_{R}^{+}=\{(x,0)\in \partial \mathbb{R}^{N+1}_{+}: |x|<R \}, \\
&\Gamma_{R}^{+}:=\partial''B_{R}^{+}=\{(x,y)\in \mathbb{R}^{N+1}: y\geq 0, \, |(x, y)|= R \}.\end{aligned}$$ Now, we introduce the Lebesgue spaces with weight (see [@FKS; @JLX] for more details). Let $D\subset {\mathbb{R}}^{N+1}_{+}$ be an open set and $r\in (1, \infty)$. Denote by $L^{r}(D, y^{1-2s})$ the weighted Lebesgue space of all measurable functions $v:D\rightarrow {\mathbb{R}}$ such that $$\|v\|_{L^{r}(D, y^{1-2s})}:=\left(\iint_{D} y^{1-2s} |v|^{r}\, dx dy\right)^{\frac{1}{r}}<\infty.$$ We say that $v\in H^{1}(D, y^{1-2s})$ if $v\in L^{2}(D, y^{1-2s})$ and its weak derivatives, collectively denoted by $\nabla v$, exist and belong to $L^{2}(D, y^{1-2s})$. The norm of $v$ in $H^{1}(D, y^{1-2s})$ is given by $$\|v\|_{H^{1}(D, y^{1-2s})}^{2}:=\iint_{D} y^{1-2s} (|\nabla v|^{2}+ v^{2}) \, dx dy<\infty.$$ It is clear that $H^{1}(D, y^{1-2s})$ is a Hilbert space with the inner product $$\iint_{D} y^{1-2s} (\nabla v \nabla w+ v w) \, dx dy.$$
Let $H^{s}({\mathbb{R}}^{N})$ be the fractional Sobolev space defined as the completion of $C^{\infty}_{c}({\mathbb{R}}^{N})$ with respect to the norm $$|u|_{H^{s}({\mathbb{R}}^{N})}:=\left( \int_{{\mathbb{R}}^{N}} (|k|^{2}+m^{2})^{s} |\mathcal{F}u(k)|^{2} dk \right)^{\frac{1}{2}}.$$ Then, $H^{s}({\mathbb{R}}^{N})$ is continuously embedded in $L^{p}({\mathbb{R}}^{N})$ for all $p\in [2, {2^{*}_{s}})$ and compactly in $L^{p}_{loc}({\mathbb{R}}^{N})$ for all $p\in [1, {2^{*}_{s}})$; see [@Adams; @ArS; @DPV; @LL]. Next we define $X^{s}({\mathbb{R}}^{N+1}_{+}):=H^{1}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})$ as the completion of $C^{\infty}_{c}(\overline{{\mathbb{R}}^{N+1}_{+}})$ with respect to the norm $$\|u\|_{X^{s}({\mathbb{R}}^{N+1}_{+})}:= \left(\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (|\nabla u|^{2}+m^{2}u^{2})\, dx dy \right)^{\frac{1}{2}}.$$ By Lemma 3.1 in [@FF], we deduce that ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ is continuously embedded in $L^{2\gamma}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})$, that is $$\label{weightedE}
\|u\|_{L^{2\gamma}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}\leq C_{*}\|u\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}} \quad \forall u\in {X^{s}({\mathbb{R}}^{N+1}_{+})},$$ where $\gamma:=1+\frac{2}{N-2s}$, and $L^{r}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})$ is the weighted Lebesgue space, with $r\in (1, \infty)$, endowed with the norm $$\|u\|_{L^{r}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}:=\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} |u|^{r}\, dx dy.$$ Moreover, by Lemma 3.1.2 in [@DMV], we also have that ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ is compactly embedded in $L^{2}(B_{R}^{+}, y^{1-2s})$ for all $R>0$. From Proposition 5 in [@FF], we know that there exists a linear trace operator ${\rm Tr}: {X^{s}({\mathbb{R}}^{N+1}_{+})}{\rightarrow}{H^{s}({\mathbb{R}}^{N})}$ such that $$\label{traceineq}
\sqrt{\sigma_{s}} |{\rm Tr}(u)|_{H^{s}({\mathbb{R}}^{N})}\leq \|u\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}} \mbox{ for any } u\in {X^{s}({\mathbb{R}}^{N+1}_{+})},$$ where $\sigma_{s}:=2^{1-2s}\Gamma(1-s)/\Gamma(s)$. We also note the and the definition of $H^{s}$-norm imply that $$\label{m-ineq}
\sigma_{s}m^{2s} \int_{{\mathbb{R}}^{N}} |{\rm Tr}(u)|^{2}\, dx\leq \iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (|\nabla u|^{2}+m^{2}u^{2})\, dx dy.$$ In what follows, in order to simplify the notation, we denote ${\rm Tr}(u)$ by $u(\cdot, 0)$. Since $Tr(X^{s}({\mathbb{R}}^{N+1}_{+}))\subset {H^{s}({\mathbb{R}}^{N})}$ and ${H^{s}({\mathbb{R}}^{N})}\subset L^{q}({\mathbb{R}}^{N})$ for all $q\in [2, {2^{*}_{s}})$ and the embedding is locally compact for all $q\in [1, {2^{*}_{s}})$ (see [@Adams; @ArS; @DPV]), we obtain the following result:
\[Sembedding\] $Tr(X^{s}({\mathbb{R}}^{N+1}_{+}))$ is continuously embedded in $L^{q}({\mathbb{R}}^{N})$ for all $q\in [2, {2^{*}_{s}})$ and compactly embedded in $L^{q}_{loc}({\mathbb{R}}^{N})$ for all $q\in [1, {2^{*}_{s}})$.
In order to circumvent the nonlocal character of the pseudo-differential operator $(-\Delta+m^{2})^{s}$, we make use of a variant of the extension method [@CS] given in [@FF; @StingaT]. More precisely, for any $u\in {H^{s}({\mathbb{R}}^{N})}$ there exists a unique function $U\in {X^{s}({\mathbb{R}}^{N+1}_{+})}$ solving the following problem $$\begin{aligned}
\left\{
\begin{array}{ll}
-\operatorname{div}(y^{1-2s} \nabla U)+m^{2}y^{1-2s}U=0 &\mbox{ in } {\mathbb{R}}^{N+1}_{+}, \\
U(\cdot, 0)=u &\mbox{ on } \partial {\mathbb{R}}^{N+1}_{+}\cong{\mathbb{R}}^{N}.
\end{array}
\right.\end{aligned}$$ The function $U$ is called the $s$-extension of $u$ and possesses the following properties:
$$\frac{\partial U}{\partial \nu^{1-2s}}:=-\lim_{y{\rightarrow}0} y^{1-2s} \frac{\partial U}{\partial y}(x,y)=\sigma_{s}(-\Delta+m^{2})^{s}u(x) \mbox{ in distribution sense, }$$
$\sqrt{\sigma_{s}}|u|_{{H^{s}({\mathbb{R}}^{N})}}=\|U\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\leq \|V\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}$ for all $V\in {X^{s}({\mathbb{R}}^{N+1}_{+})}$ such that $V(\cdot,0)=u$.
$U\in C^{\infty}({\mathbb{R}}^{N+1}_{+})$ and can be expressed as $$U(x, y)=\int_{{\mathbb{R}}^{N}} P_{s, m}(x-z,y) u(z)\, dz$$ with $$P_{s, m}(x,y):=c'_{N,s} y^{2s} m^{\frac{N+2s}{2}} |(x, y)|^{-\frac{N+2s}{2}} K_{\frac{N+2s}{2}}(m|(x, y)|),$$ and $$c'_{N, s}:=p_{N, s}\frac{2^{\frac{N+2s}{2}-1}}{\Gamma(\frac{N+2s}{2})},$$ where $p_{N, s}$ is the constant for the (normalized) Poisson kernel with $m=0$; see [@StingaT].\
We note that $P_{s, m}$ is the Fourier transform of $k\mapsto \vartheta(\sqrt{|k|^{2}+m^{2}})$ and that $$\begin{aligned}
\label{Nkernel}
\int_{{\mathbb{R}}^{N}} P_{s, m}(x, y)\, dx=\vartheta(m y),\end{aligned}$$ where $\vartheta\in H^{1}({\mathbb{R}}_{+}, y^{1-2s})$ solves the following ordinary differential equation $$\begin{aligned}
\left\{
\begin{array}{ll}
\vartheta''+\frac{1-2s}{y} \vartheta'-\vartheta=0 &\mbox{ in } {\mathbb{R}}_{+}, \\
\vartheta(0)=1.
\end{array}
\right.\end{aligned}$$ We also recall that $\vartheta$ can be expressed via modified Bessel functions, more precisely $\vartheta(r)=\frac{2}{\Gamma(s)} (\frac{r}{2})^{s}K_{s}(r)$; see [@FF] for more details.
Taking into account the previous facts, problem can be realized in a local manner through the following nonlinear boundary value problem: $$\begin{aligned}
\label{EP}
\left\{
\begin{array}{ll}
-\operatorname{div}(y^{1-2s} \nabla w)+m^{2}y^{1-2s}w=0 &\mbox{ in } {\mathbb{R}}^{N+1}_{+}, \\
\frac{\partial w}{\partial \nu^{1-2s}}=\sigma_{s} [-V_{\operatorname{\varepsilon}}(x) w(\cdot, 0)+f(w(\cdot, 0))] &\mbox{ in } {\mathbb{R}}^{N},
\end{array}
\right.\end{aligned}$$ where $V_{\operatorname{\varepsilon}}(x):=V(\operatorname{\varepsilon}x)$. For simplicity of notation, we will omit the constant $\sigma_{s}$ from the second equation in . For all $\operatorname{\varepsilon}>0$, we define $$X_{\operatorname{\varepsilon}}:=\left\{u\in X^{s}({\mathbb{R}}^{N+1}_{+}): \int_{{\mathbb{R}}^{N}} V_{\operatorname{\varepsilon}}(x) u^{2}(x, 0)\, dx<\infty\right\}$$ endowed with the norm $$\|u\|_{\operatorname{\varepsilon}}:=\left(\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}+ \int_{{\mathbb{R}}^{N}} V_{\operatorname{\varepsilon}}(x) u^{2}(x, 0)\, dx \right)^{\frac{1}{2}}.$$ We note that $\|\cdot\|_{\operatorname{\varepsilon}}$ is actually a norm. Indeed, $$\|u\|_{\operatorname{\varepsilon}}^{2}=\left[\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}\int_{{\mathbb{R}}^{N}}u^{2}(x, 0)\, dx\right]+\int_{{\mathbb{R}}^{N}} (V_{\operatorname{\varepsilon}}(x)+V_{1})u^{2}(x, 0)\, dx,$$ and using and $(V_1)$ we can see that $$\left(1-\frac{V_{1}}{m^{2s}}\right)\|u\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}\leq\left[\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}\int_{{\mathbb{R}}^{N}}u^{2}(x, 0)\, dx\right]\leq \|u\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}$$ that is $$\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}\int_{{\mathbb{R}}^{N}}u^{2}(x, 0)\, dx$$ is a norm equivalent to $\|\cdot\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}$. This observation yields the required claim. Clearly, $X_{\operatorname{\varepsilon}}\subset X^{s}({\mathbb{R}}^{N+1}_{+})$, and using $(V_1)$ we have $$\label{equivalent}
\|u\|_{X^{s}({\mathbb{R}}^{N+1}_{+})}^{2}\leq \left(\frac{m^{2s}}{m^{2s}-V_{1}}\right) \|u\|^{2}_{\operatorname{\varepsilon}} \quad \forall u\in X_{\operatorname{\varepsilon}}.
$$
Elliptic estimates
------------------
For reader’s convenience, we list some results about local Schauder estimates for degenerate elliptic equations involving the operator $$-{\rm div}(y^{1-2s} \nabla v)+m^{2}y^{1-2s}v.$$ Firstly we give the following definition:
Let $D\subset {\mathbb{R}}^{N+1}_{+}$ be a bounded domain with $\partial'D\neq \emptyset$. Let $f\in L^{\frac{2N}{N+2s}}_{loc}(\partial'D)$ and $g\in L^{1}_{loc}(\partial'D)$. Consider $$\label{JLXproblem}
\left\{
\begin{array}{ll}
-{\rm div}(y^{1-2s} \nabla v)+m^{2} y^{1-2s}v=0 &\mbox{ in } D, \\
\frac{\partial v}{\partial \nu^{a}}=f(x)v+g(x) &\mbox{ on } \partial'D.
\end{array}
\right.$$ We say that $v\in H^{1}(D, y^{1-2s})$ is a weak supersolution (resp. subsolution) to in $D$ if for any nonnegative $\varphi \in C^{1}_{c}(D\cup \partial'D)$, $$\iint_{D} y^{1-2s} (\nabla v \nabla \phi+m^{2} v\phi) \, dx dy\geq (\leq) \int_{\partial'D} [f(x) v(\cdot, 0)+g(x)] \varphi(\cdot,0) \, dx.$$ We say that $v\in H^{1}(D, y^{1-2s})$ is a weak solution to in $D$ if it is both a weak supersolution and a weak subsolution.
We denote by $Q_{R}=B_{R}\times (0, R)$ where $B_{R}\subset {\mathbb{R}}^{N}$ is the ball in ${\mathbb{R}}^{N}$ centered at $0$ and with radius $R>0$. Then we recall the following version of De Giorgi-Nash-Moser type theorems established in [@FF] (see also [@FKS; @JLX] for the case $m=0$).
[@FF]\[PROPFF\] Let $f, g\in L^{q}(B_{1})$ for some $q>\frac{N}{2s}$.
Let $v\in H^{1}(Q_{1},y^{1-2s})$ be a weak subsolution to in $Q_{1}$. Then $$\sup_{Q_{1/2}} v^{+}\leq C(\|v^{+}\|_{L^{2}(Q_{1}, y^{1-2s})}+|g^{+}|_{L^{q}(B_{1})}),$$ where $v^{+}:=\max\{v, 0\}$, and $C>0$ depends only on $N$, $s$, $q$, $|f^{+}|_{L^{q}(B_{1})}$.
(weak Harnack inequality) Let $v\in H^{1}(Q_{1},y^{1-2s})$ be a nonnegative weak supersolution to in $Q_{1}$. Then for some $p_{0}>0$ and any $0<\mu<\tau<1$ we have $$\inf_{\bar{Q}_{\mu}} v+|g_{-}|_{L^{q}(B_{1})}\geq C\| v \|_{L^{p_{0}}(Q_{\tau}, y^{1-2s})},$$ where $v_{-}:=\max\{-v, 0\}$, and $C>0$ depends only on $N$, $s$, $q$, $\mu$, $\tau$, $|f_{-}|_{L^{q}(B_{1})}$.
Let $v\in H^{1}(Q_{1},y^{1-2s})$ be a nonnegative weak solution to in $Q_{1}$. Then $v\in C^{0,\alpha}(\overline{Q_{1/2}})$ and in addition $$\|v\|_{C^{0,\alpha}(\overline{Q_{1/2}})}\leq C(\|v\|_{L^{2}(Q_{1})}+|g|_{L^{q}(B_{1})}),$$ with $C>0$, $\alpha\in (0, 1)$ depending only on $N$, $s$, $p$, $|f|_{L^{q}(B_{1})}$.
The penalization argument
=========================
In order to find solutions to , we follow the penalization approach in [@DF] which permits to study our problem via variational arguments. Fix $\kappa>\frac{V_{1}}{m^{2s}-V_{1}}$ and $a>0$ such that $\frac{f(a)}{a}=\frac{V_{1}}{\kappa}$. Define $$\tilde{f}(t):=\left\{
\begin{array}{ll}
f(t) &\mbox{ for } t<a, \\
\frac{V_{1}}{\kappa} t &\mbox{ for } t\geq a.
\end{array}
\right.$$ Let us consider the following Carathéodory function $$g(x, t):=\chi_{\Lambda}(x) f(t)+(1-\chi_{\Lambda}(x)) \tilde{f}(t) \quad \mbox{ for } (x,t)\in {\mathbb{R}}^{N}\times {\mathbb{R}},$$ where $\chi_{\Lambda}$ denotes the characteristic function of $\Lambda$. Set $G(x,t):=\int_{0}^{t} g(x, \tau)\, d\tau$. By $(f_1)$-$(f_4)$ it follows that
$\lim_{t{\rightarrow}0} \frac{g(x, t)}{t}=0$ uniformly in $x\in {\mathbb{R}}^{N}$,
$g(x, t)\leq f(t)$ for all $x\in {\mathbb{R}}^{N}$, $t>0$,
$(i)$ $0< \theta G(x,t)\leq tg(x, t)$ for all $x\in \Lambda$ and $t>0$,\
$(ii)$ $0\leq 2 G(x,t)\leq tg(x, t)\leq \frac{V_{1}}{\kappa} t^{2}$ for all $x\in \Lambda^{c}={\mathbb{R}}^{N}\setminus \Lambda$ and $t>0$,
for each $x\in \Lambda$ the function $t\mapsto \frac{g(x, t)}{t}$ is increasing in $(0, \infty)$, and for each $x\in \Lambda^{c}$ the function $t\mapsto \frac{g(x, t)}{t}$ is increasing in $(0, a)$.
Consider the following modified problem: $$\begin{aligned}
\label{MEP}
\left\{
\begin{array}{ll}
-\operatorname{div}(y^{1-2s} \nabla u)+m^{2}y^{1-2s}u=0 &\mbox{ in } {\mathbb{R}}^{N+1}_{+}, \\
\frac{\partial u}{\partial \nu^{1-2s}}=-V_{\operatorname{\varepsilon}}(x) u(\cdot, 0)+g_{\operatorname{\varepsilon}}(\cdot, u(\cdot, 0)) &\mbox{ in } {\mathbb{R}}^{N},
\end{array}
\right.\end{aligned}$$ where we set $g_{\operatorname{\varepsilon}}(x,t):=g(\operatorname{\varepsilon}x, t)$. Obviously, if $u_{\operatorname{\varepsilon}}$ is a positive solution of satisfying $u_{\operatorname{\varepsilon}}(x,0)< a$ for all $x\in \Lambda_{\operatorname{\varepsilon}}^{c}$, then $u_{\operatorname{\varepsilon}}$ is indeed a solution of . The corresponding energy functional is defined as $$J_{\operatorname{\varepsilon}}(u):=\frac{1}{2}\|u\|^{2}_{\operatorname{\varepsilon}}-\int_{{\mathbb{R}}^{N}} G_{\operatorname{\varepsilon}}(x,u(x,0))\, dx.
$$ Clearly, $J_{\operatorname{\varepsilon}}\in C^{1}(X_{\operatorname{\varepsilon}}, {\mathbb{R}})$ and its differential is given by: $$\begin{aligned}
\langle J'_{\operatorname{\varepsilon}}(u), v\rangle&=\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla u\nabla v+m^{2}u v)\, dx dy+\int_{{\mathbb{R}}^{N}} V_{\operatorname{\varepsilon}}(x) u(x, 0) v(x, 0)\, dx\\
&\quad -\int_{{\mathbb{R}}^{N}} g_{\operatorname{\varepsilon}}(x, u(x, 0)) v(x, 0)\, dx\end{aligned}$$ for all $u, v\in X_{\operatorname{\varepsilon}}$.
We start by proving that $J_{\operatorname{\varepsilon}}$ satisfies all the assumptions of the mountain pass theorem [@AR].
\[lemma1\] $J_{\operatorname{\varepsilon}}$ has a mountain pass geometry, that is:
- there exist $\alpha, \rho>0$ such that $J_{\operatorname{\varepsilon}}(u)\geq \alpha$ for all $u\in X$ such that $\|u\|=\rho$,
- there exists $v\in X$ such that $\|v\|>\rho$ and $J_{\operatorname{\varepsilon}}(v)<0$.
By $(f_1)$, $(f_2)$, $(g_1)$, $(g_2)$, we deduce that for all $\eta>0$ there exists $C_{\eta}>0$ such that $$\label{growthg}
|g_{\operatorname{\varepsilon}}(x, t)|\leq \eta |t|+C_{\eta}|t|^{{2^{*}_{s}}-1} \mbox{ for all } x\in {\mathbb{R}}^{N}, t>0,$$ and $$\label{growthG}
|G_{\operatorname{\varepsilon}}(x, t)|\leq \frac{\eta}{2} |t|^{2}+\frac{C_{\eta}}{{2^{*}_{s}}}|t|^{{2^{*}_{s}}} \mbox{ for all } x\in {\mathbb{R}}^{N}, t>0.$$ Fix $\eta\in (0, m^{2s}-V_{1})$. Using , , and Theorem \[Sembedding\] we have $$\begin{aligned}
J_{\operatorname{\varepsilon}}(u)&\geq \frac{1}{2} \|u\|^{2}_{\operatorname{\varepsilon}}-\frac{\eta}{2} |u(\cdot, 0)|_{2}^{2}-\frac{C_{\eta}}{{2^{*}_{s}}} |u(\cdot, 0)|^{{2^{*}_{s}}}_{{2^{*}_{s}}} \\
&=\frac{1}{2} \|u\|^{2}_{\operatorname{\varepsilon}}-\frac{\eta}{2m^{2s}} m^{2s} |u(\cdot, 0)|_{2}^{2}-\frac{C_{\eta}}{{2^{*}_{s}}} |u(\cdot, 0)|^{{2^{*}_{s}}}_{{2^{*}_{s}}} \\
&\geq \frac{1}{2}\|u\|^{2}_{\operatorname{\varepsilon}}-\frac{\eta}{2m^{2s}} \|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-C_{\eta}C\|u\|^{{2^{*}_{s}}}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}} \\
&\geq \left(\frac{1}{2}- \frac{\eta}{2(m^{2s}-V_{1})} \right) \|u\|^{2}_{\operatorname{\varepsilon}}-C_{\eta}C\|u\|^{{2^{*}_{s}}}_{\operatorname{\varepsilon}}
$$ from which we deduce the thesis. Now, let $v\in C^{\infty}_{c}({\mathbb{R}}^{N+1}_{+})$ such that $\operatorname{supp}(v(\cdot, 0))\subset \Lambda_{\operatorname{\varepsilon}}$. Then, by $(f_3)$, we have $$\begin{aligned}
J_{\operatorname{\varepsilon}}(tv)&=\frac{t^{2}}{2} \|v\|^{2}_{\operatorname{\varepsilon}}-\int_{{\mathbb{R}}^{N}} F(t v(x, 0))\, dx\\
&\leq \frac{t^{2}}{2} \|v\|^{2}_{\operatorname{\varepsilon}}-t^{\theta}\int_{\Lambda_{\operatorname{\varepsilon}}} |v(x, 0)|^{\theta}\, dx+C|\Lambda_{\operatorname{\varepsilon}}|{\rightarrow}-\infty \mbox{ as } t{\rightarrow}\infty.\end{aligned}$$
By Lemma \[lemma1\] and using a variant of mountain pass lemma without Palais-Smale condition [@W], we can find a Palais-Smale sequence $(u_{n})\subset X_{\operatorname{\varepsilon}}$ such that $$\begin{aligned}
J_{\operatorname{\varepsilon}}(u_{n}){\rightarrow}c_{\operatorname{\varepsilon}} \quad J'_{\operatorname{\varepsilon}}(u_{n}){\rightarrow}0,\end{aligned}$$ where $$c_{\operatorname{\varepsilon}}:=\inf_{\gamma\in \Gamma_{\operatorname{\varepsilon}}}\max_{t\in [0, 1]}J_{\operatorname{\varepsilon}}(\gamma(t))$$ and $$\Gamma_{\operatorname{\varepsilon}}:=\{\gamma\in C([0, 1], X_{\operatorname{\varepsilon}}): \gamma(0)=0, \, J_{\operatorname{\varepsilon}}(\gamma(1))\leq 0\}.$$ In view of $(f_4)$, it is standard to check (see [@W]) that $$c_{\operatorname{\varepsilon}}=\inf_{u\in \mathcal{N}_{\operatorname{\varepsilon}}} J_{\operatorname{\varepsilon}}(u)=\inf_{u\in X_{\operatorname{\varepsilon}}\setminus\{0\}} \sup_{t\geq 0}J_{\operatorname{\varepsilon}}(t u),$$ where $$\mathcal{N}_{\operatorname{\varepsilon}}:=\{u\in X_{\operatorname{\varepsilon}}: \langle J'_{\operatorname{\varepsilon}}(u), u\rangle=0\}$$ is the Nehari manifold associated with $J_{\operatorname{\varepsilon}}$. Now we prove the following fundamental compactness result:
\[lemma2\] $J_{\operatorname{\varepsilon}}$ satisfies the Palais-Smale condition at any level $c\in {\mathbb{R}}$.
Let $c\in {\mathbb{R}}$ and $(u_{n})\subset X_{\operatorname{\varepsilon}}$ be such that $$\begin{aligned}
J_{\operatorname{\varepsilon}}(u_{n}){\rightarrow}c \quad J'_{\operatorname{\varepsilon}}(u_{n}){\rightarrow}0.\end{aligned}$$ Then, using $(g_3)$, and , we have $$\begin{aligned}
J_{\operatorname{\varepsilon}}(u_{n})&-\frac{1}{\theta}\langle J'_{\operatorname{\varepsilon}}(u_{n}), u_{n}\rangle\\
&=\left(\frac{1}{2}-\frac{1}{\theta}\right) \|u_{n}\|^{2}_{\operatorname{\varepsilon}}+\frac{1}{\theta} \int_{{\mathbb{R}}^{N}} g_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))u_{n}(x, 0)-\theta G_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))\, dx\\
&\geq \left(\frac{1}{2}-\frac{1}{\theta}\right) \|u_{n}\|^{2}_{\operatorname{\varepsilon}}-\left(\frac{1}{2}-\frac{1}{\theta}\right) \frac{V_{1}}{\kappa} \int_{{\mathbb{R}}^{N}} u_{n}^{2}(x, 0)\, dx \\
&= \left(\frac{1}{2}-\frac{1}{\theta}\right) \|u_{n}\|^{2}_{\operatorname{\varepsilon}}-\left(\frac{1}{2}-\frac{1}{\theta}\right) \frac{V_{1}}{\kappa m^{2s}} m^{2s}\int_{{\mathbb{R}}^{N}} u_{n}^{2}(x, 0)\, dx \\
&\geq \left(\frac{1}{2}-\frac{1}{\theta}\right) \|u_{n}\|^{2}_{\operatorname{\varepsilon}}-\left(\frac{1}{2}-\frac{1}{\theta}\right) \frac{V_{1}}{\kappa m^{2s}}\|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}} \\
&\geq \left(\frac{1}{2}-\frac{1}{\theta}\right) \left(1- \frac{V_{1}}{\kappa(m^{2s}-V_{1})}\right) \|u_{n}\|^{2}_{\operatorname{\varepsilon}}.\end{aligned}$$ Since $\theta>2$ and $\kappa>\frac{V_{1}}{m^{2s}-V_{1}}$, we deduce that $(u_{n})$ is bounded in $X_{\operatorname{\varepsilon}}$. Hence, up to a subsequence, we may assume that $u_{n}\rightharpoonup u$ in $X_{\operatorname{\varepsilon}}$. Now we prove that this convergence is indeed strong. Using the fact that $g$ has subcritical growth and applying Theorem \[Sembedding\], it is easy to check that $\langle J'_{\operatorname{\varepsilon}}(u), \varphi\rangle=0$ for all $\varphi\in X_{\operatorname{\varepsilon}}$. In particular, $$\begin{aligned}
\label{AM1}
&\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|u(\cdot, 0)|^{2}_{2}+\int_{\Lambda_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)\, dx+\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u(x, 0))u(x,0)]\, dx \nonumber\\
&=\int_{\Lambda_{\operatorname{\varepsilon}}} f(u(x, 0)) u(x, 0)\, dx.\end{aligned}$$ On the other hand, by $\langle J'_{\operatorname{\varepsilon}}(u_{n}), u_{n}\rangle=o_{n}(1)$, we get $$\begin{aligned}
\label{AM2}
&\|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|u_{n}(\cdot, 0)|^{2}_{2}+\int_{\Lambda_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}_{n}(x,0)\, dx+\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))u_{n}(x,0)]\, dx \nonumber\\
&=\int_{\Lambda_{\operatorname{\varepsilon}}} f(u_{n}(x, 0)) u_{n}(x, 0)\, dx+o_{n}(1).\end{aligned}$$ Since $\Lambda_{\operatorname{\varepsilon}}$ is bounded, by the compactness of Sobolev embeddings in Theorem \[Sembedding\] we have $$\begin{aligned}
\label{AM3}
\lim_{n{\rightarrow}\infty} \int_{\Lambda_{\operatorname{\varepsilon}}} f(u_{n}(x, 0)) u_{n}(x, 0)\, dx=\int_{\Lambda_{\operatorname{\varepsilon}}} f(u(x, 0)) u(x, 0)\, dx,\end{aligned}$$ $$\begin{aligned}
\label{AM4}
\lim_{n{\rightarrow}\infty} \int_{\Lambda_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)\, dx=\int_{\Lambda_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_1) u^{2}(x,0)\, dx.\end{aligned}$$ In view of , , , , we obtain $$\begin{aligned}
&\limsup_{n{\rightarrow}\infty} \left(\|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u_{n}(\cdot, 0)|^{2}_{2}+\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))u_{n}(x,0)]\, dx \right)\\
&=\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u(\cdot, 0)|^{2}_{2}+\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u(x, 0))u(x,0)]\, dx.\end{aligned}$$ On the other hand, by $(g_2)$ and Fatou’s Lemma, we get $$\begin{aligned}
&\liminf_{n{\rightarrow}\infty} \left( \|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u_{n}(\cdot, 0)|^{2}_{2}+\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))u_{n}(x,0)]\, dx \right)\\
&\geq \|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u(\cdot, 0)|^{2}_{2} +\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u(x, 0))u(x,0)]\, dx.\end{aligned}$$ Hence, $$\begin{aligned}
\label{AM5}
\lim_{n{\rightarrow}\infty} \|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u_{n}(\cdot, 0)|^{2}_{2}=\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1} |u(\cdot, 0)|^{2}_{2}\end{aligned}$$ and $$\begin{aligned}
\lim_{n{\rightarrow}\infty} \int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u_{n}(x, 0))u_{n}(x,0)]\, dx=\int_{\Lambda^{c}_{\operatorname{\varepsilon}}} [(V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)-g_{\operatorname{\varepsilon}}(x, u(x, 0))u(x,0)]\, dx.\end{aligned}$$ The last limit combined with definition of $g_{\operatorname{\varepsilon}}$ yields $$\begin{aligned}
\lim_{n{\rightarrow}\infty} \int_{\Lambda^{c}_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)\, dx=\int_{\Lambda_{\operatorname{\varepsilon}}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)\, dx\end{aligned}$$ which together with implies that $$\begin{aligned}
\label{AM6}
\lim_{n{\rightarrow}\infty} \int_{{\mathbb{R}}^{N}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u_{n}^{2}(x,0)\, dx=\int_{{\mathbb{R}}^{N}} (V_{\operatorname{\varepsilon}}(x)+V_{1}) u^{2}(x,0)\, dx.\end{aligned}$$ From and we have $$\lim_{n{\rightarrow}\infty} \|u_{n}\|^{2}_{\operatorname{\varepsilon}}=\|u\|^{2}_{\operatorname{\varepsilon}}$$ and since $X_{\operatorname{\varepsilon}}$ turns out to be a Hilbert space we deduce that $u_{n}{\rightarrow}u$ in $X_{\operatorname{\varepsilon}}$.
In light of Lemma \[lemma1\] and Lemma \[lemma2\], it follows from the mountain pass theorem [@AR] that for all $\operatorname{\varepsilon}>0$ there exists $u_{\operatorname{\varepsilon}}\in X_{\operatorname{\varepsilon}}\setminus\{0\}$ such that $$\begin{aligned}
\label{AM2.8}
J_{\operatorname{\varepsilon}}(u_{\operatorname{\varepsilon}})=c_{\operatorname{\varepsilon}} \mbox{ and } J'_{\operatorname{\varepsilon}}(u_{\operatorname{\varepsilon}})=0.\end{aligned}$$ Since $f(t)=0$ for $t\leq 0$, it follows from $\langle J'_{\operatorname{\varepsilon}}(u_{\operatorname{\varepsilon}}), u_{\operatorname{\varepsilon}}^{-}\rangle=0$, where $u_{\operatorname{\varepsilon}}^{-}:=\min\{u_{\operatorname{\varepsilon}}, 0\}$, that $u_{\operatorname{\varepsilon}}\geq 0$ in ${\mathbb{R}}^{N+1}_{+}$.
Next, for all $\mu>-m^{2s}$, we consider the following family of autonomous problems related to , namely $$\begin{aligned}
\label{AEP}
\left\{
\begin{array}{ll}
-\operatorname{div}(y^{1-2s} \nabla w)+m^{2}y^{1-2s}w=0 &\mbox{ in } {\mathbb{R}}^{N+1}_{+}, \\
\frac{\partial w}{\partial \nu^{1-2s}}= -\mu w(\cdot, 0)+f(w(\cdot, 0)) &\mbox{ in } {\mathbb{R}}^{N}.
\end{array}
\right.\end{aligned}$$ Define the energy functionals $L_{\mu}: {X^{s}({\mathbb{R}}^{N+1}_{+})}{\rightarrow}{\mathbb{R}}$ given by $$L_{\mu}(u):=\frac{1}{2}\|u\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}+\frac{\mu}{2}|u(\cdot, 0)|^{2}_{2}-\int_{{\mathbb{R}}^{N}} F(u(x, 0))\, dx,$$ and denote by $\mathcal{M}_{\mu}$ the Nehari manifold associated with $L_{\mu}$ , that is $\mathcal{M}_{\mu}:=\{u\in {X^{s}({\mathbb{R}}^{N+1}_{+})}: \langle L'_{\mu}(u), u\rangle=0\}$. Arguing as in the case $\operatorname{\varepsilon}>0$, one can check that $L_{\mu}$ has a mountain pass geometry [@AR], so we can find a Palais-Smale sequence $(u_{n})\subset {X^{s}({\mathbb{R}}^{N+1}_{+})}$ at the mountain pass level $d_{\mu}$ of $L_{\mu}$. As before, $(u_{n})$ is bounded in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$. We also note that, by $(f_4)$, it holds $$d_{\mu}=\inf_{u\in \mathcal{M}_{\mu}} L_{\mu}(u)=\inf_{u\in {X^{s}({\mathbb{R}}^{N+1}_{+})}\setminus\{0\}} \sup_{t\geq 0}L_{\mu}(t u).$$ Next we show the existence of a ground state solution to . We first prove some useful technical lemmas. The first one is a vanishing Lions type result.
\[Lions\] Let $t\in [2, {2^{*}_{s}})$ and $R>0$. If $(u_{n})\subset {X^{s}({\mathbb{R}}^{N+1}_{+})}$ is a bounded sequence such that $$\lim_{n{\rightarrow}\infty} \sup_{z\in {\mathbb{R}}^{N}} \int_{B_{R}(z)} |u_{n}(x, 0)|^{t}\, dx=0,$$ then $u_{n}(\cdot, 0){\rightarrow}0$ in $L^{q}({\mathbb{R}}^{N})$ for all $q\in (t, {2^{*}_{s}})$.
Take $q\in (t, {2^{*}_{s}})$. Given $R>0$ and $z\in {\mathbb{R}}^{N}$, by using Hölder inequality, we get $$\begin{aligned}
|u_{n}(\cdot, 0)|_{L^{q}(B_{R}(z))}&\leq |u_{n}(\cdot, 0)|^{1-\lambda}_{L^{t}(B_{R}(z))} |u_{n}(\cdot, 0)|^{\lambda}_{L^{{2^{*}_{s}}}(B_{R}(z))} \quad \forall n\in \mathbb{N},\end{aligned}$$ where $$\frac{1-\lambda}{t}+\frac{\lambda}{{2^{*}_{s}}}=\frac{1}{q}.$$ Now, covering ${\mathbb{R}}^{N}$ by balls of radius $R$, in such a way that each point of ${\mathbb{R}}^{N}$ is contained in at most $N+1$ balls, we find $$\begin{aligned}
|u_{n}(\cdot, 0)|^{q}_{q}\leq (N+1) |u_{n}(\cdot, 0)|^{(1-\lambda)q}_{L^{t}(B_{R}(z))} |u_{n}(\cdot, 0)|^{\lambda q}_{{2^{*}_{s}}},\end{aligned}$$ which combined with Theorem \[Sembedding\] and the assumptions yields $$\begin{aligned}
|u_{n}(\cdot, 0)|^{q}_{q}&\leq C(N+1) |u_{n}(\cdot, 0)|^{(1-\lambda)q}_{L^{t}(B_{R}(z))} \|u_{n}\|^{\lambda q}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}} \\
&\leq C(N+1) \sup_{z\in {\mathbb{R}}^{N}}|u_{n}(\cdot, 0)|^{(1-\lambda)q}_{L^{t}(B_{R}(z))}{\rightarrow}0 \mbox{ as } n{\rightarrow}\infty.\end{aligned}$$ This ends the proof of lemma.
\[Lions2\] Let $(u_{n})\subset {X^{s}({\mathbb{R}}^{N+1}_{+})}$ be a Palais-Smale sequence at level $c\in {\mathbb{R}}$ and such that $u_{n}\rightharpoonup 0$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$. Then we have either
$u_{n}{\rightarrow}0$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$, or
there exists a sequence $(z_{n})\subset {\mathbb{R}}^{N}$ and constants $R, \beta>0$ such that $$\liminf_{n{\rightarrow}\infty} \int_{B_{R}(z_{n})} |u_{n}(x, 0)|^{2}\, dx\geq \beta.$$
Assume that $(b)$ does not occur. Then, for all $R>0$, we have $$\lim_{n{\rightarrow}\infty} \sup_{z\in {\mathbb{R}}^{N}} \int_{B_{R}(z)} |u_{n}(x, 0)|^{2}\, dx=0.$$ Using Lemma \[Lions\] (with $t=2$), we can see that $u_{n}(\cdot, 0){\rightarrow}0$ in $L^{q}({\mathbb{R}}^{N})$ for all $q\in (2, {2^{*}_{s}})$. This fact and $(f_1)$-$(f_2)$ imply that $$\int_{{\mathbb{R}}^{N}}f(u_{n}(x, 0))u_{n}(x, 0)\, dx{\rightarrow}0 \mbox{ as } n{\rightarrow}\infty.$$ Hence, using $\langle L'_{\mu}(u_{n}), u_{n}\rangle=o_{n}(1)$, $\mu>-m^{2s}$ and , we get $$\left(1+\frac{\mu}{m^{2s}}\right)\|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\leq \|u_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}+\mu |u_{n}(\cdot, 0)|^{2}_{2}\leq o_{n}(1)$$ that is $u_{n}{\rightarrow}0$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ as $n{\rightarrow}\infty$.
Now we prove the following existence result for .
\[EGS\] Let $\mu>-m^{2s}$. Then has a positive ground state solution.
Since $L_{\mu}$ has a mountain pass geometry [@AR], we can find a Palais-Smale sequence $(u_{n})\subset {X^{s}({\mathbb{R}}^{N+1}_{+})}$ at the level $d_{\mu}$. Thus $(u_{n})$ is bounded in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ and there exists $u\in {X^{s}({\mathbb{R}}^{N+1}_{+})}$ such that $u_{n}\rightharpoonup u$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$. Clearly, $L'_{\mu}(u)=0$. If $u=0$, then we can use Lemma \[Lions2\] to deduce that for some sequence $(z_{n})\subset {\mathbb{R}}^{N}$, $v_{n}(x, y):=u_{n}(x+z_{n}, y)$ is a bounded Palais-Smale sequence at the level $d_{\mu}$ and having a nontrivial weak limit $v$. Hence, $v\in \mathcal{M}_{\mu}$, where $\mathcal{M}_{\mu}$ denotes the Nehari manifold associated with $L_{\mu}$. Moreover, using the weak lower semicontinuity of $\|\cdot\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}+\mu |\cdot|_{2}$, $(f_4)$ and Fatou’s lemma, it is easy to see that $L_{\mu}(v)=d_{\mu}$. When $u\neq 0$, as before, we can deduce that $u$ is a ground state solution to . In conclusion, for any $\mu>-m^{2s}$, there exists a ground state solution $w=w_{\mu}\in {X^{s}({\mathbb{R}}^{N+1}_{+})}\setminus\{0\}$ such that $$\begin{aligned}
L_{\mu}(w)=d_{\mu} \mbox{ and } L'_{\mu}(w)=0.\end{aligned}$$ Since $f(t)=0$ for $t\leq 0$, we deduce that $w\geq 0$ in ${\mathbb{R}}^{N+1}_{+}$ and $w\not\equiv 0$. A Moser iteration argument (see Lemma \[moser\] below) shows that $w(\cdot, 0)\in L^{p}({\mathbb{R}}^{N})$ for all $p\in [2, \infty]$ and $w\in L^{\infty}({\mathbb{R}}^{N+1}_{+})$. Using Proposition \[PROPFF\]-$(iii)$ we obtain that $w\in C^{0, \alpha}(\overline{{\mathbb{R}}^{N+1}_{+}})$ for some $\alpha\in (0, 1)$. By Proposition \[PROPFF\]-$(ii)$ we conclude that $w$ is positive.
In the next lemma we establish an important connection between $c_{\operatorname{\varepsilon}}$ and $d_{V(0)}=d_{-V_{0}}$ (we remark that $V(0)=-V_{0}>-m^{2s}$):
\[lem2.3AM\] The numbers $c_{\operatorname{\varepsilon}}$ and $d_{V(0)}$ verify the following inequality: $$\limsup_{\operatorname{\varepsilon}{\rightarrow}0} c_{\operatorname{\varepsilon}}\leq d_{V(0)}.$$
By Theorem \[EGS\], we know that there exists a positive ground state solution $w$ to with $\mu=V(0)$. Let $\eta\in C^{\infty}_{c}({\mathbb{R}})$ be such that $0\leq \eta\leq 1$, $\eta=1$ in $[-1, 1]$ and $\eta=0$ in ${\mathbb{R}}\setminus (-2, 2)$. Suppose that $B_{2}(0)\subset \Lambda$. Define $w_{\operatorname{\varepsilon}}(x, y)=\eta(\operatorname{\varepsilon}|(x, y)|) w(x, y)$ and note that $\operatorname{supp}(w_{\operatorname{\varepsilon}}(\cdot, 0))\subset \Lambda_{\operatorname{\varepsilon}}$. It is easy to prove that $w_{\operatorname{\varepsilon}}{\rightarrow}w$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ and $L_{V(0)}(w_{\operatorname{\varepsilon}}){\rightarrow}L_{V(0)}(w)$ as $\operatorname{\varepsilon}{\rightarrow}0$. On the other hand, by definition of $c_{\operatorname{\varepsilon}}$, we have $$\begin{aligned}
\label{15ADOM}
c_{\operatorname{\varepsilon}}\leq \max_{t\geq 0} J_{\operatorname{\varepsilon}}(t w_{\operatorname{\varepsilon}})=J_{\operatorname{\varepsilon}}(t_{\operatorname{\varepsilon}} w_{\operatorname{\varepsilon}})=\frac{t^{2}_{\operatorname{\varepsilon}}}{2} \|w_{\operatorname{\varepsilon}}\|^{2}_{\operatorname{\varepsilon}}-\int_{{\mathbb{R}}^{N}} F(t w_{\operatorname{\varepsilon}}(x, 0))\, dx\end{aligned}$$ for some $t_{\operatorname{\varepsilon}}>0$. Recalling that $w\in \mathcal{M}_{V(0)}$ and using $(f_4)$, it is easy to check that $t_{\operatorname{\varepsilon}}{\rightarrow}1$ as $\operatorname{\varepsilon}{\rightarrow}0$. Note that $$\begin{aligned}
\label{16ADOM}
J_{\operatorname{\varepsilon}}(t_{\operatorname{\varepsilon}}w_{\operatorname{\varepsilon}})=L_{V(0)}(t_{\operatorname{\varepsilon}}w_{\operatorname{\varepsilon}})+\frac{t^{2}_{\operatorname{\varepsilon}}}{2}\int_{{\mathbb{R}}^{N}} (V_{\operatorname{\varepsilon}}(x)-V(0)) w_{\operatorname{\varepsilon}}^{2}(x, 0)\, dx.\end{aligned}$$ Since $V_{\operatorname{\varepsilon}}(x)$ is bounded on the support of $w_{\operatorname{\varepsilon}}(\cdot, 0)$, and $V_{\operatorname{\varepsilon}}(x){\rightarrow}V(0)$ as $\operatorname{\varepsilon}{\rightarrow}0$, we can apply the dominated convergence theorem and use and to conclude the proof.
Now we come back to study and consider the mountain pass solutions $u_{\operatorname{\varepsilon}}$ satisfying .
\[lem2.4AM\] There exist $R, \beta, \operatorname{\varepsilon}^{*}>0$ and $(y_{\operatorname{\varepsilon}})\subset {\mathbb{R}}^{N}$ such that $$\int_{B_{r}(y_{\operatorname{\varepsilon}})} u_{\operatorname{\varepsilon}}^{2}(x, 0)\, dx\geq \beta, \, \forall \operatorname{\varepsilon}\in (0, \operatorname{\varepsilon}^{*}).$$
Since $u_{\operatorname{\varepsilon}}$ verifies , it follows from the growth assumptions on $f$ that there exist $\alpha>0$ independent of $\operatorname{\varepsilon}>0$ such that $$\label{contradiction}
\|u_{\operatorname{\varepsilon}}\|_{\operatorname{\varepsilon}}^{2}\geq \alpha \quad \forall \operatorname{\varepsilon}>0.$$ Let $(\operatorname{\varepsilon}_{n})\subset (0, \infty)$ be such that $\operatorname{\varepsilon}_{n}{\rightarrow}0$. If by contradiction there exists $r>0$ such that $$\lim_{n{\rightarrow}\infty}\sup_{y\in {\mathbb{R}}^{N}} \int_{B_{r}(y)} u_{\operatorname{\varepsilon}_{n}}^{2}(x, 0)\, dx=0,$$ we can use Lemma \[Lions\] to deduce that $u_{\operatorname{\varepsilon}_{n}}(\cdot, 0){\rightarrow}0$ in $L^{q}({\mathbb{R}}^{N})$ for all $q\in (2, {2^{*}_{s}})$. Then, and the growth assumptions on $f$ imply that $\|u_{\operatorname{\varepsilon}_{n}}\|_{\operatorname{\varepsilon}_{n}}{\rightarrow}0$ as $n{\rightarrow}\infty$ which contradicts .
\[lem2.5AM\] For any $\operatorname{\varepsilon}_{n}{\rightarrow}0$, consider the sequence $(y_{\operatorname{\varepsilon}_{n}})\subset {\mathbb{R}}^{N}$ given in Lemma \[lem2.4AM\] and $w_{n}(x, y)=u_{\operatorname{\varepsilon}_{n}}(x+y_{\operatorname{\varepsilon}_{n}}, y)$. Then there exists a subsequence of $w_{n}$, still denoted by itself, and $w\in {X^{s}({\mathbb{R}}^{N+1}_{+})}\setminus \{0\}$ such that $$\begin{aligned}
w_{n}{\rightarrow}w \mbox{ in } {X^{s}({\mathbb{R}}^{N+1}_{+})}.\end{aligned}$$ Moreover, there exists $x_{0}\in \Lambda$ such that $$\begin{aligned}
\operatorname{\varepsilon}_{n}y_{\operatorname{\varepsilon}_{n}}{\rightarrow}x_{0} \mbox{ and } V(x_{0})=-V_{0}.\end{aligned}$$
In what follows, we denote by $(y_{n})$ and $(u_{n})$, the sequences $(y_{\operatorname{\varepsilon}_n})$ and $(u_{\operatorname{\varepsilon}_n})$ respectively. Since each $u_{n}$ satisfies , we can argue as in the proof of Lemma \[lemma2\] to deduce that $(u_{n})$ is bounded in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$. Thus $(w_{n})$ is bounded in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ and there exists $w\in {X^{s}({\mathbb{R}}^{N+1}_{+})}\setminus \{0\}$ such that $$\begin{aligned}
\label{2.15AM}
w_{n}\rightharpoonup w \mbox{ in } {X^{s}({\mathbb{R}}^{N+1}_{+})}\mbox{ as } n{\rightarrow}\infty.\end{aligned}$$ Moreover, by Lemma \[lem2.4AM\], we know that $$\begin{aligned}
\label{2.16AM}
\int_{B_{r}(0)} w^{2}(x, 0)\, dx\geq \beta>0.\end{aligned}$$ Next we show that $(\operatorname{\varepsilon}_{n}y_{n})$ is bounded in ${\mathbb{R}}$. First of all, we prove that $$\begin{aligned}
\label{Claim1}
{\mathit{dist}}(\operatorname{\varepsilon}_{n}y_{n}, \overline{\Lambda}){\rightarrow}0 \mbox{ as } n{\rightarrow}\infty.\end{aligned}$$ If does not hold, there exists $\delta>0$ and a subsequence of $(\operatorname{\varepsilon}_{n}y_{n})$, still denoted by itself, such that $${\mathit{dist}}(\operatorname{\varepsilon}_{n}y_{n}, \overline{\Lambda})\geq \delta \quad \forall n\in \mathbb{N}.$$ Then there is $R>0$ such that $B_{R}(\operatorname{\varepsilon}_{n}y_{n})\subset \Lambda^{c}$ for all $n\in \mathbb{N}$. By the definition of ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ and using the fact that $w\geq 0$, we know that there exists $(\psi_{j})\subset {X^{s}({\mathbb{R}}^{N+1}_{+})}$ such that $\psi_{j}$ has compact support in $\overline{\mathbb{R}^{N+1}_{+}}$ and $\psi_{j}{\rightarrow}w$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ as $j{\rightarrow}\infty$. Fix $j\in \mathbb{N}$. Taking $\psi_{j}$ as test function in $\langle J'_{\operatorname{\varepsilon}}(u_{n}), \phi\rangle=0$ for all $\phi\in X_{\operatorname{\varepsilon}}$, we get $$\begin{aligned}
\label{2.17AM}
&\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w_{n}\nabla \psi_{j}+m^{2}w_{n}\psi_{j})\, dx dy+\int_{{\mathbb{R}}^{N}} V(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}) w_{n}(x, 0) \psi_{j}(x, 0)\, dx \nonumber\\
&=\int_{{\mathbb{R}}^{N}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\psi_{j}(x, 0)\, dx.\end{aligned}$$ By the definition of $g_{\operatorname{\varepsilon}}$ there holds $$\begin{aligned}
\int_{{\mathbb{R}}^{N}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\psi_{j}(x, 0)\, dx&\leq \int_{B_{\frac{R}{\operatorname{\varepsilon}_{n}}}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\psi_{j}(x, 0)\, dx \\
&+\int_{B^{c}_{\frac{R}{\operatorname{\varepsilon}_{n}}}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\psi_{j}(x, 0)\, dx \\
&\leq \frac{V_{1}}{\kappa} \int_{B_{\frac{R}{\operatorname{\varepsilon}_{n}}}} w_{n}(x, 0) \psi_{j}(x, 0)\, dx+\int_{B^{c}_{\frac{R}{\operatorname{\varepsilon}_{n}}}} f(w_{n}(x, 0))\psi_{j}(x, 0)\, dx.\end{aligned}$$ Using $(V_1)$ and we can see that $$\begin{aligned}
&\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w_{n}\nabla \psi_{j}+m^{2}w_{n}\psi_{j})\, dx dy-V_{1}\left(1+\frac{1}{\kappa}\right) \int_{{\mathbb{R}}^{N}} w_{n}(x, 0) \psi_{j}(x, 0)\, dx \\
&\leq \int_{B_{\frac{R}{\operatorname{\varepsilon}_{n}}}^{c}} f(w_{n}(x, 0))\psi_{j}(x, 0)\, dx.\end{aligned}$$ Taking into account that $\psi_{j}$ has compact support, $\operatorname{\varepsilon}_{n}{\rightarrow}0$, the growth assumptions on $f$, and , we deduce that as $n{\rightarrow}\infty$ $$\begin{aligned}
\int_{B_{\frac{R}{\operatorname{\varepsilon}_{n}}}^{c}} f(w_{n}(x, 0))\psi_{j}(x, 0)\, dx{\rightarrow}0,\end{aligned}$$ and $$\begin{aligned}
&\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w_{n}\nabla \psi_{j}+m^{2}w_{n}\psi_{j})\, dx dy-V_{1}\left(1+\frac{1}{\kappa}\right) \int_{{\mathbb{R}}^{N}} w_{n}(x, 0) \psi_{j}(x, 0)\, dx \\
&{\rightarrow}\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w\nabla \psi_{j}+m^{2}w\psi_{j})\, dx dy-V_{1}\left(1+\frac{1}{\kappa}\right) \int_{{\mathbb{R}}^{N}} w(x, 0) \psi_{j}(x, 0)\, dx.\end{aligned}$$ The previous relations of limits combined with give $$\begin{aligned}
\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w\nabla \psi_{j}+m^{2}w\psi_{j})\, dx dy-V_{1}\left(1+\frac{1}{\kappa}\right) \int_{{\mathbb{R}}^{N}} w(x, 0) \psi_{j}(x, 0)\, dx\leq 0\end{aligned}$$ and passing to the limit as $j{\rightarrow}\infty$ we find $$\begin{aligned}
\|w\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}\left(1+\frac{1}{\kappa}\right) |w(\cdot, 0)|^{2}_{2} dx\leq 0.\end{aligned}$$ Thus and $\kappa>\frac{V_{1}}{m^{2s}-V_{1}}$ yield $$0\leq \left(1-\frac{V_{1}}{m^{2s}} \left(1+\frac{1}{\kappa}\right) \right)\|w\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\leq 0$$ that is $w\equiv 0$ in ${\mathbb{R}}^{N}$ and this is in contrast with . Consequently, there exist a subsequence of $(\operatorname{\varepsilon}_{n}y_{n})$, still denoted by itself, and $x_{0}\in \overline{\Lambda}$ such that $\operatorname{\varepsilon}_{n}y_{n}{\rightarrow}x_{0}$ as $n{\rightarrow}\infty$. Next we prove that $x_{0}\in \Lambda$.
Using $(g_2)$ and , we know that $$\begin{aligned}
\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w_{n}\nabla \psi_{j}+m^{2}w_{n}\psi_{j})\, dx dy+\int_{{\mathbb{R}}^{N}} V(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}) w_{n}(x, 0) \psi_{j}(x, 0)\, dx \leq \int_{{\mathbb{R}}^{N}} f(w_{n}(x, 0))\psi_{j}(x, 0)\, dx.\end{aligned}$$ Letting $n{\rightarrow}\infty$ and using and the continuity of $V$ we find $$\begin{aligned}
\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (\nabla w\nabla \psi_{j}+m^{2}w\psi_{j})\, dx dy+\int_{{\mathbb{R}}^{N}} V(x_{0}) w(x, 0) \psi_{j}(x, 0)\, dx \leq \int_{{\mathbb{R}}^{N}} f(w(x, 0)) \psi_{j}(x, 0)\, dx.\end{aligned}$$ By passing to the limit as $j{\rightarrow}\infty$ we obtain $$\begin{aligned}
\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} (|\nabla w|^{2}+m^{2}w^{2})\, dx dy+\int_{{\mathbb{R}}^{N}} V(x_{0}) w^{2}(x, 0)\, dx \leq \int_{{\mathbb{R}}^{N}} f(w(x, 0)) w(x, 0)\, dx.\end{aligned}$$ Hence there exists $t_{1}\in (0, 1)$ such that $t_{1}w\in \mathcal{M}_{V(x_{0})}$. In view of Lemma \[lem2.3AM\] we have $$\begin{aligned}
d_{V(x_{0})}\leq L_{V(x_{0})}(t_{1}w)\leq \liminf_{n{\rightarrow}\infty} J_{\operatorname{\varepsilon}_{n}}(u_{n})=\liminf_{n{\rightarrow}\infty} c_{\operatorname{\varepsilon}_{n}}\leq d_{V(0)}\end{aligned}$$ from which $d_{V(x_{0})}\leq d_{V(0)}$ and thus $V(x_{0})\leq V(0)=-V_{0}$. Since $-V_{0}=\inf_{x\in \overline{\Lambda}} V(x)$, we achieve that $V(x_{0})=-V_{0}$. Using $(V_2)$, we conclude that $x_{0}\in \Lambda$.
Finally, we show that $w_{n}{\rightarrow}w$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$. Set $$\tilde{\Lambda}_{n} := \frac{\Lambda - \operatorname{\varepsilon}_{n}\tilde{y}_{n}}{\operatorname{\varepsilon}_{n}}$$ and define $$\begin{aligned}
&\tilde{\chi}_{n}^{1}(x):= \left\{
\begin{array}{ll}
1 \, &\mbox{ if } x\in \tilde{\Lambda}_{n},\\
0 \, &\mbox{ if } x\in {\mathbb{R}}^{N}\setminus \tilde{\Lambda}_{n},
\end{array}
\right.\\
&\tilde{\chi}_{n}^{2}(x):= 1- \tilde{\chi}_{n}^{1}(x).\end{aligned}$$ Let us also consider the following functions for all $x\in {\mathbb{R}}^{N}$ and $n\in \mathbb{N}$ $$\begin{aligned}
&h_{n}^{1}(x):= \left(\frac{1}{2}-\frac{1}{\theta}\right) (V(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n})+V_{1}) w^{2}_{n}(x, 0) \tilde{\chi}_{n}^{1}(x)\\
&h^{1}(x):=\left(\frac{1}{2}-\frac{1}{\theta}\right) (V(x_{0})+V_{1}) w^{2}(x, 0) \\
&h_{n}^{2}(x)\!\!:=\!\!\left[ \left(\frac{1}{2}-\frac{1}{\theta}\right) (V(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n})+V_{1}) w^{2}_{n}(x, 0) + \frac{1}{\theta} g(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0)) w_{n}(x, 0) - G(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\right] \tilde{\chi}_{n}^{2}(x) \\
&\quad \quad \, \, \, \geq \left( \left(\frac{1}{2}-\frac{1}{\theta}\right) -\frac{1}{\kappa}\right) (V(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}\tilde{y}_{n})+V_{1}) w_{n}^{2}(x, 0) \tilde{\chi}_{n}^{2}(x) \\
&h_{n}^{3}(x):= \left(\frac{1}{\theta} g(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0)) w_{n}(x, 0) - G(\operatorname{\varepsilon}_{n}x+ \operatorname{\varepsilon}_{n}y_{n}, w_{n}(x, 0))\right) \tilde{\chi}_{n}^{1}(x) \\
&\quad \quad \, \, \, =\left[\frac{1}{\theta} \left(f(w_{n}(x, 0))w_{n}(x, 0)- F(w_{n}(x, 0))\right) \right] \tilde{\chi}_{n}^{1}(x) \\
&h^{3}(x):= \frac{1}{\theta} \left(f(w(x, 0))w(x, 0)- F(w(x, 0))\right). \end{aligned}$$ From $(f_3)$ and $(g_3)$, we see that the above functions are nonnegative in ${\mathbb{R}}^{N}$. Since $$\begin{aligned}
&w_{n}(x, 0) {\rightarrow}w(x, 0)\quad \mbox{ a.e. } x\in {\mathbb{R}}^{N}, \\
&\operatorname{\varepsilon}_{n}y_{n}{\rightarrow}x_{0}\in \Lambda,\end{aligned}$$ we get $$\begin{aligned}
&\tilde{\chi}_{n}^{1}(x){\rightarrow}1, \, h_{n}^{1}(x){\rightarrow}h^{1}(x), \, h_{n}^{2}(x){\rightarrow}0 \, \mbox{ and } \, h_{n}^{3}(x){\rightarrow}h^{3}(x) \, \mbox{ a.e. } x\in {\mathbb{R}}^{N}. \end{aligned}$$ Hence, observing that $\|\cdot\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-\mu |\cdot|_{2}^{2}$ is weakly lower semicontinuous for all $\mu\in (0, m^{2s})$, and using Fatou’s Lemma and the invariance of ${\mathbb{R}}^{N}$ by translation, we have $$\begin{aligned}
d_{V(0)} &\geq \limsup_{n{\rightarrow}\infty} c_{\operatorname{\varepsilon}_{n}} = \limsup_{n{\rightarrow}\infty} \left( J_{\operatorname{\varepsilon}_{n}}(u_{n}) - \frac{1}{\theta} \langle J'_{\operatorname{\varepsilon}_{n}}(u_{n}), u_{n}\rangle \right)\\
&\geq \limsup_{n{\rightarrow}\infty} \left\{\left(\frac{1}{2}-\frac{1}{\theta} \right)\left[\|w_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|w_{n}(\cdot, 0)|^{2}_{2}\right]+ \int_{{\mathbb{R}}^{N}} (h_{n}^{1}+ h_{n}^{2}+ h_{n}^{3}) \, dx\right\}\\
&\geq \liminf_{n{\rightarrow}\infty} \left\{\left(\frac{1}{2}-\frac{1}{\theta} \right)\left[\|w_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|w_{n}(\cdot, 0)|^{2}_{2}\right]+ \int_{{\mathbb{R}}^{N}} (h_{n}^{1}+ h_{n}^{2}+ h_{n}^{3}) \, dx\right\} \\
&\geq \left(\frac{1}{2}-\frac{1}{\theta} \right)\left[\|w\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|w(\cdot, 0)|^{2}_{2}\right]+ \int_{{\mathbb{R}}^{N}} (h^{1}+ h^{3}) \, dx
= d_{V(0)}.\end{aligned}$$ Accordingly, $$\begin{aligned}
\label{2.19AM}
\lim_{n{\rightarrow}\infty}\|w_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|w_{n}(\cdot, 0)|^{2}_{2}=\|w\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}-V_{1}|w(\cdot, 0)|^{2}_{2},\end{aligned}$$ and $$\begin{aligned}
h_{n}^{1}{\rightarrow}h^{1}, \, h_{n}^{2}{\rightarrow}0 \, \mbox{ and }\, h_{n}^{3}{\rightarrow}h^{3} \, \mbox{ in } \, L^{1}({\mathbb{R}}^{3}). \end{aligned}$$ Then $$\begin{aligned}
\lim_{n{\rightarrow}\infty} \int_{{\mathbb{R}}^{N}} (V(\operatorname{\varepsilon}_{n} x+ \operatorname{\varepsilon}_{n}y_{n})+V_{1}) w^{2}(x, 0) \, dx = \int_{{\mathbb{R}}^{N}} (V(x_{0})+V_{1}) w^{2}(x, 0) \, dx, \end{aligned}$$ which implies that $$\begin{aligned}
\label{2.20AM}
\lim_{n{\rightarrow}\infty} |w_{n}(\cdot, 0)|_{2}^{2}= |w(\cdot, 0)|_{2}^{2}. \end{aligned}$$ Putting together and , and using the fact that ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ is a Hilbert space, we attain $$\begin{aligned}
\|w_{n}-w\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}{\rightarrow}0 \quad \mbox{ as } n{\rightarrow}\infty.\end{aligned}$$ This ends the proof of lemma.
The proof of Theorem \[thm1\]
=============================
In this last section we give the proof of Theorem \[thm1\]. We start by proving the following lemma which will be crucial to study the behavior of maximum points of the solutions. The proof is based on a variant of the Moser iteration argument [@Moser] (see also [@Ajmp; @Armi; @CZN1]).
\[moser\] Let $(w_{n})$ be the sequence of Lemma \[lem2.5AM\]. Then, $w_{n}(\cdot, 0)\in L^{\infty}({\mathbb{R}}^{N})$ and there exists $M>0$ such that $$\begin{aligned}
|w_{n}(\cdot, 0)|_{\infty}\leq M \quad \mbox{ for all } n\in \mathbb{N}.\end{aligned}$$ Moreover, $w_{n}\in L^{\infty}({\mathbb{R}}^{N+1}_{+})$ and there exists $R>0$ such that $$\|w_{n}\|_{L^{\infty}({\mathbb{R}}^{N+1}_{+})}\leq R \quad \mbox{ for all } n\in \mathbb{N}.$$
We note that $w_{n}$ is a weak solution to $$\begin{aligned}
\label{traslato}
\left\{
\begin{array}{ll}
-\operatorname{div}(y^{1-2s} \nabla w_{n})+m^{2}y^{1-2s}w_{n}=0 &\mbox{ in } {\mathbb{R}}^{N+1}_{+}, \\
\frac{\partial w_{n}}{\partial \nu^{1-2s}}=-V(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n})w_{n}(\cdot, 0)+g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(\cdot, 0)) &\mbox{ in } {\mathbb{R}}^{N}.
\end{array}
\right.\end{aligned}$$ Let $z_{n, L}:=w_{n}w_{n, L}^{2\beta}$ where $w_{n,L}:=\min\{w_{n},L\}$, $L>0$ and $\beta>0$ will be chosen later. Taking $z_{L, n}$ as test-function in we deduce that $$\begin{aligned}
\label{conto1JMP}
\iint_{\mathbb{R}^{N+1}_{+}} &y^{1-2s}w^{2\beta}_{n,L}(|\nabla w_{n}|^{2}+m^{2}w^{2}_{n}) \, dxdy+\iint_{D_{n, L}} 2\beta y^{1-2s}w^{2\beta}_{n, L} |\nabla w_{n}|^{2} \, dx dy \nonumber \\
&= -\int_{\mathbb{R}^{N}} V(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}) w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0) \,dx+ \int_{\mathbb{R}^{N}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(\cdot, 0)) w_{n}(x,0)w^{2\beta}_{n, L}(x,0) \,dx \end{aligned}$$ where $D_{n, L}:=\{(x,y)\in \mathbb{R}^{N+1}_{+}: |w_{n}(x, y)|\leq L\}$. It is easy to check that $$\begin{aligned}
\label{conto2JMP}
\iint_{\mathbb{R}^{N+1}_{+}} &y^{1-2s}|\nabla (w_{n}w_{n, L}^{\beta})|^{2} \,dxdy =\iint_{\mathbb{R}^{N+1}_{+}} y^{1-2s}w_{n, L}^{2\beta} |\nabla w_{n}|^{2} \,dxdy \nonumber \\
&\quad+\iint_{D_{n, L}} (2\beta+\beta^{2}) y^{1-2s}w_{n, L}^{2\beta} |\nabla w_{n}|^{2} \,dxdy.\end{aligned}$$ Then, putting together (\[conto1JMP\]), (\[conto2JMP\]), $(V_{1})$, $(f_1)$-$(f_2)$, $(g_1)$-$(g_2)$, we get $$\begin{aligned}
\label{S1JMP}
&\|ww_{n, L}^{\beta}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}=\iint_{\mathbb{R}^{N+1}_{+}} y^{1-2s}(|\nabla (w_{n}w_{n, L}^{\beta})|^{2}+m^{2}w^{2}_{n}w^{2\beta}_{n, L}) \,dxdy \nonumber \\
&=\iint_{\mathbb{R}^{N+1}_{+}} y^{1-2s}w_{n, L}^{2\beta} (|\nabla w_{n}|^{2}+m^{2}w^{2}_{n}) \,dxdy+\iint_{D_{n, L}} 2\beta \left(1+\frac{\beta}{2}\right) y^{1-2s}w_{n, L}^{2\beta} |\nabla w_{n}|^{2} \,dxdy \nonumber \\
&\leq c_{\beta} \left[\iint_{\mathbb{R}^{N+1}_{+}} y^{1-2s}w_{n, L}^{2\beta} (|\nabla w_{n}|^{2}+m^{2}w^{2}_{n}) \,dxdy+\iint_{D_{n, L}} 2\beta y^{1-2s}w_{n, L}^{2\beta} |\nabla w_{n}|^{2} \,dxdy\right] \nonumber \\
&=c_{\beta} \left[-\int_{\mathbb{R}^{N}} V(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}) w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0) \,dx+ \int_{\mathbb{R}^{N}} g(\operatorname{\varepsilon}_{n}x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(\cdot, 0)) w_{n}(x,0)w^{2\beta}_{n, L}(x,0) \,dx\right] \nonumber \\
&\leq c_{\beta} \left[\int_{{\mathbb{R}}^{N}} (V_{1}+1)w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0)+ C_{1} w^{p+1}_{n}(x,0)w^{2\beta}_{n, L}(x,0) \,dx \right]\end{aligned}$$ where $$\displaystyle{c_{\beta}:=1+\frac{\beta}{2}}.$$ Now, we prove that there exist a constant $c>0$ independent of $n$, $L$, $\beta$, and $h_{n}\in L^{N/2s}(\mathbb{R}^{N})$, $h_{n}\geq 0$ and independent of $L$ and $\beta$, such that $$\label{S2JMP}
(V_{1}+1)w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0)+ C_{1}w^{p+1}_{n}(x,0)w^{2\beta}_{n, L}(x,0)\leq (c+h_{n})w^{2}_{n}(\cdot,0) w_{n, L}^{2\beta}(\cdot,0) \quad \mbox{ on } \mathbb{R}^{N}.$$ Firstly, we notice that $$\begin{aligned}
&(V_{1}+1)w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0)+ C_{1} w^{p+1}_{n}(x,0)w^{2\beta}_{n, L}(x,0) \\
&\quad \leq (V_{1}+1)w^{2}_{n}(x,0) w^{2\beta}_{n, L}(x,0)+C_{1} w_{n}^{p-1}(\cdot,0)w^{2}_{n}(\cdot,0) w_{n, L}^{2\beta}(\cdot,0) \quad \mbox{ on } \mathbb{R}^{N}.\end{aligned}$$ Moreover, $$w_{n}^{p-1}(\cdot,0)\leq 1+h_{n} \quad \mbox{ on } \mathbb{R}^{N},$$ where $h_{n}:=\chi_{\{|w_{n}(\cdot,0)|>1\}}w_{n}^{p-1}(\cdot,0)\in L^{N/2s}(\mathbb{R}^{N})$. In fact, we can observe that $$w_{n}^{p-1}(\cdot,0)=\chi_{\{|w_{n}(\cdot,0)|\leq 1\}}w_{n}^{p-1}(\cdot,0)+\chi_{\{|w_{n}(\cdot,0)|>1\}}w_{n}^{p-1}(\cdot,0)\leq 1+\chi_{\{|w_{n}(\cdot,0)|>1\}}w_{n}^{p-1}(\cdot,0) \quad \mbox{ on } \mathbb{R}^{N},$$ and that if $(p-1)\frac{N}{2s}<2$ then, recalling that $(w_{n}(\cdot, 0))$ is bounded in ${H^{s}({\mathbb{R}}^{N})}$, $$\int_{\mathbb{R}^{N}} \chi_{\{|w_{n}(\cdot,0)|>1\}}|w_{n}(x,0)|^{\frac{N}{2s}(p-1)} \,dx \leq \int_{\mathbb{R}^{N}} \chi_{\{|w_{n}(\cdot,0)|>1\}}|w_{n}(x,0)|^{2} \, dx\leq C, \quad \mbox{ for all } n\in \mathbb{N},$$ while if $2\leq (p-1)\frac{N}{2s}$ we deduce that $(p-1)\frac{N}{2s}\in [2,2^{*}_{s}]$, and by Theorem \[Sembedding\] and the boundedness of $(w_{n})$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ we find $$\int_{{\mathbb{R}}^{N}} \chi_{\{|w_{n}(\cdot,0)|>1\}} |w_{n}(x,0)|^{\frac{N}{2s}(p-1)} \,dx\leq C\|w_{n}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{\frac{N}{2s}(p-1)}\leq C,$$ for some $C>0$ depending only on $N$, $s$ and $p$. Taking into account (\[S1JMP\]) and (\[S2JMP\]) we obtain that $$\|w_{n}w_{n, L}^{\beta}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}\leq c_{\beta} \int_{\mathbb{R}^{N}} (c+h_{n}(x))w^{2}_{n}(x,0)w^{2\beta}_{n, L}(x,0) dx,$$ and by the monotone convergence theorem ($(w_{n, L})$ is nondecreasing with respect to $L$) we have as $L\rightarrow \infty$ $$\label{i1JMP}
\||w_{n}|^{\beta+1}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}\leq cc_{\beta} \int_{\mathbb{R}^{N}} |w_{n}(x,0)|^{2(\beta +1)} \,dx + c_{\beta}\int_{\mathbb{R}^{N}} h_{n}(x) |w_{n}(x,0)|^{2(\beta +1)}\,dx.$$ Fix $M>1$ and let $A_{1,n}:=\{h_{n}\leq M\}$ and $A_{2,n}:=\{h_{n}>M\}$. Then, by Hölder inequality, $$\begin{aligned}
\label{i2JMP}
\int_{\mathbb{R}^{N}} h_{n}(x) |w_{n}(x,0)|^{2(\beta +1)} dx&=\int_{A_{1, n}} h_{n}(x) |w_{n}(x,0)|^{2(\beta +1)} dx+\int_{A_{2, n}} h_{n}(x) |w_{n}(x,0)|^{2(\beta +1)} dx \nonumber \\
&\leq M ||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2}+\varepsilon(M)||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}\end{aligned}$$ where $\displaystyle{\varepsilon(M):=\sup_{n\in \mathbb{N}}\left(\int_{A_{2,n}} h^{N/2s}_{n} dx \right)^{\frac{2s}{N}}\rightarrow 0}$ as $M\rightarrow \infty$ due to the fact that $w_{n}(\cdot, 0){\rightarrow}w(\cdot, 0)$ in ${H^{s}({\mathbb{R}}^{N})}$. In view of (\[i1JMP\]) and (\[i2JMP\]) we get $$\label{regvJMP}
\||w_{n}|^{\beta+1}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}\leq c_{\beta}(c+M)||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2}+c_{\beta}\varepsilon(M)||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}.$$ We note that Theorem \[Sembedding\] yields $$\label{S3JMP}
||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}\leq C^{2}_{2^{*}_{s}}\||w_{n}|^{\beta+1}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}.$$ Then, choosing $M$ large so that $$\varepsilon(M) c_{\beta} C^{2}_{2^{*}_{s}}<\frac{1}{2},$$ and using $(\ref{regvJMP})$ and $(\ref{S3JMP})$ we obtain that $$\label{iterJMP}
||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}\leq 2 C^{2}_{2^{*}_{s}} c_{\beta}(c+M)||w_{n}(\cdot,0)|^{\beta+1}|^{2}_{2}.$$ Then we can start a bootstrap argument: since $w_{n}(\cdot,0)\in L^{2^{*}_{s}}(\mathbb{R}^{N})$ and $|w_{n}(\cdot, 0)|_{{2^{*}_{s}}}\leq C$ for all $n\in \mathbb{N}$, we can apply (\[iterJMP\]) with $\beta_{1}+1=\frac{N}{N-2s}$ to deduce that $w_{n}(\cdot,0)\in L^{(\beta_{1}+1){2^{*}_{s}}}(\mathbb{R}^{N})=L^{\frac{2N^{2}}{(N-2s)^{2}}}(\mathbb{R}^{N})$. Applying again (\[iterJMP\]), after $k$ iterations, we find $w_{n}(\cdot,0)\in L^{\frac{2N^{k}}{(N-2s)^{k}}}(\mathbb{R}^{N})$, and so $w_{n}(\cdot,0)\in L^{q}(\mathbb{R}^{N})$ for all $q\in[2,\infty)$ and $|w_{n}(\cdot, 0)|_{q}\leq C$ for all $n\in \mathbb{N}$.
Now we prove that actually $w_{n}(\cdot,0)\in L^{\infty}(\mathbb{R}^{N})$. Since $w_{n}(\cdot,0)\in L^{q}(\mathbb{R}^{N})$ for all $q\in[2,\infty)$ we have that $h_{n}\in L^{\frac{N}{s}}(\mathbb{R}^{N})$ and $|h_{n}|_{\frac{N}{s}}\leq D$ for all $n\in \mathbb{N}$. Then, by the generalized Hölder inequality and Young’s inequality with $\lambda>0$, we can see that for all $\lambda>0$ $$\begin{aligned}
\int_{\mathbb{R}^{N}} h_{n}(x) |w_{n}(x,0)|^{2(\beta+1)} dx &\leq |h_{n}|_{\frac{N}{s}} ||w_{n}(\cdot,0)|^{\beta+1}|_{2} ||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}} \nonumber \\
&\leq D\left(\lambda ||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2}+\frac{1}{\lambda} ||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}\right).
$$ Consequently, using (\[i1JMP\]) and (\[S3JMP\]), we deduce that $$\begin{aligned}
\label{i3JMP}
&||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}\leq C^{2}_{2^{*}_{s}}\||w_{n}|^{\beta+1}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2} \nonumber\\
&\leq c_{\beta} C^{2}_{2^{*}_{s}} (c+ D \lambda) ||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2}+C^{2}_{2^{*}_{s}}\frac{c_{\beta}D}{\lambda} ||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2}.
$$ Taking $\lambda>0$ such that $$\frac{c_{\beta}DC^{2}_{2^{*}_{s}}}{\lambda}=\frac{1}{2}
$$ we obtain that $$\begin{aligned}
||w_{n}(\cdot,0)|^{\beta+1}|_{{2^{*}_{s}}}^{2} \leq 2c_{\beta}(c+D\lambda)C^{2}_{2^{*}_{s}} ||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2}= M_{\beta} ||w_{n}(\cdot,0)|^{\beta+1}|_{2}^{2},\end{aligned}$$ where $$\begin{aligned}
M_{\beta}:=2c_{\beta}(c+D\lambda)C^{2}_{2^{*}_{s}}.\end{aligned}$$ Now we can control the dependence on $\beta$ of $M_{\beta}$ as follows: $$M_{\beta}\leq Cc^{2}_{\beta}\leq C(1+\beta)^{2}\leq M_{0}^{2}e^{2\sqrt{\beta+1}},$$ for some $M_{0}>0$ independent of $\beta$, and we get $$|w_{n}(\cdot,0)|_{2^{*}_{s}(\beta+1)} \leq M_{0}^{\frac{1}{\beta+1}} e^{\frac{1}{\sqrt{\beta+1}}}|w_{n}(\cdot,0)|_{2(\beta+1)}.$$ As before, iterating this last relation and choosing $\beta_{0}=0$ and $2(\beta_{j+1}+1)=2^{*}_{s}(\beta_{j}+1)$ we have that $$|w_{n}(\cdot,0)|_{2^{*}_{s}(\beta_{j}+1)} \leq M_{0}^{\sum_{i=0}^{j} \frac{1}{\beta_{i}+1}} e^{\sum_{i=0}^{j} \frac{1}{\sqrt{\beta_{i}+1}}}|w_{n}(\cdot,0)|_{2(\beta_{0}+1)}.$$ We note that $$\label{BETAj}
\beta_{j}=\left(\frac{N}{N-2s}\right)^{j}-1,$$ so the series $$\sum_{i=0}^{\infty} \frac{1}{\beta_{i}+1} \quad \mbox{ and } \quad \sum_{i=0}^{\infty} \frac{1}{\sqrt{\beta_{i}+1}}$$ are convergent. Recalling that $|w_{n}(\cdot, 0)|_{q}\leq C$ for all $n\in \mathbb{N}$ and $q\in [2, \infty)$, we get $$|w_{n}(\cdot,0)|_{\infty}=\lim_{j\rightarrow \infty} |w_{n}(\cdot,0)|_{2^{*}_{s}(\beta_{j}+1)}\leq M \quad \mbox{ for all } n\in \mathbb{N}.$$ This proves the $L^{\infty}$-desired estimate for the trace. At this point, we prove that there exists $R>0$ such that $$\begin{aligned}
\label{COLACOLA1}
\|w_{n}\|_{L^{\infty}({\mathbb{R}}^{N+1}_{+})}\leq R \quad \mbox{ for all } n\in \mathbb{N}.\end{aligned}$$ Using with $\lambda=1$ and that $|w_{n}(\cdot, 0)|_{q}\leq C$ for all $q\in [2, \infty]$, we deduce that $$\begin{aligned}
\||w_{n}|^{\beta+1}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\leq \tilde{c}c_{\beta}C^{2(\beta+1)} \quad \mbox{ for all } n\in \mathbb{N},\end{aligned}$$ for some $\tilde{c}, C>0$ independent on $\beta$ and $n$. On the other hand, from , we obtain that $$\begin{aligned}
\left(\iint_{{\mathbb{R}}^{N+1}_{+}} y^{a} |w_{n}(x, y)|^{2\gamma(\beta+1)}\, dx dy\right)^{\frac{\beta+1}{2\gamma(\beta+1)}}=\||w_{n}|^{\beta+1}\|_{L^{2\gamma}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}\leq C_{*}\||w_{n}|^{\beta+1}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\end{aligned}$$ which combined with the previous inequality yields $$\begin{aligned}
\|w_{n}\|_{L^{2\gamma(\beta+1)}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}\leq C'(\tilde{C}_{*}c_{\beta})^{\frac{1}{2(\beta+1)}} \quad \mbox{ for all } n\in \mathbb{N}.\end{aligned}$$ Since $$\frac{1}{\beta+1}\log\left(1+\frac{\beta}{2}\right)\leq \frac{\beta}{2(\beta+1)}\leq \frac{1}{2} \quad \mbox{ for all } \beta>0,$$ we can see that there exists $\bar{C}>0$ such that $C'(\tilde{C}_{*}c_{\beta})^{\frac{1}{2(\beta+1)}}\leq \bar{C}$ for all $\beta>0$, and so $$\|w_{n}\|_{L^{2\gamma(\beta+1)}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}\leq \bar{C} \quad \mbox{ for all } n\in \mathbb{N}, \beta>0.$$ Now, fix $R>\bar{C}$ and define $A_{n}:=\{(x, y)\in {\mathbb{R}}^{N+1}_{+}: w_{n}(x, y)>R\}$. Hence, for all $n, j\in \mathbb{N}$, we have $$\bar{C}\geq \left(\iint_{{\mathbb{R}}^{N+1}_{+}} y^{1-2s} |w_{n}(x, y)|^{2\gamma(\beta_{j}+1)}\, dx dy\right)^{\frac{1}{2\gamma(\beta_{j}+1)}}\geq R^{\frac{\gamma(\beta_{j}+1)-1}{\gamma(\beta_{j}+1)}} \left(\iint_{A_{n}} y^{1-2s} |w_{n}(x, y)|^{2}\, dx dy\right)^{\frac{1}{2\gamma(\beta_{j}+1)}}$$ which yields $$\left( \frac{\bar{C}}{R} \right)^{\beta_{j}+1-\frac{1}{\gamma}}\geq \frac{1}{\bar{C}^{\frac{1}{\gamma}}} \left(\iint_{A_{n}} y^{1-2s} |w_{n}(x, y)|^{2}\, dx dy\right)^{\frac{1}{2\gamma}},$$ where $\beta_{j}$ is given in . Letting $j\rightarrow \infty$, we have that $\beta_{j}{\rightarrow}\infty$ and then $$\iint_{A_{n}} y^{1-2s} |w_{n}(x, y)|^{2}\, dx dy=0 \quad\mbox{ for all } n\in \mathbb{N},
$$ which implies that $|A_{n}|=0$ for all $n\in \mathbb{N}$. Consequently, holds true.
\[lem2.6AM\] The sequence $(w_{n})$ satisfies $w_{n}(\cdot, 0){\rightarrow}0$ as $|x|{\rightarrow}\infty$ uniformly in $n\in \mathbb{N}$.
By Lemma \[moser\] and Proposition \[PROPFF\]-$(iii)$, we obtain that $w_{n}$ is continuous in $\overline{{\mathbb{R}}^{N+1}_{+}}$. On the other hand, from and $w_{n}{\rightarrow}w$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$, we have that $w_{n}{\rightarrow}w$ in $L^{2\gamma}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})$. Fix $\bar{x}\in {\mathbb{R}}^{N}$. Using $(V_1)$ and we see that $w_{n}$ is a weak subsolution to $$\left\{
\begin{array}{ll}
-{\rm div}(y^{1-2s} \nabla w_{n})+m^{2}y^{1-2s}w_{n}= 0 &\mbox{ in } Q_{1}(\bar{x},0):=B_{1}(\bar{x})\times (0,1), \\
\frac{\partial w_{n}}{\partial \nu^{1-2s}}= (V_{1}+\eta)w_{n}+C_{\eta}w_{n}^{{2^{*}_{s}}-1} &\mbox{ on } B_{1}(\bar{x}),
\end{array}
\right.$$ where $\eta\in (0, m^{2s}-V_{1})$ is fixed. Applying Proposition \[PROPFF\]-$(i)$ and observing that $L^{2\gamma}(A, y^{1-2s})\subset L^{2}(A, y^{1-2s})$ for any bounded set $A\subset {\mathbb{R}}^{N}$, we get $$0\leq \sup_{Q_{1}(\bar{x},0)}w_{n}\leq C(\|w_{n}\|_{L^{2\gamma}(Q_{1}(\bar{x},0), y^{1-2s})}+|w_{n}^{2^{*}_{s}-1}(\cdot, 0)|_{L^{q}(B_{1}(\bar{x}))}) \quad \mbox{ for all } n\in \mathbb{N},$$ where $q>\frac{N}{2s}$ is fixed and $C>0$ is a constant depending only on $N, m, s, q, \gamma$ and independent of $n\in \mathbb{N}$ and $\bar{x}$. Note that $q(2^{*}_{s}-1)\in (2, \infty)$ because $N>2s$ and $q>\frac{N}{2s}$. Taking the limit as $|\bar{x}|\rightarrow \infty$ we infer that $w_{n}(\bar{x}, 0)\rightarrow 0$ as $|\bar{x}|\rightarrow \infty$ uniformly in $n\in \mathbb{N}$.
Now we have all tools to give the proof of the main result of this work.
We begin by proving that there exists $\operatorname{\varepsilon}_{0}>0$ such that for any $\operatorname{\varepsilon}\in (0, \operatorname{\varepsilon}_{0})$ and any mountain pass solution $u_{\operatorname{\varepsilon}} \in X_{\operatorname{\varepsilon}}$ of , it holds $$\label{inftyAMPA}
|u_{\operatorname{\varepsilon}}(\cdot, 0)|_{L^{\infty}(\mathbb{R}^{N}\setminus \Lambda_{\operatorname{\varepsilon}})}<a.$$ Assume by contradiction that for some subsequence $(\operatorname{\varepsilon}_{n})$ such that $\operatorname{\varepsilon}_{n}\rightarrow 0$, we can find $u_{n}:=u_{\operatorname{\varepsilon}_{n}}\in X_{\operatorname{\varepsilon}_{n}}$ such that $J_{\operatorname{\varepsilon}_{n}}(u_{n})=c_{\operatorname{\varepsilon}_{n}}$, $J'_{\operatorname{\varepsilon}_{n}}(u_{n})=0$ and $$\label{eeeAMPA}
|u_{n}(\cdot, 0)|_{L^{\infty}(\mathbb{R}^{N}\setminus \Lambda_{\operatorname{\varepsilon}_{n}})}\geq a.$$ In view of Lemma \[lem2.5AM\], we can find $(y_{n})\subset \mathbb{R}^{N}$ such that $w_{n}(x, y):=u_{n}(x+y_{n}, y)\rightarrow w$ in ${X^{s}({\mathbb{R}}^{N+1}_{+})}$ and $\operatorname{\varepsilon}_{n}y_{n}\rightarrow x_{0}$ for some $x_{0}\in \Lambda$ such that $V(x_{0})=-V_{0}$.
Now, if we choose $r>0$ such that $B_{r}(x_{0})\subset B_{2r}(x_{0})\subset \Lambda$, we can see that $B_{\frac{r}{\operatorname{\varepsilon}_{n}}}(\frac{x_{0}}{\operatorname{\varepsilon}_{n}})\subset \Lambda_{\operatorname{\varepsilon}_{n}}$. Then, for any $x\in B_{\frac{r}{\operatorname{\varepsilon}_{n}}}(y_{n})$ it holds $$\left|x - \frac{x_{0}}{\operatorname{\varepsilon}_{n}}\right| \leq |x- y_{n}|+ \left|y_{n} - \frac{x_{0}}{\operatorname{\varepsilon}_{n}}\right|<\frac{1}{\operatorname{\varepsilon}_{n}}(r+o_{n}(1))<\frac{2r}{\operatorname{\varepsilon}_{n}}\quad \mbox{ for } n \mbox{ sufficiently large. }$$ Therefore, $$\label{ernAMPA}
\mathbb{R}^{N}\setminus \Lambda_{\operatorname{\varepsilon}_{n}}\subset \mathbb{R}^{N} \setminus B_{\frac{r}{\operatorname{\varepsilon}_{n}}}(y_{n})$$ for any $n$ big enough. Using Lemma \[lem2.6AM\] we see that $$\label{freddiAMPA}
w_{n}(x, 0)\rightarrow 0 \quad \mbox{ as } |x|\rightarrow \infty$$ uniformly in $n\in \mathbb{N}$. Therefore there exists $R>0$ such that $$w_{n}(x, 0)<a \quad \mbox{ for any } |x|\geq R, n\in \mathbb{N}.$$ Hence, $u_{n}(x, 0)=w_{n}(x-y_{n}, 0)<a$ for any $x\in \mathbb{R}^{N}\setminus B_{R}(y_{n})$ and $n\in \mathbb{N}$. On the other hand, by , there exists $n_{0} \in \mathbb{N}$ such that for any $n\geq n_{0}$ we have $$\mathbb{R}^{N}\setminus \Lambda_{\operatorname{\varepsilon}_{n}}\subset \mathbb{R}^{N} \setminus B_{\frac{r}{\operatorname{\varepsilon}_{n}}}(y_{n})\subset \mathbb{R}^{N}\setminus B_{R}(y_{n}),$$ which implies that $u_{n}(x, 0)<a$ for any $x\in \mathbb{R}^{N}\setminus \Lambda_{\operatorname{\varepsilon}_{n}}$ and $n\geq n_{0}$. This is impossible according to . Since $u_{\operatorname{\varepsilon}}\in X_{\operatorname{\varepsilon}}$ satisfies , by the definition of $g$ it follows that $u_{\operatorname{\varepsilon}}$ is a solution of for $\operatorname{\varepsilon}\in (0, \operatorname{\varepsilon}_{0})$. From the Harnack inequality we conclude that $u_{\operatorname{\varepsilon}}(x, 0)>0$ in ${\mathbb{R}}^{N}$. In what follows, we study the behavior of the maximum points of solutions to problem . Take $\operatorname{\varepsilon}_{n}\rightarrow 0$ and let $(u_{n})\subset X_{\operatorname{\varepsilon}_{n}}$ be a sequence of solutions to as above. Consider the translated sequence $w_{n}(x, y)=u_{n}(x+y_{n}, y)$ where $(y_{n})$ is given by Lemma \[lem2.5AM\]. Let us prove that there exists $\delta>0$ such that $$\label{DELTA}
|w_{n}(\cdot, 0)|_{\infty}\geq \delta \mbox{ for all } n\in \mathbb{N}.$$ Assume by contradiction that $|w_{n}(\cdot, 0)|_{\infty}{\rightarrow}0$. Using $(f_1)$ we can see that there exists $\nu\in \mathbb{N}$ such that $$\frac{f(|w_{n}(\cdot, 0)|_{\infty})}{|w_{n}(\cdot, 0)|_{\infty}}<\frac{V_{1}}{\kappa} \quad \mbox{ for all } n\geq \nu.$$ From $\langle J'_{\operatorname{\varepsilon}_{n}}(u_{n}), u_{n}\rangle=0$, $(g_2)$ and $(f_4)$, we can see that for all $n\geq \nu$ $$\begin{aligned}
\|w_{n}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}-V_{1}|w_{n}(\cdot, 0)|_{2}^{2}&= \|u_{n}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}^{2}-V_{1}|u_{n}(\cdot, 0)|_{2}^{2}\\
&\leq\int_{\mathbb{R}^{N}} f(u_{n}(x, 0)) u_{n}(x, 0) \,dx \\
&\leq \int_{\mathbb{R}^{N}} \frac{f(|w_{n}(\cdot, 0)|_{\infty})}{|w_{n}(\cdot, 0)|_{\infty}} w^{2}_{n}(x, 0) \,dx \\
&\leq \frac{V_{1}}{\kappa} \int_{\mathbb{R}^{N}} w_{n}^{2}(x, 0) \, dx\end{aligned}$$ which combined with yields $$\left(1-\frac{V_{1}}{m^{2s}}\left(1+\frac{1}{\kappa}\right) \right)\|w_{n}\|^{2}_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}\leq 0.$$ Since $\kappa>\frac{V_{1}}{m^{2s}-V_{1}}$ we get $\|w_{n}\|_{{X^{s}({\mathbb{R}}^{N+1}_{+})}}=0$ for all $n\geq \nu$ which is a contradiction. Therefore, if $q_{n}$ is a global maximum point of $w_{n}(\cdot, 0)$, we deduce from Lemma \[moser\] and that there exists $R>0$ such that $|q_{n}|<R$ for all $n\in \mathbb{N}$. Thus $x_{n}:=q_{n}+y_{n}$ is a global maximum point of $u_{n}(\cdot, 0)$, and $\operatorname{\varepsilon}_{n}x_{n}{\rightarrow}x_{0}\in \mathcal{M}$. Using the continuity of $V$ we deduce that $$\lim_{n\rightarrow \infty} V(\operatorname{\varepsilon}_{n} x_{n})=V(x_{0})=-V_{0}.$$
Finally, we prove a decay estimate for $u_{n}(\cdot, 0)$. Using $(f_1)$, the definition of $g$ and , we can find $R_{1}>1$ sufficiently large such that $$\begin{aligned}
\label{TERESA}
g(\operatorname{\varepsilon}_{n} x+\operatorname{\varepsilon}_{n}y_{n}, w_{n}(x,0))w_{n}(x, 0)\leq \delta w_{n}^{2}(x, 0) \mbox{ for } |x|\geq R_{1},\end{aligned}$$ where $\delta\in (0, m^{2s}-V_{1})$ is fixed. Pick a smooth cut-off function $\phi$ such that $0\leq \phi\leq 1$, $\phi(x)=0$ for $|x|\geq 1$, and $\phi\not\equiv 0$. By Riesz representation theorem, there exists a unique function $\bar{w}\in {H^{s}({\mathbb{R}}^{N})}$ such that $$\label{EqVD}
(-\Delta+m^{2})^{s}\bar{w}-(V_{1}+\delta) \bar{w}=\phi \mbox{ in } {\mathbb{R}}^{N}.$$ Since $\bar{w}=\mathcal{B}_{2s,m}*\phi$, for some positive kernel $\mathcal{B}_{2s, m}$ whose expression is given below, we can see that $\bar{w}\geq 0$ in ${\mathbb{R}}^{N}$. Denote by $\bar{W}$ the $s$-extension of $\bar{w}$, namely $\bar{W}(x, y)=(P_{s, m}(\cdot, y)*\bar{w})(x)$. Fix $x\in {\mathbb{R}}^{N}$. Then, from Young’s inequality and , we can see that $$\begin{aligned}
\|\bar{W}\|_{L^{2}(B_{1}(x)\times [0, 1], y^{1-2s})}&\leq \|\bar{W}\|_{L^{2}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}=\|P_{s, m}(\cdot, y)*\bar{w}\|_{L^{2}({\mathbb{R}}^{N+1}_{+}, y^{1-2s})}\\
&\leq \left(\int_{0}^{1} |P_{s, m}(\cdot, y)*\bar{w}|_{2}^{2} \,y^{1-2s} \, dy\right)^{\frac{1}{2}} \\
&\leq \left(\int_{0}^{1} |P_{s, m}(\cdot, y)|_{1}^{2} |\bar{w}|_{2}^{2} \, y^{1-2s} \, dy\right)^{\frac{1}{2}} \\
&\leq |\bar{w}|_{2} \left(\int_{0}^{1} \vartheta^{2}(my) y^{1-2s} \, dy\right)^{\frac{1}{2}}\\
&\leq c_{s, m} |\bar{w}|_{2},\end{aligned}$$ for some constant $c_{s, m}>0$. Therefore, by Proposition \[PROPFF\]-$(i)$ and translation invariance of with respect to $x\in {\mathbb{R}}^{N}$, we deduce that $\bar{w}\in L^{\infty}({\mathbb{R}}^{N})$, and thus, by interpolation, $\bar{w}\in L^{q}({\mathbb{R}}^{N})$ for any $q\in [2, \infty]$. Hence, $(-\Delta+m^{2})^{s}\bar{w}=(V_{1}+\delta) \bar{w}+\phi\in L^{\infty}({\mathbb{R}}^{N})$ and applying Theorem \[SilvestreLinfty\] (see also Theorem \[Besselembedding\]-$(v)$ and Proposition \[PROPFF\]-$(iii)$) we obtain that $\bar{w}$ is Hölder continuous in ${\mathbb{R}}^{N}$. Using Harnack’s inequality (see Proposition \[PROPFF\]-$(ii)$) we conclude that $\bar{w}>0$ in ${\mathbb{R}}^{N}$. Moreover, we can prove that there exist $c, C>0$ such that $$\begin{aligned}
\label{HZ1AMPA}
0<\bar{w}(x)\leq Ce^{-c|x|} \quad \mbox{ for all } x\in {\mathbb{R}}^{N}.\end{aligned}$$ Assume for the moment that holds and we postpone the proof of it after proving the decay estimate for $\hat{u}_{n}(\cdot, 0)$. We know that $$\begin{aligned}
\label{HZ2AMPA}
(-\Delta+m^{2})^{s} \bar{w}-(V_{1}+\delta)\bar{w}= 0 \quad \mbox{ in } {\mathbb{R}}^{N}\setminus B_{R_{1}}. \end{aligned}$$ On the other hand, by $(V_1)$ and , we can see that $$\begin{aligned}
\label{HZ3AMPA}
(-\Delta+m^{2})^{s} w_{n}(\cdot, 0)-(V_{1}+\delta) w_{n}(\cdot,0) \leq 0 \mbox{ in } {\mathbb{R}}^{N}\setminus B_{R_{1}}. \end{aligned}$$ Set $b:=\inf_{B_{R_{1}}} \bar{w}>0$ and $z_{n}:=(\ell+1)\bar{w}-bw_{n}(\cdot, 0)$, where $\ell:=\sup_{n\in \mathbb{N}} |w_{n}(\cdot, 0)|_{\infty}$. We aim to show that $z_{n}\geq 0$ in ${\mathbb{R}}^{N}$. Assume by contradiction that there exists a sequence $(x_{j, n})$ such that $$\begin{aligned}
\label{4.25HZ}
\inf_{x\in {\mathbb{R}}^{N}} z_{n}(x)=\lim_{j{\rightarrow}\infty} z_{n}(x_{j, n})<0.\end{aligned}$$ In view of and , we notice that $$\lim_{|x|{\rightarrow}\infty} \bar{w}(x)=\lim_{|x|{\rightarrow}\infty} w_{n}(x, 0)=0$$ where the second limit is uniform in $n\in \mathbb{N}$, and so $$\lim_{|x|{\rightarrow}\infty} z_{n}(x)=0 \mbox{ uniformly in } n\in \mathbb{N}.$$ Consequently, $(x_{j, n})$ is bounded and, up to a subsequence, we may assume that $x_{j, n}{\rightarrow}x_{n}^{*}$ for some $x_{n}\in {\mathbb{R}}^{N}$. Hence, becomes $$\begin{aligned}
\label{4.26HZ}
\inf_{x\in {\mathbb{R}}^{N}} z_{n}(x)=z_{n}(x_{n}^{*})<0.\end{aligned}$$ Then, from the minimality of $x_{n}^{*}$ and using Theorem \[FINALE\], we deduce that $$\begin{aligned}
\label{4.27HZ}
(-\Delta+m^{2})^{s}z_{n}(x_{n}^{*})&=m^{2s}z_{n}(x_{n}^{*})+\frac{C(N,s)}{2} m^{\frac{N+2s}{2}} \int_{{\mathbb{R}}^{N}} \frac{2z_{n}(x_{n}^{*})-z_{n}(x_{n}^{*}+y)-z_{n}(x_{n}^{*}-y)}{|y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|y|)\, dy \nonumber\\
&\leq m^{2s}z_{n}(x_{n}^{*}).\end{aligned}$$ Since $z_{n}\geq 0$ in $B_{R_{1}}$, it follows from that $$\begin{aligned}
\label{ELENA}
x^{*}_{n}\in {\mathbb{R}}^{N}\setminus B_{R_{1}}.\end{aligned}$$ Gathering and we can see that $$(-\Delta+m^{2})^{s}z_{n}-(V_{1}+\delta)z_{n}\geq 0 \mbox{ in } {\mathbb{R}}^{N}\setminus B_{R_{1}},$$ and this is a contradiction because , , and $V_{1}+\delta<m^{2s}$ imply that $$(-\Delta+m^{2})^{s}z_{n}(x^{*}_{n})-(V_{1}+\delta)z_{n}(x^{*}_{n})\leq (m^{2s}-(V_{1}+\delta))z_{n}(x^{*}_{n})<0.$$ Hence, $z_{n}\geq 0$ in ${\mathbb{R}}^{N}$. In the light of (\[HZ1AMPA\]) we obtain that there exist $c, C>0$ such that $$\begin{aligned}
0\leq w_{n}(x)\leq Ce^{-c|x|} \quad \mbox{ for all } x\in {\mathbb{R}}^{N}, n\in \mathbb{N},\end{aligned}$$ which combined with $u_{n}(x, 0)=w_{n}(x-y_{n}, 0)$ yields $$\begin{aligned}
u_{n}(x, 0)=w_{n}(x-y_{n},0)\leq Ce^{-c|x-y_{n}|} \quad \mbox{ for all } x\in {\mathbb{R}}^{N}, n\in \mathbb{N}.\end{aligned}$$ In what follows, we focus our attention on the estimate . Note that $\bar{w}=\mathcal{B}_{2s, m}*\phi$, where $$\mathcal{B}_{2s, m}(x)=\mathcal{F}^{-1}([(|k|^{2}+m^{2})^{s}-(V_{1}+\delta)]^{-1}).$$ Since $\phi$ has compact support, the exponential decay of $\bar{w}$ at infinity follows if we show the exponential decay of $\mathcal{B}_{2s, m}(x)$ for big values of $|x|$. After that, due to the fact that $\bar{w}$ is continuous in ${\mathbb{R}}^{N}$, we can deduce the exponential decay of $\bar{w}$ in the whole of ${\mathbb{R}}^{N}$. Next we prove the exponential decay of $\mathcal{B}_{2s, m}(x)$ for $|x|$ large. Then $$\begin{aligned}
\label{RY1}
\mathcal{B}_{2s, m}(x)&=\frac{1}{(2\pi)^{\frac{N}{2}}} \int_{{\mathbb{R}}^{N}} e^{\imath k\cdot x} \frac{1}{[(|k|^{2}+m^{2})^{s}-(V_{1}+\delta)]}\, dk \nonumber\\
&=\frac{1}{(2\pi)^{\frac{N}{2}}} \int_{{\mathbb{R}}^{N}} e^{\imath k\cdot x}\left(\int_{0}^{\infty} e^{-t[(|k|^{2}+m^{2})^{s}-(V_{1}+\delta)]}\, dt\right)\, dk \nonumber\\
&= \int_{0}^{\infty} e^{-\gamma t} \left(\frac{1}{(2\pi)^{\frac{N}{2}}}\int_{{\mathbb{R}}^{N}} e^{\imath k\cdot x} e^{-t[(|k|^{2}+m^{2})^{s}-m^{2s}]}\, dk\right) \, dt \nonumber\\
&=\int_{0}^{\infty} e^{-\gamma t} \, p_{s, m}(x, t)\, dt \end{aligned}$$ where $$\gamma:=m^{2s}-(V_{1}+\delta)>0,$$ and $$p_{s, m}(x, t):=e^{m^{2s}t} \int_{0}^{\infty} \frac{1}{(4\pi z)^{\frac{N}{2}}} e^{-\frac{|x|^{2}}{4t}} e^{-m^{2}z} \vartheta_{s}(t, z)\, dz$$ is the $2s$-stable relativistic density with parameter $m$ (see pag. 4 formula (7) in [@ryznar], and pag. 4875 formula (2.12) and Lemma 2.2 in [@BMR]), and $\vartheta_{s}(t, z)$ is the density function of the $s$-stable process whose Laplace transform is $e^{-t\lambda^{s}}$. Using the scaling property $p_{s, m}(x, t)=m^N p_{s, 1}(mx, m^{2s} t)$ (see pag. 4876 formula (2.15) in [@BMR]) and Lemma $2.2$ in [@GR], we can see that $$\begin{aligned}
\label{GRestimate}
p_{s, m}(x, t)\leq C\left(g_{m^{2s}t}\left(\frac{mx}{\sqrt{2}}\right)+m^{2s} t \nu^{m}\left(\frac{mx}{\sqrt{2}}\right) \right) \quad \mbox{ for all } x\in {\mathbb{R}}^{N}, t>0,\end{aligned}$$ where $$g_{t}(x):=\frac{1}{(4\pi t)^{\frac{N}{2}}} e^{-\frac{|x|^{2}}{4t}}$$ and $\nu^{m}$ is the Lévy measure of relativistic process with parameter $m>0$ given by $$\nu^{m}(x):=\frac{2s2^{\frac{2s-N}{2}}}{\pi^{\frac{N}{2}} \Gamma(1-s)} \left( \frac{m}{|x|} \right)^{\frac{N+2s}{2}} K_{\frac{N+2s}{2}}(m|x|)$$ (see pag. 4877 formula (2.17) in [@BMR]). Therefore, and yield $$\begin{aligned}
\label{RY2}
\mathcal{B}_{2s, m}(x)&\leq C\int_{0}^{\infty} e^{-\gamma t} g_{m^{2s}t}\left(\frac{mx}{\sqrt{2}}\right)\, dt+C\int_{0}^{\infty} e^{-\gamma t} \, t \nu^{m}\left(\frac{mx}{\sqrt{2}}\right) \,dt \nonumber\\
&=:I_{1}(x)+I_{2}(x).\end{aligned}$$ We start with the estimate of $I_{1}(x)$ for $|x|\geq 2$. Observing that $$\begin{aligned}
\gamma t+\frac{m^{2-2s}}{8t}|x|^{2}\geq \gamma t+\frac{m^{2-2s}}{2t} \quad \mbox{ for all } |x|\geq 2, t>0,\end{aligned}$$ and that $ab\leq \operatorname{\varepsilon}a^{2}+\frac{1}{4\operatorname{\varepsilon}} b^{2}$ for all $a, b\geq 0$ and $\operatorname{\varepsilon}>0$ gives $$\begin{aligned}
\gamma t+\frac{m^{2-2s}}{8t}|x|^{2}\geq \frac{m^{1-s}}{\sqrt{2}} |x|\sqrt{\gamma} \quad \mbox{ for all } x\in {\mathbb{R}}^{N}, t>0,\end{aligned}$$ we deduce that for all $|x|\geq 2$ and $t>0$ $$\begin{aligned}
\gamma t+\frac{m^{2-2s}}{8t}|x|^{2}\geq \gamma \frac{t}{2}+\frac{m^{2-2s}}{4t}+\frac{m^{1-s}}{2\sqrt{2}} |x|\sqrt{\gamma}.\end{aligned}$$ Thus, using the definition of $g_{t}$, we can see that for all $|x|\geq 2$ $$\begin{aligned}
\label{RY3}
I_{1}(x)&\leq C\int_{0}^{\infty} \frac{e^{-\gamma\frac{t}{2}}}{t^{\frac{N}{2}}} e^{-\frac{m^{2-2s}}{4t}} e^{-\frac{m^{1-s}}{2\sqrt{2}} |x|\sqrt{\gamma}}\, dt \nonumber \\
&\leq C e^{-c|x|} \int_{0}^{\infty} \frac{e^{-\gamma\frac{t}{2}}}{t^{\frac{N}{2}}} e^{-\frac{m^{2-2s}}{4t}} \,dt\leq C e^{-c|x|}\end{aligned}$$ where we used the fact that $$\int_{0}^{\infty} \frac{e^{-\alpha t}}{t^{p}} e^{-\frac{\beta}{t}} \,dt<\infty \quad \forall \alpha, \beta, p>0.$$ Here $C, c>0$ depend only on $N, s, m$. Now we estimate $I_{2}(x)$ for large values of $|x|$. Recalling formula concerning the asymptotic behavior of $K_{\nu}$ at infinity, we deduce that there exists $r_{0}>0$ such that $$\frac{K_{\nu}(mr)}{r^{\nu}}\leq C\frac{e^{-mr}}{r^{\nu+\frac{1}{2}}} \quad \mbox{ for all } r\geq r_{0},$$ and then $$\begin{aligned}
\frac{K_{\frac{N+2s}{2}}(m|x|)}{|x|^{\frac{N+2s}{2}}}\leq C \frac{e^{-m|x|}}{|x|^{\frac{N+2s+1}{2}}} \quad \mbox{ for all } |x|\geq r_{0}.\end{aligned}$$ Consequently, using the definition of $\nu^{m}$, for all $|x|\geq r_{0}$ we get $$\begin{aligned}
\label{RY4}
I_{2}(x)\leq C\frac{e^{-m|x|}}{|x|^{\frac{N+2s+1}{2}}} \int_{0}^{\infty} t \,e^{-\gamma t}\, dt\leq C\frac{e^{-m|x|}}{|x|^{\frac{N+2s+1}{2}}}. \end{aligned}$$ Gathering , , , we find that for any $|x|\geq \max\{r_{0}, 2\}$ $$\mathcal{B}_{2s, m}(x)\leq C_{1} e^{-C_{2}|x|}+C_{3}\frac{e^{-C_{4}|x|}}{|x|^{\frac{N+2s+1}{2}}}\leq C_{5}e^{-C_{6}|x|}.$$ Thus holds true and this ends the proof of Theorem \[thm1\].
When $s=\frac{1}{2}$, $p_{\frac{1}{2},m}(x, t)$ can be calculated explicitly (see [@BMR; @LL]) and is given by $$p_{\frac{1}{2},m}(x, t)=2\left(\frac{m}{2\pi}\right)^{\frac{N+1}{2}} t\, e^{m t} (|x|^{2}+t^{2})^{-\frac{N+1}{4}} K_{\frac{N+1}{2}}(m\sqrt{|x|^{2}+t^{2}}).$$
By the definitions of $\mathcal{B}_{2s, m}$ and $p_{s, m}$, it follows that $\mathcal{B}_{2s, m}$ is radial, positive, strictly decreasing in $|x|$, and smooth on ${\mathbb{R}}^{N}\setminus \{0\}$.
Appendix: Bessel potentials
===========================
In this appendix we collect some useful results concerning Bessel potentials (with $m=1$). For more details we refer to [@ArS; @Calderon; @Grafakos; @stein].
Let $\alpha>0$. The Bessel potential of order $\alpha$ of $u\in \mathcal{S}({\mathbb{R}}^{N})$ is defined as $$\begin{aligned}
\mathcal{J}_{\alpha}u(x):=(1-\Delta)^{-\frac{\alpha}{2}}u(x)=(\mathcal{G}_{\alpha}*u)(x)=\int_{{\mathbb{R}}^{N}} \mathcal{G}_{\alpha}(x-y)u(y)\, dy,
$$ where $$\mathcal{G}_{\alpha}(x):=(2\pi)^{-\frac{N}{2}}\mathcal{F}^{-1}((1+|k|^{2})^{-\frac{\alpha}{2}})(x)
$$ is called Bessel kernel.
If $\alpha\in \mathbb{R}$ (or $\alpha\in \mathbb{C}$), then we may define the Bessel potential of a temperate distribution $u\in \mathcal{S}'({\mathbb{R}}^{N})$ (see [@Calderon]) by setting $$\mathcal{F}\mathcal{J}_{\alpha}u(k):=(2\pi)^{-\frac{N}{2}} (1+|k|^{2})^{-\frac{\alpha}{2}} \mathcal{F}u(k).$$
From definition it is evident that $\mathcal{G}_{\alpha+\beta}=\mathcal{G}_{\alpha}*\mathcal{G}_{\beta}$. It is possible to prove (see [@ArS]) that $$\mathcal{G}_{\alpha}(x)=\frac{1}{2^{\frac{N+\alpha-2}{2}}\pi^{\frac{N}{2}}\Gamma(\frac{\alpha}{2})} K_{\frac{N-\alpha}{2}}(|x|) |x|^{\frac{\alpha-N}{2}}.$$ Thus $\mathcal{G}_{\alpha}(x)$ is positive, decreasing function of $|x|$, analytic except at $x=0$, and for $x\in {\mathbb{R}}^{N}\setminus \{0\}$, $\mathcal{G}_{\alpha}(x)$ is an entire function of $\alpha$. Moreover, from and , we have $$\begin{aligned}
&\mathcal{G}_{\alpha}(x)=\frac{|x|^{-N+\alpha}}{\gamma(\alpha)}+o(|x|^{-N+\alpha}) \mbox{ as } |x|{\rightarrow}0, \mbox{ if } 0<\alpha<N, \\
&\mathcal{G}_{\alpha}(x)=O(e^{-c|x|}) \mbox{ as } |x|{\rightarrow}\infty, \mbox{ for some } c>0,\end{aligned}$$ $\mathcal{G}_{\alpha}\in L^{1}({\mathbb{R}}^{N})$ for all $\alpha>0$, and $\int_{{\mathbb{R}}^{N}} \mathcal{G}_{\alpha}(x) \, dx=1$. We also have the following integral formula (see [@stein]): $$\mathcal{G}_{\alpha}(x)=\frac{1}{(4\pi)^{\frac{\alpha}{2}}}\frac{1}{\Gamma\left(\frac{\alpha}{2}\right)} \int_{0}^{\infty} e^{-\frac{\pi|x|^{2}}{\delta}} e^{-\frac{\delta}{4\pi}} \delta^{\frac{\alpha-N}{2}} \frac{d\delta}{\delta}.
$$ One the most interesting facts concerning Bessel potentials is they can be employed to define the Bessel potential spaces; see [@Adams; @ArS; @Calderon; @Grafakos; @stein]. For $p\in [1, \infty]$ and $\alpha\in {\mathbb{R}}$ we define the Banach space $$\begin{aligned}
\mathscr{L}^{p}_{\alpha}:=\mathcal{G}_{\alpha}(L^{p}({\mathbb{R}}^{N}))=\{u: u=\mathcal{G}_{\alpha}*f, f\in L^{p}({\mathbb{R}}^{N})\}
$$ endowed with the norm $$\|u\|_{\mathscr{L}^{p}_{\alpha}}:=|f|_{p} \quad \mbox{ if } u=\mathcal{G}_{\alpha}*f.
$$ Thus $\mathscr{L}^{p}_{\alpha}$ is a subspace of $L^{p}({\mathbb{R}}^{N})$ for all $\alpha\geq 0$. We also have the following useful result:
[@Adams; @Calderon; @stein]\[Besselembedding\]
If $\alpha\geq 0$ and $1\leq p< \infty$, then $\mathcal{D}({\mathbb{R}}^{N})$ is dense in $\mathscr{L}^{p}_{\alpha}$.
If $1<p<\infty$ and $p'$ its conjugate exponent, then the dual of $\mathscr{L}^{p}_{\alpha}$ is isometrically isomorphic to $\mathscr{L}^{p'}_{-\alpha}$.
If $\beta< \alpha$, then $\mathscr{L}^{p}_{\alpha}$ is continuously embedded in $\mathscr{L}^{p}_{\beta}$.
If $\beta\leq \alpha$ and if either $1<p\leq q\leq \frac{Np}{N-(\alpha-\beta) p}<\infty$ or $p=1$ and $1\leq q<\frac{N}{N-\alpha+\beta}$, then $\mathscr{L}^{p}_{\alpha}$ is continuously embedded in $\mathscr{L}^{q}_{\beta}$.
If $0\leq \mu\leq \alpha-\frac{N}{p}<1$, then $\mathscr{L}^{p}_{\alpha}$ is continuously embedded in $C^{0, \mu}({\mathbb{R}}^{N})$.
$\mathscr{L}^{p}_{k}=W^{k,p}({\mathbb{R}}^{N})$ for all $k\in \mathbb{N}$ and $1<p<\infty$, $\mathscr{L}^{2}_{\alpha}=W^{\alpha, 2}({\mathbb{R}}^{N})$ for any $\alpha$.
If $1<p<\infty$ and $\operatorname{\varepsilon}>0$, then for every $\alpha$ we have the following continuous embeddings: $$\mathscr{L}^{p}_{\alpha+\operatorname{\varepsilon}}\subset W^{\alpha, p}({\mathbb{R}}^{N})\subset \mathscr{L}^{p}_{\alpha-\operatorname{\varepsilon}}.$$
In order to accomplish some useful regularity results for equations driven by $(-\Delta+m^{2})^{s}$, with $m>0$, we introduce the Hölder-Zygmund (or Lipschitz) spaces $\Lambda_{\alpha}$; see [@Calderon; @K; @stein]. If $\alpha>0$ and $\alpha\notin \mathbb{N}$ then we set $\Lambda_{\alpha}:=C^{[\alpha], \alpha-[\alpha]}({\mathbb{R}}^{N})$. If $\alpha=k\in \mathbb{N}$ then we set $\Lambda_{\alpha}:=\Lambda^{*}_{k}$ where $$\Lambda^{*}_{1}:=\left\{u\in L^{\infty}({\mathbb{R}}^{N})\cap C({\mathbb{R}}^{N}): \sup_{x, h\in {\mathbb{R}}^{N}, |h|>0} \frac{|u(x+h)+u(x-h)-2u(x)|}{|h|}<\infty\right\} \,\mbox{ if } k=1,$$ and $$\Lambda^{*}_{k}:=\left\{u\in C^{k-1}({\mathbb{R}}^{N}): D^{\gamma}u\in \Lambda^{*}_{1} \mbox{ for all } |\gamma|\leq k-1\right\} \,\mbox{ if } k\in \mathbb{N}, k\geq 2.$$ Then we have the following useful result:
[@Calderon; @stein]\[regBesseloperators\] If $\alpha, \beta\geq 0$, $(1-\Delta)^{-\alpha}$ is an isomorphism of $\mathscr{L}^{p}_{\beta}$ to $\mathscr{L}^{p}_{\beta+2\alpha}$. If $\alpha\geq 0$ and $\beta>0$, $(1-\Delta)^{-\alpha}$ is an isomorphism of $\Lambda_{\beta}$ to $\Lambda_{\beta+2\alpha}$.
As a consequence of Theorem \[regBesseloperators\] and the definition of Hölder-Zygmund spaces, we easily deduce the following result:
\[COROLLARIO\] Let $s\in (0, 1)$ and $0<\alpha\leq 1$. Assume that $f\in C^{0, \alpha}({\mathbb{R}}^{N})$ and that $u\in L^{\infty}({\mathbb{R}}^{N})$ is a solution to $(-\Delta+1)^{s}u=f$ in ${\mathbb{R}}^{N}$.
- If $\alpha+2s<1$ then $u\in C^{0, \alpha+2s}({\mathbb{R}}^{N})$.
- If $1<\alpha+2s<2$ then $u\in C^{1, \alpha+2s-1}({\mathbb{R}}^{N})$.
- If $2<\alpha+2s<3$ then $u\in C^{2, \alpha+2s-2}({\mathbb{R}}^{N})$.
- If $\alpha+2s=k\in \{1, 2\}$ then $u\in \Lambda_{k}^{*}$.
Bearing in mind the asymptotic estimates and for $K_{\nu}$, we are able to gain an integral representation formula for $(-\Delta+m^{2})^{s}$, with $s\in (0, 1)$, in the spirit of [@DPV].
\[INTEGRALE\] Let $s\in (0, 1)$. Then, for all $u\in \mathcal{S}({\mathbb{R}}^{N})$, $$(-\Delta+m^{2})^{s}u(x)=m^{2s}u(x)+\frac{C(N,s)}{2} m^{\frac{N+2s}{2}} \int_{{\mathbb{R}}^{N}} \frac{2u(x)-u(x+y)-u(x-y)}{|y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|y|)\, dy.$$
Choosing the substitution $z=y-x$ in , we obtain $$\begin{aligned}
\label{3.4DPV}
(-\Delta+m^{2})^{s}u(x)&=m^{2s}u(x)+C(N,s) m^{\frac{N+2s}{2}} P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x)-u(y)}{|x-y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|x-y|)\, dy \nonumber\\
&=m^{2s}u(x)+C(N,s) m^{\frac{N+2s}{2}} P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x)-u(x+z)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz.\end{aligned}$$ By substituting $\tilde{z}=-z$ in the last term in , we get $$\begin{aligned}
\label{3.5DPV}
P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x+z)-u(x)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz=P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x-\tilde{z})-u(x)}{|\tilde{z}|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|\tilde{z}|)\, d\tilde{z},\end{aligned}$$ and so, after relabeling $\tilde{z}$ as $z$, $$\begin{aligned}
\label{3.6DPV}
&2P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x+z)-u(x)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz \nonumber\\
&=P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x+z)-u(x)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz \nonumber\\
&\quad+P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x-z)-u(x)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz \nonumber\\
&=P. V. \int_{{\mathbb{R}}^{N}} \frac{u(x+z)+u(x-z)-2u(x)}{|z|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|z|)\, dz.\end{aligned}$$ Hence, if we rename $z$ as $y$ in and , we can write $(-\Delta+m^{2})^{s}$ as $$\begin{aligned}
\label{FINALE}
(-\Delta+m^{2})^{s}u(x)=m^{2s}u(x)+\frac{C(N,s)}{2} m^{\frac{N+2s}{2}} P. V. \int_{{\mathbb{R}}^{N}} \frac{2u(x)-u(x+y)-u(x-y)}{|y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|y|)\, dy.\end{aligned}$$ Now, by using a second order Taylor expansion, we see that $$\begin{aligned}
\left|\frac{2u(x)-u(x+y)-u(x-y)}{|y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|y|) \right|\leq \frac{|D^{2}u|_{\infty}}{|y|^{\frac{N+2s-4}{2}}} K_{\frac{N+2s}{2}}(m|y|). \end{aligned}$$ From we deduce that $$\frac{|D^{2}u|_{\infty}}{|y|^{\frac{N+2s-4}{2}}} K_{\frac{N+2s}{2}}(m|y|)\sim \frac{C}{|y|^{N+2s-2}} \mbox{ as } |y|{\rightarrow}0$$ which is integrable near $0$. On the other hand, using , we get $$\left|\frac{2u(x)-u(x+y)-u(x-y)}{|y|^{\frac{N+2s}{2}}}K_{\frac{N+2s}{2}}(m|y|) \right|\leq \frac{3|u|_{\infty}}{|y|^{\frac{N+2s}{2}}} K_{\frac{N+2s}{2}}(m|y|)\sim \frac{C}{|y|^{\frac{N+2s+1}{2}}}e^{-m|y|} \mbox{ as } |y|{\rightarrow}\infty$$ which is integrable near $\infty$. Therefore, we can remove the $P. V.$ in .
By Theorem \[INTEGRALE\] and using the fact that $|z|^{-\nu}K_{\nu}(m|z|)\leq C_{\nu}|z|^{-2\nu}$ (see pag. 5865 in [@FF]), we can argue as in [@Silvestre] and use the $C^{k, \alpha}$ estimates, with $k\in \{1, 2\}$, for the elliptic equation $-\Delta+m^{2}u=g$ in ${\mathbb{R}}^{N}$, to obtain Schauder-Hölder estimates for $(-\Delta+m^{2})^{s}$. In the light of this observation, we can give an alternative proof of Corollary \[COROLLARIO\] and deduce the next helpful result. Since the proofs are similar to the ones performed in [@Silvestre] for the case $m=0$, we skip the details.
\[SilvestreLinfty\] Let $s\in (0, 1)$ and $m>0$. Assume that $f\in L^{\infty}({\mathbb{R}}^{N})$ and that $u\in L^{\infty}({\mathbb{R}}^{N})$ is a solution to $(-\Delta+m^{2})^{s}u=f$ in ${\mathbb{R}}^{N}$.
- If $2s\leq 1$ then $u\in C^{0, \alpha}({\mathbb{R}}^{N})$ for any $\alpha<2s$.
- If $2s>1$ then $u\in C^{1, \alpha}({\mathbb{R}}^{N})$ for any $\alpha<2s-1$.
[777]{}
R. A. Adams, [*Sobolev spaces*]{}, Pure and Applied Mathematics, Vol. [**65**]{} Academic Press, New York-London, 1975. xviii+268 pp.
C. O. Alves, O. H. Miyagaki, [*Existence and concentration of solution for a class of fractional elliptic equation in ${\mathbb{R}}^N$ via penalization method*]{}, Calc. Var. Partial Differential Equations [**55**]{} (2016), no. 3, Art. 47, 19 pp.
A. Ambrosetti, P. H. Rabinowitz, [*Dual variational methods in critical point theory and applications*]{}, J. Functional Analysis [**14**]{} (1973), 349–381.
V. Ambrosio, [*Ground states solutions for a non-linear equation involving a pseudo-relativistic Schrödinger operator*]{}, J. Math. Phys. [**57**]{} (2016), no. 5, 051502, 18 pp.
V. Ambrosio, [*Multiplicity of positive solutions for a class of fractional Schrödinger equations via penalization method*]{}, Ann. Mat. Pura Appl. (4) [**196**]{} (2017), no. 6, 2043–2062.
V. Ambrosio, [*Concentrating solutions for a class of nonlinear fractional Schrödinger equations in ${\mathbb{R}}^{N}$*]{}, Rev. Mat. Iberoam. [**35**]{} (2019), no. 5, 1367–1414.
V. Ambrosio, [*Concentration phenomena for a class of fractional Kirchhoff equations in ${\mathbb{R}}^{N}$ with general nonlinearities*]{}, Nonlinear Anal. [**195**]{} (2020), 111761.
V. Ambrosio, [*Multiple concentrating positive solutions for a relativistic fractional Schrödinger equation with a general nonlinearity*]{}, in preparation.
N. Aronszajn, K. T. Smith, [*Theory of Bessel potentials. I*]{}, Ann. Inst. Fourier (Grenoble) [**11**]{} (1961), 385–475.
H. Berestycki, P. -L. Lions, [*Nonlinear scalar field equations. I. Existence of a ground state*]{}, Arch. Ration. Mech. Anal. [**82**]{} (1983), no. 4, 313–345.
C. Bucur and E. Valdinoci, [*Nonlocal diffusion and applications*]{}. Lecture Notes of the Unione Matematica Italiana, 20. Springer, \[Cham\]; Unione Matematica Italiana, Bologna, 2016. xii+155 pp. ISBN: 978-3-319-28738-6; 978-3-319-28739-3.
H. Bueno, O.H. Miyagaki, G.A. Pereira, [*Remarks about a generalized pseudo-relativistic Hartree equation*]{}, J. Differential Equations [**266**]{} (2019), no. 1, 876–909.
T. Byczkowski, J. Malecki, M. Ryznar, [*Bessel potentials, hitting distributions and Green functions*]{}, Trans. Amer. Math. Soc. [**361**]{} (2009), no. 9, 4871–4900.
L. Caffarelli, L. Silvestre, [*An extension problem related to the fractional Laplacian*]{}, Comm. Partial Differential Equations [**32**]{} (2007), no. 7-9, 1245–1260
A.-P. Calderón, [*Lebesgue spaces of differentiable functions and distributions*]{}, 1961 Proc. Sympos. Pure Math., Vol. IV pp. 33–49 American Mathematical Society, Providence, R.I.
R. Carmona, W.C. Masters, B. Simon, [*Relativistic Schrödinger operators: Asymptotic behavior of the eigenfunctions*]{}, J. Func. Anal [**91**]{} (1990), 117–142.
S. Cingolani, S. Secchi, [*Ground states for the pseudo-relativistic Hartree equation with external potential*]{}, Proc. Roy. Soc. Edinburgh Sect. A [**145**]{} (2015), no. 1, 73–90.
S. Cingolani, S. Secchi, [*Semiclassical analysis for pseudo-relativistic Hartree equations*]{}, J. Differential Equations [**258**]{} (2015), no. 12, 4156–4179.
V. Coti Zelati, M. Nolasco, [*Existence of ground states for nonlinear, pseudo-relativistic [S]{}chr[ö]{}dinger equations*]{}, Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. [**22** ]{} (2011), 51–72.
V. Coti Zelati, M. Nolasco, [*Ground states for pseudo-relativistic Hartree equations of critical type*]{}, Rev. Mat. Iberoam. [**29**]{} (2013), no. 4, 1421–1436.
J. Dávila, M. del Pino, J. Wei, [*Concentrating standing waves for the fractional nonlinear Schrödinger equation*]{}, J. Differential Equations [**256**]{} (2014), no. 2, 858–892.
M. Del Pino, P.L. Felmer, [*Local mountain passes for semilinear elliptic problems in unbounded domains*]{}, Calc. Var. Partial Differential Equations, [**4**]{} (1996), 121–137.
E. Di Nezza, G. Palatucci, E. Valdinoci, [*Hitchhiker’s guide to the fractional Sobolev spaces*]{}, Bull. Sci. Math. [**136**]{} (2012), no. 5, 521–573.
S. Dipierro, M. Medina, E. Valdinoci, [*Fractional elliptic problems with critical growth in the whole of ${\mathbb{R}}^n$*]{}, Appunti. Scuola Normale Superiore di Pisa (Nuova Serie) \[Lecture Notes. Scuola Normale Superiore di Pisa (New Series)\], 15. Edizioni della Normale, Pisa, 2017. viii+152 pp.
A. Elgart, B. Schlein, [*Mean field dynamics of Boson stars*]{}, Comm. Pure Appl. Math. [**60**]{} (2007) 500–545.
A. Erdélyi, W. Magnus, F. Oberhettinger, F. G. Tricomi, [*Higher transcendental functions. Vol. II*]{}, Based on notes left by Harry Bateman. Reprint of the 1953 original. Robert E. Krieger Publishing Co., Inc., Melbourne, Fla., 1981. xviii+396 pp.
E. B. Fabes, C. E. Kenig, R. P. Serapioni, [*The local regularity of solutions of degenerate elliptic equations*]{}, Comm. Partial Differential Equations [**7**]{} (1982), no. 1, 77–116.
M. M. Fall, V. Felli, [*Unique continuation properties for relativistic Schrödinger operators with a singular potential*]{}, Discrete Contin. Dyn. Syst. [**35**]{} (2015), no. 12, 5827–5867.
M. M. Fall, F. Mahmoudi, E. Valdinoci, [*Ground states and concentration phenomena for the fractional Schrödinger equation*]{}, Nonlinearity [**28**]{} (2015), no. 6, 1937–1961.
P. Felmer, A. Quaas, J. Tan, [*Positive solutions of the nonlinear Schrödinger equation with the fractional Laplacian*]{}, Proc. Roy. Soc. Edinburgh Sect. A [**142**]{} (2012), no. 6, 1237–1262.
P. Felmer, I. Vergara, [*Scalar field equation with non-local diffusion*]{}, NoDEA Nonlinear Differential Equations Appl. [**22**]{} (2015), no. 5, 1411–1428.
G.M. Figueiredo, G. Siciliano, [*A multiplicity result via Ljusternick-Schnirelmann category and Morse theory for a fractional Schrödinger equation in ${\mathbb{R}}^{N}$*]{}, NoDEA Nonlinear Differential Equations Appl. [**23**]{} (2016), art. 12, 22 pp.
L. Grafakos, [*Modern Fourier analysis*]{}, Third edition. Graduate Texts in Mathematics, 250. Springer, New York, 2014. xvi+624 pp.
T. Grzywny, M. Ryznar, [*Two-sided optimal bounds for Green functions of half-spaces for relativistic $\alpha$-stable process*]{}, Potential Anal. [**28**]{} (2008), no. 3, 201–239.
I. W. Herbst, [*Spectral theory of the operator $(p^{2}+m^{2})^{1/2}-Ze^{2}/r$*]{}, Comm. Math. Phys. [**53**]{} (1977), no. 3, 285–294.
N. Ikoma, [*Existence of solutions of scalar field equations with fractional operator*]{}, J. Fixed Point Theory Appl. [**19**]{} (2017), no. 1, 649–690.
T. Jin, Y. Li, J. Xiong, [*On a fractional Nirenberg problem, part I: blow up analysis and compactness of solutions*]{}, J. Eur. Math. Soc. (JEMS) [**16**]{} (2014), no. 6, 1111–1171.
S. G. Krantz, [*Lipschitz spaces, smoothness of functions, and approximation theory*]{}, Exposition. Math. [**1**]{} (1983), no. 3, 193–260.
N. Laskin, [*Fractional quantum mechanics*]{}, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2018. xv+341 pp. E. H. Lieb, M. Loss, [*Analysis*]{}, Graduate Studies in Mathematics, 14. American Mathematical Society, Providence, RI, 1997. xviii+278 pp.
E. H. Lieb, H. T. Yau, [*The Chandrasekhar theory of stellar collapse as the limit of quantum mechanics*]{}, Comm. Math. Phys. 112 (1987), no. 1, 147–174.
E. H. Lieb, H. T. Yau, [*The stability and instability of relativistic matter*]{}, Comm. Math. Phys., [**118**]{} (1988), no. 2, 177–213.
G. Molica Bisci, V. Rădulescu, R. Servadei, [*Variational Methods for Nonlocal Fractional Problems*]{}, [*Cambridge University Press*]{}, **162** Cambridge, 2016.
J. Moser, [*A new proof of De Giorgi’s theorem concerning the regularity problem for elliptic differential equations*]{}, [Comm. Pure Appl. Math.]{}, [**13**]{} (1960), 457–468.
M. Ryznar, [*Estimate of Green function for relativistic $\alpha$-stable processes*]{}, Potential Analysis, [**17**]{}, (2002), 1–23.
S. Secchi, [*On some nonlinear fractional equations involving the Bessel potential*]{}, J. Dynam. Differential Equations [**29**]{} (2017), no. 3, 1173–1193.
L. Silvestre, [*Regularity of the obstacle problem for a fractional power of the Laplace operator*]{}, Comm. Pure Appl. Math., [**60**]{} (2007), no. 1, 67–112.
E. Stein, [*Singular integrals and differentiability properties of functions*]{}, Princeton Mathematical Series, No. 30 Princeton University Press, Princeton, N.J. 1970 xiv+290 pp.
P. R. Stinga, J. L. Torrea, [*Extension problem and Harnack’s inequality for some fractional operators*]{}, Comm. Partial Differential Equations [**35**]{} (2010), no. 11, 2092–2122.
M. Willem, [*Minimax theorems*]{}, Progress in Nonlinear Differential Equations and their Applications, 24. Birkhäuser Boston, Inc., Boston, MA, 1996. x+162 pp.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The concept of freeze out volume used in many statistical approaches for disassembly of hot nuclei leads to ambiguities. The fragmentation pattern and the momentum distribution (temperature) of the emanated fragments are determined by the phase space at the freeze-out volume where the interaction among the fragments is supposedly frozen out. However, to get coherence with the experimental momentum distribution of the charged particles, one introduces Coulomb acceleration beyond this freeze-out. To be consistent, we investigate the effect of the attractive nuclear force beyond this volume and find that the possible recombination of the fragments alters the physical observables significantly casting doubt on the consistency of the statistical model.'
---
=8.1 true in =5.9 true in
\
1.0cm S. K. Samaddar$^1$, J. N. De$^{1,2}$ and A. Bonasera$^{3}$\
$^1$Saha Institute of Nuclear Physics, 1/AF Bidhannagar, Kolkata 700064, India\
$^2$The Cyclotron Institute, Texas A$\&$M University, College Station,\
Texas 77843, USA\
$^3$Laboratori Nazionali del Sud, INFN, via S. Sofia 44 95125 Catania,Italy\
1.0cm PACS Number(s): 25.70.Pq 24.10.Pa
Multifragmentation of nuclear systems in energetic nuclear collisions serves a novel window for understanding the properties of hot nuclear matter. It has a sensitive bearing on the nuclear equation of state (EoS) [@bona; @sto], focusses our attention on the possibility of liquid-gas phase transition in finite and infinite nuclear systems [@poc; @bor] and from the production of rare isotopes helps for a better understanding of the nucleosynthesis in astrophysical context [@ish]. Various statistical models have been suggested to explain the phenomenon of nuclear multifragmentation; dynamical models have also been proposed which we will not discuss further in this work[@bona; @sto; @bon1]. Broadly the statistical approaches are classified in two groups, namely, sequential binary decay (SBD) [@fri; @ric] and one-step prompt multifragmentation (PM) [@gro1; @bon2]. It is generally believed that at low excitation energy, fragmentation proceeds through SBD whereas at relatively higher energies, it is possibly a one-step break-up process. Nuclear disassembly in the PM picture has been viewed as a statistical process and different genres of statistical hierarchy have been employed to explain the physical process, from grand canonical [@koo], canonical [@bon2] to micro-canonical [@gro1]. In all these statistical calculations, a freeze-out volume, around $3V_0$ to $8V_0$ ($V_0$ being the normal nuclear volume of the fragmenting system) is employed when the PM process takes place. The fragments so generated are the primary fragments which are in general in the excited states. Secondary decay from these hot fragments have been taken into account [@kol; @hua]. Furthermore, in the PM models the fragments are generally distributed in the freeze-out volume and Coulomb trajectories are calculated for an improved description of the momentum distribution of the charged fragments. At high excitation energies a collective motion further need to be added to the fragments [@bona; @sto]. These are inconsistent with the assumption of [*freeze-out*]{} and is a pointer to the inadequacy of the statistical model. On the same footing, the action of interfragment nuclear force beyond freeze-out needs to be considered. The evolution of the fragments under the combined action of the Coulomb and nuclear fields has not been considered so far except the one reported in ref.[@pal1]. In the exit channel motion, two fragments, when close enough to be under the influence of the nuclear force may recombine to produce an excited heavier fragment which may or may not decay further. In [@pal1], these effects were taken into account and it was found that the yield of relatively heavier fragments is enhanced significantly. Of course this implies that the original yield and momentum distributions given by the statistical model are altered and in order to get, for instance, roughly the original yield, one has to adjust the parameters including the collective flow. The large success of statistical models (without recombination) in reproducing experimental data, has been tacitly assumed as a proof of equilibration in nuclear collisions. Including the nuclear force (which is a must) for such small freeze out volumes might significantly alter this ’idola tribus’ (or belief) and the role of dynamics, so far hidden under the carpet of a few parameters and ideal assumptions, must be reconsidered. In this work we will discuss the role of nuclear forces in the model beyond the freeze-out; because we are adding some minimum of dynamical effects, we will dub our approach as Dynamical Statistical Fragmentation Model (DSFM). In a later work we will discuss the flow effects[@pal1; @pal2]. The situation now is somewhat similar to fission, where one utilizes statistical models to determine the mass distribution at the saddle point and evolves the system including Coulomb, nuclear and even friction forces. At variance with fragmentation, the system evolves after the saddle point thus the mass distributions given by the statistical models are not altered. Including radial flow in DSFM will roughly cancel the effect of recombination and give a picture similar to the fission one.
Isotopic yields from multifragmentation have been employed to infer about important physical observables like the temperature of the fragmenting system [@alb] and the associated liquid-gas phase transition in finite nuclei. The effect of recombination has so far been ignored in drawing these inferences. The results reported in[@pal1] are restricted to a fixed freeze-out volume $V_f$ and excitation energy $E^*$; no attempt was made to look into the consequences of the changed isotopic yield on the physical observables after recombination. In the present communication, the effect of recombination with the variation in $V_f$ as well as in $E^*$ has been addressed in some detail and we find that the importance of recombination on the multifragmentation scenario can not be ignored, further adding to the ambiguities of the statistical approaches discussed above.
The model employed in the present calculation is the same as that in[@pal1]. For the sake of completeness, only the salient features of the methodology are discussed here. In the first step, the fragment multiplicities $n_i$ for the various fragments are evaluated in the grand canonical model (GCM). They are given by $$\begin{aligned}
n_i = V_f\left(\frac {mA_i}{2\pi\hbar
^2\beta}\right)^{3/2} \phi_i(\beta)exp[-\beta (B-B_i+V_i-\mu_n N_i-\mu_p
Z_i)],\end{aligned}$$ where $\beta$ is the inverse of the temperature $T$, $m$ the nucleon mass, $A_i, N_i$ and $Z_i$ are the mass, neutron and charge numbers of the fragment species $i$, $B$’s are the ground state binding energies of the fragmenting system and the generated species, $\mu$’s are the nucleonic chemical potentials and $\phi_i(\beta)$’s are the internal partition function. The internal partition function is calculated with the assumption that the excitation of the fragment is below the particle emission threshold. The single particle potential $V_i$ is the sum of the Coulomb and nuclear interaction of the $i$th fragment with the rest of the fragments and is evaluated in the complementary fragment approximation [@gro2; @sat]. Employing the GCM fragment formation probability $p_i=n_i/\sum n_i$, microcanonical events are generated following the method similar to that given by Fai and Randrup [@fai]. After generation of fragments in a microcanonical event, the fragments are placed in a nonoverlapping manner within the freeze-out volume. The [*microcanonical temperature*]{} is evaluated from energy conservation. The fragment velocities are generated from a Maxwell-Boltzmann distribution commensurate with the microcanonical temperature. At this stage the role of the statistical model is over, but a further Coulomb acceleration is now considered which is in contrast to the statistical assumption. Even if the introduction of dynamics in the model is accepted, one should be consistent and include the nuclear forces as well since the nuclear surfaces in the freeze out volume are rather close to each other. Evaluation of the Coulomb interaction is straightforward; the nuclear part of the interfragment interaction in the exit channel is broadly classified in three groups depending on the masses of the fragments. The details are given in ref.[@pal1]. Two fragments in the exit channel are assumed to coalesce when they touch each other. If the excitation energy of the coalesced fragment is above the particle emission threshold (taken as 8 $MeV$), the fragment is assumed to undergo binary decay; the decay probability is calculated in the transition state model of Swiatecki [@swi].
To study the effect of recombination in nuclear multifragmentation we have considered $^{197}Au$ as a representative system. In order to see the effect of recombination on excitation energy, the calculations have been performed at $E^*/A$=3, 4 and 5 $MeV$ with a fixed freeze-out volume $V_f=6V_0$. Volume effects have also been considered with $E^*/A$ fixed at 4 $MeV$. For generation of microcanonical ensemble, typically $10^5$ events have been used. Since we have assumed that the fragments are produced in the particle stable states, the charge or mass distribution is decided at the very onset of fragmentation if there is no recombination. The recombined complex may have excitation above the particle emission threshold and they may undergo sequential binary decay in flight till a particle-stable state is reached. In the panels (a), (b) and (c) of Fig.1, the charge distributions at different excitation energies at $V_f = 6V_0$ are displayed. Except for the very light charge particles, the fragment yield is substantially enhanced. At the lowest excitation energy considered (3 $MeV/A$), the yield of very heavy fragments is found to be somewhat reduced. The neutron yield is enhanced at all the excitation energies considered. This behavior results from a delicate interplay between fragment recombination and subsequent binary decay. It is expected that with reduction in freeze-out volume, the recombination effect would be more prominent. This is apparent from Fig.1(b) and 1(d). This is further evident from the left panel of Fig.2 where the charge distribution has been displayed for a very large freeze-out volume (16$V_0$) at the same excitation energy of 4 $MeV$ per particle. One would expect the recombination effect to be minimal at this large freeze-out volume, however, we find that though it is significantly reduced, it is not negligible, particularly for fragments with $Z > 10$. In order to understand the persistence of the recombination effect at this large freeze-out volume, we have calculated the surface to surface separation ($S$) of the different fragment pairs ($N_{pair}$) produced in a disassembly event. In the right panel of Fig.2, the average number of fragment pairs ($\langle N_{pair} \rangle$) present within the separation distance $S$ is displayed for different freeze-out volumes at $E^*/A$ = 4 $MeV$. The fragment pairs within the nuclear force range (taken as 1.4 $fm$ shown as the horizontal dotted line) are potential candidates to undergo recombination. It is seen from the figure that even at $V_f =
16V_0$, there are significant number of fragment pairs within the nuclear force field.
The knowledge of temperature of the disassembling system is crucial in drawing many important physical inferences such as liquid-gas phase transition. There is no direct way to measure the temperature in such processes; a number of thermometers have been proposed to that end. Experimentally, it has been the usual practice to resort to the isotopic double-ratio [@alb] to extract the temperature which is based on the statistical multifragmentation model with certain approximations. If the isotopic yield changes due to recombination, the extracted temperature based on this model is bound to be erroneous. To investigate this aspect, we have calculated temperatures from different isotopic double-ratios at a number of excitation energies with and without the effects of recombination. This is displayed in panels (a), (b) and (c) of Fig.3. It is found that the temperatures extracted without recombinaton are consistent with the excitation energies, however, with inclusion of recombination effects, the extracted temperatures from the isotopic double-ratios decrease dramatically. Recombination introduces a multitude of low temperature sources in the system which may be responsible for the reduction in the temperature observed. An anomalous fall in temperature at $E^*/A =$ 4 $MeV$ is also seen for all the double-ratio thermometers. The temperature extracted after recombination are, however, found to be not too sensitive to the excitation energy(3-6 $MeV$ per nucleon) that we have considered. The dependence of the double-ratio temperature on the freeze-out volume is displayed in Fig.3(d). We have chosen a representative thermometer $(d/t)/(^3He/^4He)$ at an excitation energy $E^*/A$ = 4 $MeV$. Even at the very large freeze-out volume of 16$V_0$, the temperatures extracted without and with recombination effects are appreciably different, but a very slow approach to a common temperature with increasing $V_f$ is apparent from the figure.
To sum up, the effect of recombination of fragments on the charge distributions and isotopic double-ratio temperatures in a nuclear disassembly process in the statistical model has been investigated in this paper at different excitation energies and freeze-out volumes. The effect is found to be very significant for both the observables. With recombination, the yields for relatively heavy fragments are appreciably enhanced at all the excitation energies we have considered. This persistence of larger yield continues even at a freeze-out volume as large as $16V_0$. With recombination considered, the isotopic double-ratio temperature is reduced dramatically. The extracted temperatures without recombination are found to be not too different from those obtained in the Fermi gas model, however, with inclusion of recombination, in the excitation energy range of 3-6 $MeV$ per nucleon that we have investigated, the temperatures are found to be $\sim 4$ $MeV$ and not too sensitive to the excitation energy. At relatively higher excitations, the collective flow and an improved dynamics in the fusion process are likely to play a very important role and should be taken into account. We, however, stress that the statistical approaches in a freeze-out scenario should not need any dynamics. The large effects seen with the introduction of dynamics (though has been done in an [*ad-hoc*]{} manner) even in a large freeze-out volume is counter-intuitive and cast doubt on the applicability of the freeze-out concept in the fragmentation process; nonequilibrium dynamical features [@bona; @bon1] should possibly be incorporated at the very outset. 0.7cm
[**Acknowledgement**]{}
S.K.S. acknowledges the Council of Scientific and Industrial Research of the Government of India, for the financial support. J.N.D. gratefully acknowledges the kind hospitality at the Cyclotron Institute at Texas A $\&$ M University where the work was partially done.
Ths DSFM-code discussed here is available and may be requested to one of the authors.
[99]{} A.Bonasera, M.Bruno, C.Dorso, P.F. Mastinu, Rivista del Nuovo Cimento 23(2000)1. H. Stocker, W. Greiner, Phys. Rep. 137(1986)277. J. Pochodzalla [*et al*]{}, Phys. Rev. Lett. 75(1995)1040. B. Borderie, J. Phys. G: Nucl. Part. Phys. 28(2002)R217. C. Ishizuka, A. Ohnishi, and K. Sumiyoshi,Nucl. Phys. A723(2003)517. J. P. Bondorf, D. Idier, I. N. Mishustin, Phys. Lett B359(1995)261. W. A. Friedman, W. G. Lynch, Phys. Rev. C28(1983)16. J. Richert ,P. Wagner, Nucl. Phys. A517(1990)399. D. H. E. Gross, Phys. Rep. 279(1997)119. J. P. Bondorf, A. S. Botvina, A. S. Iljinov, I. N. Mishustin, K. Sneppen, Phys. Rep. 257(1995)130. S. E. Koonin, J. Randrup, Nucl. Phys. A356(1981)223. A. Kolomiets [*et al*]{}, Phys. Rev. C54(1996)R472. M. J. Huang [*et al*]{}, Phys. Rev. Lett. 78(1997)1648. S. Pal, S. K. Samaddar, A. Das, J. N. De, Nucl. Phys. A586(1995)466. S. Pal, S. K. Samaddar, J. N. De, Nucl. Phys. A608(1996)49. S. Albergo, S. Costa, E. Costanzo, A. Rubbino, Nuovo Cimento A89(1985)1. D. H. E. Gross, L. Satpathy, M. T. Chung, M. Satpathy, Z. Phys. A309(1982)41. L. Satpathy, M. Mishra, A. Das, M. Satpathy, Phys. Lett. B237(1990)181. G. Fai, J. Randrup, Nucl. Phys. A404(1983)551. W. J. Swiatecki, Aust. J. Phys. 36(1983)641.
**Figure Captions**
- Charge distributions from the fragmenting system$^{197}Au$ with and without recombination at different excitation energies and freeze-out volumes as indicated in the figure.
- In the left panel the charge distribution from$^{197}Au$ at an excitation energy of 4 $MeV$ per nucleon and$V_f =
16V_0$ is displayed. The right panel shows the average number of fragment pairs within a separation distance $S$ at different freeze-out volumes at the same excitation energy.
- In panels (a), (b) and (c), different isotopic double-ratio temperatures are shown at various excitation energies at $V_f = 6V_0$ with and without recombination. In panel (d), the volume dependence of the double-ratio temperature $(d/t)/(^3He/^4He)$ at $E^*/A$ =4 $MeV$ is displayed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using 58$\times 10^{6}$ ${J/\psi}$ events collected with the Beijing Spectrometer (BESII) at the Beijing Electron Positron Collider (BEPC), the branching fractions of ${J/\psi}$ to $p\bar{p}\eta$ and $p\bar{p}{\eta^{\prime}}$ are determined. The ratio $\frac{\Gamma({J/\psi}{\rightarrow}{p\bar{p}}\eta)}{\Gamma({J/\psi}{\rightarrow}{p\bar{p}})}$ obtained by this analysis agrees with expectations based on soft-pion theorem calculations.'
author:
- |
M. Ablikim$^{1}$, J. Z. Bai$^{1}$, Y. Bai$^{1}$, Y. Ban$^{11}$, X. Cai$^{1}$, H. F. Chen$^{16}$, H. S. Chen$^{1}$, H. X. Chen$^{1}$, J. C. Chen$^{1}$, Jin Chen$^{1}$, X. D. Chen$^{5}$, Y. B. Chen$^{1}$, Y. P. Chu$^{1}$, Y. S. Dai$^{18}$, Z. Y. Deng$^{1}$, S. X. Du$^{1}$$^{a}$, J. Fang$^{1}$, C. D. Fu$^{1}$, C. S. Gao$^{1}$, Y. N. Gao$^{14}$, S. D. Gu$^{1}$, Y. T. Gu$^{4}$, Y. N. Guo$^{1}$, Z. J. Guo$^{15}$$^{b}$, F. A. Harris$^{15}$, K. L. He$^{1}$, M. He$^{12}$, Y. K. Heng$^{1}$, H. M. Hu$^{1}$, T. Hu$^{1}$, G. S. Huang$^{1}$$^{c}$, X. T. Huang$^{12}$, Y. P. Huang$^{1}$, X. B. Ji$^{1}$, X. S. Jiang$^{1}$, J. B. Jiao$^{12}$, D. P. Jin$^{1}$, S. Jin$^{1}$, G. Li$^{1}$, H. B. Li$^{1}$, J. Li$^{1}$, L. Li$^{1}$, R. Y. Li$^{1}$, W. D. Li$^{1}$, W. G. Li$^{1}$, X. L. Li$^{1}$, X. N. Li$^{1}$, X. Q. Li$^{10}$, Y. F. Liang$^{13}$, B. J. Liu$^{1}$$^{d}$, C. X. Liu$^{1}$, Fang Liu$^{1}$, Feng Liu$^{6}$, H. M. Liu$^{1}$, J. P. Liu$^{17}$, H. B. Liu$^{4}$$^{e}$, J. Liu$^{1}$, Q. Liu$^{15}$, R. G. Liu$^{1}$, S. Liu$^{8}$, Z. A. Liu$^{1}$, F. Lu$^{1}$, G. R. Lu$^{5}$, J. G. Lu$^{1}$, C. L. Luo$^{9}$, F. C. Ma$^{8}$, H. L. Ma$^{2}$, Q. M. Ma$^{1}$, M. Q. A. Malik$^{1}$, Z. P. Mao$^{1}$, X. H. Mo$^{1}$, J. Nie$^{1}$, S. L. Olsen$^{15}$, R. G. Ping$^{1}$, N. D. Qi$^{1}$, J. F. Qiu$^{1}$, G. Rong$^{1}$, X. D. Ruan$^{4}$, L. Y. Shan$^{1}$, L. Shang$^{1}$, C. P. Shen$^{15}$, X. Y. Shen$^{1}$, H. Y. Sheng$^{1}$, H. S. Sun$^{1}$, S. S. Sun$^{1}$, Y. Z. Sun$^{1}$, Z. J. Sun$^{1}$, X. Tang$^{1}$, J. P. Tian$^{14}$, G. L. Tong$^{1}$, G. S. Varner$^{15}$, X. Wan$^{1}$, L. Wang$^{1}$, L. L. Wang$^{1}$, L. S. Wang$^{1}$, P. Wang$^{1}$, P. L. Wang$^{1}$, Y. F. Wang$^{1}$, Z. Wang$^{1}$, Z. Y. Wang$^{1}$, C. L. Wei$^{1}$, D. H. Wei$^{3}$, N. Wu$^{1}$, X. M. Xia$^{1}$, G. F. Xu$^{1}$, X. P. Xu$^{6}$, Y. Xu$^{10}$, M. L. Yan$^{16}$, H. X. Yang$^{1}$, M. Yang$^{1}$, Y. X. Yang$^{3}$, M. H. Ye$^{2}$, Y. X. Ye$^{16}$, C. X. Yu$^{10}$, C. Z. Yuan$^{1}$, Y. Yuan$^{1}$, Y. Zeng$^{7}$, B. X. Zhang$^{1}$, B. Y. Zhang$^{1}$, C. C. Zhang$^{1}$, D. H. Zhang$^{1}$, H. Q. Zhang$^{1}$, H. Y. Zhang$^{1}$, J. W. Zhang$^{1}$, J. Y. Zhang$^{1}$, X. Y. Zhang$^{12}$, Y. Y. Zhang$^{13}$, Z. X. Zhang$^{11}$, Z. P. Zhang$^{16}$, D. X. Zhao$^{1}$, J. W. Zhao$^{1}$, M. G. Zhao$^{1}$, P. P. Zhao$^{1}$, Z. G. Zhao$^{16}$, B. Zheng$^{1}$, H. Q. Zheng$^{11}$, J. P. Zheng$^{1}$, Z. P. Zheng$^{1}$, B. Zhong$^{9}$ L. Zhou$^{1}$, K. J. Zhu$^{1}$, Q. M. Zhu$^{1}$, X. W. Zhu$^{1}$, Y. S. Zhu$^{1}$, Z. A. Zhu$^{1}$, Z. L. Zhu$^{3}$, B. A. Zhuang$^{1}$, B. S. Zou$^{1}$\
(BES Collaboration)\
title: '**Measurement of the Branching Fractions for $J/\psi$ ${\rightarrow}$ $p\bar{p}\eta$ and $p\bar{p}{\eta^{\prime}}$** '
---
Introduction {#introd}
============
The ${J/\psi}$ meson has hadronic, electromagnetic, and radiative decays to light hadrons, and a radiative transition to the ${\eta_{c}}$. In Ref. [@PCX], direct hadronic, electromagnetic and radiative decays are estimated to account for 69.2$\%$, 13.4$\%$, and 4.3$\%$, respectively, of all ${J/\psi}$ decays. However, individual exclusive ${J/\psi}$ decays are more difficult to analyze quantitatively in QCD. To date, two-body decay modes such as ${J/\psi \rightarrow }B_{8}\bar{B_{8}}$ or $P_{9}V_{9}$, where $B_{8}$, $P_{9}$ and $V_{9}$ refer to baryon octet, pseudoscalar nonet, and vector nonet particle, respectively, have been studied with some success using an effective model, and other similar methods [@SPIT].
Studies of three-body decays of ${J/\psi}$ are a natural extension of studies of two-body decays. Since most ${J/\psi}$ decays proceed via two-body intermediate states, including wide resonances, it is hard to experimentally extract the non-resonant three-body contribution [@JPDC]. Specific models based on proton and $N^{*}$ pole diagrams have been introduced to deal with these problems [@SPIT]. In the calculation, the soft-pion theorem [@LABW] has been applied to the decay ${J/\psi \rightarrow }{p\bar{p}}{\pi^0}$ successfully. This method has also been used for ${J/\psi \rightarrow }{p\bar{p}}\eta$ and ${J/\psi \rightarrow }{p\bar{p}}{\eta^{\prime}}$ decays [@SPIT].
This paper reports measurement of the branching fractions for $p\bar{p}\eta$ and $p\bar{p}{\eta^{\prime}}$, and tests of the soft-pion theorem for ${J/\psi \rightarrow }{p\bar{p}}\eta$, which states [@SPIT]: $$\begin{aligned}
\frac{\Gamma({J/\psi}{\rightarrow}{p\bar{p}}\eta)}{\Gamma({J/\psi}{\rightarrow}{p\bar{p}})}\simeq 0.64\pm0.52 .\end{aligned}$$
The BES detector and Monte Carlo simulation {#BESD}
===========================================
BESII is a conventional solenoidal magnet detector that is described in detail in Refs. [@JZB]. A 12-layer vertex chamber (VC) surrounding the beam pipe provides trigger and track information. A forty-layer main drift chamber (MDC), located radially outside the VC, provides trajectory and energy loss ($dE/dx$) information for tracks over $85\%$ of the total solid angle. The momentum resolution is $\sigma _p/p = 0.017 \sqrt{1+p^2}$ ($p$ in $\hbox{\rm GeV}/c$), and the $dE/dx$ resolution for hadron tracks is $\sim 8\%$. An array of 48 scintillation counters surrounding the MDC measures the time-of-flight (TOF) of tracks with a resolution of $\sim 200$ ps for hadrons. Radially outside the TOF system is a 12 radiation length, lead-gas barrel shower counter (BSC). This measures the energies of electrons and photons over $\sim 80\%$ of the total solid angle with an energy resolution of $\sigma_E/E=22\%/\sqrt{E}$ ($E$ in GeV). Outside of the solenoidal coil, which provides a 0.4 Tesla magnetic field over the tracking volume, is an iron flux return that is instrumented with three double layers of counters that identify muons of momentum greater than 0.5 GeV/$c$.
In the analysis, a GEANT3-based Monte Carlo (MC) simulation program (SIMBES) [@GEANT] with detailed consideration of detector performance is used. The consistency between data and MC has been validated using many high purity physics channels [@NIM].
In this analysis, the detection efficiency for each decay mode is determined by a MC simulation that takes into account the angular distributions. For ${J/\psi}{\rightarrow}{p\bar{p}}\eta$, the angle ($\theta$) between the directions of $e^{+}$ and $p$ in the laboratory frame is generated according to $1+\alpha\cdot\cos^{2}\theta$ distribution, where $\alpha$ is obtained by fitting the data from ${J/\psi}{\rightarrow}{p\bar{p}}\eta$. A uniform phase space distribution is used for ${J/\psi}$ decaying into ${p\bar{p}}{\eta^{\prime}}$.
General Selection Criteria
==========================
Candidate events are required to satisfy the following common selection criteria:
Charged track selection
-----------------------
Each charged track must: (1) have a good helix fit in order to ensure a correct error matrix in the kinematic fit; (2) originate from the interaction region, $\sqrt{V^{2}_{x}+V^{2}_{y}}<2$ cm and $|V_{z}|<20$ cm, where $V_{x}$, $V_{y}$, and $V_{z}$ are the $x$, $y$ and $z$ coordinates of the point of closest approach of the track to the beam axis; (3) have a transverse momentum greater than 70 MeV/$c$; and (4) have $|\cos \theta|\le 0.80$, where $\theta$ is the polar angle of the track.
Photon selection
----------------
A neutral cluster in the BSC is assumed to be a photon candidate if the following requirements are satisfied: (1) the energy deposited in the BSC is greater than 0.05 GeV; (2) energy is deposited in more than one layer; (3) the angle between the direction of photon emission and the direction of shower development is less than $30^{\circ}$; and (4) the angle between the photon and the nearest charged track is greater than $5^{\circ}$ (if the charged track is an antiproton, the angle is required to be great than $25^{\circ}$).
Particle Identification (PID)
-----------------------------
For each charged track in an event, $\chi^{2}_{PID}(i)$ is determined using both $dE/dx$ and TOF information:
$\chi^{2}_{PID}(i)=\chi^{2}_{dE/dx}(i)+\chi^{2}_{TOF}(i)$,\
where $i$ corresponds to the particle hypothesis. A charged track is identified as a pion if $\chi^{2}_{PID}$ for the $\pi$ hypothesis is less than those for the kaon and proton hypotheses. For $p$ or $\bar{p}$ identification, the same method is used. In this analysis, all charged tracks are required to be positively identified.
Analysis of ${J/\psi}{\rightarrow}{p\bar{p}}\eta$
=================================================
The decay modes for the ${J/\psi}{\rightarrow}{p\bar{p}}\eta$ measurement are $\eta
{\rightarrow}\gamma \gamma$ and $\eta {\rightarrow}{\pi^+\pi^-\pi^0}$. The use of different decay modes allows us to cross check our measurements, as well as to obtain higher statistical precision.
$\eta {\rightarrow}\gamma\gamma$
--------------------------------
Events with two charged tracks and two photons are selected. A four-constraint (4C) kinematic fit is performed to the hypothesis ${J/\psi}{\rightarrow}{p\bar{p}}\gamma\gamma$. For events with more than two photons, all combinations are tried, and the combination with the smallest $\chi^{2}$ is retained. $\chi^{2}_{\gamma\gamma{p\bar{p}}}$ is required to be less than 20.
The $\gamma\gamma$ invariant mass $(m_{\gamma\gamma})$ distribution for selected events is shown in Fig. \[mggfit\]. A peak around the $\eta$ mass is evident. The curves in the figure indicate the best fit to the signal and background. The shaded part is the background estimated from a MC simulation of inclusive ${J/\psi}$ events [@LUND]. The main background comes from ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^0}{\pi^0}$ and $\Sigma^{+}\bar{\Sigma}^{-}$. By fitting the $\eta$ signal with a MC-simulated signal histogram plus a third order polynomial background function, the number of $\eta$ signal events is determined to be $(12220\pm149)$.
For the signal MC simulation, the events are generated with a proton angle distribution of $1+\alpha\cos^{2}\theta$, where $\alpha$ is taken to be -0.6185 in order to describe the data. In the decay, intermediate resonances, N(1440), N(1535), N(1650), and N(1800) and antiparticles, with fractional contribution of $(8\pm4)$%, $(56\pm15)$%, $(24^{+5}_{-15})$%, and $(12\pm7)$% [@LHBD], respectively, are included. The resulting detection efficiency for ${J/\psi}{\rightarrow}{p\bar{p}}\eta$ $(\eta{\rightarrow}\gamma\gamma)$ is determined to be $28.70$%. The ${p\bar{p}}\eta$ branching fraction, calculated using $$\begin{aligned}
B({J/\psi}{\rightarrow}{p\bar{p}}\eta)=\frac{N_{obs}}{\epsilon\cdot N_{{J/\psi}}\cdot
B(\eta{\rightarrow}2\gamma)\cdot f_{1}},\end{aligned}$$ is $(1.93\pm0.02)\times 10^{-3}$, where the error is statistical only. Here $N_{obs}$ represents the number of observed events, $\epsilon$ is the detection efficiency for ${J/\psi}{\rightarrow}{p\bar{p}}\eta (\eta\to
2\gamma)$, $f_{1} = 0.9739$ is the efficiency correction factor (see Section \[SYSE\]), and $N_{{J/\psi}}$ is the total number of ${J/\psi}$ events.
$\eta{\rightarrow}{\pi^+}{\pi^-}{\pi^0}$
----------------------------------------
Similar to the above analysis, events with four charged tracks and two photons are selected. A 4C kinematic fit is performed to the ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^+}{\pi^-}\gamma\gamma$ hypothesis, and the $\chi^{2}_{\gamma\gamma{p\bar{p}}{\pi^+}{\pi^-}}$ value is required to be less than 20. In order to suppress multi-photon backgrounds, the number of photons is required to be two. The invariant mass of the $\gamma\gamma$ is required to be between 0.095 and 0.175 GeV/$c^{2}$.
The ${\pi^+\pi^-\pi^0}$ invariant mass $(m_{{\pi^+\pi^-\pi^0}})$ distribution is shown in Fig. \[m3pifit\], where a peak at the $\eta$ mass is observed. The curves in the figure are the results of a fit to the signal and background. The shaded part is background estimated from MC simulation of inclusive ${J/\psi}$ decay events [@LUND]. Here the main background comes from ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^+\pi^-\pi^0}$ and ${p\bar{p}}{\pi^+}{\pi^-}\gamma$ decays. By fitting the distribution with a MC simulated signal histogram plus a third order polynomial background function, $(954\pm45)$ signal events are obtained. Similar to the $\eta {\rightarrow}2 \gamma$ decay, contributions from the baryon excited states N(1440), N(1535), N(1650), and N(1800), as well as their anti-particles [@LHBD], are considered. The detection efficiency of ${J/\psi}{\rightarrow}{p\bar{p}}\eta$ $(\eta{\rightarrow}{\pi^+\pi^-\pi^0}$, ${\pi^0}{\rightarrow}\gamma\gamma)$ is determined to be $4.20$%. The branching fraction is determined from the calculation $$\begin{aligned}
B({J/\psi}{\rightarrow}{p\bar{p}}\eta)&=&\frac{N_{obs}}{\epsilon\cdot N_{{J/\psi}}\cdot
B(\eta{\rightarrow}{\pi^+\pi^-\pi^0})}\\
&\cdot& \frac{1}{B({\pi^0}{\rightarrow}\gamma\gamma)\cdot f_{2}},\end{aligned}$$ where $f_{2} = 0.9582$ is a correction factor for the efficiency that is described below in Section \[SYSE\]. We determine a branching fraction for ${J/\psi}{\rightarrow}{p\bar{p}}\eta$ of $(1.83\pm0.09)\times 10^{-3}$, where the error is statistical only.
Analysis of ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$
============================================================
There are three main decay modes of the ${\eta^{\prime}}$: ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, ${\eta^{\prime}}{\rightarrow}\gamma{\rho^0}$ and ${\eta^{\prime}}{\rightarrow}{\pi^0}{\pi^0}\eta$. Here the first two decay modes are used.
${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, $\eta{\rightarrow}\gamma\gamma$
---------------------------------------------------------------------------------
In the search for ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$ decays, events with four charged tracks and two photons are selected. A five-constraint (5C) kinematic fit is performed to the hypothesis of ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^+}{\pi^-}\gamma\gamma$, in which the $2\gamma$ invariant mass is constrained to equal the $\eta$ mass, and the $\chi^{2}_{\gamma\gamma{p\bar{p}}{\pi^+}{\pi^-}}$ value is required to be less than 20.
The ${\pi^+}{\pi^-}\eta$ invariant mass $(m_{{\pi^+}{\pi^-}\eta})$ distribution for events that survive the selection criteria is shown in Fig. \[metp1fit\]. A clear ${\eta^{\prime}}$ signal is observed. The curves in the figure are the best fit to the signal and background. The shaded part is background estimated from MC simulation of inclusive ${J/\psi}$ decay events [@LUND]. The main background comes from ${J/\psi}{\rightarrow}\Delta^{+}\bar{\Delta}^{-}\eta$, and $\Delta^{0}\bar{\Delta}^{0}\eta$ decays. By fitting the distribution with a MC simulated signal histogram plus a third order polynomial background function, a signal yield of $(65\pm12)$ events is observed. According to a MC simulation, in which the events are generated with uniform phase space, the detection efficiency of ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ $({\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, $\eta{\rightarrow}\gamma\gamma)$ is $3.38$%. The effect of intermediate resonances is considered as a source of systematic error. Using $$\begin{aligned}
B({J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}})&=&\frac{N_{obs}}{\epsilon\cdot N_{{J/\psi}}\cdot
B({\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta)}\\
&\cdot& \frac{1}{B(\eta{\rightarrow}\gamma\gamma)\cdot f_{3}}\end{aligned}$$ with $f_{3} = 0.8228$ being the efficiency correction factor (see Section \[SYSE\]), we determine the branching fraction for ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ to be $(2.31\pm0.43)\times 10^{-4}$, where the error is statistical only.
${\eta^{\prime}}{\rightarrow}\gamma{\rho^0},{\rho^0}{\rightarrow}{\pi^+}{\pi^-}$
--------------------------------------------------------------------------------
In order to select ${\eta^{\prime}}{\rightarrow}\gamma{\rho^0}$, a 4C kinematic fit is performed under the hypothesis of ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^+}{\pi^-}\gamma$. The $\chi^{2}_{\gamma{p\bar{p}}{\pi^+}{\pi^-}}$ value is required to be less than 20. To ensure the events are from $\gamma\rho^{0}$, a $|m_{{\pi^+}{\pi^-}}-m_{\rho}|<0.20$ GeV/$c^{2}$ requirement is imposed, where $m_{{\pi^+}{\pi^-}}$ is the ${\pi^+}{\pi^-}$ invariant mass, and $m_{\rho}$ is the $\rho$ mass. In order to exclude the background from ${J/\psi}{\rightarrow}p\bar{p}{\pi^+}{\pi^-}$, it is required that the invariant mass of the four charged tracks is less than 3.02 GeV/$c^{2}$.
The $\gamma{\rho^0}$ invariant mass $(m_{\gamma{\rho^0}})$ distribution for selected events, where a clear ${\eta^{\prime}}$ signal is observed, is shown in Fig. \[metp2fit\]. The curves in the figure are the best fit to the signal and background. The shaded part is the background estimated from MC simulation of inclusive ${J/\psi}$ decay events [@LUND]. The main background comes from ${J/\psi}{\rightarrow}{p\bar{p}}{\pi^+}{\pi^-}\gamma$, ${\Delta}^{++}\bar{\Delta}^{--}{\pi^0}$ and ${p\bar{p}}{\pi^+\pi^-\pi^0}$ decays. By fitting the $m_{\gamma{\pi^+}{\pi^-}}$ distribution with a MC simulated signal shape and a third order polynomial background function, we determine the number of ${\eta^{\prime}}$ signal events to be $(200\pm29)$. The detection efficiency for ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ $({\eta^{\prime}}{\rightarrow}\gamma{\rho^0})$ is determined to be $7.48$%, assuming phase space production, where the ${\pi^+}{\pi^-}$ mass distribution is generated according to measurements from${J/\psi}{\rightarrow}\phi{\eta^{\prime}},{\eta^{\prime}}{\rightarrow}\gamma{\pi^+}{\pi^-}$ [@FPRD71]. Using $$\begin{aligned}
B({J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}})&=&\frac{N_{obs}}{\epsilon\cdot
N_{{J/\psi}}\cdot B({\eta^{\prime}}{\rightarrow}\gamma{\rho^0})\cdot f_{4}}\end{aligned}$$ with the $f_{4}$ correction factor of 0.8522 (see Section \[SYSE\]). The resulting branching fraction for ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ is $(1.85\pm0.27)\times 10^{-4}$, where the error is statistical only.
\[TABFAC\]
Decay mode $N_{obs}$ $\epsilon(\%)$ Branching fraction
------------------------------------------------------------------------------------------------- ---------------- ---------------- -----------------------------------------------------------------------------------------
${J/\psi}{\rightarrow}{p\bar{p}}\eta,\eta{\rightarrow}\gamma\gamma$ $12220\pm 149$ $28.70$ $B({J/\psi}{\rightarrow}{p\bar{p}}\eta)=(1.93\pm 0.02\pm 0.18)\times10^{-3}$
${J/\psi}{\rightarrow}{p\bar{p}}\eta,\eta{\rightarrow}{\pi^+\pi^-\pi^0}$ $954\pm 45$ $4.20$ $B({J/\psi}{\rightarrow}{p\bar{p}}\eta)=(1.83\pm 0.09\pm 0.24)\times10^{-3}$
${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}},{\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$ $65\pm 12$ $3.38$ $B({J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}})=(2.31\pm 0.43\pm 0.34)\times10^{-4}$
${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}},{\eta^{\prime}}{\rightarrow}\gamma{\rho^0}$ $200\pm 29$ $7.48$ $B({J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}})=(1.85\pm 0.27\pm 0.31)\times10^{-4}$
Systematic errors {#SYSE}
=================
In our analysis, the systematic errors on the branching fractions come from the uncertainties in the MDC tracking, photon efficiency, particle identification, photon identification, kinematic fit, background shapes, hadronic interaction model, intermediate decay branching fraction, the ${\pi^0}$ and $\rho$ selection requirements, intermediate resonance states, and the total number of ${J/\psi}$ events. The errors from the different sources are listed in Table \[TABSYS\].
The MDC tracking efficiency has been measured using ${J/\psi}{\rightarrow}\rho\pi$, $\Lambda\bar{\Lambda}$, and $\psi(2S){\rightarrow}\pi^{+}\pi^{-}J/\psi$, $J/\psi$ to $\mu^{+}\mu^{-}$. The MC simulation agrees with data within 1 to 2$\%$ for each charged track [@NIM]. Thus $4\%$ is regarded as the systematic error for the two charged-track mode, and $8\%$ for the four charged-track final states.
The photon detection efficiency has been studied using a sample of $J/\psi {\rightarrow}\rho\pi$ [@NIM] decays; the difference between data and MC simulation is about $2\%$ for each photon. In this analysis, $2\%$ is included in the systematic error for one-photon modes and $4\%$ for two-photon modes.
The charged pion PID efficiency has been studied using $J/\psi{\rightarrow}\rho\pi$ decays [@NIM]. The PID efficiency from data is in good agreement with that from MC simulation with an average difference that is less than $1\%$ for each charged pion. Here $2\%$ is taken as the systematic error for identifying two pions.
The proton PID efficiencies have been studied using ${J/\psi}{\rightarrow}p
\bar{p}{\pi^+}{\pi^-}$ decays. The main difference between data and MC simulation occurs for tracks with momentum less than 0.35 GeV/$c$. We determine a weighting factor for identifying a proton or anti-proton as a function of momentum from studies of the ${J/\psi}{\rightarrow}p\bar{p}{\pi^+}{\pi^-}$ channel. After considering the weight of each particle in an event, the difference between data and MC is determined to be $\frac{\epsilon_{DT}}{\epsilon_{MC}}=0.9739\pm0.0078$ for $\eta{\rightarrow}2\gamma$, $0.9582\pm0.0199$ for $\eta{\rightarrow}3\pi$, $0.8228\pm0.0211$ for ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, and $0.8522\pm0.0140$ for ${\eta^{\prime}}{\rightarrow}\gamma\rho^{0}$. We take $f_{1}=0.9739$, $f_{2}=0.9582$, $f_{3}=0.8228$, and $f_{4}=0.8522$ as efficiency correction factors for the corresponding decay channel, and $0.8\%$, $2.1\%$, $2.6\%$, and $1.6\%$ are taken as the errors associated with identifying protons and anti-protons, respectively. The PID systematic errors for the four decay modes are listed in Table \[TABSYS\].
For the systematic error of photon ID, which arises mainly from the simulation of fake photons, ${p\bar{p}}$ and ${J/\psi}\to {p\bar{p}}{\pi^+}{\pi^-}$ data samples were selected and $10^{5}$ simulated ${p\bar{p}}$ and ${J/\psi}\to
{p\bar{p}}{\pi^+}{\pi^-}$ events were generated, with real and fake photons. The decay mode ${J/\psi}\to {p\bar{p}}$ is used for the photon ID systematic error of ${J/\psi}\to {p\bar{p}}\eta$ $(\eta\to2\gamma)$, and the decay mode ${J/\psi}\to
{p\bar{p}}{\pi^+}{\pi^-}$ for ${J/\psi}\to {p\bar{p}}\eta$ $(\eta\to 3\pi)$ and ${J/\psi}\to
{p\bar{p}}{\eta^{\prime}}$. From the decay mode ${J/\psi}\to {p\bar{p}}$, the fake photon differences between data and MC is about $2.0\%$, while for the decay mode ${J/\psi}\to{p\bar{p}}{\pi^+}{\pi^-}$, the difference is $1.6\%$. Here $2.0\%$ is taken as the systematic error associated with photon ID for the decay mode determined to be ${J/\psi}\to{p\bar{p}}\eta$ $(\eta\to\gamma\gamma)$, and $1.6\%$ for the decay modes ${J/\psi}\to{p\bar{p}}\eta$ $(\eta\to 3\pi)$ and ${J/\psi}\to{p\bar{p}}{\eta^{\prime}}$.
In Ref. [@PRD7], the uncertainty of the 4C kinematic fit is $4\%$, which we include here in the systematic error. The uncertainty of the 5C kinematic fit is $4.1\%$ in Ref. [@RHOP]. Here we conservatively take $5\%$ as the systematic error from the 5C kinematic fit for the decay mode ${\eta^{\prime}}\to {\pi^+}{\pi^-}\eta$.
The systematic errors of the background uncertainty is obtained by changing the range of the fit and varying the order of the polynomial background. The errors range from 0.8$\%$ to 7.3$\%$ in all decay modes (see Table \[TABSYS\] for detail).
There are two models, FLUKA and GCALOR, used for simulating hadronic interactions; the different models lead to different detection efficiencies. The difference between them is regarded as a systematic error. For the decay ${J/\psi}\to {p\bar{p}}\eta$ $(\eta {\rightarrow}2\gamma)$, the difference is very small and negligible. For the other decay modes, it is about 1.4$\%$ for ${J/\psi}{\rightarrow}{p\bar{p}}\eta$ $(\eta\to{\pi^+\pi^-\pi^0})$, ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ $({\eta^{\prime}}\to{\pi^+}{\pi^-}\eta)$, and 5.2$\%$ for ${J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}}$ $({\eta^{\prime}}\to\gamma{\rho^0})$.
The branching fractions for the decays ${\pi^0}{\rightarrow}2\gamma$, $\eta {\rightarrow}2\gamma$, $\eta {\rightarrow}{\pi^+\pi^-\pi^0}$, ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, and ${\eta^{\prime}}{\rightarrow}\gamma\rho$ are taken from the PDG [@PDG]. The errors on these branching fractions are systematic errors in our measurements.
For the $\eta\to 3\pi$ mode, the ${\pi^0}$ mass is required to satisfy $|M_{\gamma\gamma}-M_{{\pi^0}}|<0.04$ GeV/$c^{2}$. To study the systematic error associated with this requirement, ${\pi^0}$ samples are selected and simulated using ${J/\psi}\to \rho\pi$, and the data and MC efficiencies in the 3$\sigma$ signal region are compared with using the requirement or not, the difference is about $1\%$. Here it is taken as the systematic error caused by the ${\pi^0}$ requirement.
For the ${\eta^{\prime}}\to \gamma\rho$ mode, we require that $|M_{{\pi^+}{\pi^-}}-M_{\rho}|<0.20$ GeV/$c^{2}$. According to Ref. [@FANGP], the uncertainty associated with this requirement is $5.9\%$. Here we take this as the systematic error for the $\rho$ mass requirement.
In the signal MC simulation, we assume the presence of N(1440), N(1535), N(1650), and N(1800) in the ${p\bar{p}}\eta$ channel. If some of these resonances are not included, the efficiency of this channel changes. These differences are taken as the systematic error associated with possible intermediate states. The total systematic error associated with this is taken as the sum added in quadrature. For the decay modes with an ${\eta^{\prime}}$, we take the difference in efficiency determined assuming the decay proceeds via an intermediate $N(2090)$ resonance compared with phase space generation as the systematic error (see Table \[TABSYS\] for detail).
The uncertainty of the total number of ${J/\psi}$ events is $4.7\%$ [@FSS]. Combining all errors in quadrature gives total systematic errors of $9.3\%$ for $\eta{\rightarrow}\gamma\gamma$, $12.9\%$ for $\eta{\rightarrow}{\pi^+\pi^-\pi^0}$, $14.8\%$ for ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$, and $16.6\%$ for ${\eta^{\prime}}{\rightarrow}\gamma\rho$.
\[TABSYS\]
Sources Relative error ($\%$)
----------------------------- ---------------------------- -------------------------------------- -------------------------------------------------- ----------------------------------------------
Decay modes $\eta{\rightarrow}2\gamma$ $\eta{\rightarrow}{\pi^+\pi^-\pi^0}$ ${\eta^{\prime}}{\rightarrow}{\pi^+}{\pi^-}\eta$ ${\eta^{\prime}}{\rightarrow}\gamma{\rho^0}$
MDC tracking 4 8 8 8
Photon detection efficiency 4 4 4 2
Particle ID $\sim$1 4.1 4.6 3.6
Photon ID 2.0 1.6 1.6 1.6
Kinematic fit 4.0 4.0 5.0 4.0
Background uncertainty $\sim$1 3.1 7.3 5.8
Hadronic Interaction Model $\sim$0 1.4 1.4 5.2
Intermediate decay Br. Fr. $\sim$ 1 1.2 3.1 3.1
${\pi^0}$ selection - $\sim$1 -
$\rho$ selection - - - 5.9
Intermediate resonances 3.0 4.0 2.0 7.1
Number of ${J/\psi}$ events 4.7 4.7 4.7 4.7
Total systematic error 9.3 12.9 14.8 16.6
Results
=======
Table \[TABFAC\] shows the branching fractions of the two channels into their different decay modes; the first error is statistical and the second is systematic. The results for the different decay modes in the same channel are consistent within errors and are combined after taking out the common systematic errors (8.37 % for the $\eta$ mode and 10.8% for the $\eta^{'}$ mode): $$\begin{aligned}
Br({J/\psi}{\rightarrow}{p\bar{p}}\eta)=(1.91\pm0.17)\times 10^{-3},\\
Br({J/\psi}{\rightarrow}{p\bar{p}}{\eta^{\prime}})=(2.00\pm0.36)\times 10^{-4}.\end{aligned}$$ In comparison with previous measurements of ${J/\psi}\to{p\bar{p}}\eta$ and ${J/\psi}\to{p\bar{p}}{\eta^{\prime}}$, the present results are of higher precision.
Using the result of ${J/\psi}\to{p\bar{p}}\eta$ from this analysis and that of ${J/\psi}\to{p\bar{p}}$ in Ref. [@LXLM], we determine: $$\begin{aligned}
\frac{\Gamma({J/\psi}{\rightarrow}{p\bar{p}}\eta)}{\Gamma({J/\psi}{\rightarrow}{p\bar{p}})}=0.85\pm0.08.\end{aligned}$$ This is consistent with the calculation based on the soft-pion theorem, and indicates that the contribution of $N^{*}$- pole diagram is dominant for the ${J/\psi}\to{p\bar{p}}\eta$ mode.
Acknowledgments
===============
The BES collaboration thanks the staff of BEPC and computing center for their hard efforts. This work is supported in part by the National Natural Science Foundation of China under contracts Nos. 10491300, 10225524, 10225525, 10425523, 10625524, 10521003, 10821063, 10825524, the Chinese Academy of Sciences under contract No. KJ 95T-03, the 100 Talents Program of CAS under Contract Nos. U-11, U-24, U-25, and the Knowledge Innovation Project of CAS under Contract Nos. U-602, U-34 (IHEP), the National Natural Science Foundation of China under Contract No. 10225522 (Tsinghua University), and the Department of Energy under Contract No. DE-FG02-04ER41291 (U. Hawaii).
[\*\*]{} P. Wang, C. Z. Yuan, X. H. Mo, Phys. Rev. D [**70**]{} (2004) 114014. Rahul Sinha and Susumu Okubo, Phys. Rev. D [**30**]{} (1984) 2333. L. Köpke and N. Wermes, [*$J/\psi$ Decays*]{}, CERN, CH-1211 Geneva 23 Switzerland. L. Adler and R. F. Dashen, [*Current Algebra and Application to Particle Physics*]{} (Benjamin, New York, 1968); B. W. Lee [*Chiral Dynamics*]{} (Gordon and Breach, New York, 1972). J. Z. Bai [*et al.*]{}, Nucl. Instrum. Meth. A [**458**]{} (2001) 627. CERN Application Software Group, GEANT Detector Description and simulation Tool, CERN Program Library Writeup W5013, Geneva (1994). M. Ablikim [*et al.*]{}, Nucl. Instrum. Meth. A [**552**]{} (2005) 344. J. C. Chen [*et al.*]{}, Phys. Rev. D [**62**]{} (2000) 034003. J. Z. Bai [*et al.*]{}, Phys. Lett. B [**510**]{} (2001) 75. M.Ablikim [*et al.*]{}, Phys. Rev. D [**71**]{} (2005) 032003. M.Ablikim [*et al.*]{}, Phys. Rev. D [**74**]{} (2006) 012004. J. Z. Bai [*et al.*]{}, Phys. Rev. D [**70**]{} (2004) 012005. Particle Data Group, C. Amsler [*et al.*]{}, Phys. Lett. B [**667**]{} (2008) issue 1-5. M.Ablikim [*et al.*]{}, Phys. Rev. Lett. [**95**]{} (2005) 262001. S. S. Fang [*et al.*]{}, High Energy Phys. Nucl. Phys. [**27**]{} (2003) 277 (in Chinese). J. Z. Bai [*et al.*]{}, Phys. Lett. B [**591**]{} (2004) 42.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study a nonequilibrium model with up-down symmetry and a noise parameter $q$ known as majority-vote model of M.J. Oliveira $1992$ on opinion-dependent network or Stauffer-Hohnisch-Pittnauer networks. By Monte Carlo simulations and finite-size scaling relations the critical exponents $\beta/\nu$, $\gamma/\nu$, and $1/\nu$ and points $q_{c}$ and $U^*$ are obtained. After extensive simulations, we obtain $\beta/\nu=0.230(3)$, $\gamma/\nu=0.535(2)$, and $1/\nu=0.475(8)$. The calculated values of the critical noise parameter and Binder cumulant are $q_{c}=0.166(3)$ and $U^*=0.288(3)$. Within the error bars, the exponents obey the relation $2\beta/\nu+\gamma/\nu=1$ and the results presented here demonstrate that the majority-vote model belongs to a different universality class than the equilibrium Ising model on Stauffer-Hohnisch-Pittnauer networks, but to the same class as majority-vote models on some other networks.'
address: |
Dietrich Stauffer Computational Physics Lab, Departamento de Física,Universidade Federal do Piauí, Teresina, Piauí, 64049-550, Brasil[^1]\
[email protected]
author:
- 'F. W. S. Lima'
title: 'Majority-vote model on Opinion-dependent Network'
---
[^2]
Introduction
============
The equilibrium Ising model [@a3; @onsager] has become an excellent tool to study models of social application [@latané]. Many of these works are well described in a thorough review [@SPSD], a more recent summary by Stauffer [@stauffer1] and the following papers in these special issues on sociophysics in this journal. The majority-vote model (MVM) of Oliveira [@mario] is a nonequilibrium model of social interaction: individuals of a certain population make their decisions based on the opinion of the majority of their neighbors. This model has been studied for several years by various researchers in order to model social and economic systems [@zaklan; @zaklan1; @limanew; @lima1; @lima2] in regular structures [@MVM-regular; @lima-malarz; @santos; @lima3] and various other complex networks [@MVM-SW0; @MVM-SW1; @MVM-ERU; @MVM-VD; @MVM-ABD; @MVM-ABU; @MVM-ERD; @MVM-APN].
There are applications to real elections in which similar models of opinion dynamics have been explored in the literature, such as [@TVP].
In the present work, we study the critical properties of MVM on Stauffer-Hohnisch-Pittnauer (SHP) networks. Hohnisch bonds of SHP networks [@LHS; @TH; @DS] are links connecting nodes with different values (spins, opinions, etc.) on them; they are at each time step with a low probability $0.0001$ replaced by a link to another randomly selected node. Links connecting agreeing nodes are not replaced. In the present work, we start with each node having links to four randomly selected neighbors. Thus our SHP networks are similar to Small-World (Watts-Strogatz) networks but start from a random network instead of a square lattice and use opinion-dependent (instead of random) rewiring. All links are directed[@LHS; @TH]. The critical exponents and noise parameter were obtained using Monte Carlo simulation (MC) and with a finite size scaling analysis. The effective dimension of the SHP network is also determined for MVM. Finally, the critical exponents calculated for SHP networks are compared with the results obtained for [*undirected*]{} and [*directed*]{} Barabási-Albert networks (UBA and DBA) [@MVM-ABD; @MVM-ABU] and Erdös-Rènyi random graphs (UER and DER) [@MVM-ERU; @MVM-ERD].
Model and simulation
====================
Our network is SHP type composed of $N$ sites and $k=4$ neighbors. On the MVM model, the system dynamics is as follows. Initially, we assign a spin variable $\sigma$ with values $\pm 1$ at each node of the network. At each step we try to spin flip a node. The flip is accepted with probability $$w_i=\frac{1}{2}\left[ 1-(1-2q)\sigma_{i}\cdot\text{S}
\left(\sum_{j=1}^k\sigma_j\right)\right],
\label{eq_1}$$ where $S(x)$ is the sign $\pm 1$ of $x$ if $x\neq0$, $S(x)=0$ if $x=0$. To calculate $w_i$ our sum runs over the $k$ nearest neighbors of spin $i$. Eq. (\[eq\_1\]) means that with probability $(1-q)$ the spin will adopt the same state as the majority of its neighbors.
Here, the control parameter $0\le q\le 1$ plays a role similar to the temperature in equilibrium systems: the smaller $q$, the greater the probability of parallel aligning with the local majority.
To study the critical behavior of the model we define the variable $m\equiv\sum_{i=1}^{N}\sigma_{i}/N$. In particular, we are interested in the magnetization $M$, susceptibility $\chi$ and the reduced fourth-order cumulant $U$
\[eq-def\] $$M_{N}(q)\equiv \langle|m|\rangle,$$ $$\chi_{N}(q)\equiv N\left(\langle m^2\rangle-\langle m \rangle^2\right),$$ $$U_{N}(q)\equiv 1-\dfrac{\langle m^{4}\rangle}{3\langle m^2 \rangle^2},$$
where $\langle\cdots\rangle$ stands for a thermodynamics average.
The results are averaged over the $N_{\text{run}}$ independent simulations. These quantities are functions of the noise parameter $q$ and obey the finite-size scaling relations
\[eq-scal\] $$\label{eq-scal-M}
M_{N}(q)=N^{-\beta/\nu}f_m(x),$$ $$\label{eq-scal-chi}
\chi_{N}(q)=N^{\gamma/\nu}f_\chi(x),$$ $$\label{eq-scal-dUdq}
\frac{dU_{N}(q)}{dq}=N^{1/\nu}f_U(x),$$ where $\nu$, $\beta$, and $\gamma$ are the usual critical exponents, $f_{m,\chi,U}(x)$ are the finite size scaling functions with $$\label{eq-scal-x}
x=(q-q_c)N^{1/\nu}$$
being the scaling variable. Therefore, from the size dependence of $M$ and $\chi$ we obtained the exponents $\beta/\nu$ and $\gamma/\nu$, respectively. The maximum value of susceptibility also scales as $N^{\gamma/\nu}$. Moreover, the value of $q^*$ for which $\chi$ has a maximum is expected to scale with the system size $N$ as $$\label{eq-q-max}
q^*=q_c+bN^{-1/\nu} \text{ with } b\approx 1.$$
Therefore, the relations and may be used to get the exponent $1/\nu$. We also have applied the calculated exponents to the hyperscaling hypothesis $$\label{eq-q-def}
2\beta/\nu + \gamma/\nu=D_{eff}$$ in order to get the effective dimensionality, $D_{eff}$, for connectivity $k$.
We performed Monte Carlo simulation on the SHP networks with various systems sizes $N$ (250, 500, 1000, 2000, 4000, 8000 and 16000 sites). It takes $2\times 10^5$ Monte Carlo steps (MCS) to make the system reach the steady state, and then the time averages are estimated over the next $2\times 10^5$ MCS. One MCS is accomplished after all the $N$ spins are investigated whether they flip or not.
The results are averaged over $N_{\text{run}}$ $(50\le N_{\text{run}} \le 100)$ independent simulation runs for each network and for given set of parameters $(q,N)$. Here, were used $50$ independent networks for each system size $N$ cited earlier.
Results and Discussion
======================
In Figs. \[f1\], \[f2\], and \[f3\] we show the dependence of the magnetization $M$, susceptibility $\chi$, and Binder cumulant $U$ on the noise parameter $q$, obtained from simulations on $SHP$ networks with $L$ ranging from $N=250$ to $16000$ sites. The shape of $M(q)$, $\chi(q)$, and $U(q)$ curve, for a given value of $N$, suggests the presence of a second-order phase transition in the system. The phase transition occurs at the critical value $q_c$ of the noise parameter $q$. This parameter $q_c$ is estimated as the point where the $U_N(q)$ curves for different system sizes $N$ intercept each other [@binder]. Then, we obtain $q_c=0.166(3)$ and $U^*=0.288(3)$ for $SHP$ networks.
In Fig. \[f4\] we plot the dependence of the magnetization $M^*=M(q_c)$ vs. the system size $N$. The slope of curve corresponds to the exponent ratio $\beta/\nu$ according to Eq. (3a). The obtained exponent is $\beta/\nu= 0.230(3)$ for our $SHP$ network.
The exponent ratio $\gamma/\nu$ at $q_{c}$ is obtained from the slope of the straight line with $\gamma/\nu= 0.535(2)$ , as presented in Fig. \[f5\]. The exponents ratio $\gamma/\nu$ at $q_{\chi_{max}}(N)$ is $\gamma/\nu=0.523(5)$ for $SHP$ networks as presented in Fig. \[f6\].
To obtain the critical exponent $1/\nu$, we used the scaling relation (4). The calculated value of the exponent $1/\nu$ are $1/\nu=0.475(8)$ for SHP networks (see Fig. \[f7\]). We plot $MN^{\beta/\nu}$ versus $(q-q_{c})N^{1/\nu}$ in Fig. \[f8\] using the critical exponents $1/\nu=0.475(8)$ and $\beta/\nu=0.230(3)$ for system size $N=1000$,$2000$, $4000$,$8000$, and $16000$ for $SHP$ network. The excellent collapse of the curves for five different system sizes corroborates the estimate for $q_c$ and the critical exponents $\beta/\nu$ and $1/\nu$.
In Fig. \[f9\] we plot $\chi N^{-\gamma/\nu}$ versus $(q-q_{c})N^{1/\nu}$ using the critical exponents $\gamma/\nu=0.523(5)$ and $1/\nu=0.475(8)$ for system size $N=1000$, 2000, 4000, 8000, and 16000 for $SHP$ network. Again, the excellent collapse of the curves for five different system size corroborates the extimation for $q_c$ and the critical exponents $\gamma/\nu$ and $1/\nu$. The results of simulations are collected in Tab. \[table1\].
Conclusion
==========
The determination of the universality class of the MVM model on differents non-regular structure as Small-Worlds, scale-free networks, random graphs, and others has been studied by many researchers in recent years [@MVM-SW0; @MVM-SW1; @MVM-ERU; @MVM-VD; @MVM-ABD; @MVM-ABU; @MVM-ERD; @MVM-APN]. Finally, here, we remark that our MC results obtained on $SHP$ network for MVM model show that critical exponent ratios $\beta/\nu$ and $\gamma/\nu$ are similar to those for $UBA$ networks, $UER$ and $DER$ random graphs for values of conectivity $k=4$, but different from the results of $DBA$ networks and MVM model for regular lattice [@mario] and equilibrium 2D Ising model [@onsager]. Here, we show also that the critical exponent $1/\nu$ is different from random Erdös-Rényi graphs. Therefore, unfortunately, because of the critical exponent ratio $1/\nu$, we cannot assert that MVM models on different structures as scale-free ($UBA$) network and random graphs ($UER$ and $DER$) belong to the same universality class of the MVM model on $SHP$ network; only the critical exponent ratios $\beta/\nu$ and $\gamma/\nu$ are similar to $UBA$ networks and to $UER$ and $DER$ random graphs. Here, we also showed that the effective dimension $D_{eff}$ is close to $1$ for all networks and graphs studied in this work. The agreement in $D_{eff}$ and the two exponent ratios but not in $1/\nu$ (weak universality) remains to be explained.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks D. Stauffer for many suggestion and fruitful discussions during the development this work and also for reading this paper. We also acknowledge the Brazilian agency CNPQ for its financial support. This work also was supported the system SGI Altix 1350 in the computational park CENAPAD.UNICAMP-USP, SP-BRAZIL.
[0]{}
L. Onsager, Phys. Rev. 65, 117 (1944); B. Kaufmann, Phys. Rev. [**76**]{}, 1232 (1949). R. J. Baxter, [*Exactly solved models in statistical mechanics*]{}, London, Academic Press (1982). B. Latané, [*Am. Psychologist*]{} [**36**]{}, 343 (1981). C. Castellano, S. Fortunato and V. Loreto, Rev. Mod. Physics [**81**]{}, 591 (2009) = http://rmp.aps.org/abstract/RMP/v81/i2/p591-1. D. Stauffer, J. Stat. Phys. [**151**]{}, 9 (2013). M. J. Oliveira, [*J. Stat. Phys.*]{} [**66**]{} 273 (1992). G. Zaklan, F. Westerhoff and D. Stauffer, [*arXiv:0801.2980.*]{} (2008), = [*J. Econ. Interact. Coordination*]{} [**4**]{}, 1 (2008). G. Zaklan, F. W. S. Lima and F. Westerhoff, [*Physica A*]{} [**387**]{}, 5857 (2008). F. W. S. Lima, [*Journal of Physics: Conference Series*]{} [**246**]{}, 012033 (2010). F. W .S. Lima, [*Theoretical Economics Letters*]{} [**02**]{}, 87 (2012). F. W. S. Lima, [*International Journal of Modern Physics C*]{} [**23**]{}, 1250079 (2012). M. A. Santos, S. Teixeira, [*J. Stat. Phys.*]{} [**78**]{}, 963 (1995). F. W. S. Lima and K. Malarz, [*Int. J. Mod. Phys. C*]{} [**17**]{}, 1273 (2006). J. C Santos, F. W. S. Lima, and K. Malarz, [*Physica*]{} A, [**390**]{}, 359 (2011). F. W. S. Lima, [*Physica. A*]{}, [**391**]{}, 1753 (2012). P. R. Campos, V. M. Oliveira, and F. G. B. Moreira, [*Phys. Rev. E*]{} [**67**]{}, 026104 (2003). E. M. S. Luz, F. W. S. Lima, [*Int. J. Mod. Phys. C*]{} [**18**]{}, 1251 (2007). L. F. C. Pereira and F. G. B. Moreira, [*Phys. Rev. E*]{} [**71**]{}, 016123 (2005). F. W. S. Lima, U. L. Fulco, and R. N. C. Filho, [*Phys. Rev. E*]{} [**71**]{}, 036105 (2005). F. W. S. Lima, [*Int. J. Mod. Phys. C*]{} [**17**]{}, 1257 (2006). F. W. S. Lima, [*Communications in Computational Physics*]{} [**2**]{}, 358 (2007). F. W. S. Lima, A. O. Sousa, and M. A. Sumour, [*Physica A*]{} [**387**]{}, 3503 (2008) F. W. S. Lima, A. A. Moreira, A. D. Araújo, [*Phys. Rev. E*]{} [**86**]{}, 056109 (2012). N. A. M. Araújo, J. S. Andrade Jr., H. J. Herrmann, “Tactical Voting in Plurality Elections”, PLoS ONE [**5**]{}(9) (2010): e12446. doi:10.1371/journal.pone.0012446. F. W. S. Lima, T. Hadzibeganovic and D. Stauffer, [*Physica A*]{} [**388**]{}, 4999 (2009). T. Qiu, T. Hadzibeganovic, G. Chen, L.-X. Zhong and X.-R. Wu, [*Computer Phys. Comm.*]{} [**181**]{}, 2057 (2010). D. Stauffer, M. Hohnisch and S. Pittnauer, [*Physica A*]{} [**370**]{}, 734 (2006). See also P. Holme and M. E. J. Newman, [*Phys. Rev.*]{} [**74**]{}, 056108 (2006); A. E. Allahveryan and K. G. Petrosyan, [*Europhysics Letters*]{} [**75**]{}, 908 (2006). In preparation, “Majority-vote model with 2 and 3 states on SHP networks”. K. Binder and D. W. Heermann, [*Monte Carlo Simulation in Statistical Phyics*]{}, (Springer Verlag, 1988).
[^1]: *Piauí, the Universidade Federal do Piauí, Brasil.*
[^2]: Dietrich Stauffer Computational Physics Lab, Departamento de Física,Universidade Federal do Piauí, Teresina, Piauí, 64049-550, Brasil
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $\g$ be a finite-dimensional simple Lie algebra of rank $\rg$ over an algebraically closed field $\k$ of characteristic zero, and let $(e,h,f)$ be an $\mathfrak{sl}_2$-triple of $\g$. Denote by $\g^{e}$ the centralizer of $e$ in $\g$ and by $\ai g{e}{}$ the algebra of symmetric invariants of $\g^{e}$. We say that $e$ is good if the nullvariety of some $\rg$ homogeneous elements of $\ai g{e}{}$ in $({\goth g}^{e})^{*}$ has codimension $\rg$. If $e$ is good then $\ai g{e}{}$ is a polynomial algebra. In this paper, we prove that the converse of the main result of [@CM1] is true. Namely, we prove that $e$ is good if and only if for some homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the initial homogeneous components of their restrictions to $e+\g^{f}$ are algebraically independent over $\k$.'
address:
- |
Jean-Yves Charbonnel, Université Paris Diderot - CNRS\
Institut de Mathématiques de Jussieu - Paris Rive Gauche\
UMR 7586\
Groupes, représentations et géométrie\
Bâtiment Sophie Germain\
Case 7012\
75205 Paris Cedex 13, France
- |
Anne Moreau, Laboratoire de Mathématiques et Applications\
Téléport 2 - BP 30179\
Boulevard Marie et Pierre Curie\
86962 Futuroscope Chasseneuil Cedex, France
author:
- 'Jean-Yves Charbonnel'
- Anne Moreau
title: The symmetric invariants of centralizers and Slodowy grading II
---
Introduction {#i}
============
{#i1}
Let $\g$ be a finite-dimensional simple Lie algebra of rank $\rg$ over an algebraically closed field $\k$ of characteristic zero, let $\dv ..$ be the Killing form of $\g$ and let $G$ be the adjoint group of $\g$. If ${\goth a}$ is a subalgebra of $\g$, we denote by S$({\goth a})$ the symmetric algebra of ${\goth a}$. For $x\in\g$, we denote by $\g^{x}$ the centralizer of $x$ in $\g$ and by $G^{x}$ the stabilizer of $x$ in $G$. Then ${\rm Lie}(G^{x})={\rm Lie}(G^{x}_{0})=\g^x$ where $G^{x}_{0}$ is the identity component of $G^{x}$. Moreover, $\es S{\g^{x}}$ is a $\g^x$-module and $\ai gx{}=\es S{\g^{x}}^{G^{x}_{0}}$.
In [@CM1], we continued the works of [@PPY] and we studied the question on whether the algebra $\es S{\g^{x}}^{\g^{x}}$ is polynomial in $\rg$ variables; see [@Y3; @CM; @JS; @Y4] for other references related to the topic.
{#section}
Let us first summarize the main results of [@CM1].
\[di\] An element $x\in{\goth g}$ is called a [*good element of ${\goth g}$*]{} if for some homogeneous sequence $(\poi p1{,\ldots,}{\rg}{}{}{})$ in $\ai g{x}{}$, the nullvariety of $\poi p1{,\ldots,}{\rg}{}{}{}$ in $({\goth g}^{x})^{*}$ has codimension $\rg$ in $({\goth g}^{x})^{*}$.
Thus an element $x\in{\goth g}$ is good if the nullcone of $\es S{\g^{x}}$, that is, the nullvariety in $({\goth g}^{x})^{*}$ of the augmentation ideal $\ai g{x}{}_+$ of $\ai g{x}{}$, is a complete intersection in $({\goth g}^{x})^{*}$ since the transcendence degree over $\k$ of the fraction field of $\ai g{x}{}$ is $\ell$ by the main result of [@CM].
For example, regular nilpotent elements are good; see the introduction of [@CM1] for more details and other examples.
\[ti3\] Let $x$ be a good element of ${\goth g}$. Then $\ai gx{}$ is a polynomial algebra and $\es S{{\goth g}^{x}}$ is a free extension of $\ai gx{}$.
An element $x$ is good if and only if so is its nilpotent component in the Jordan decomposition [@CM1 Proposition 3.5]. As a consequence, we can restrict the study to the case of nilpotent elements.
Let $e$ be a nilpotent element of $\g$. By the Jacobson-Morosov Theorem, $e$ is embedded into an ${\goth {sl}}_2$-triple $(e,h,f)$ of $\g$. Identify $\g$ with $\g^*$, and ${\goth g}^{f}$ with $({\g}^{e})^*$, through the Killing isomorphism $\g \to \g^*, \, x \mapsto \dv{x}{.}$. Thus we have the following algebra isomorphisms: $\e Sg \simeq\k[\g^*]\simeq\k[\g]$ and $\es S{{\goth g}^{e}}\simeq\k[(\g^{e})^*]\simeq\k[\g^{f}]$. Denote by ${\cal S}_e:=e+\g^{f}$ the [*Slodowy slice associated with $e$*]{}, and let $T_e \colon \g \to \g, \, x \mapsto e+x$ be the translation map. It induces an isomorphism of affine varieties $\g^{f} \simeq
{\cal S}_e$, and the comorphism $T_e^{*}$ induces an isomorphism between the coordinate algebras $\k[{\cal S}_e]$ and $\k[\g^{f}]$.
Let $p$ be a homogeneous element of $\e Sg \simeq\k[\g]$. Then its restriction to ${\cal S}_e$ is an element of $\k[{\cal S}_e] \simeq \k[\g^{f}] \simeq \es S{{\goth g}^{e}}$ through the above isomorphisms. For $p$ in $\e Sg$, we denote by $\kappa (p)$ its restriction to ${\cal S}_e$ so that $\kappa(p) \in {\rm S}(\g^{e})$. Denote by $\ie{p}$ the initial homogeneous component of $\kappa (p)$. According to [@PPY Proposition 0.1], if $p$ is in $\ai g{}{}$, then $\ie{p}$ is in $\ai ge{}{}$.
\[ti4\] Suppose that for some homogeneous generators $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the polynomial functions $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. Then $e$ is a good element of ${\goth g}$. In particular, $\ai ge{}$ is a polynomial algebra and $\es S{{\goth g}^{e}}$ is a free extension of $\ai ge{}$. Moreover, $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ is a regular sequence in $\es S{{\goth g}^{e}}$.
In other words, Theorem \[ti4\] provides a sufficient condition for that $\ai ge{}$ is polynomial. By [@PPY], one knows that for homogeneous elements $q_1,\ldots,q_\ell$ of $\ai g{}{}$, the polynomial functions $\ie{q_1},\ldots,\ie{q_\ell}$ are algebraically independent if and only if $$\begin{aligned}
\label{eq:alg_indep}
\sum_{i=1}^{\ell} \deg \ie{q_i}=\frac{\dim\g^{e}+\ell}{2}.\end{aligned}$$ So we have a practical criterion to verify the sufficient condition of Theorem \[ti4\]. However, even if the condition of Theorem \[ti4\] holds, that is, if (\[eq:alg\_indep\]) holds, $\ai ge{}$ is not necessarily generated by the polynomial functions $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$. As a matter of fact, there are nilpotent elements $e$ satisfying this condition and for which ${\rm S}(\g^{e})^{\g^{e}}$ is not generated by some $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$, for any choice of homogeneous generators $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$ (cf. [@CM1 Remark 2.25]).
Theorem \[ti4\] can be applied to a great number of nilpotent orbits in the simple classical Lie algebras, and for some nilpotent orbits in the exceptional Lie algebras, see [@CM1 Sections 5 and 6]. We also provided in [@CM1 Example 7.8] an example of a nilpotent element $e$ for which ${\rm S}(\g^{e})^{\g^{e}}$ is not polynomial, with $\g$ of type D$_7$.
{#section-1}
In this note, we prove that the converse of Theorem \[ti4\] also holds. Namely, our main result is the following theorem.
\[ti5\] The nilpotent element $e$ of ${\goth g}$ is good if and only if for some homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the elements $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$.
Theorem \[ti5\] was conjectured in [@CM1 Conjecture 7.11]. Notice that it may happen that for some $r_1,\ldots,r_\ell$ in $\ai g{}{}$, the elements $\poi {\ie r}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$, and that however $e$ is not good. This is the case for instance for the nilpotent elements in $\mathfrak{so}(\k^{12})$ associated with the partition $(5,3,2,2)$, cf. [@CM1 Example 7.6]. In fact, according to [@PPY Corollary 2.3], for any nilpotent element $e$ of $\g$, there exist $\poi r1{,\ldots,}{\rg}{}{}{}$ in $\ai g{}{}$ such that $\poi {\ie r}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. So the assumption that $\poi q1{,\ldots,}{\rg}{}{}{}$ generate $\ai g{}{}$ is crucial.
{#ss:plan}
We introduce in this subsection the main notations of the paper and we outline our strategy to prove Theorem \[ti5\].
First of all, recall that ${\goth g}^{f}$ identifies with the dual of ${\goth g}^{e}$ through the Killing isomorphism so that $\es S{{\goth g}^{e}}$ is the algebra $\k[\g^{f}]$ of polynomial functions on ${\goth g}^{f}$, and that $\k[\g^{f}]$ identifies with the coordinate algebra of the Slodowy slice ${\cal S}_e=e+{\goth g}^{f}$.
Let $\poi x1{,\ldots,}{r}{}{}{}$ be a basis of ${\goth g}^{e}$ such that for $i=1,\ldots,r$, $[h,x_{i}] = n_{i}x_{i}$ with $n_{i}$ a nonnegative integer. For ${\bf j}=(\poi j1{,\ldots,}{r}{}{}{})$ in ${\Bbb N}^{r}$, set: $$\vert {\bf j} \vert := \poi j1{+\cdots +}{r}{}{}{}, \quad
\vert {\bf j} \vert _{e} := j_{1}(n_{1}+2) + \cdots + j_{r}(n_{r}+2), \quad
x^{{\bf j}} := \poie x1{\cdots }{r}{}{}{}{j_{1}}{j_{r}}.$$ There are two gradings on $\es S{{\goth g}^{e}}$: the standard one and the Slodowy grading. For all ${\bf j}$ in ${\Bbb N}^{r}$, $x^{{\bf j}}$ has standard degree $\vert {\bf j} \vert$ and, by definition, it has Slodowy degree $\vert {\bf j} \vert_{e}$. Denoting by $t\mapsto \rho (t)$ the one-parameter subgroup of $G$ generated by $\ad h$, the Slodowy slice $e+{\goth g}^{f}$ is invariant under the one-parameter subgroup $t \mapsto t^{-2}\rho (t)$ of $G$. Hence the one-parameter subgroup $t\mapsto t^{-2}\rho (t)$ induces an action on $\k[{\cal S}_e]$. Let $j \in\{1,\ldots,r\}$, $y$ in ${\goth g}^{f}$ and $t$ in $\k^{*}$. Viewing the element $x_j$ of $\g^{e} \subset \es S{{\goth g}^{e}}$ as an element $\k[{\cal S}_e]$, we have: $$x_{j}(t^{-2}\rho (t)(e+y)) = x_{j}(e+t^{-2}\rho (t)(y)) =
t^{-2}\rho (t^{-1})(x_{j})(e+y) = t^{-2-n_{j}}x_{j}(e+y) ,$$ whence for all ${\bf j}$ in ${\Bbb N}^{r}$ and for all $y$ in ${\goth g}^{f}$, $$x^{{\bf j}}(t^{-2}\rho(t)(e+y)) = t^{-\vert {\bf j} \vert_{e}} x^{{\bf j}}(e+y) .$$ This means that $x^{{\bf j}}$, as a regular function on ${\cal S}_e$, is homogeneous of degree $\vert {\bf j} \vert_{e}$ for the Slodowy grading.
Let $t$ be an indeterminate and let $R$ be the polynomial algebra $\k[t]$. The polynomial algebra $$\es S{{\goth g}^{e}}[t] :=\tk \k{\k[t]}\es S{{\goth g}^{e}}$$ identifies with the algebra of polynomial functions on ${\goth g}^{f}\times \k$. The grading of $\es S{{\goth g}^{e}}$ induces a grading of $\es S{{\goth g}^{e}}[t]$ such that $t$ has degree $0$. Denote by $\varepsilon $ the evaluation map at $t=0$ so that $\varepsilon $ is a graded morphism from $\es S{{\goth g}^{e}}[t]$ onto $\es S{{\goth g}^{e}}$. Let $\tau $ be the embedding of $\es S{{\goth g}^{e}}$ into $\es S{{\goth g}^{e}}[t]$ such that $\tau (x_{i}):=tx_{i}$ for $i=1,\ldots,r$.
Recall that for $p$ in $\e Sg$, $\kappa (p)$ denotes the restriction to ${\cal S}_e$ of $p$ so that $\kappa(p) \in {\rm S}(\g^{e})$. Denote by $A$ the intersection of $\es S{{\goth g}^{e}}[t]$ with the sub-$\k[t,t^{-1}]$-module of $$\es S{{\goth g}^{e}}[t,t^{-1}]:= \tk \k{\k[t,t^{-1}]}\es S{{\goth g}^{e}}$$ generated by $\tau \rond \kappa(\e Sg^\g)$, and let $A_+$ be its augmentation ideal. Let ${\cal V}$ be the nullvariety of $A_{+}$ in ${\goth g}^{f}\times \k$ and ${\cal V}_{*}$ the union of the irreducible components of ${\cal V}$ which are not contained in ${\goth g}^{f}\times \{0\}$. Let ${\cal N}$ be the nullvariety of $\varepsilon (A)_{+}$ in ${\goth g}^{f}$, with $\varepsilon (A)_+$ the augmentation ideal of $\varepsilon (A)$. Then ${\cal V}$ is the union of ${\cal V}_{*}$ and ${\cal N}\times \{0\}$.
The properties of the varieties ${\cal V}$ and ${\cal V}_{*}$ allow us to prove the following result.
\[ti6\] Suppose that ${\cal N}$ has dimension $r-\rg$. Then for some homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the elements $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$.
The key point is to show that, under the hypothesis of Theorem \[ti6\], $\varepsilon(A)$ is the subalgebra of $\es S{{\goth g}^{e}}$ generated by the family $\ie p,\; p \in \ai g{}{}$, and hence that ${\cal N}$ coincides with the nullvariety in $\g^{f}$ of $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$. So, if ${\cal N}$ has dimension $r -\ell$, then the elements $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ must be algebraically independent over $\k$.
The remainder of the paper is organized as follows. In Section \[gf\], we state useful results on commutative algebra of independent interest. Some of these results are probably well-known. Since we have not found appropriate references, proofs are provided. Moreover, we formulate them as they are used in the paper. We study in Section \[sg\] properties of the varieties ${\cal V}$ and ${\cal V}_{*}$. The proof of Theorem \[ti6\] is achieved in Section \[sg\]. Theorem \[ti5\] is a consequence of Theorem \[ti6\], and it is proven in Section \[proof\].
Acknowledgments {#acknowledgments .unnumbered}
---------------
The second author is partially supported by the ANR Project GeoLie Grant number ANR-15-CE40-0012. We thank the referee for valuable comments and suggestions.
Some results on commutative algebra {#gf}
===================================
In this section $t$ is an indeterminate and the base ring $R$ is $\k$, $\k[t]$ or $\k[[t]]$. For $M$ a graded space over ${\Bbb N}$ and for $j$ in ${\Bbb N}$, denote by $M^{[j]}$ the space of degree $j$ and by $M_{+}$ the sum of $M^{[j]},j>0$. Let $A$ be a finitely generated graded $R$-algebra over ${\Bbb N}$ such that $A^{[0]}=R$ and such that $A^{[j]}$ is a free $R$-module of finite rank for any $j \in \N$. Moreover, $A$ is an integral domain. Denote by $\dim A$ the Krull dimension of $A$ and set[^1]: $$\rg := \left \{ \begin{array}{lcl} \dim A & \mbox{ if } & R = \k \\
\dim A -1 & \mbox{ if } & R = \k[t] \text{ or }\k[t]]. \end{array} \right.$$ As a rule, for $B$ an integral domain, we denote by $K(B)$ its fraction field.
The one-dimensional multiplicative group of $\k$ is denoted by $\gm$.
{#gf1}
Let $B$ be a graded subalgebra of $A$.
\[lgf1\]
1. Let $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$ be pairwise different graded prime ideals contained in $A_{+}$. If they are the minimal prime ideals containing their intersection, then for some homogeneous element $p$ of $A_{+}$, the element $p$ is not in the union of $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$.
2. For some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, $A_{+}$ is the radical of the ideal generated by $\poi p1{,\ldots,}{\rg}{}{}{}$.
3. Suppose that $A_{+}$ is the radical of $AB_{+}$. Then for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $B_{+}$, $A_{+}$ is the radical of the ideal generated by $\poi p1{,\ldots,}{\rg}{}{}{}$.
\(i) Prove by induction on $j$ that for some homogeneous element $p_{j}$ of $A_{+}$, $p_{j}$ is not in the union of $\poi {{\goth p}}1{,\ldots,}{j}{}{}{}$. Since ${\goth p}_{1}$ is a graded ideal strictly contained in $A_{+}$, it is true for $j=1$. Suppose that it is true for $j-1$. If $p_{j-1}$ is not in ${\goth p}_{j}$, there is nothing to prove. Suppose that $p_{j-1}$ is in ${\goth p}_{j}$. According to the hypothesis, ${\goth p}_{j}$ is stricly contained in $A_{+}$ and it does not contain the intersection of $\poi {{\goth p}}1{,\ldots,}{j-1}{}{}{}$. So, since $\poi {{\goth p}}1{,\ldots,}{j}{}{}{}$ are graded ideals, for some homogeneous sequence $r,q$ in $A_{+}$, $$r \in \bigcap _{k=1}^{j-1} {\goth p}_{k}\setminus {\goth p}_{j},
\quad \text{and} \quad q \in A_{+}\setminus {\goth p}_{j} .$$ Denoting by $m$ and $n$ the respective degrees of $p_{j-1}$ and $rq$, $p_{j-1}^{n}+(rq)^{m}$ is homogeneous of degree $mn$ and it is not in $\poi {{\goth p}}1{,\ldots,}{j}{}{}{}$ since these ideals are prime.
\(ii) Prove by induction on $i$ that for some homogeneous sequence $\poi p1{,\ldots,}{i}{}{}{}$ in $A_{+}$, the minimal prime ideals of $A$ containing $\poi p1{,\ldots,}{i}{}{}{}$ have height $i$. Let $p_{1}$ be in $A_{+}\setminus \{0\}$. By [@Mat Ch. 5, Theorem 13.5], all minimal prime ideal containing $p_{1}$ has height $1$. Suppose that it is true for $i-1$. Let $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$ be the minimal prime ideals containing $\poi p1{,\ldots,}{i-1}{}{}{}$. Since $A_{+}$ has height $\rg > i-1$, $A_{+}$ strictly contains $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$. By (i), there exists a homogeneous element $p_{i}$ in $A_{+}$ not in the union of $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$. Then, by [@Mat Ch. 5, Theorem 13.5], the minimal prime ideals containing $\poi p1{,\ldots,}{i}{}{}{}$ have height $i$. For $i=\rg$, the minimal prime ideals containing $\poi p1{,\ldots,}{\rg}{}{}{}$ have height $\rg$. Hence they are equal to $A_{+}$ since $A_{+}$ is a prime ideal of height $\rg$ containing $\poi p1{,\ldots,}{\rg}{}{}{}$, whence the assertion.
\(iii) The ideal $AB_{+}$ is generated by a homogeneous sequence $\poi a1{,\ldots,}{m}{}{}{}$ in $B_{+}$. Denote by $B'$ the subalgebra of $A$ generated by $\poi a1{,\ldots,}{m}{}{}{}$. Then $B'$ is a finitely generated graded subalgbera of $A$ such that $A_{+}$ is the radical of $AB'_{+}$. If $R=\k$, denote by $d$ its dimension and if $t\in R$, denote by $d+1$ its dimension. By (ii), for some homogeneous sequence $\poi p1{,\ldots,}{d}{}{}{}$ in $B'_{+}$, $B'_{+}$ is the radical of the ideal generated by $\poi p1{,\ldots,}{d}{}{}{}$. Then $A_{+}$ is the radical of the ideal of $A$ generated by $\poi p1{,\ldots,}{d}{}{}{}$. Since $A_{+}$ has height $\rg$, $\rg \leq d$ by [@Mat Ch. 5, Theorem 3.5]. Since $B'$ is a subalgebra of $A$, its dimension is at most $\dim A$. Hence $d=\rg$.
\[pgf1\] Suppose that $A_{+}$ is the radical of $AB_{+}$. Then $B$ is finitely generated and $A$ is a finite extension of $B$.
Since $A$ is a noetherian ring, for some homogeneous sequence $\poi a1{,\ldots,}{m}{}{}{}$ in $B_{+}$, $AB_{+}$ is the ideal generated by this sequence. Denote by $C$ the $R$-subalgebra of $A$ generated by $\poi a1{,\ldots,}{m}{}{}{}$. Then $C$ is a graded subalgebra of $A$. Denote by $\pi $ the morphism $$\xymatrix{ {\mathrm {Specm}}(A) \ar[r]^{\pi } &
{\mathrm {Specm}}(C) }$$ whose comorphism is the canonical injection $C
\hookrightarrow A$. Let $\overline{A}$ and $\overline{C}$ be the respective integral closures of $A$ and $C$ in $K(A)$. Since $C$ is contained in $A$, $\overline{C}$ is contained in $\overline{A}$. Let $\alpha $ and $\beta $ be the morphisms $$\xymatrix{ {\mathrm {Specm}}(\overline{A}) \ar[r]^{\alpha } &
{\mathrm {Specm}}(A) } \quad \text{and} \quad
\xymatrix{ {\mathrm {Specm}}(\overline{C}) \ar[r]^{\beta } &
{\mathrm {Specm}}(C) }$$ whose comorphisms are the canonical injections $A\hookrightarrow \overline{A}$ and $C\hookrightarrow \overline{C}$ respectively. Then there is a commutative diagram $$\xymatrix{ {\mathrm {Specm}}(\overline{A}) \ar[rr]^{\overline{\pi }}
\ar[d] _{\alpha } & &
{\mathrm {Specm}}(\overline{C}) \ar[d]^{\beta } \\
{\mathrm {Specm}}(A) \ar[rr]^{\pi } & & {\mathrm {Specm}}(C) }$$ with $\overline{\pi }$ the morphism whose comorphism is the canonical injection $\overline{C}\rightarrow \overline{A}$.
The action of $\gm$ in $A$ extends to an action of $K(A)$, and $\overline{A}$ is invariant under this action. Denoting by $\overline{R}$ the integral closure of $R$ in $K(A)$, $\overline{R}$ is the set of fixed points under the action of $\gm$ in $\overline{A}$. Since $C$ is invariant under $\gm$ so is $\overline{C}$. For ${\goth m}$ a maximal ideal of $\overline{R}$, the ideal ${\goth m}+\overline{C}_{+}$ is the maximal ideal of $\overline{C}$ containing ${\goth m}$ and invariant under $\gm$. Then, for ${\goth p}$ a maximal ideal of $\overline{C}$, ${\goth p}\cap \overline{R}+\overline{C}_{+}$ is in the closure of the orbit of ${\goth p}$ under $\gm$. Moreover, $$\{{\goth m}+\overline{A}_{+}\} = \overline{\pi }^{-1}\{{\goth m}+\overline{C}_{+}\}$$ for all maximal ideal ${\goth m}$ of $\overline{R}$. Hence $\overline{\pi }$ is quasi finite. Moreover $\overline{\pi }$ is birational. Then, by Zariski’s main theorem [@Mu], $\overline{\pi }$ is an open immersion. The image of $\overline{\pi }$ contains fixed points for the $\gm$-action, and the closure of each $\gm$-orbit contains fixed points. As a result, $\overline{\pi }$ is surjective since it is $\gm$-equivariant. Hence $\overline{\pi }$ is an isomorphism and $\overline{A}=\overline{C}$. As a result, $\overline{A}$ is a finite extension of $C$ since $\beta $ is a finite morphism. As submodules of the finite module $\overline{A}$ over the noetherian ring $C$, $A$ and $B$ are finite $C$-modules. Hence $A$ is a finite extension of $B$. Denoting by $\poi {\omega }1{,\ldots,}{d}{}{}{}$ a generating family of the $C$-module $B$, $B$ is the subalgebra of $A$ generated by $\poi a1{,\ldots,}{m}{}{}{}$, $\poi {\omega }1{,\ldots,}{d}{}{}{}$.
Denote by $\k[t]_{*}$ the localization of $\k[t]$ at the prime ideal $t\k[t]$ and set: $$R_{*} := \left \{ \begin{array}{lcl} \k & \mbox{ if } & R = \k \\
\k[t]_{*} & \mbox{ if } & R = \k[t] \\
\k[[t]] & \mbox{ if } & R = \k[[t]] \end{array}
\right. \qquad
\widehat{R} := \left \{ \begin{array}{lcl} \k & \mbox{ if } & R = \k \\
\k[[t]] & \mbox{ if } & R = \k[t] \\
\k[[t]] & \mbox{ if } & R = \k[[t]]. \end{array}
\right.$$ For $M$ a $R$-module, set $\widehat{M} := \tk R{\widehat{R}}M$.
\[l2gf1\] Suppose $R=\k[t]$. Let $M$ be a torsion free $R$-module and let $N$ be a submodule of $M$. Then for $a$ in $\widehat{N}\cap M$, $ra$ is in $N$ for some $r$ in $R$ such that $r(0)\neq 0$.
Since $M$ is torsion free, the canonical map $M\rightarrow \widehat{M}$ is an embedding. Moreover, the canonical map $\widehat{N}\rightarrow \widehat{M}$ is an embedding since $\widehat{R}$ is flat over $R$. Let $a$ be in $\widehat{N}\cap M$ and let $\overline{a}$ be its image in $M/N$ by the quotient map. Denote by $J_{a}$ the annihilator of $\overline{a}$ in $R$, whence a commutative diagram $$\xymatrix{
0 \ar[r] & N \ar[r]^{\dd} & M \ar[rr]^{\dd} && M/N \ar[r] & 0 \\
0 \ar[r] & J_{a} \ar[r]^{\dd} & R \ar[rr]^{\dd} \ar[u]^{\delta } &&
R\overline{a} \ar[r] \ar[u]^{\delta } & 0 \\
& & 0 \ar[u] && 0 \ar[u] & }$$ with exact lines and columns. Since $\widehat{R}$ is a flat extension of $R$, tensoring this diagram by $R$ gives the following diagram with exact lines and columns: $$\xymatrix{
0 \ar[r] & \widehat{N} \ar[r]^{\dd} & \widehat{M}
\ar[rr]^{\dd} && \tk {R}{\widehat{R}}M/N \ar[r] & 0 \\
0 \ar[r] & \widehat{R}J_{a} \ar[r]^{\dd} & \widehat{R} \ar[rr]^{\dd} \ar[u]^{\delta }
&& \widehat{R}\overline{a} \ar[r] \ar[u]^{\delta } & 0 \\
& & 0 \ar[u] && 0 \ar[u] & }$$ For $b$ in $\widehat{R}$, $(\delta \rond \dd) b = (\dd \rond \delta) b = 0$ since $a$ is in $\widehat{N}$, whence $\dd b = 0$. As a result, $\widehat{R}J_{a}=\widehat{R}$. Hence $J_{a}$ contains an element $r$, invertible in $\widehat{R}$, that is $r(0)\neq 0$, whence the lemma.
Set $$A_{*} := \tk R{R_{*}}A \qquad
\text{and}\qquad \widehat{A} := \tk R{\widehat{R}}A.$$ Since $A^{[0]}=R$, the grading on $A$ extends to gradings on $A_{*}$ and $\widehat{A}$ such that $A_{*}^{[0]} = R_{*}$ and $\widehat{A}^{[0]}=\widehat{R}$. When $R=\k$ or $R=\k[[t]]$, $A_{*}=A$ and $\widehat{A}=A$.
For $\poi p1{,\ldots,}{\rg}{}{}{}$ a homogeneous sequence in $A_{+}$ set: $$\begin{aligned}
\underline{p} := \left \{ \begin{array}{lcl}
\poi p1{,\ldots,}{\rg}{}{}{} & \mbox{ if } &
R = \k \\ t,\poi p1{,\ldots,}{\rg}{}{}{} & \mbox{ if } & R=\k[[t]] ,
\end{array} \right. \end{aligned}$$ and denote by $J_{\underline{p}}$ the ideal of $A$ generated by the sequence $\underline{p}$.
\[l3gf1\] Suppose that $A$ is Cohen-Macaulay. Let $\poi p1{,\ldots,}{\rg}{}{}{}$ be a homogeneous sequence in $A_{+}$ such that $A_{+}$ is the radical of the ideal of $A$ generated by $\poi p1{,\ldots,}{\rg}{}{}{}$ and let $V$ be a graded complement in $A$ to the $\k$-subspace $J_{\underline{p}}$.
1. The space $V$ has finite dimension.
2. The space $A_{*}$ is equal to $VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$.
3. The algebra $A$ is a flat extension of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$.
4. For all homogeneous elements $\poi a1{,\ldots,}{n}{}{}{}$ in $A$, linearly independent over $\k$ modulo $J_{\underline{p}}$, $\poi a1{,\ldots,}{n}{}{}{}$ are linearly independent over $R[\poi p1{,\ldots,}{\rg}{}{}{}]$.
5. The linear map $$\tk {\k}VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow A_{*}, \qquad
v\tens a \longmapsto va$$ is an isomorphism.
According to Lemma \[lgf1\](ii), the sequence $p$ does exist.
\(i) Let $J_{p}$ be the ideal of $A$ generated by $\poi p1{,\ldots,}{\rg}{}{}{}$. Since $A_{+}$ is the radical of $J_{p}$, $A^{[d]}=J_{p}^{[d]}$ for $d$ sufficiently big. When $t\in R$, for all $d$, then $tA^{[d]}$ has finite codimension in $A^{[d]}$ since $A^{[d]}$ is a finite free $R$-module. Hence $J_{\underline{p}}$ has finite codimension in $A$ so that $V$ has finite dimension.
\(ii) Suppose that $t$ is in $R$. First of all, we prove by induction on $d$ the inclusion $$A^{[d]} \subset (VR[\poi p1{,\ldots,}{\rg}{}{}{}])^{[d]} + tA^{[d]} .$$ Since $A^{[0]}$ is the direct sum of $V^{[0]}$ and $J_{\underline{p}}^{[0]}$, $V^{[0]}$ is contained in $\k + tR$, whence the inclusion for $d=0$. Suppose that it is true for all $j$ smaller than $d$. Since $\poi p1{,\ldots,}{\rg}{}{}{}$ have positive degrees, by induction hypothesis, $$J_{\underline{p}}^{[d]} \subset (VR[\poi p1{,\ldots,}{\rg}{}{}{}])^{[d]} + tA^{[d]} ,$$ whence the inclusion for $d$. Then, by induction on $m$, $$A^{[d]} \subset (VR[\poi p1{,\ldots,}{\rg}{}{}{}])^{[d]} + t^{m}A^{[d]} .$$ As a result, since $A^{[d]}$ is a finite $R$-module, $$A^{[d]} \subset (V\widehat{R}[\poi p1{,\ldots,}{\rg}{}{}{}])^{[d]} ,$$ whence $\widehat{A} = V\widehat{R}[\poi p1{,\ldots,}{\rg}{}{}{}]$. This equality remains true when $R=\k$ by an analogous and simpler argument.
When $R=\k[t]$, according to Lemma \[l2gf1\], for $a$ in $A$, $ra$ is in $VR[\poi p1{,\ldots,}{\rg}{}{}{}]$ for some $r$ in $R$ such that $r(0)\neq 0$. As a result, $A_{*}=VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$.
\(iii) By Proposition \[pgf1\], $A$ is a finite extension of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$. In particular, $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ has dimension $\rg+\dim R$ so that $\poi p1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $R$. Hence $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ is a regular algebra, whence the assertion by [@Mat Ch. 8, Theorem 23.1].
\(iv) Prove the assertion by induction on $n$. Since $A$ is an integral domain, the assertion is true for $n=1$. Suppose the assertion true for $n-1$. Let $(\poi b1{,\ldots,}{n}{}{}{})$ be a homogeneous sequence in $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ such that $$b_{1}a_{1} + \cdots + b_{n}a_{n} = 0 .$$ Let $K$ and $I$ be the kernel and the image of the linear map $$R[\poi p1{,\ldots,}{\rg}{}{}{}]^{n} \longrightarrow R[\poi p1{,\ldots,}{\rg}{}{}{}],
\qquad (\poi c1{,\ldots,}{n}{}{}{}) \longmapsto c_{1}b_{1} + \cdots + c_{n}b_{n} ,$$ whence the short exact sequence of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ modules $$0 \longrightarrow K \longrightarrow R[\poi p1{,\ldots,}{\rg}{}{}{}]^{n}
\longrightarrow I \longrightarrow 0 .$$ The grading of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ induces a grading of $R[\poi p1{,\ldots,}{\rg}{}{}{}]^{n}$ and $K$ is a graded submodule of $R[\poi p1{,\ldots,}{\rg}{}{}{}]^{n}$ since $\poi b1{,\ldots,}{n}{}{}{}$ is a homogeneous sequence in $R[\poi p1{,\ldots,}{\rg}{}{}{}]$. Denote by $\poi y1{,\ldots,}{m}{}{}{}$ a generating homogeneous sequence of the $R[\poi p1{,\ldots,}{\rg}{}{}{}]$-module $K$. By (iii), the short sequence of $A$-modules $$0 \longrightarrow \tk {R[\poi p1{,\ldots,}{\rg}{}{}{}]}AK \longrightarrow
A^{n} \longrightarrow \tk {R[\poi p1{,\ldots,}{\rg}{}{}{}]}AI \longrightarrow 0$$ is exact. So, for some homogeneous sequence $\poi x1{,\ldots,}{m}{}{}{}$ in $A$, $$a_{i} = \sum_{j=1}^{m} x_{j} y_{j,i}$$ for $i=1,\ldots,n$. Since $a_{n}$ is not in $J_{\underline{p}}$, for some $j_{*}$, the element $y_{j_{*},i}$ is an invertible element of $R_{*}$, whence $$b_{n}y_{j_{*},n} = - \sum_{i=1}^{n-1} b_{i}y_{j_{*},i} \quad \text{and} \quad
\sum_{i=1}^{n-1} b_{i} (y_{j_{*},n}a_{i} - a_{n}y_{j_{*},i}) = 0 .$$ So, by induction hypothesis, $$\poi b1{ = \cdots = }{n-1}{}{}{} = 0$$ since the elements $$y_{j_{*},n}a_{1} - a_{n}y_{j_{*},1} ,\ldots,
y_{j_{*},n}a_{n-1} - a_{n}y_{j_{*},n-1}$$ are linearly independent over $\k$ modulo $J_{\underline{p}}$. Then $b_{n} = 0 $ since $y_{j_{*},n}$ is invertible.
\(v) Let $(\poi v1{,\ldots,}{n}{}{}{})$ be a homogeneous basis of $V$. Since the space of relations of linear dependence over $R[\poi p1{,\ldots,}{\rg}{}{}{}]$ of $\poi v1{,\ldots,}{n}{}{}{}$ is graded, it is equal to $\{0\}$ by (iv), whence the assertion by (ii).
\[cgf1\]
1. The algebra $A_{*}$ is Cohen-Macaulay if and only if for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, the algebra $A_{*}$ is a finite free extension of $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$.
2. Suppose that $A_{*}$ is Cohen-Macaulay. For a homogeneous sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, $A_{*}$ is a finite free extension of $R_{*}[\poi q1{,\ldots,}{\rg}{}{}{}]$ if and only if $R_{*}A_{+}$ is the radical of the ideal of $A_{*}$ generated by $\poi q1{,\ldots,}{\rg}{}{}{}$.
\(i) The “only if” part results from Lemma \[l3gf1\](v). Suppose that for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, the algebra $A_{*}$ is a finite free extension of $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$. In particular, $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$ is a polynomial algebra over $R_{*}$ since $A_{*}$ has dimension $\dim A$. Let ${\goth p}$ be a prime ideal of $A_{*}$ and let ${\goth q}$ be its intersection with $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$. Denote by $A_{{\goth p}}$ and $R[\poi p1{,\ldots,}{\rg}{}{}{}]_{{\goth q}}$ the localizations of $A_{*}$ and $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$ at ${\goth p}$ and ${\goth q}$ respectively. Since $A_{*}$ is a finite extension of $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$, these local rings have the same dimension. Denote by $d$ this dimension. By flatness, any regular sequence $\poi a1{,\ldots,}{d}{}{}{}$ in $R[\poi p1{,\ldots,}{\rg}{}{}{}]_{{\goth q}}$ is regular in $A_{{\goth p}}$ so that $A_{{\goth p}}$ is Cohen-Macaulay. Hence $A_{*}$ is Cohen-Macaulay.
\(ii) The “only if” part results from (i) and Proposition \[pgf1\]. Suppose that $A_{*}$ is a finite free extension of $R_{*}[\poi q1{,\ldots,}{\rg}{}{}{}]$. Let ${\goth p}$ be a minimal prime ideal of $A_{*}$ containing $\poi q1{,\ldots,}{\rg}{}{}{}$ and let ${\goth q}$ be its intersection with $R_{*}[\poi q1{,\ldots,}{\rg}{}{}{}]$. Then ${\goth q}$ is generated by $\poi q1{,\ldots,}{\rg}{}{}{}$. In particular it has height $\rg$. So ${\goth p}$ has height $\rg$ since $A_{*}$ is a finite extension of $R_{*}[\poi q1{,\ldots,}{\rg}{}{}{}]$. As a result, ${\goth p}=R_{*}A_{+}$ since $R_{*}A_{+}$ is a prime ideal of height $\rg$, containing $\poi q1{,\ldots,}{\rg}{}{}{}$, whence the assertion.
Recall that $B$ is a graded subalgebra of $A$. Set $B_{*} := \tk R{R_{*}}B$ and for ${\goth p}$ a prime ideal of $B$, denote by $B_{{\goth p}}$ its localization at ${\goth p}$.
\[p2gf1\] Suppose that the following conditions are satisfied:
1. $B$ is normal,
2. $A_{+}$ is the radical of $AB_{+}$,
3. $A$ is Cohen-Macaulay.
<!-- -->
1. Let $\poi p1{,\ldots,}{\rg}{}{}{}$ be a homogeneous sequence in $B_{+}$ such that $B_{+}$ is the radical of the ideal of $B$ generated by this sequence. Then for some graded subspace $V$ of $A$, having finite dimension, the linear morphisms $$\tk {\k}VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow A_{*}, \qquad
v\tens a \longmapsto va,$$ $$\tk {\k}{(V\cap B)}R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow B_{*},
\qquad v\tens a \longmapsto va$$ are isomorphisms.
2. If $R=\k$ or $R=\k[[t]]$, the algebra $B_{*}$ is Cohen-Macaulay.
3. For ${\goth p}$ prime ideal of $B$, containing $t$, the local ring $B_{{\goth p}}$ is Cohen-Macaulay.
\(i) By Proposition \[pgf1\] and by Condition (2), $B$ is finitely generated and $A$ is a finite extension of $B$. By Condition (2) and by Lemma \[lgf1\](iii), for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $B_{+}$, $A_{+}$ is the radical of the ideal generated by $\poi p1{,\ldots,}{\rg}{}{}{}$.
Let $\underline{p}$ be as in Lemma \[l3gf1\]. Denote by $m$ the degree of the extension $K(A)$ of $K(B)$. For $a$ in $A_{*} \subset K(A)$, set: $$a^{\#} := \frac{1}{m} \tr a$$ with $\tr:=\tr_{K(A)/K(B)}$ the trace map. By Condition (1), $B_{*}$ is normal and the map $a\mapsto a^{\#}$ is a projection from $A_{*}$ onto $B_{*}$ whose restriction to $A$ is a projection onto $B$. Moreover, it is a graded morphism of $B$-modules. Let $M$ be its kernel. Let $J_{0}$ and $J$ be the ideals of $B$ and $A$ generated by $\underline{p}$ respectively. Since $t,\poi p1{,\ldots,}{\rg}{}{}{}$ are in $B$, $J$ is the direct sum of $J_{0}$ and $MJ_{0}$. Let $V_{0}$ be a graded complement in $B$ to the $\k$-space $J_{0}$ and let $V_{1}$ be a graded complement in $M$ to the $\k$-space $MJ_{0}$. Setting $V:=V_{0}+V_{1}$, $V$ is a graded complement in $A$ to the $\k$-space $J$. By Condition (3) and Lemma \[l3gf1\], $V$ has finite dimension and the linear map $$\tk {\k}VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow A_{*}, \qquad
v\tens a \longmapsto va$$ is an isomorphism. So, since $V_{0}=V^{\#}$, the linear map $$\tk {\k}{V_{0}}R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow B_{*},
\qquad v\tens a \longmapsto va$$ is an isomorphism, whence the assertion.
\(ii) results from (i) and Corollary \[cgf1\].
\(iii) By (i) and Corollary \[cgf1\], $A_{*}$ is Cohen-Macaulay. For ${\goth p}$ a prime ideal of $B$, containing $t$, $B_{{\goth p}}$ is the localization of $B_{*}$ at the prime ideal $B_{*}{\goth p}$, whence the assertion by (ii).
{#gf2}
In this subsection $R=\k[t]$. Then $\widehat{R}=\k[[t]]$. For $M$ a graded module over $R$ such that $M^{[j]}$ is a free submodule of finite rank for all $j$, we denote by $P_{M,R}(T)$ its Hilbert series: $$P_{M,R}(T) := \sum_{j\in {\Bbb N}} \rk M^{[j]} T^{j} .$$ For $V$ a graded space over $\k$ such that $V^{[j]}$ has finite dimension, we denote by $P_{V,\k}(T)$ its Hilbert series: $$P_{V,\k}(T) := \sum_{j\in {\Bbb N}} \dim V^{[j]} T^{j} .$$ Let $S$ be a graded polynomial algebra over $\k$ such that $S^{[0]}=\k$ and $S^{[j]}$ has finite dimension for all $j$. Consider on $S[t]$ and $S[[t]]$ the gradings extending that of $S$ and such that $t$ has degree $0$. Consider the following conditions on $A$:
- $A$ is graded subalgebra of $S[t]$,
- for some homogeneous sequence $\poi a1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, $A=\k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}]\cap S[t]$,
- $A$ is Cohen-Macaulay.
If the condition (C2) holds, then $A[t^{-1}]=R[\poi a1{,\ldots,}{\rg}{}{}{}][t^{-1}]$. Moreover, if so, since $A$ has dimension $\rg +1$, then the elements $t,\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. Set $\widehat{A} := \tk R{\widehat{R}}A$.
\[lgf2\] Assume that the conditions [(C1)]{} and [(C2)]{} hold.
1. The element $t$ is a prime element of $A$.
2. The algebra $A$ is a factorial ring.
3. The Hilbert series of the $R$-module $A$ is equal to $$P_{A,R}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{d_{i}}},$$ with $\poi d1{,\ldots,}{\rg}{}{}{}$ the degrees of $\poi a1{,\ldots,}{\rg}{}{}{}$ respectively.
\(i) Let $a$ and $b$ be in $A$ such that $ab$ is in $tA$. Since $tS[t]$ is a prime ideal of $S[t]$, $a$ or $b$ is in $tS[t]$. Suppose $a=ta'$ for some $a'$ in $S[t]$. Then $a'$ is in $A[t^{-1}]$. By Condition (C2), $A[t^{-1}]=R[\poi a1{,\ldots,}{\rg}{}{}{}][t^{-1}]$. Hence $a'$ is in $A$ by Condition (C2) again. As a result, $At$ is a prime ideal of $A$.
\(ii) Since $A$ is finitely generated, it suffices to prove that all prime ideal of height $1$ is principal by [@Mat Ch. 7, Theorem 20.1]. Let ${\goth p}$ be a prime ideal of height $1$. If $t$ is in ${\goth p}$, then ${\goth p}=At$ by (i). Suppose that $t$ is not in ${\goth p}$ and set $\overline{{\goth p}}=A[t^{-1}]{\goth p}$. Then $\overline{{\goth p}}$ is a prime ideal of height $1$ of $R[\poi a1{,\ldots,}{\rg}{}{}{}][t^{-1}]$ by Condition (C2). For $a$ in $\overline{{\goth p}}$, $t^{m}a$ is in ${\goth p}$ for some nonnegative integer $m$. Hence $${\goth p}=\overline{{\goth p}} \cap A$$ since ${\goth p}$ is prime. As a polynomial ring over the principal ring $\k[t,t^{-1}]$, the ring $R[\poi a1{,\ldots,}{\rg}{}{}{}][t^{-1}]$ is a factorial ring. Then $\overline{{\goth p}}$ is generated by an element $a$ in ${\goth p}$. Since $S$ is a polynomial ring, $S[t]$ is a factorial ring. So, for some nonnegative integer $m$ and for some $a'$ in $S[t]$, prime to $t$, $a=t^{m}a'$. By Condition (C2), $a'$ is in $A$. Then $a'$ is an element of ${\goth p}$, generating $\overline{{\goth p}}$ and not divisible by $t$ in $A$. Let $b$ and $c$ be in $A$ such that $bc$ is in $Aa'$. Then $b$ or $c$ is in $A[t^{-1}]a'$. Suppose $b$ in $A[t^{-1}]a'$. So, for some $l$ in ${\Bbb N}$, $t^{l}b=b'a'$ for some $b'$ in $A$. We choose $l$ minimal satisfying this condition. By (i), since $a'$ is not divisible by $t$ in $A$, $b'$ is divisible by $t$ in $A$ if $l>0$. By minimality of $l$, $l=0$ and $b$ is in $Aa'$. As a result, $Aa'$ is a prime ideal and ${\goth p}=Aa'$ since ${\goth p}$ has height $1$.
\(iii) By Condition (C2), $$A[t^{-1}] = \tk {\k}{\k[t,t^{-1}]}\k[\poi a1{,\ldots,}{\rg}{}{}{}] \quad
\text{whence} \quad \rk A^{[d]} = \dim \k[\poi a1{,\ldots,}{\rg}{}{}{}]^{[d]}$$ for all nonnegative integer $d$. Since $\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$, $$P_{\k[\poi a1{,\ldots,}{\rg}{}{}{}],\k}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{d_{i}}}
,$$ whence the assertion.
Let $\poi p1{,\ldots,}{\rg}{}{}{}$ be a homogeneous sequence in $A$ such that $A_{+}$ is the radical of the ideal of $A$ generated by this sequence. By Lemma \[lgf1\](ii), such a sequence does exist. Denote by $C$ the integral closure of $\k[\poi p1{,\ldots,}{\rg}{}{}{}]$ in $\k(t,\poi a1{,\ldots,}{\rg}{}{}{})$.
\[l2gf2\] Assume that the conditions [(C1)]{}, [(C2)]{} and [(C3)]{} hold.
1. The algebra $C$ is a graded subalgebra of $A$ and $t$ is not algebraic over $C$.
2. The algebra $C$ is Cohen-Macaulay. Moreover, $C$ is a finite free extension of $\k[\poi p1{,\ldots,}{\rg}{}{}{}]$.
3. The algebra $C+tA$ is normal.
\(i) By Lemma \[lgf2\](ii), $A$ is a normal ring such that $K(A)=\k(t,\poi a1{,\ldots,}{\rg}{}{}{})$ by Condition (C2). Then $C$ is contained in $A$ since $\k[\poi p1{,\ldots,}{\rg}{}{}{}]$ is contained in $A$. Moreover, $C$ is a graded algebra since so is $\k[\poi p1{,\ldots,}{\rg}{}{}{}]$. By Proposition \[pgf1\], $A$ is a finite extension of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$. So, since $A$ has dimension $\rg+1$, the elements $t,\poi p1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. As a result, $t$ is not algebraic over $C$.
\(ii) By (i), $C[[t]]=\tk {\k}{C}\k[[t]]$ so that $C[[t]]$ is a flat extension of $\k[[t]]$. Moreover, $C$ is the quotient of $C[[t]]$ by $tC[[t]]$. As $C$ and $\k[[t]]$ are normal rings, $C[[t]]$ is a normal ring by [@Mat Ch. 8, Corollary of Theorem 23.9]. By definition, $A_{+}$ is the radical of the ideal of $A$ generated by $\poi p1{,\ldots,}{\rg}{}{}{}$. As $\k[[t]]$ is a flat extension of $\k[t]$, from the short exact sequence $$\xymatrix{ 0 \ar[r] & A_{+} \ar[r] & A \ar[r] & \k[t] \ar[r] & 0}$$ we deduce the short exact sequence $$\xymatrix{ 0 \ar[r] & \widehat{A}_{+} \ar[r] & \widehat{A} \ar[r] &
\k[[t]] \ar[r] & 0}.$$ Hence $\widehat{A}_{+}$ is a prime ideal. As $A_{+}$ is the radical of the ideal generated by the sequence $\poi p1{,\ldots,}{\rg}{}{}{}$, $\widehat{A}_{+}$ is contained in the radical of $AC[[t]]_{+}$. Then, by (i), $\widehat{A}_{+}$ is the radical of $AC[[t]]_{+}$. Since $\widehat{R}$ is a flat extension of $R$, the algebra $\widehat{A}$ is Cohen-Macaulay by Condition (C3). Then, by Proposition \[p2gf1\](ii), $C[[t]]$ is Cohen-Macaulay. Let $V$ be a graded complement in $C$ to the ideal of $C$ generated by $\poi p1{,\ldots,}{\rg}{}{}{}$. Since $t$ is not algebraic over $C$, the space $V$ is a complement in $C[t]$ to the ideal of $C[t]$ generated by $t,\poi p1{,\ldots,}{\rg}{}{}{}$. Then, by Lemma \[l3gf1\], $V$ has finite dimension and the linear morphism $$\tk {\k}VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow R_{*}C, \qquad
v\tens a \longmapsto va$$ is an isomorphism. As a result, the linear morphism $$\tk {\k}V\k[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow C, \qquad
v\tens a \longmapsto va$$ is an isomorphism, whence the assertion by Corollary \[cgf1\](ii).
\(iii) Set $\tilde{A} := C+tA$. At first, $\tilde{A}$ is a graded subalgebra of $A$ since $C$ is a graded algebra and $tA$ is a graded ideal of $A$. According to Proposition \[p2gf1\](i), for some graded subspace $V$ of $A$, having finite dimension, the linear morphisms $$\tk {\k}VR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow A_{*},
\qquad
v\tens a \longmapsto va,$$ $$\tk {\k}{(V\cap C[t])}R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] \longrightarrow R_{*}C,
\qquad v\tens a \longmapsto va$$ are isomorphisms. Let $\poi v1{,\ldots,}{n}{}{}{}$ be a basis of $V$ such that $\poi v1{,\ldots,}{m}{}{}{}$ is a basis of $V\cap C[t]$. For $a$ in $A_{*}$, the element $a$ has unique expansion $$a = v_{1}a_{1} + \cdots + v_{n}a_{n}$$ with $\poi a1{,\ldots,}{n}{}{}{}$ in $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$. If $a$ is in $tA_{*}$, $\poi a1{,\ldots,}{n}{}{}{}$ are in $tR_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$ and if $a$ is in $R_{*}C$, $\poi a1{,\ldots,}{m}{}{}{}$ are in $\k[\poi p1{,\ldots,}{\rg}{}{}{}]$ and $\poi a{m+1}{,\ldots,}{n}{}{}{}$ are equal to $0$, whence $R_{*}C\cap tA_{*}=tR_{*}C$ and $C\cap tA = \{0\}$. In particular, $C$ is the quotient of $\tilde{A}$ by $t\tilde{A}$.
For ${\goth p}$ a prime ideal of $\tilde{A}$, denote by $\tilde{A}_{{\goth p}}$ the localization of $\tilde{A}$ at ${\goth p}$. If $t$ is not in ${\goth p}$, then $A[t^{-1}]$ is contained in $\tilde{A}_{{\goth p}}$ so that $\tilde{A}_{{\goth p}}$ is a localization of the regular algebra $R[\poi a1{,\ldots,}{\rg}{}{}{}][t^{-1}]$ by Condition (C2). Hence $\tilde{A}_{{\goth p}}$ is a regular local algebra. Suppose that $t$ is in ${\goth p}$. Denote by $\overline{{\goth p}}$ the image of ${\goth p}$ in $C$ by the quotient map. Then $\tilde{A}_{{\goth p}}/t\tilde{A}_{{\goth p}}$ is the localization $C_{\overline{{\goth p}}}$ of $C$ at the prime ideal $\overline{{\goth p}}$. Since $C$ is Cohen-Macaulay, so are $C_{\overline{{\goth p}}}$ and $\tilde{A}_{{\goth p}}$. As a result, $\tilde{A}$ is Cohen-Macaulay.
Let ${\goth p}$ be a prime ideal of height $1$ of $\tilde{A}$. If $t$ is not in ${\goth p}$, $\tilde{A}_{{\goth p}}$ is a regular local algebra as it is already mentioned. Suppose that $t$ is in ${\goth p}$. By Lemma \[lgf2\](i), $t\tilde{A}={\goth p}$ so that all element of $C\setminus \{0\}$ is invertible in $\tilde{A}_{{\goth p}}$, whence $$\tilde{A}_{{\goth p}}=K(C)+t\tilde{A}_{{\goth p}} \quad \text{and} \quad
t\tilde{A}_{{\goth p}} = tK(C) + t^{2}\tilde{A}_{{\goth p}} .$$ Hence $\tilde{A}_{{\goth p}}$ is a regular local ring of dimension $1$. As a result, $\tilde{A}$ is regular in codimension $1$. Then, by Serre’s normality criterion [@Bou §1, 10, Théorème 4], $\tilde{A}$ is normal since $\tilde{A}$ is Cohen-Macaulay.
\[cgf2\] Assume that the conditions [(C1)]{}, [(C2)]{} and [(C3)]{} hold.
1. The algebra $\widehat{A}$ is equal to $C[[t]]$.
2. For $a$ in $A$, the element $ra$ is in $C[t]$ for some $r$ in $\k[t]$ such that $r(0) \neq 0$.
\(i) Since $tA$ is contained in $A$, we have $K(A)=K(\tilde{A})$. Since $C_{+}$ is contained in $\tilde{A}_{+}$, $A_{+}$ is the radical of $A\tilde{A}_{+}$. Then, by Proposition \[pgf1\], $A$ is a finite extension of $\tilde{A}$. So, by Lemma \[l2gf2\](iii), $A=\tilde{A}$ and by induction on $m$, $$A \subset C[t] + t^{m}A$$ for all positive integer $m$. Since $A$ and $C[t]$ are graded and since the $R$-module $A^{[d]}$ is finitely generated for all $d$, $\widehat{A}=C[[t]]$.
\(ii) The assertion results from (i) and Lemma \[l2gf1\].
\[pgf2\] Assume that the conditions [(C1)]{}, [(C2)]{} and [(C3)]{} hold. Then the algebra $A_{*}$ is polynomial over $R_{*}$. Moreover, for some homogeneous sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$ such that $\poi q1{,\ldots,}{\rg}{}{}{}$ have degree $\poi d1{,\ldots,}{\rg}{}{}{}$ respectively, $A_{*}=R_{*}[\poi q1{,\ldots,}{\rg}{}{}{}]$.
According to Corollary \[cgf2\] and Lemma \[l2gf2\](i), it suffices to prove that $C$ is a polynomial algebra over $\k$ generated by a homogeneous sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ such that $\poi q1{,\ldots,}{\rg}{}{}{}$ have degree $\poi d1{,\ldots,}{\rg}{}{}{}$ respectively. According to Corollary \[cgf2\](i) Lemma \[l2gf2\](i) and Lemma \[lgf2\](iii), $$P_{C,\k}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{d_{i}}} .$$ By Corollary \[cgf2\](ii), for $i=1,\ldots,\rg$, for some $r_{i}$ in $R$ such that $r_{i}(0)\neq 0$, $r_{i}a_{i}$ has an expansion $$r_{i}a_{i} = \sum_{m\in {\Bbb N}} c_{i,m} t^{m}$$ with $c_{i,m},m\in {\Bbb N}$ in $C^{[d_{i}]}$, with finite support. For $z$ in $\k$ and $i=1,\ldots,\rg$, set: $$b_{i}(z) = \sum_{m\in {\Bbb N}} c_{i,m} z^{m}$$ so that $b_{i}(z)$ is in $C^{[d_{i}]}$ for all $z$. As already mentioned, $t,\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$ by Condition (C2) since $A$ has dimension $\rg +1$. Then, so are $t,r_{1}a_{1},\ldots,r_{\rg}a_{\rg}$ and for some $z$ in $\k$, $\poi z{}{,\ldots,}{}{b}{1}{\rg}$ are algebraically independent over $\k$. Denoting by $C'$ the subalgebra of $C$ generated by this sequence, $$P_{C',\k}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{d_{i}}},$$ whence $C=C'$ so that $C$ is a polynomial algebra.
Proof of Theorem \[ti6\] {#sg}
========================
In this section, unless otherwise specified, the grading on $\es S{{\goth g}^{e}}$ is the Slodowy grading.
For $m$ a nonnegative integer, $\es S{{\goth g}^{e}}^{[m]}$ denotes the space of degree $m$ of $\es S{{\goth g}^{e}}$. We retain the notations of the introduction, in particular of Subsection \[ss:plan\].
{#sg1}
Let $R$ be the ring $\k[t]$. As in Section \[gf\], for $M$ a graded subspace of $\es S{{\goth g}^{e}}[t]=\tk \k{R}\es S{{\goth g}^{e}}$, its subspace of degree $m$ is denoted by $M^{[m]}$. In particular, $\es S{{\goth g}^{e}}[t]^{[m]}$ is equal to $\es S{{\goth g}^{e}}^{[m]}[t]$ and it is a free $R$-module of finite rank. As a result, for all graded $R$-submodule $M$ of $\es S{{\goth g}^{e}}[t]$, its Hilbert series is well defined.
For $m$ a nonnegative integer, denote by $F_{m}$ the space of elements of $\kappa (\ai g{}{})$ whose component of minimal standard degree is at least $m$. Then $\poi F0{,}{1}{}{}{},\ldots$ is a decreasing filtration of the algebra $\kappa (\ai g{}{})$. Let $\poi d1{,\ldots,}{\rg}{}{}{}$ be the standard degrees of a homogeneous generating sequence of $\ai g{}{}$. We assume that the sequence $\poi d1{,\ldots,}{\rg}{}{}{}$ is increasing.
Recall that $A$ is the intersection of $\es S{{\goth g}^{e}}[t]$ with the sub-$\k[t,t^{-1}]$-module of $\es S{{\goth g}^{e}}[t,t^{-1}]$ generated by $\tau \rond \kappa(\e Sg^\g)$, and that $A_+$ is the augmentation ideal of $A$.
\[lsg1\]
1. For $p$ a homogeneous element of standard degree $d$ in $\ai g{}{}$, the element $\kappa (p)$ and $\ie p$ have degree $2d$.
2. For some homogeneous sequence $\poi a1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, the elements $t,\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$, and $A$ is the intersection of ${\rm S}({\goth g}^{e})[t]$ with $\k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}]$.
3. The Hilbert series of the $R$-algebra $A$ is equal to $$P_{A,R}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{2d_{i}}} .$$
4. The Hilbert series of the $\k$-algebra $\varepsilon (A)$ is equal to $$P_{\varepsilon (A),\k}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{2d_{i}}} .$$
5. The subalgebra $\varepsilon (A)$ is the graded algebra associated with the filtration $\poi F0{,}{1}{}{}{},\ldots$.
\(i) Let $\rho$ be as in Subsection \[ss:plan\]. For $y$ in ${\goth g}^{f}$ and $s$ in $\k^{*}$, $$p(s^{-2}\rho (s)(e+y)) = s^{-2d}p(\rho (s)(e+y)) = s^{-2d}p(e+y)$$ since $p$ is invariant under the one-parameter subgroup $\rho$. Hence $\kappa (p)$ is homogeneous of degree $2d$. Since the monomials $x^{{\bf j}}$ are homogeneous, $\ie p$ has degree $2d$.
\(ii) Let $\poi q1{,\ldots,}{\rg}{}{}{}$ be a homogeneous generating sequence of $\ai g{}{}$. By a well known fact (cf. e.g. [@CM1 Lemma 4.4(i)]), the morphism $$G \times (e+{\goth g}^{f}) \longrightarrow {\goth g}, \qquad (g,x) \longmapsto g(x)$$ is dominant. Then $\kappa (\ai g{}{})$ is a polynomial algebra generated by $\poi q1{,\ldots,}{\rg}{\kappa }{}{}$. So, setting $a_{i}:=\tau \rond \kappa (q_{i})$ for $i=1,\ldots,\rg$, the sequence $\poi a1{,\ldots,}{\rg}{}{}{}$ is a homogeneous sequence in $A_{+}$ such that $$\tau \rond \kappa (\ai g{}{})[t,t^{-1}] = \k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}] .$$ Let $\overline{\tau }$ be the automorphism of $\es S{{\goth g}^{e}}[t,t^{-1}]$ extending $\tau $ and such that $\overline{\tau }(t)=t$. Then $$\tau \rond \kappa (\ai g{}{})[t,t^{-1}] =
\overline{\tau }(\kappa (\ai g{}{})[t,t^{-1}]) .$$ Since $\kappa (\ai g{}{})[t,t^{-1}]$ has dimension $\rg +1$, $\tau \rond \kappa (\ai g{}{})[t,t^{-1}]$ has dimension $\rg +1$ too, and $t,\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. By definition, $A=\es S{{\goth g}^{e}}[t]\cap \tau \rond \kappa (\ai g{}{})[t,t^{-1}]$. Hence $$A[t^{-1}] = \k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}] \quad \text{and} \quad
A = \es S{{\goth g}^{e}}[t] \cap \k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}] .$$
\(iii) Since $t$ has degree $0$, the grading of $\es S{{\goth g}^{e}}[t]$ extends to a grading of $\es S{{\goth g}^{e}}[t,t^{-1}]$ such that for all $m$, its space of degree $m$ is equal to $\es S{{\goth g}^{e}}^{[m]}[t,t^{-1}]$. Then for all $\k[t,t^{-1}]$-submodule $M$ of $\es S{{\goth g}^{e}}[t,t^{-1}]$, $M$ has a Hilbert series: $$P_{M,\k[t,t^{-1}]}(T) := \sum_{m\in {\Bbb N}} \rk M^{[m]} T^{m}$$ with $M^{[m]}$ the subspace of degree $m$ of $M$. From the equality $A[t^{-1}] = \k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}]$, we deduce $$P_{A[t^{-1}],\k[t,t^{-1}]}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{2d_{i}}}$$ since for $i=1,\ldots,\rg$, the element $a_{i}$ has degree $2d_{i}$ by (i). For all $m$, the rank of the $R$-module $A^{[m]}$ is equal to the rank of the $\k[t,t^{-1}]$-module $A[t^{-1}]^{[m]}$, whence $$P_{A,R}(T) = \prod_{i=1}^{\rg} \frac{1}{1-T^{2d_{i}}} .$$
\(iv) Let $m$ be a nonnegative integer. The $R$-module $A^{[m]}$ is free of finite rank and for $(\poi v1{,\ldots,}{n}{}{}{})$ a basis of this module, $(\poi {tv}1{,\ldots,}{n}{}{}{})$ is a basis of the $R$-module $tA^{[m]}$. Since $\varepsilon (A)^{[m]}$ is the quotient of $A^{[m]}$ by $tA^{[m]}$, $$\dim \varepsilon (A)^{[m]} = n = \rk A^{[m]},$$ whence the assertion by (iii).
\(v) Let ${\rm gr}_FA$ be the graded algebra associated with the filtration $\poi F0{,}{1}{}{}{},\ldots$ of $\kappa (\ai g{}{})$. Denote by $a\mapsto a(1)$ the evaluation map at $t=1$ from $\es S{{\goth g}^{e}}[t]$ to $\es S{{\goth g}^{e}}$. For $a$ in $A$ such that $\varepsilon (a)\neq 0$, $a(1)$ is in $\kappa (\ai g{}{})$ and $\varepsilon (a)$ is the component of minimal degree of $a(1)$ with respect to the standard grading, whence $\varepsilon (A)\subset {\rm gr}_F A$. Conversely, let $\overline{a}$ be a homogeneous element of degree $m$ of ${\rm gr}_F A$ and let $a$ be a representative of $\overline{a}$ in $F_{m}$. Then $\tau (a)=t^{m}b$ with $b$ in $A$ such that $\varepsilon (b)=\overline{a}$, whence ${\rm gr}_F A \subset \varepsilon (A)$ and the assertion.
Let $R_{*}$ be the localization of $R$ at the prime ideal $tR$ and set $$\widehat{R} := \k[[t]], \qquad A_{*} := \tk R{R_{*}}A, \qquad
\widehat{A} := \tk R{\widehat{R}}A .$$ The grading of $A$ extends to gradings on $A_{*}$ and $\widehat{A}$ such that $A_{*}^{[0]}=R_{*}$ and $\widehat{A}^{[0]}=\widehat{R}$.
\[psg1\]
1. The algebra $\varepsilon (A)$ is polynomial if and only if for some standard homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the elements $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. Moreover, in this case, $A$ is a polynomial algebra.
2. If $A_{*}$ is a polynomial algebra over $R_{*}$, then for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, we have $A_{*} = R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$, the elements $t,\poi p1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$ and $\poi p1{,\ldots,}{\rg}{}{}{}$ have degree $\poi {2d}1{,\ldots,}{\rg}{}{}{}$ respectively.
\(i) Let $\poi q1{,\ldots,}{\rg}{}{}{}$ be a homogeneous generating sequence of $\ai g{}{}$ such that $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. We can assume that for $i=1,\ldots,\rg$, $q_i$ has standard degree $d_i$. For $i=1,\ldots,\rg$, $\ie{q_i}$ has degree $2 d_i$ by Lemma \[lsg1\](i), and we set $$Q_{i} := t^{- 2 d_i}\tau \rond \kappa (q_{i}).$$ Then $Q_i$, for $i=1,\ldots,\rg$, is in $A$ by definition of $A$. For ${\bf i}=(\poi i1{,\ldots,}{\rg}{}{}{})$ in ${\Bbb N}^{\rg}$, set: $$q^{{\bf i}} := \poie q1{\cdots}{\rg}{}{}{}{i_{1}}{i_{\rg}}, \qquad
Q^{{\bf i}} := \poie Q1{\cdots}{\rg}{}{}{}{i_{1}}{i_{\rg}}, \qquad
\ie q^{{\bf i}} := \poie {\ie q}1{\cdots}{\rg}{}{}{}{i_{1}}{i_{\rg}},$$ $$\vert {\bf i} \vert_{{\mathrm {min}}} :=
2 i_{1} d _{1} + \cdots + 2 i_{\rg} d _{\rg} .$$ Then, for all ${\bf i}$ in ${\Bbb N}^{\rg}$, $$\tau \rond \kappa (q^{{\bf i}}) = t^{\vert {\bf i} \vert_{{\mathrm {min}}}}
Q^{{\bf i}}.$$ Moreover, $$\tau \rond \kappa (\ai g{}{})[t,t^{-1}] = \k[t,t^{-1},\poi Q1{,\ldots,}{\rg}{}{}{}].$$ Let $a$ be in $A$. For some $l$ in ${\Bbb N}$ and for some sequence $c_{{\bf i},m}, \; ({\bf i},m)\in {\Bbb N}^{\rg}\times {\Bbb N}$ in $\k$, of finite support, $$t^{l}a = \sum_{({\bf i},m) \in {\Bbb N}^{\rg}\times {\Bbb N}}
c_{{\bf i},m} t^{m}Q^{{\bf i}} \quad \text{whence} \quad
\sum_{{\bf i}\in {\Bbb N}^{\rg}} c_{{\bf i},m} \ie q^{{\bf i}} = 0$$ for $m<l$. Hence $a$ is in $R[\poi Q1{,\ldots,}{\rg}{}{}{}]$ since the elements $\ie q^{{\bf i}}$, ${\bf i}\in {\Bbb N}^{\rg}$ are linearly independent over $\k$. As a result, $$A = R[\poi Q1{,\ldots,}{\rg}{}{}{}] \quad \text{and} \quad
\varepsilon (A) = \k[\poi {\ie q}1{,\ldots,}{\rg}{}{}{}]$$ so that $A$ and $\varepsilon (A)$ are polynomial algebras over $\k$ since $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$.
Conversely, suppose that $\varepsilon (A)$ is a polynomial algebra. By Lemma \[lsg1\], (i) and (iv), the algebra $\varepsilon (A)$ is graded for both Slodowy grading and standard grading. Let $d$ be the dimension of $\varepsilon (A)$. As $\varepsilon (A)$ is a polynomial algebra, it is regular so that the $\k$-space $\varepsilon (A)_{+}/\varepsilon (A)_{+}^{2}$ has dimension $d$. Moreover, the two gradings on $\varepsilon (A)$ induce gradings on $\varepsilon (A)_{+}/\varepsilon (A)_{+}^{2}$. Hence $\varepsilon (A)_{+}/\varepsilon (A)_{+}^{2}$ has a bihomogeneous basis. Then some bihomogeneous sequence $\poi u1{,\ldots,}{d}{}{}{}$ in $\varepsilon (A)_{+}$ represents a basis of $\varepsilon (A)_{+}/\varepsilon (A)_{+}^{2}$. As a result, the $\k$-algebra $\varepsilon (A)$ is generated by the bihomogeneous sequence $\poi u1{,\ldots,}{d}{}{}{}$. For $i=1,\ldots,d$, denote by $\delta _{i}$ the Slodowy degree of $u_{i}$. As $\varepsilon $ is homogeneous with respect to the Slodowy grading, $u_{i}=\varepsilon (r_{i})$ for some homogeneous element $r_{i}$ of degree $\delta _{i}$ of $A$. Let $m_{i}$ be the smallest nonnegative integer such that $t^{m_{i}}r_{i}$ is in $\tau \rond \kappa (\ai g{}{})$. According to Lemma \[lsg1\](i), $\delta _{i}$ is even and for some standard homogeneous element $p_{i}$ of standard degree $\delta _{i}/2$ of $\ai g{}{}$, $t^{m_{i}}r_{i}=\tau \rond \kappa (p_{i})$. Then $u_{i}=\ie {p_{i}}$ since $p_{i}$ is standard homogeneous.
Let ${\goth P}$ be the subalgebra of $\e Sg$ generated by $\poi p1{,\ldots,}{d}{}{}{}$. Suppose that ${\goth P}$ is strictly contained in $\ai g{}{}$. A contradiction is expected. For some positive integer $m$, the space $\ai g{}{}_{m}$ of standard degree $m$ of $\ai g{}{}$ is not contained in ${\goth P}$. Let $q$ be in $(\ai g{}{})_{m}\setminus {\goth P}$ such that $\ie q$ has maximal standard degree. By Lemma \[lsg1\](i), $\ie q$ is a polynomial in $\poi u1{,\ldots,}{d}{}{}{}$, of degree $2m$. So, for some polynomial $q'$ of degree $m$ in ${\goth P}$, $\ie (q-q')$ has standard degree bigger than the standard degree of $\ie q$. So, by maximality of the standard degree of $\ie q$, the elements $q-q'$ and $q$ are in ${\goth P}$, whence the contradiction. As a result, ${\goth P}=\ai g{}{}$ and $d=\rg$.
\(ii) Suppose that $A_{*}$ is a polynomial algebra. Denoting by $J$ the ideal of $A_{*}$ generated by $t$ and $A_{+}$, the $\k$-space $J/J^{2}$ is a graded space of dimension $\rg +1$ since $A_{*}$ is a regular algebra of dimension $\rg+1$. Then for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, $(t,\poi p1{,\ldots,}{\rg}{}{}{})$ is a basis of $J$ modulo $J^{2}$. Since $\poi p1{,\ldots,}{\rg}{}{}{}$ have positive degree, we prove by induction on $d$ that $$A_{*}^{[d]} \subset R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]^{[d]} + tA_{*}^{[d]} .$$ Then by induction on $m$, we get $$A_{*}^{[d]} \subset R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] + t^{m}A_{*}^{[d]} .$$ So, since the $R_{*}$-module $A_{*}^{[d]}$ is finitely generated, $$A_{*}^{[d]} \subset \widehat{R}[\poi p1{,\ldots,}{\rg}{}{}{}]^{[d]}.$$ Apply Lemma \[l2gf1\] to $N=A$ and $M=\es S{{\goth g}^{e}}[t]$. Since $\widehat{N}=\widehat{R}[\poi p1{,\ldots,}{\rg}{}{}{}]$, for $a \in N$, there exists $r \in R$ such that $r(0)\neq 0$ and $ra \in R[\poi p1{,\ldots,}{\rg}{}{}{}]$ by Lemma \[l2gf1\]. So $A_{*}$ is contained in $R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$, whence $A_{*}=R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}]$.
Denote by $\poi {\delta }1{,\ldots,}{\rg}{}{}{}$ the respective degrees of $\poi p1{,\ldots,}{\rg}{}{}{}$. We can suppose that $\poi p1{,\ldots,}{\rg}{}{}{}$ is ordered so that $\poi {\delta }1{\leq \cdots \leq }{\rg}{}{}{}$. Prove by induction on $i$ that $\delta _{j}=2d_{j}$ for $j=1,\ldots,i$. By Lemma \[lsg1\](iii), $2d_{1}$ is the smallest positive degree of the elements of $A$. Moreover, $\delta _{1}$ is the smallest positive degree of the elements of $R[\poi p1{,\ldots,}{\rg}{}{}{}]$, whence $\delta _{1}=2d_{1}$. Suppose $\delta _{j}=2d_{j}$ for $j=1,\ldots,i-1$. Set $A_{i} := R[\poi pi{,\ldots,}{\rg}{}{}{}]$. Then, by induction hypothesis and Lemma \[lsg1\](iii), $$P_{A_{i},R}(T) = \prod_{j=i}^{\rg} \frac{1}{1-T^{\delta _{j}}} =
\prod_{j=i}^{\rg} \frac{1}{1-T^{2d_{j}}} .$$ By the first equality, $\delta _{i}$ is the smallest positive degree of the elements of $A_{i}$ and by the second equality, $2d_{i}$ is the smallest positive degree of the elements of $A_{i}$ too, whence $\delta _{i}=2d_{i}$. Then with $i=\rg$, we get that $\delta _{j}=2d_{j}$ for $j=1,\ldots,\rg$.
Recall that $\widehat{R}=\k[[t]]$.
\[csg1\] Suppose that $A_{*}$ is a polynomial algebra. Then for some standard homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ in $\ai g{}{}$, $$A_{*}=R_{*}[t^{- 2 d _{1}}\tau \rond \kappa (q_{1}),\ldots,
t^{- 2 d _{\rg}}\tau \rond \kappa (q_{\rg})] .$$
For $m$ nonnegative integer, denote by $\ai g{}{}_{m}$ the space of standard degree $m$ of $\ai g{}{}$. By Proposition \[psg1\](ii), for some homogeneous sequence $\poi p1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$ such that $\poi p1{,\ldots,}{\rg}{}{}{}$ have degree $\poi {2d}1{,\ldots,}{\rg}{}{}{}$ respectively, $$A_{*} = R_{*}[\poi p1{,\ldots,}{\rg}{}{}{}] .$$ For $i=1,\ldots,\rg$, let $m_{i}$ be the smallest integer such that $t^{m_{i}}p_{i}$ is in $\tau \rond \kappa (\ai g{}{})$. By Lemma \[lsg1\](i), $t^{m_{i}}p_{i}$ has an expansion $$t^{m_{i}}p_{i} = \sum_{j\in {\Bbb N}} t^{j}\tau \rond \kappa (q_{i,j})$$ with $q_{i,j}$, $j\in {\Bbb N}$, in $\ai g{}{}_{d_{i}}$ of finite support. Denoting by $\delta _{i,j}$ the standard degree of $\ie q_{i,j}$, set: $$J'_{i} := \{j \in {\Bbb N} \; ; \; m_{i} = j+\delta _{i,j}\}, \qquad
\delta _{i} := \inf \{ \delta _{i,j} \; ; \; j\in J'_{i}\},$$ $$j_{i} := m_{i}- 2 d_i , \qquad
Q_{i} := t^{-2 d _{i}} \tau \rond \kappa (q_{i,j_{i}}).$$ For $i=1,\ldots,\rg$, since $p_{i}$ is not divisible by $t$ in $A$, $$p_{i} - Q_{i} \in tA ,$$ whence $$A_{*} \subset R_{*}[\poi Q1{,\ldots,}{\rg}{}{}{}] + tA_{*} .$$ Then, by induction $m$, $$A_{*} \subset R_{*}[\poi Q1{,\ldots,}{m}{}{}{}] + t^{m}A_{*}$$ for all $m$. As a result, $$\widehat{A} = \widehat{R}[\poi Q1{,\ldots,}{\rg}{}{}{}] ,$$ since for all $d$, the $R_{*}$-module $A_{*}^{[d]}$ is finitely generated. Then, by Lemma \[l2gf1\], $$A_{*} = R_{*}[\poi Q1{,\ldots,}{\rg}{}{}{}] .$$ As a result, since $A$ has dimension $\rg +1$, the elements $t,\poi Q1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$ and so are $\poi q{1,j_{1}}{,\ldots,}{\rg,j_{\rg}}{}{}{}$. Moreover the algebra $\ai g{}{}$ is generated by $\poi q{1,j_{1}}{,\ldots,}{\rg,j_{\rg}}{}{}{}$ since they have degree $\poi d1{,\ldots,}{\rg}{}{}{}$ respectively.
{#sg2}
Denote by ${\cal V}$ the nullvariety of $A_{+}$ in ${\goth g}^{f}\times \k$. Let ${\cal V}_{*}$ be the union of the irreducible components of ${\cal V}$ which are not contained in ${\goth g}^{f}\times \{0\}$. The following result is proven in [@CM1 Corollary 4.4(i)]. Indeed, the proof of this result does not use the assumption of [@CM1 Section 4] that for some homogeneous generators $q_1,\ldots,q_\ell$ of ${\rm S}(\g)^\g$, the elements $\ie{q_1},\ldots,\ie{q_\ell}$ are algebraically independent.
\[lsg2\]
1. The variety ${\cal V}_{*}$ is equidimensional of dimension $r+1-\rg$.
2. For all irreducible component $X$ of ${\cal V}_{*}$ and for all $z$ in $\k$, $X$ is not contained in ${\goth g}^{f}\times \{z\}$.
Let ${\cal N}$ be the nullvariety of $\varepsilon (A)_{+}$ in ${\goth g}^{f}$. Then ${\cal V}$ is the union of ${\cal V}_{*}$ and ${\cal N}\times \{0\}$.
\[l2sg2\]
1. All irreducible component of ${\cal N}$ have dimension at least $r-\rg$ and all irreducible component of ${\cal V}$ have dimension at least $r+1-\rg$.
2. Assume that ${\cal N}$ has dimension $r-\rg$. Then for some homogeneous sequence $\poi p1{,\ldots,}{r-\rg}{}{}{}$ in $\es S{{\goth g}^{e}}_{+}$, the nullvariety of $t,\poi p1{,\ldots,}{r-\rg}{}{}{}$ in ${\cal V}$ is equal to $\{0\}$.
\(i) By Lemma \[lsg1\](ii), for some homogeneous sequence $\poi a1{,\ldots,}{\rg}{}{}{}$ in $A_{+}$, the elements $t,\poi a1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$. Let $\poi b1{,\ldots,}{m}{}{}{}$ be a homogeneous sequence in $A_{+}$, generating the ideal $\es S{{\goth g}^{e}}[t]A_{+}$ of $\es S{{\goth g}^{e}}[t]$. Set: $$B := \k[\poi a1{,\ldots,}{\rg }{}{}{},\poi b1{,\ldots,}{m}{}{}{}],
\qquad B_{+} := \poi {Ba}1{+\cdots +}{\rg }{}{}{}+\poi {Bb}1{+\cdots +}{m}{}{}{},$$ $$C := B[t], \qquad C_{++} := B_{+}[t] + Ct .$$ Then $B$ and $C$ are graded subalgebras of $A$ and $B_{+}$ and $C_{++}$ are maximal ideals of $B$ and $C$ respectively. Moreover, $C$ has dimension $\rg +1$. We have a commutative diagram $$\xymatrix{ & {\goth g}^{f}\times \k \ar[ld]_{\alpha } \ar[rd]^{\beta } & \\
{\mathrm {Specm}}(C) \ar[rr]^{\pi } & & {\mathrm {Specm}}(B) }$$ with $\alpha $, $\beta $, $\pi $ the morphisms whose comorphisms are the canonical injections $$C\hookrightarrow \es S{{\goth g}^{e}}[t], \quad B \hookrightarrow
\es S{{\goth g}^{e}}[t], \quad B \hookrightarrow C$$ respectively. Since $C$ has dimension $\rg +1$, the irreducible components of the fibers of $\alpha $ have dimension at least $r-\rg$, whence the result for ${\cal N}$ since ${\cal N}\times \{0\}=\alpha ^{-1}(C_{++})$. Moreover, ${\cal V}=\beta ^{-1}(B_{+})$ and $\pi ^{-1}(B_{+})$ is a subvariety of dimension $1$ of ${\mathrm {Specm}}(C)$. Hence all irreducible component of ${\cal V}$ has dimension at least $r+1-\rg$.
\(ii) Prove by induction on $i$ that there exists a homogeneous sequence $\poi p1{,\ldots,}{i}{}{}{}$ in $\es S{{\goth g}^{e}}_{+}$ such that the minimal prime ideals of $\es S{{\goth g}^{e}}$ containing $\varepsilon (A)_{+}$ and $\poi p1{,\ldots,}{i}{}{}{}$ have height $\rg +i$. First of all, $\es S{{\goth g}^{e}}\varepsilon (A)_{+}$ is graded. Then the minimal prime ideals of $\es S{{\goth g}^{e}}$ containing $\varepsilon (A)_{+}$ are graded too. By, (i), they have height $\rg$ since ${\cal N}$ has dimension $r-\rg$ by hypothesis. In particular, they are strictly contained in $\es S{{\goth g}^{e}}_{+}$. Hence, by Lemma \[lgf1\](ii), for some homogeneous element $p_{1}$ in $\es S{{\goth g}^{e}}$, $p_{1}$ is not in the union of these ideals so that the statement is true for $i=1$ by [@Mat Ch. 5, Theorem 13.5]. Suppose that it is true for $i-1$. Then the minimal prime ideals containing $\varepsilon (A)_{+}$ and $\poi p1{,\ldots,}{i-1}{}{}{}$ are graded and strictly contained in $\es S{{\goth g}^{e}}_{+}$ by the induction hypothesis. So, by Lemma \[lgf1\](ii), for some homogeneous element $p_{i}$ in $\es S{{\goth g}^{e}}$, $p_{i}$ is not in the union of these ideals and the sequence $\poi p1{,\ldots,}{i}{}{}{}$ satisfy the condition of the statement by [@Mat Ch. 5, Theorem 13.5]. For $i=r-\rg$, the nullvariety of $\poi p1{,\ldots,}{r-\rg}{}{}{}$ in ${\cal N}$ has dimension $0$. Then it is equal to $\{0\}$ as the nullvariety of a graded ideal, whence the assertion since ${\cal N}\times \{0\}$ is the nullvariety of $t$ in ${\cal V}$.
{#sg3}
We assume in this subsection that ${\cal N}$ has dimension $r-\ell$. Let $\poi p1{,\ldots,}{r-\rg}{}{}{}$ be as in Lemma \[l2sg2\](ii), and set $$C := A[\poi p1{,\ldots,}{r-\rg}{}{}{}] .$$ Then $\poi p1{,\ldots,}{r-\rg}{}{}{}$ are algebraically independent over $A$ since ${\cal N}$ has dimension $r-\ell$.
\[lsg3\] The ideal $\es S{{\goth g}^{e}}[t]_{+}$ of $\es S{{\goth g}^{e}}[t]$ is the radical of $\es S{{\goth g}^{e}}[t]C_{+}$.
Let $Y$ be an irreducible component of the nullvariety of $C_{+}$ in ${\goth g}^{f}\times \k$. Then $Y$ has dimension at least $1$. By definition the nullvariety of $t$ in $Y$ is equal to $\{0\}$. Hence $Y$ has dimension $1$. The grading on $\es S{{\goth g}^{e}}[t]$ induces an action of the one-dimensional multiplicative group $\gm$ on ${\goth g}^{f}\times \k$ such that for all $(x,z)$ in ${\goth g}^{f}\times \k$, $(0,z)$ is in the closure of the orbit of $(x,z)$ under $\gm$. Since $C_{+}$ is graded, $Y$ is invariant under $\gm$. As a result, $Y=\{0\}\times \k$ or for some $x$ in ${\goth g}^{f}\times \k$, $Y$ is the closure of the orbit of $(x,0)$ under $\gm$ since $0$ is the nullvariety of $t$ in $Y$. In the last case, $x$ is a zero of $\poi p1{,\ldots,}{r-\rg}{}{}{}$ in ${\cal N}$, that is $x=0$. Hence $Y=\{0\}\times \k$. As a result, the nullvariety of $C_{+}$ in ${\goth g}^{f}\times \k$ is equal to $\{0\}\times \k$ that is the nullvariety of $\es S{{\goth g}^{e}}[t]_{+}$, whence the assertion since $\es S{{\goth g}^{e}}[t]_{+}$ is a prime ideal of $\es S{{\goth g}^{e}}[t]$.
For ${\goth p}$ a prime ideal of $A$, denote by $A_{{\goth p}}$ the localization of $A$ at ${\goth p}$ and by $\overline{{\goth p}}$ the ideal of $C$ generated by ${\goth p}$. Since $C$ is a polynomial algebra over $A$, $\overline{{\goth p}}$ is a prime ideal of $C$ and $A\setminus {\goth p}$ is the intersection of $A$ and $C\setminus \overline{{\goth p}}$. Hence the localization $C_{\overline{{\goth p}}}$ of $C$ at $\overline{{\goth p}}$ is a localization of the polynomial algebra $A_{{\goth p}}[\poi p1{,\ldots,}{r-\rg}{}{}{}]$. Moreover, $A_{{\goth p}}$ is the quotient of $C_{\overline{{\goth p}}}$ by the ideal generated by $\poi p1{,\ldots,}{r-\rg}{}{}{}$. According to [@Mat Ch. 6, Theorem 17.4], if $C_{\overline{{\goth p}}}$ is Cohen-Macaulay, $\poi p1{,\ldots,}{r-\rg}{}{}{}$ is a regular sequence in $C_{\overline{{\goth p}}}$ since $A_{{\goth p}}$ has dimension $\dim C_{\overline{{\goth p}}}-r+\rg$. Then, again by [@Mat Ch. 6, Theorem 17.4], $A_{{\goth p}}$ is Cohen-Macaulay if so is $C_{\overline{{\goth p}}}$.
By Lemma \[lsg3\] and Proposition \[pgf1\], the algebra $C$ is finitely generated. Then $A$ is finitely generated as a quotient of $C$. Hence by Lemma \[lgf2\](ii), $A$ is a factorial ring and so is $C$ as a polynomial ring over $A$. As a result, $C$ is normal so that $\es S{{\goth g}^{e}}[t]$ and $C$ satisfy the conditions (1), (2), (3) of Proposition \[p2gf1\]. Hence by Proposition \[p2gf1\], for all prime ideal ${\goth p}$ of $A$, containing $t$, $C_{\overline{{\goth p}}}$ is Cohen-Macaulay, whence $A_{{\goth p}}$ is Cohen-Macaulay. By Lemma \[lsg1\](ii), for ${\goth p}$ a prime ideal of $A$, not containing $t$, $A_{{\goth p}}$ is the localization of $\k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}]$ at the prime ideal generated by ${\goth p}$. Therefore $A_{{\goth p}}$ is Cohen-Macaulay since the algebra $\k[t,t^{-1},\poi a1{,\ldots,}{\rg}{}{}{}]$ is regular. As a result $A$ is Cohen-Macaulay. In particular, $A$ satisfies the conditions (1), (2), (3) of Subsection \[gf2\]. So, by Proposition \[pgf2\], $A_{*}$ is a polynomial algebra over $R_*$. Then by Corollary \[csg1\], for some homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ in $\ai g{}{}$, $$A_{*} = R_{*}[t^{- 2 d _{1}}\tau \rond \kappa (q_{1}),\ldots,
t^{- 2 d_{\rg}}\tau \rond \kappa (q_{\rg})].$$ Form the above equality, we deduce that any element of $A$ is the product of an element of the algebra $R[t^{- 2 d _{1}}\tau \rond \kappa (q_{1}),\ldots,
t^{- 2 d_{\rg}}\tau \rond \kappa (q_{\rg})]$ by a polynomial in $t$ with nonzero constant term, whence $$A= R[t^{- 2 d _{1}}\tau \rond \kappa (q_{1}),\ldots,
t^{- 2 d_{\rg}}\tau \rond \kappa (q_{\rg})]
\quad \text{ and so }\quad
\varepsilon (A) = \k[\poi {\ie q}1{,\ldots,}{\rg}{}{}{}]$$ since for $i=1,\ldots,\rg$, $$\ie q_{i}:=\varepsilon (t^{- 2 d _{i}}\tau \rond \kappa (q_{i})) .$$ Since ${\cal N}\times \{0\}$ is the nullvariety of $t$ and $A_{+}$ in ${\goth g}^{f}\times \k$, ${\cal N}$ is the nullvariety in ${\goth g}^{f}$ of $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$. Hence $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$ since ${\cal N}$ has dimension $r-\rg$.
Proof of Theorem \[ti5\] {#proof}
========================
Let $(e,h,f)$ be an ${\goth {sl}}_{2}$-triple in ${\goth g}$. We use the notations $\kappa $ and $\ie p$, $p\in \ai g{}{}$, as in the introduction. In this section, we use the standard gradings on $\e Sg$ and $\es S{{\goth g}^{e}}$. Let $A_0$ be the subalgebra of $\es S{{\goth g}^{e}}$ generated by the family $\ie p, \; p \in \ai g{}{}$, and let ${\cal N}_0$ be the nullvariety of $A_{0,+}$ in ${\goth g}^{f}$ where $A_{0,+}$ denotes the augmentation ideal of $A_0$.
Let $\poi a1{,\ldots,}{m}{}{}{}$ be a homogeneous sequence in $A_{0,+}$ generating the ideal of $\es S{{\goth g}^{e}}$ generated by $A_{0,+}$. According to [@PPY Corollary 2.3], $A_0$ contains homogeneous elements $\poi b1{,\ldots,}{\rg}{}{}{}$ algebraically independent over $\k$.
\[lproof\] Let ${\goth A}$ be the integral closure of $\k[\poi a1{,\ldots,}{m}{}{}{},\poi b1{,\ldots,}{\rg}{}{}{}]$ in the fraction field of $\es S{{\goth g}^{e}}$.
1. The algebra ${\goth A}$ is contained in $\ai ge{}$ and its fraction field is the fraction field of $\ai ge{}$.
2. Let $a$ in $\ai ge{}_{+}$. If $a$ is equal to $0$ on ${\cal N}_0$, then $a$ is in ${\goth A}_{+}$.
3. The algebra ${\goth A}$ is the integral closure of $A_0$ in the fraction field of $\es S{{\goth g}^{e}}$.
\(i) Let $K_{0}$ be the field of invariant elements under the adjoint action of ${\goth g}^{e}$ in the fraction field of $\es S{{\goth g}^{e}}$. According to [@CM1 Lemma 3.1], $K_{0}$ is the fraction field of $\ai ge{}$. Since $\poi a1{,\ldots,}{m}{}{}{},\poi b1{,\ldots,}{\rg}{}{}{}$ are in $\ai ge{}$, ${\goth A}$ is contained in $K_{0}$. Moreover, ${\goth A}$ is contained in $\ai ge{}$ since $\ai ge{}$ is integrally closed in $K_{0}$. Since $K_{0}$ has transcendence degree $\rg$ over $\k$ and since $\poi b1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$, $K_{0}$ is the fraction field of ${\goth A}$.
\(ii) Since ${\cal N}_0$ is the nullvariety of $\poi a1{,\ldots,}{m}{}{}{},\poi b1{,\ldots,}{\rg}{}{}{}$ in ${\goth g}^{f}$, ${\cal N}_0$ is the nullvariety of ${\goth A}_{+}$ in ${\goth g}^{f}$. Let $a$ be in $\ai ge{}_{+}$ such that $a$ is equal to $0$ on ${\cal N}_0$. Since ${\cal N}_0$ is a cone, all homgogeneous components of $a$ is equal to $0$ on ${\cal N}_0$. So it suffices to prove the assertion for $a$ homogeneous. We have a commutative diagram $$\xymatrix{ {\goth g}^{f} \ar[rr]^{\pi } \ar[rd]_{\alpha } &&
{\mathrm {Specm}}({\goth A}[a]) \ar[ld]^{\beta } \\ & {\mathrm {Specm}}({\goth A}) & }$$ with $\pi $, $\alpha $, $\beta $ the comorphisms of the canonical injections $${\goth A}[a] \hookrightarrow \es S{{\goth g}^{e}}, \quad
{\goth A} \hookrightarrow \es S{{\goth g}^{e}}, \quad
{\goth A} \hookrightarrow {\goth A}[a] .$$ Since ${\cal N}_0$ is the nullvariety of ${\goth A}[a]_{+}$ and ${\goth A}_{+}$ in ${\goth g}^{f}$, $\beta ^{-1}({\goth A}_{+})={\goth A}[a]_{+}$. The gradings of ${\goth A}$ and ${\goth A}[a]$ induce actions of $\gm$ on ${\mathrm {Specm}}({\goth A})$ and ${\mathrm {Specm}}({\goth A}[a])$ such that $\beta $ is equivariant. Moreover, ${\goth A}_{+}$ is in the closure of all orbit under $\gm$ in ${\mathrm {Specm}}({\goth A})$. Hence $\beta $ is a quasi finite morphism. Moreover, $\beta $ is a birational since ${\goth A}$ and ${\goth A}[a]$ have the same fraction field by (i). Hence, by Zariski’s main theorem [@Mu], $\beta $ is an open immersion from ${\mathrm {Specm}}({\goth A}[a])$ into ${\mathrm {Specm}}({\goth A})$. So, $\beta $ is surjective since ${\goth A}_{+}$ is in the image of $\beta $ and since it is in the closure of all $\gm$-orbit in ${\mathrm {Specm}}({\goth A})$. As a result, $\beta $ is an isomorphism and $a$ is in ${\goth A}$, whence the assertion.
\(iii) By (ii), $A_0$ is contained in ${\goth A}$. Moreover, since $\poi a1{,\ldots,}{m}{}{}{},\poi b1{,\ldots,}{\rg}{}{}{}$ are in $A_0$, ${\goth A}$ is contained in the integral closure of $A_0$ in the fraction field of $\es S{{\goth g}^{e}}$, whence the assertion.
\[cproof\] Suppose that the algebra $\ai ge{}$ is finitely generated. Then ${\goth A}$ is equal to $\ai ge{}$.
Let $C$ be the quotient of $\ai ge{}$ by the ideal $\ai ge{}{\goth A}_{+}$. By hypothesis, $C$ is finitely generated. Then it has finitely many minimal prime ideals. Denote them by $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$. For $a$ in the radical of $\ai ge{}{\goth A}_{+}$, $a$ is equal to $0$ on ${\cal N}_0$. Moreover, it is in $\ai ge{}_{+}$. Then, by Lemma \[lproof\](ii), $a$ is in ${\goth A}_{+}$. As a result, $C$ is a reduced algebra and the canonical map $$C \longrightarrow \poi {C/{\goth p}}1{\times \cdots \times }{m}{}{}{}$$ is injective. Since ${\goth A}$ and $\ai ge{}$ have the same fraction field, they have the same Krull dimension. Denote by $d$ this dimension and by ${\goth p}'_{j}$, for $j=1,\ldots,m$, the inverse image of ${\goth p}_{j}$ in $\ai ge{}$ by the quotient map $\ai ge{} \to C$.
\[clproof\] Let $j=1,\ldots,m$. For $i=1,\ldots,d$, there exists a sequence $\poi c1{,\ldots,}{i}{}{}{}$ of elements of ${\goth A}_{+}$ and an increasing sequence $$\{0\} = \poi {{\goth q}}0{ \subsetneq \cdots \subsetneq }{i}{}{}{}
\subset {\goth p}'_{j}$$ of prime ideals of of $\ai ge{}$ such that $c_{i}$ is not ${\goth q}_{i-1}$ and $\poi c1{,\ldots,}{j}{}{}{}$ are in ${\goth q}_{j}$ for $j=1,\ldots,i$.
Prove the claim by induction on $i$. Let $c_{1}$ be in ${\goth A}_{+}\setminus \{0\}$. As ${\goth A}_{+}$ is contained in ${\goth p}'_{j}$, there exists a minimal prime ideal ${\goth q}_{1}$ of $\ai ge{}$, contained in ${\goth p}'_{j}$ and containing $c_{1}$. Suppose $i>1$ and the claim true for $i-1$. As the sequence $$\{0\} = \poi {{\goth A}_{+}\cap {\goth q}}1{\subsetneq \cdots \subsetneq}{i-1}{}{}{}
\subset {\goth A}_{+}$$ is an increasing sequence of prime ideals of ${\goth A}_{+}$ and ${\goth A}_{+}$ has height $d>i-1$, ${\goth A}_{+}$ is not contained in ${\goth q}_{i-1}$. Let $c_{i}$ be in ${\goth A}_{+}\setminus {\goth q}_{i-1}$ and ${\goth q}_{i}$ the minimal prime ideal of $\ai ge{}$ contained in ${\goth p}'_{j}$ and containing $c_{i}$ and ${\goth q}_{i-1}$. So by the induction hypothesis, the sequence $c_1,\ldots,c_i$ satisfies the conditions of the claim. This concludes the proof.
By the claim, ${\goth p}'_{j}$ has height at least $d$ for $j=1,\ldots,m$. Hence $\poi {{\goth p}}1{,\ldots,}{m}{}{}{}$ are maximal ideals of $C$. As a result, the $\k$-algebra $C$ is finite dimensional. Let $V$ be a graded complement to $\ai ge{}{\goth A}_{+}$ in $\es S{{\goth g}^{e}}$. From the equality ${\rm S}(\g^{e}) = V + {\rm S}(\g^{e})^{\g^{e}} {\goth A}_{+}$, we get that ${\rm S}(\g^{e}) = V{\goth A} +
{\rm S}(\g^{e})^{\g^{e}}{\goth A}_{+} ^{m}$ for any nonnegative integer $m$ by induction on $m$. Hence $\ai ge{}=V{\goth A}$ so that $\ai ge{}$ is a finite extension of ${\goth A}$. Since ${\goth A}$ is integrally closed in the fraction field of $\ai ge{}$, ${\goth A}=\ai ge{}$.
The “if” part results from [@CM1 Theorem 1.5] (or, here, Theorem \[ti4\]).
Suppose that $e$ is good. By Definition \[di\] and Theorem \[ti3\], $\ai ge{}$ is a polynomial algebra and the nullvariety of $\ai ge{}_{+}$ in ${\goth g}^{f}$ is equidimensional of dimension $r-\rg$. On the other hand, by Lemma \[lproof\](iii), $\mathfrak{A}$ is the integral closure of $A_0$ in the fraction field of ${\rm S}(\g^{e})$. Hence the nullvarieties of $\mathfrak{A}_+$ and $A_{0,+}$ in $\g^{f}$ are the same. But by Corollary \[cproof\], $\mathfrak{A}=\ai ge{}$, so ${\cal N}_0$ has dimension $r-\rg$ since $e$ is good. On the other hand, $A_0$ is contained in $\varepsilon(A)$ by construction of $\varepsilon(A)$, and $\varepsilon(A)$ is contained in ${\rm S}(\g^{e})^{\g^{e}}$ by [@PPY Proposition 0.1], whence ${\cal N}= {\cal N} _0$.
As a result, ${\cal N}$ has dimension $r-\ell$ and so by Theorem \[ti6\], for some homogeneous generating sequence $\poi q1{,\ldots,}{\rg}{}{}{}$ of $\ai g{}{}$, the element $\poi {\ie q}1{,\ldots,}{\rg}{}{}{}$ are algebraically independent over $\k$.
[DK00]{} N. Bourbaki, [*Algèbre commutative, Chapitre 10, Éléments de mathématiques*]{}, Masson (1998), Paris.
J.-Y. Charbonnel and A. Moreau, [*The index of centralizers of elements of reductive Lie algebras*]{}, Documenta Mathematica, [**15**]{} (2010), 387-421.
J.-Y. Charbonnel and A. Moreau, [*The symmetric invariants of centralizers and Slodowy grading*]{}, Math. Zeitschrift [**282**]{} (2016), [**n$^{\circ}$1-2**]{}, 273-339.
A. Joseph and D. Shafrir, [*Polynomiality of invariants, unimodularity and adapted pairs*]{}, Transformation Groups, [**15**]{}, (2010), 851–882.
H. Matsumura, [*Commutative ring theory*]{} Cambridge studies in advanced mathematics (1986), [**n$^{\circ}$8**]{}, Cambridge University Press, Cambridge, London, New York, New Rochelle, Melbourne, Sydney.
D. Mumford, [*The Red Book of Varieties and Schemes*]{}, Lecture Notes in Mathematics (1988), [**n$^{\circ}$ 1358**]{}, Springer-Verlag, Berlin, Heidelberg, New York, London, Paris, Tokyo.
D.I. Panyushev, A. Premet and O. Yakimova, [*On symmetric invariants of centralizers in reductive Lie algebras*]{}, Journal of Algebra [**313**]{} (2007), 343–391.
O. Yakimova, [*A counterexample to Premet’s and Joseph’s conjecture*]{}, Bulletin of the London Mathematical Society [**39**]{} (2007), 749–754.
O. Yakimova, [*Symmetric invariants of $\mathbb{Z}_2$-contractions and other semi-direct products*]{}, preprint 2016.
[^1]: Since the Lie algebra $\g$ does not appear in this section, there will be no possible confusion between $\rg$ and the rank of $\g$, denoted $\rg$, in the introduction too. However, the notation will be justified in the next sections.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $A$ be a commutative associative integrally closed $\kk$-algebra without zero divisors effectively graded by a lattice. We obtain a criterion of local nilpotency of the sum of two homogeneous locally nilpotent derivations (LNDs) of fiber type on $A$ in terms of their degrees. The same problem is solved for commutators of two homogeneous LNDs.'
address: 'Department of Higher Algebra, Faculty of Mechanics and Mathematics, Moscow State University, Leninskie Gory 1, GSP-1, Moscow, 119991, Russia'
author:
- Elena Romaskevich
title: On sums of homogeneous locally nilpotent derivations
---
Introduction {#introduction .unnumbered}
============
Let $\kk$ be an algebraically closed field of characteristic zero. We consider an algebraic torus $\TT\simeq(\kk^{\times})^n$ acting effectively on a normal affine variety $X$ and the corresponding grading of the algebra $A=\kk[X]$ by the lattice $M$ of characters of $\TT$.
In this paper we study some properties of homogeneous locally nilpotent derivations (LNDs). A derivation ${\partial}$ on $A$ is said to be [*locally nilpotent*]{} if for each $a\in A$ there exists $n\in \ZZ_{>0}$ such that ${\partial}^n(a)=0$. LNDs on $A$ are in one-to-one correspondence with regular actions of the group $\GG_a(\kk)=(\kk,+)$ on $X$, see [@GF]. It is easy to see that a derivation ${\partial}$ on $A$ is homogeneous if and only if the corresponding $\GG_a(\kk)$-action is normalized by the torus $\TT$.
We use a description of homogeneous LNDs on an $M$-graded algebra $A$. Recall that a homogeneous LND on $\kk[X]$ is said to be of [*fiber type*]{} if ${\partial}(\kk(X)^{\TT})=0$, see Definition $\ref{def-fiber/horizontal}$. In geometric terms, ${\partial}$ is of fiber type if and only if generic orbits of the corresponding $\GG_a(\kk)$-action are contained in the closures of $\TT$-orbits. A complete classification of homogeneous LNDs of fiber type is due to A. Liendo, see [@AL-2].
A problem that one faces when dealing with locally nilpotent derivations is that the set of all LNDs on an algebra $A$ admits no obvious algebraic structure. In Epilogue to [@GF] G. Freudenburg poses several natural questions concerning the structure of $\LND(A)$. Namely, given ${\partial}_1,{\partial}_2\in \LND(A)$, under what conditions are $[{\partial}_1,{\partial}_2]$ and ${\partial}_1+{\partial}_2$ locally nilpotent? These questions are still open in general. Some results were otained by M. Ferrero, Y. Lequain, and A. Nowicki in [@MF+YL+AN]. Namely, given two commuting locally nilpotent derivations $d, \delta$ $\in \Der(R)$ and an element $a \in R$, it is proven that the derivation $ad + \delta$ is locally nilpotent if and only if $d(a) = 0$.
In this note we give a complete answer to the above-mentioned questions for homogeneous locally nilpotent derivations of fiber type. Namely, ${\partial}_1+{\partial}_2$ (resp. $[{\partial}_1,{\partial}_2]$) is locally nilpotent if and only if the sum $\deg\,{\partial}_1+\deg\,{\partial}_2$ of their degrees is not in the weight cone $\omega_M(A)$, see Theorem $\ref{main_theorem}$ (resp. Proposition $\ref{commutator}$).
It should be noted that study of algebraic properties of homogeneous locally nilpotent derivations plays an important role in recent works on automorphisms of algebraic varieties, see, e.g. [@IA+AL], [@IA+HF+SK+FK+MZ], [@IA+JH+EH+AL]. One more motivation comes from the question posed by V. Popov, see [@VP Problem 3.1]. He considers two locally nilpotent derivations ${\partial}_1$ and ${\partial}_2\in \LND(\kk[x_1,\dots,x_n])$ and asks when the minimal closed subgroup of $\Aut_{\kk}\kk[x_1,\dots,x_n]$ containing the one dimensional subgroups $\{\exp(t{\partial}_1)\ |\ t\in\kk\}$ and $\{\exp(t{\partial}_2)\ |\ t\in\kk\}$ is finite dimensional. Here it is important to know when ${\partial}_1+{\partial}_2$ and $[{\partial}_1,{\partial}_2]$ are locally finite.
In Section $\ref{sec1}$ we collect some basic definitions and facts about LNDs. In Section $\ref{sec2}$ we recall generalities on $\TT$-varieties and corresponding gradings on their coordinate algebras. Section $\ref{sec3}$ is devoted to background on homogeneous LNDs and to the classification of LNDs of fiber type from [@AL-2]. In Sections $\ref{sec4}$ and $\ref{sec5}$ we study commutators and sums of homogeneous LNDs of fiber type. Section $\ref{sec6}$ contains some corollaries of Theorem $\ref{main_theorem}$, examples and generalizations.
Locally nilpotent derivations {#sec1}
=============================
Let $A$ be a commutative associative $\kk$-algebra without zero divisors. A [*derivation*]{} ${\partial}: A\to A$ is a linear map satisfying the Leibniz rule: $${\partial}(ab)=a{\partial}b+b{\partial}a \quad \mathrm{for\ all}\ a,\,b \in A$$
Denote the set of all derivations on $A$ by $\Der(A)$.
Recall that an algebra is said to be [*graded*]{} by a commutative semigroup $S$ if there is a direct sum decomposition $A=\bigoplus_{s\in S}A_s$ such that $A_s\cdot A_{s'}\subseteq
A_{s+s'}$ for all $s,\ s'\in S$. A derivation ${\partial}$ is called [*homogeneous*]{} if it sends homogeneous elements to homogeneous ones. We will write ${\partial}(s)=s'$, for $s,\ s'\in S$, if $A_s\nsubseteq\Ker({\partial})$ and ${\partial}(A_s)\subseteq A_{s'}$. By Leibniz rule, for $a\in A_s \backslash\Ker({\partial})$ and $b\in A_{s'}\backslash
\Ker({\partial})$ we have $${\partial}(ab)=a{\partial}b+b{\partial}a \in A_{{\partial}(s+s')},$$ and so $${\partial}(s+s')=s+{\partial}(s')=s'+{\partial}(s).$$ Thus, for a homogeneous nonzero derivation ${\partial}$ there exists $s_0\in
S$ such that ${\partial}A_s \subset A_{s+s_0}$ for all $s \in S$. An element $s_0\in S$ is called the [*degree*]{} of ${\partial}$ and is denoted by $\deg\,{\partial}$.
A derivation ${\partial}\in \Der(A)$ is called [*locally nilpotent*]{} (LND for short) if for every $f\in A$ there exists $n\in \ZZ_{>0}$ such that ${\partial}^nf=0$. The set of all locally nilpotent derivations on $A$ is denoted by $\LND(A)$.
We associate to a derivation ${\partial}$ a set $\Nil({\partial})=\{f\in A\ |\
\exists\,n\in\ZZ_{>0}: {\partial}^n f=0\}$. Thus, ${\partial}\in \LND(A)$ exactly when $\Nil({\partial})=A$.
Note that if ${\partial}_1,\ {\partial}_2\in \LND(A)$ and $[{\partial}_1,{\partial}_2]=0$, then ${\partial}_1+{\partial}_2\in\LND(A)$, i.e. the sum of commuting LNDs is an LND as well.
Recall that locally nilpotent derivations on an affine algebra $A$ are in one-to-one correspondence with regular actions of $\GG_a(\kk)$ on $X=\Spec\,A$. Indeed, we have a rational representation $\eta:\,\GG_a(\kk)\hookrightarrow \Aut_{\kk}(A)$, where $\eta(t)=\exp(tD)$. In geometric terms this means that $D$ induces a regular $\GG_a(\kk)$-action on $X$. Conversely, let $\rho:\,\GG_a(\kk)\times X\to X$ be a regular $\GG_a(\kk)$-action. Then $\rho$ induces a locally nilpotent derivation $D=\frac{d}{dt}|_{t=0}\rho^*$, where $\rho^*:\, A\to A[t]$. For more detail see [@GF Section 1.5].
$\TT$-varieties {#sec2}
===============
An [*algebraic torus*]{} $\TT=\TT^n$ of dimension $n$ is the algebraic variety $(\kk^{\times})^n$ with the natural structure of algebraic group. A [*$\TT$-variety*]{} is an algebraic variety endowed with an effective $\TT$-action.
A [*character*]{} (resp. [*one-parameter subgroup*]{}) of $\TT$ is a homomorphism of algebraic groups $\chi:\,\TT\to \kk^{\times}$ (resp. $\lambda:\,~ \kk^{\times}\to\TT$). The set of all characters (resp. one-parameter subgroups) form a lattice $M$ (resp. $N$) of rank $n$. For every $m\in M$ we denote by $\chi^m$ the corresponding character of $\TT$. We also let $M_{\QQ}$ and $N_{\QQ}$ be the rational vector spaces $M\otimes_{\ZZ} \QQ$ and $N\otimes_{\ZZ} \QQ$.
Let $A$ be an [*affine*]{} algebra, i.e. a commutative associative finitely generated $\kk$-algebra with unit and without zero divisors. It is well known that effective $\TT$-actions on an affine variety $X=\Spec\,A$ are in one-to-one correspondence with effective $M$-gradings on $A$. Thus, for a $\TT$-variety $X$ we have an effective $M$-grading on $A=\kk[X]$: $$A=\bigoplus_{m\in M}\widetilde{A_m}.$$ Let $K=\Frac\,A$ be the field of fractions. We consider $$K_0=\left\{\frac{f}{g}\in K\ |\ \deg f=\deg g\right\}.$$ Notice that $K_0$ coincides with the field of $\TT$-invariant functions $\kk(X)^{\TT}$. Thus, we have a tower of field extensions $\kk\subseteq \kk(X)^{\TT} \subseteq K$. One may represent $\widetilde{A_m}=A_m\chi^m$, where $A_m\subseteq \kk(X)^{\TT}$. The [*weight cone*]{} $\omega \subseteq M_{\QQ}$ of a given $M$-grading is a cone in $M_{\QQ}$ spanned by the set $\{m\in M\ |\ A_m\ne 0\}$. For a cone $\omega \subseteq M_{\QQ}$ we denote the set $\omega\cap M$ by $\omega_M$. Finally, we have $$A=\bigoplus_{m\in \omega_M} A_m \chi^m, \quad A_m\subseteq
\kk(X)^{\TT}.$$ Since $A$ is finitely generated, the cone $\omega$ is polyhedral, and since the $M$-grading is effective, $\omega$ is of full dimension.
[*Complexity*]{} of a $\TT$-action is the transcendence degree of the field of $\TT$-invariant rational functions $\kk(X)^{\TT}$ over $\kk$. In geometric terms, complexity of a $\TT$-action equals the codimension of the generic $\TT$-orbit. In particular, for a $\TT$-variety of complexity zero, $\kk(X)^{\TT}=\kk$ and $A\subseteq
\kk[M]$, where $\kk[M]$ stays for the group algebra of the lattice $M$ and is isomorphic to the algebra of Laurent polynomials over $\kk$. A [*toric variety*]{} is a normal $\TT$-variety of complexity zero, or, equivalently, $\TT$ acts with an open orbit. From now on we will consider only normal $\TT$-varieties.
Demazure roots and homogeneous LNDs {#sec3}
===================================
Let $M$ and $N$ be the lattices of characters and one-parameter subgroups of a torus $\TT$. We consider the natural pairing $\<\cdot,\cdot\>: M\times N\to \ZZ$ given by $$\<\chi,\lambda\>=l, \quad {\rm if} \quad \chi\circ\lambda(t)=t^l.$$ This pairing extends in an obvious way to a pairing $$\<\cdot,\cdot\>: M_{\QQ}\times N_{\QQ}\to \QQ$$ between $\QQ$-vector spaces. Let $A$ be as before an $M$-graded affine algebra with the weight cone $\omega\subseteq M_{\QQ}$. Let $\omega^{\vee}=\sigma\subseteq N_{\QQ}$ be the dual cone. Since $\omega$ is of full dimension, $\sigma$ is a pointed polyhedral cone.
[@AL-1 Lemma 1.13] For any homogeneous LND ${\partial}$ on $A$ the following holds.
- The derivation ${\partial}$ extends in a unique way to a homogeneous $\kk$-derivation on $\kk(X)^{\TT}[M]$.
- If ${\partial}(\kk(X)^{\TT})=0$ then the extension of ${\partial}$ as in 1) restricts to a homogeneous locally nilpotent $\kk(X)^{\TT}$-derivation on $\kk(X)^{\TT}[\omega_M]$.
\[def-fiber/horizontal\] A homogeneous LND ${\partial}$ on $A$ is said to be of [*fiber type*]{} if ${\partial}(\kk(X)^{\TT})=0$ and of [*horizontal type*]{} otherwise. Derivations of fiber type correspond to $\GG_a(\kk)$-actions such that generic $\GG_a(\kk)$-orbits are contained in the closures of $\TT$-orbits.
In [@AL-2] A.Liendo gave a complete classification of homogeneous LNDs of fiber type on an arbitrary graded integrally closed affine algebra. We need the following notation to present his results.
For a ray $\rho$ of a cone $\sigma$ we let $\sigma_{\rho}$ denote the cone spanned by all the rays of $\sigma$ except $\rho$. From now on we denote by $\rho$ both a ray and its primitive vector. We also let $$S_{\rho}=\sigma_{\rho}^{\vee}\cap\{e\in M\ |\ \<e,\rho\>=-1\}.$$
In other words, $S_{\rho}$ is the set of lattice vectors $e\in M$ such that $\<e,\rho\>=-1$ and $\<e,\rho^{'}\>\geq 0$ for every other ray $\rho^{'} \subseteq \sigma$.
(see [@MD]) In the above notation the elements of the set $\Re=\bigcup_{\rho\subseteq\sigma} S_{\rho}$ are called [*Demazure roots*]{} of the cone $\sigma$.
Let $e\in S_{\rho}$ be a Demazure root corresponding to a ray $\rho$ of the cone $\sigma$. Set $$\Phi_e=\{f\in K\ |\ f\cdot A_m\subseteq A_{m+e}\},\ \Phi_e^{\times}=\Phi_e\backslash\{0\}.$$ We define a homogeneous derivation ${\partial}_{\rho,e}$ of degree $e$ on $\kk(X)^{\TT}[M]$ by $${\partial}_{\rho,e}(\chi^m)=\<m,\rho\>\chi^{m+e}.$$ Clearly, for $f\in\Phi_e$ the product $f{\partial}_{\rho,e}$ is a derivation on $\kk(X)^{\TT}[\omega_M]$. Finally, denote ${\partial}_{\rho,e,f}=f{\partial}_{\rho,e}|_A$, which is a homogeneous LND on $A$. The classification theorem is as follows.
\[HLND\][@AL-2 Theorem 2.4] Every nonzero homogeneous LND ${\partial}$ of fiber type on $A$ is of the form ${\partial}={\partial}_{\rho,e,f}$ for some ray $\rho\subseteq\sigma$, some Demazure root $e\in S_{\rho}$ and some function $f\in\Phi_e^{\times}$.
In particular, for any homogeneous locally nilpotent derivation ${\partial}$ of fiber type we have $\deg {\partial}\notin\omega_M$.
For a toric variety $X$ the closure of a dense $\TT$-orbit coincides with $X$, hence all derivations on $A$ are of fiber type. In this particular case Theorem $\ref{HLND}$ states that all nonzero homogeneous LNDs on $A=\kk[X]$ are of the form ${\partial}=\lambda{\partial}_{\rho,e}$, for some $\lambda\in \kk^{\times}$.
\[ex-1\] With $N=\ZZ^n$ we let $\sigma$ be the cone in $N_{\QQ}$ spanned by the basic unit vectors $e_i$. The dual cone $\omega\subseteq M_{\QQ}$ is spanned by the basic unit vectors in $M_{\QQ}$ as well. The corresponding semigroup algebra is a polynomial algebra $A=\kk[x_1,\dots,x_n],\ \deg x_i=e_i$, and the affine toric variety is $X=\AA^n$. According to Theorem $\ref{HLND}$ all nonzero homogeneous LNDs on $A$ are: $${\partial}=\lambda x_1^{i_1}\dots \widehat{x_k^{i_k}}\dots
x_n^{i_n}\frac{{\partial}}{{\partial}x_k}, \quad \lambda\in \kk^{\times},\ i_1,\dots,i_n\in
\ZZ_{\geqslant 0} \quad (k=1,\dots,n).$$
Degree of the preceding derivation is $(i_1,\dots,-1,\dots,i_n)$, which is a Demazure root of $\sigma$ corresponding to the ray $\rho_k$.
Commutators of two homogeneous LNDs {#sec4}
===================================
Suppose ${\partial}_1,{\partial}_2\in \LND(A)$ are nonzero homogeneous derivations of fiber type. It follows from Theorem $\ref{HLND}$ that they are given by $${\partial}_1(\chi^m)=f_1\<m,\rho_1\>\chi^{m+e_1},\qquad
{\partial}_2(\chi^m)=f_2\<m,\rho_2\>\chi^{m+e_2},$$ where $e_1,e_2\in\Re$ and $f_1\in \Phi_{e_1}^{\times}$, $f_2\in \Phi_{e_2}^{\times}$. According to the definition, we have $$\begin{gathered}
[{\partial}_1,{\partial}_2](\chi^m)=f_1f_2(\<m+e_2,\rho_1\>\<m,\rho_2\>-\<m+e_1,\rho_2\>\<m,\rho_1\>)\chi^{m+e_1+e_2}=\\
=f_1f_2(\<e_2,\rho_1\>\<m,\rho_2\>-\<e_1,\rho_2\>\<m,\rho_1\>)\chi^{m+e_1+e_2}.\end{gathered}$$
\[sameray\] If ${\partial}_1,{\partial}_2\in \LND(A)$ correspond to the same ray of the cone $\sigma$, then ${\partial}_1$ and ${\partial}_2$ commute.
Let $\rho_1=\rho_2=\rho$. Then $\<e_2,\rho\>\<m,\rho\>-\<e_1,\rho\>\<m,\rho\>=(-1)\<m,\rho\>-(-1)\<m,\rho\>=0$, and hence the commutator equals zero.
\[commutator\] Suppose ${\partial}_1,{\partial}_2\in \LND(A)$ correspond to different rays $\rho_1,\rho_2$ of the cone $\sigma$. Then the following holds:
- $[{\partial}_1,{\partial}_2]=0 \Leftrightarrow \<e_1,\rho_2\>=0$ and $\<e_2,\rho_1\>=0$,
- $[{\partial}_1,{\partial}_2]\in \LND(A)\Leftrightarrow\ e_1+e_2\notin\omega_M$, i.e. $\<e_1,\rho_2\>=0$ or $\<e_2,\rho_1\>=0$.
1\) Let $[{\partial}_1,{\partial}_2]=0$. Then for all $m\in\omega_M$ we have $\<e_2,\rho_1\>\<m,\rho_2\>=\<e_1,\rho_2\>\<m,\rho_1\>$. One can choose $m\in \omega_M$ such that $\<m,\rho_1\>=0$ and $\<m,\rho_2\>>0$. In this case $\<e_2,\rho_1\>\<m,\rho_2\>=0$ implies that $\<e_2,\rho_1\>=0$. Similarly $\<e_1,\rho_2\>=0$. The inverse implication holds automatically.
2\) Let us prove sufficiency. Without loss of generality assume $\<e_1,\rho_2\>=0,\ \<e_2,\rho_1\>>0$. Then $\<e_1+e_2,\rho_2\>=-1,\
\<e_1+e_2,\rho_1\>\geqslant 0$. Therefore $e_1+e_2$ is a Demazure root corresponding to the ray $\rho_2$ of the cone $\sigma$ and $[{\partial}_1,{\partial}_2]=f {\partial}_{\rho_2,e_1+e_2}$, where $f=f_1f_2\<e_2,\rho_1\>\in \Phi_{e_1+e_2}^{\times}$.
To prove necessity assume that the commutator is an LND. Since its degree equals $e_1+e_2$, there exists a ray $\rho^*$ of the cone $\sigma$ such that $\<e_1+e_2,\rho^*\>=-1$. As $e_1,e_2\in
\sigma_{\rho}^{\vee}$ for $\rho\ne\rho_1, \rho_2$, one obtains $\rho^*=\rho_1$ or $\rho_2$. This proves the assertion.
\[deg\_notin\_cone\] In general, if ${\partial}$ is a homogeneous derivation of a graded algebra $A=\bigoplus_{m\in \omega_M}\widetilde{A_m}$, and $\deg{\partial}\notin\omega_M$, then ${\partial}$ is locally nilpotent. Indeed, for any $m\in\omega_M$ one can find $k\in\ZZ_{\geqslant 1}$ such that $(m+k\deg{\partial})\notin\omega_M$ and, thus, ${\partial}^k(\widetilde{A_m})=0$.
Sums of two homogeneous LNDs {#sec5}
============================
In this section we establish a necessary and sufficient condition of local nilpotency of the sum of two homogeneous LNDs. Consider two nonzero homogeneous LNDs ${\partial}_1, {\partial}_2$ of fiber type: $${\partial}_1(\chi^m)=f_1\<m,\rho_1\>\chi^{m+e_1},\qquad
{\partial}_2(\chi^m)=f_2\<m,\rho_2\>\chi^{m+e_2},$$ where $e_1,e_2\in\Re$ and $f_1\in \Phi_{e_1}^{\times}$, $f_2\in \Phi_{e_2}^{\times}$. Note that their sum is homogeneous if and only if both ${\partial}_1$ and ${\partial}_2$ are of the same degree, i.e. $e_1=e_2$.
\[main\_theorem\] In the above notation let ${\partial}_1$ and ${\partial}_2$ be homogeneous LNDs of fiber type. Then ${\partial}_1+{\partial}_2\in \LND(A)$ if and only if $e_1+e_2\notin\omega_M$.
Let us prove sufficiency. The condition $e_1+e_2\notin\omega_M$ means that either both ${\partial}_1$ and ${\partial}_2$ correspond to the same ray, or $\<e_2,\rho_1\>=0$ (up to permutation of indices). In the first case according to Lemma $\ref{sameray}$ our derivations commute and hence their sum is an LND.
Let us consider the second case. It suffices to show that for any $m\in \omega_M$ there exists $n\in\mathbb{Z}_{\geqslant 0}$ such that $({\partial}_1+{\partial}_2)^n(\chi^m)=0$. We proceed by induction on the parameter $\<m,\rho_1\>$.
Let $\<m,\rho_1\>=0$. This means that $m\in\rho_1^\perp\cap\omega_M$ and $({\partial}_1+{\partial}_2)(\chi^m)={\partial}_2(\chi^m)$. In addition, the equality $\<e_2,\rho_1\>=0$ implies that $m+e_2\in\rho_1^\perp\cap\omega_M$. Continuing these arguments and using local nilpotency of ${\partial}_2$, we obtain the required condition.
Now consider an arbitrary point $m\in\omega_M$. Since ${\partial}_2\in \LND(A)$, there exists $n\in\mathbb{Z}_{\geqslant 0}$ such that ${\partial}_2^n(\chi^m)=0$. Therefore all the images of $m$ under powers of ${\partial}_2$, namely $m,m+e_2,\dots,m+(n-1)e_2$, lie in the hyperplane $H$ perpendicular to $\rho_1$ and given by the equation $\<x,\rho_1\>=\<m,\rho_1\>$. Now we apply $({\partial}_1+{\partial}_2)^n$ to $\chi^m$. One can easily see that applying ${\partial}_1$ to an arbitrary element of the hyperplane $H$, the image will be in the hyperplane given by the equation $\<x,\rho_1\>=\<m,\rho_1\>-1$. Since each summand (except for ${\partial}_2^n$) in $({\partial}_1+{\partial}_2)^n$ contains ${\partial}_1$ and ${\partial}_2^n(\chi^m)=0$ and using the inductive hypothesis, we obtain local nilpotency of the sum ${\partial}_1+{\partial}_2$.
Now let us prove the inverse implication. If $\mathrm{rk}\,M=1$, then $\omega_M=\ZZ$ or $\ZZ_{\geqslant 0}$ and the dual cone $\sigma$ is $0$ or $\QQ_{\geqslant 0}$ respectively. In the first case $\sigma$ contains no ray, thus there are no derivations of fiber type on $\kk[X]$. In the second case $\sigma$ consists of one ray and the condition $e_1+e_2\notin \omega_M$ follows immediately.
The case $\mathrm{rk}\,M=2$ follows from results of P. Kotenkova, see [@PK]. We recall some notation and results from this article. Consider a one-dimensional subtorus $T\subset\TT$ given by the equation $e_1-e_2=0$. The torus $T$ also acts on $X$ and every $\TT$-homogeneous LND on $\kk[X]$ is $T$-homogeneous as well. Recall that a $\TT$-root of an $M$-graded algebra $A$ is the degree of some homogeneous LND on $A$. Any $\TT$-root of $A=\kk[X]$ can be restricted to some $T$-root and we denote by $\pi$ the restriction map. Denote by $\Gamma_T\subseteq N_{\QQ}$ the hyperplane, corresponding to the subtorus $T$. Let $\<\cdot,m_T\>=0$ be the equation of the hyperplane $\Gamma_T$, where $m_T \in M$. Evidently, the locally nilpotent derivation ${\partial}_1+{\partial}_2$ restricts to a homogeneous LND with respect to the torus $T$. Our case corresponds to Case 3.3 in [@PK Proposition 6]. It follows that all $T$-homogeneous LNDs of degree $\pi(e_1)$ are of the following form (see [@PK Proposition 5]):
$${\partial}(\chi^m)=\chi^{m+e_2}(\alpha\<\rho_1,m\>\chi^{m_T}+\beta\<\rho_2,m\>)(\alpha\chi^{m_T}+\beta\<\rho_2,m_T\>)^{\<\rho_1,e_2\>},$$ for some $\alpha, \beta\in \kk$. Now one can easily see that for ${\partial}={\partial}_1+{\partial}_2$ the exponent $\<\rho_1,e_2\>$ has to vanish, so $\<\rho_1,e_1+e_2\>=-1$ and $e_1+e_2\notin\omega_M$.
Let us consider the general case $\mathrm{rk}M=n$.
\[extension to alg.closure\] Denote by $\bar{k}$ the algebraic closure of a field $k$. The derivation ${\partial}$ on a $k$-algebra $A$ extends in a unique way to a derivation $\tilde{{\partial}}$ on $A\otimes_k \bar{k}$. In addition, ${\partial}$ is locally nilpotent if and only if $\tilde{{\partial}}$ is.
To prove the first assertion of the lemma we let $x\in\bar{k}$ and assume that $p(t)$ is the minimal polynomial of $x$. Applying $\tilde{{\partial}}$ to the equation $p(x)=0$, we obtain $\tilde{{\partial}}(x)p'(x)=0$, hence $\tilde{{\partial}}(x)=0$. Further, an extension of ${\partial}$ to $A\otimes_k \bar{k}$ by linearity is unique. Since $\bar{k}\subset\Ker(\tilde{{\partial}})$, local nilpotency of $\tilde{{\partial}}$ and ${\partial}$ are equivalent.
According to the previous lemma we can replace $\kk(X)^{\TT}$ with its algebraic closure $\overline{\kk(X)^{\TT}}$.
It is required to show that $e_1+e_2\notin \omega_M$. If the roots $e_1,e_2$ correspond to the same ray, the condition holds automatically. If $e_1,e_2$ correspond to different rays $\rho_1$ and $\rho_2$, the required condition is equivalent to the following one: at least one of expressions $\<e_1,\rho_2\>$ or $\<e_2,\rho_1\>$ vanishes. We carry out the proof by contradiction.
We assume that the vectors $e_1$ and $e_2$ are not collinear. In this case for any $m\in \omega_M$ we denote by $\gamma_m$ a two-dimensional plane passing through $m$ and spanned by vectors $e_1,e_2$. It is obvious that images of $m$ under powers of the derivation ${\partial}_1+{\partial}_2$ are in $\omega_M\cap\gamma_m$. It follows from the formulae defining ${\partial}_1$ and ${\partial}_2$ that $\gamma_m\cap\rho_1^\bot\ne\varnothing$ and $\gamma_m\cap\rho_2^\bot\ne\varnothing$. If $\gamma_m\subset\rho_1^\bot$ then $\<e_2,\rho_1\>=0$ and similarly for another permutation of the indices. Otherwise $\mathrm{dim}\,(\gamma_m\cap\rho_1^\bot)=\mathrm{dim}\,(\gamma_m\cap\rho_2^\bot)=1$.
Now we assume that $m=e_1+e_2$. Denote the lines in which the plane $\gamma_m$ intersects $\rho_1^{\perp}$ and $\rho_2^{\perp}$ by $l_1$ and $l_2$ respectively. It follows from local nilpotency of ${\partial}_1,\ {\partial}_2$ that there exist $k_1,k_2\in\ZZ_{\geqslant 0}$ such that $$m_1=e_1+e_2+k_1e_1\in\rho_1^{\perp}\cap\gamma_m, \qquad
m_2=e_1+e_2+k_2e_2\in\rho_2^{\perp}\cap\gamma_m.$$ Obviously, $0\in\rho_1^{\perp}\cap\rho_2^{\perp}\cap\gamma_m$. Thus, $l_1$ and $l_2$ are precisely the lines passing through $0$ and $m_1$ or $m_2$ respectively. If $l_1$ and $l_2$ are different, we set $\omega_m=\gamma_m\cap\omega$. The two-dimensional cone $\sigma_m$ dual to $\omega_m$ can be considered as embedded in $N_{\QQ}$ and spanned by the rays $\rho_1,\ \rho_2$. One can easily see that $e_1$ and $e_2$ are Demazure roots of $\sigma_m$. A locally nilpotent derivation on $\overline{\kk(X)^{\TT}}[\omega_M]$ can be restricted to an LND on $\overline{\kk(X)^{\TT}}[\omega_m\cap M]$, hence using the assertion of the theorem in case of dimension two we obtain $e_1+e_2\notin \omega_m\cap
M$, a contradiction.
If the lines $l_1$ and $l_2$ coincide, then the vectors $e_1+e_2+k_1e_1$ and $e_1+e_2+k_2e_2$ are collinear. This means the following: $$\frac{k_1+1}{1}=\frac{1}{k_2+1} \Leftrightarrow (k_1+1)(k_2+1)=1
\Leftrightarrow k_1=k_2=0.$$ Thus, $e_1+e_2\in \rho_1^{\perp}\cap\rho_2^{\perp}$ and $\gamma_m\cap\omega$ is the line spanned by $e_1+e_2$. Indeed, let $\<ae_1+be_2,\rho_1\>\geqslant 0$. Using that $\<e_1,\rho_1\>=~-1,\
\<e_2,\rho_1\>=1$ one obtains $b\geqslant a$. Similarly we obtain $a\geqslant b$ and, hence, $a=b$. Consider now an arbitrary element $s\in\omega_M^0=\mathrm{Int}(\omega_M)$ and the corresponding plane $\gamma_s$. Since $\gamma_s
\parallel \gamma_m$, the plane $\gamma_s$ intersects both $\rho_1^{\perp}$ and $\rho_2^{\perp}$ in parallel lines $v_1$ and $v_2$ with leading vector $e_1+e_2$. Our aim now is to construct $f\in
\overline{\kk(X)^{\TT}}[\omega_M]$ such that $f\notin \Nil({\partial}_1+{\partial}_2)$.
\[fig-stripe\]
Intersection of $\gamma_s$ and $\omega_M$
Consider a two-dimensional plane containing the lines $v_1$ and $v_2$, and its intersection with the cone $\omega_M$. We obtain a stripe-shaped diagram, see Figure 1.
One can also consider a lattice $S$ passing through $s$ with generating vectors $e_1,e_2$. Since $\<e_i,\rho_j\>=(-1)^{i+j-1}\
(i,j=1,2)$, the derivations ${\partial}_1,{\partial}_2$ send elements of the lattice $S$ one level up or down respectively. Here levels are of the form $\{x\in S\ |\ \<x,\rho_2\>=\alpha\},\ \alpha=0,\dots,k$. In particular, 0-level lies on $v_2$ and $k$-level lies on $v_1$. Consider elements $m_1,\dots,m_r\in S$ as shown on Figure 1. Note that Figure 1 represents the case of even $k$, for odd $k$ arguments are similar. Let us prove that there exist coefficients $a_1,\dots,a_r\in \overline{\kk(X)^{\TT}}$ such that $$a_1\chi^{m_1}+\dots+a_r\chi^{m_r}\notin \Nil({\partial}_1+{\partial}_2).$$ Indeed, we apply $({\partial}_1+{\partial}_2)^2$ to an element with yet undetermined coefficients and write the column vector of coefficients at $\chi^{m_1+e_1+e_2},\dots,
\chi^{m_r+e_1+e_2}$ after decomposition into graded components:
$$\small
\begin{pmatrix}
a_1 \\ a_2 \\ a_3 \\ \vdots \\ a_{r-1} \\ a_r
\end{pmatrix}
\rightsquigarrow
\begin{pmatrix}
a_1\cdot f_1f_2\cdot k\cdot 1 + a_2\cdot f_1^2\cdot 2\cdot 1 \\
a_2\cdot f_1f_2\cdot ((k-2)\cdot 3 +2\cdot(k-1))+ a_3\cdot f_1^2\cdot 4\cdot 3+a_1\cdot f_2^2\cdot k\cdot (k-1)\\
a_3\cdot f_1f_2\cdot ((k-4)\cdot 5 +4\cdot(k-3))+ a_4\cdot f_1^2\cdot 6\cdot 5+a_2\cdot f_2^2\cdot (k-2)\cdot (k-3)\\
\vdots\\
a_{r-1}\cdot f_1f_2\cdot (2\cdot (k-1)+(k-2)\cdot 3)+ a_r\cdot f_1^2\cdot k\cdot (k-1)+a_{r-2}\cdot f_2^2\cdot 4\cdot 3\\
a_r\cdot f_1f_2\cdot k \cdot 1 + a_{r-1}\cdot f_2^2\cdot 2\cdot 1\\
\end{pmatrix}$$
We try to find such values $a_1,\dots,a_r$, that these two vectors are proportional. In other words, $(a_1,\dots,a_r)^T$ has to be an eigenvector of the following three-diagonal matrix $A$:
$$\small
\begin{pmatrix}
f_1f_2\cdot k\cdot 1 & * & 0 & 0 & \hdotsfor{1} & 0 \\
* & f_1f_2\cdot ((k-2)\cdot 3 +2\cdot(k-1)) & * & 0 & \hdotsfor{1} & 0 \\
0 & \hdotsfor{1} & \ddots & * & \hdotsfor{1} & \vdots \\
\vdots & \hdotsfor{1} & * & \ddots & \hdotsfor{1} & 0\\
0 & \hdotsfor{1} & 0 & * & f_1f_2\cdot (2\cdot (k-1)+(k-2)\cdot 3) & * \\
0 & \hdotsfor{1} & 0 & 0 & * & f_1f_2\cdot k \cdot 1
\end{pmatrix}$$
Therefore, it suffices to show that the matrix $A$ possesses a nonzero eigenvalue $\lambda$, because otherwise applying even powers of derivation ${\partial}_1+{\partial}_2$ we obtain the following sequence: $$\sum_{i=1}^r a_i\chi^{m_i}\longrightarrow \lambda\left(\sum_{i=1}^r
a_i\chi^{m_i+e_1+e_2}\right)\longrightarrow
\lambda^2\left(\sum_{i=1}^r
a_i\chi^{m_i+2(e_1+e_2)}\right)\longrightarrow \dots,$$ whose members do not vanish and thus $\sum_{i=1}^r
a_i\chi^{m_i}\notin\Nil({\partial}_1+{\partial}_2)$. We write the characteristic polynomial of $A$: $$\chi(\lambda)=\mathrm{det}(\lambda E -
A)=\lambda^r-\mathrm{tr}A\lambda^{r-1}+\dots+(-1)^r\mathrm{det}A$$ and compute the coefficient at $\lambda^{r-1}$. One can easily see that $\mathrm{tr}A$ is a product of two factors, where the first factor is $f_1f_2$ and the second one is a sum of positive integers. More precisely,
$$\begin{gathered}
\mathrm{tr}A=f_1f_2\sum_{j=1}^k
j(k+1-j)=f_1f_2\left((k+1)\frac{k(k+1)}{2}-\frac{k(k+1)(2k+1)}{6}\right)=\\
=f_1f_2\frac{k(k+1)(k+2)}{6}\ne0.\end{gathered}$$
Hence, $\mathrm{tr}A$ does not vanish and $\chi(\lambda)\ne
\lambda^r$. Therefore we obtain existence of a nonzero $\lambda\in\overline{\kk(X)^{\TT}}$ with $\chi(\lambda)=0$.
Now let $e_1$ and $e_2$ be collinear, i.e. $e_1=te_2,\ t\in \QQ^*$. Combining the conditions $$\<e_1,\rho_2\>=\<te_2,\rho_2\>=-t \Rightarrow t\in\mathbb{Z}_{<0}$$ and $$\<e_2,\rho_1\>=\left\langle\frac{e_1}{t},\rho_1\right\rangle=-\frac1t
\Rightarrow \frac1t\in\mathbb{Z}_{<0},$$ we see that $t=-1$ and $e_1=-e_2$. The construction from the previous case allows us to obtain an element of the semigroup algebra $\overline{\kk(X)^{\TT}}[\omega_M]$ that does not belong to $\Nil({\partial}_1+{\partial}_2)$. The proof is completed.
Concluding remarks {#sec6}
==================
The next corollary follows directly from Lemma $\ref{sameray}$, Proposition $\ref{commutator}$ and Theorem $\ref{main_theorem}$.
Let ${\partial}_1$ and ${\partial}_2$ be two nonzero homogeneous LNDs on $A$ of fiber type. Then ${\partial}_1+{\partial}_2$ is a locally nilpotent derivation if and only if $[{\partial}_1,{\partial}_2]$ is.
\[polynom\] Let us illustrate the obtained results for $A=\kk[x_1,\dots,x_n]$ in the settings of Example $\ref{ex-1}$. Using for simplicity multiindices ($x^I=x_1^{i_1}\dots x_n^{i_n}$), we set $${\partial}_1=\lambda_1 x^I \frac{{\partial}}{{\partial}x_i}, \qquad {\partial}_2=\lambda_2 x^J
\frac{{\partial}}{{\partial}x_j},$$ where $x^I$ and $x^J$ do not contain factors $x_i$ and $x_j$ respectively. Then the following conditions hold:
- $[{\partial}_1,{\partial}_2]=0 \Leftrightarrow$ $x^I$ does not depend on $x_j$ and $x^J$ does not depend on $x_i$;
- $[{\partial}_1,{\partial}_2]\in \LND(A)\Leftrightarrow {\partial}_1+{\partial}_2\in \LND(A) \Leftrightarrow$ $x^I$ does not depend on $x_j$ or $x^J$ does not depend on $x_i$.
Note that in the settings of Example $\ref{polynom}$ Theorem \[main\_theorem\] can be easily obtained from the following proposition.
\[principle\][@GF Principle 5] Let ${\partial}\in \LND(A)$ and $f_1,\dots,f_m\in A$ ($m\geqslant 1$). Suppose there exists a permutation $\sigma\in S_m$ such that ${\partial}f_i\in f_{\sigma(i)}A$ for each $i$. Then in each orbit of $\sigma$ there is an index $i$ with ${\partial}f_i=0$.
Indeed, suppose ${\partial}={\partial}_1+{\partial}_2\in \LND(A)$. We take $f_1=x_i$, $f_2=x_j$. Then ${\partial}f_1=\lambda_1 x^I\ne 0$ and ${\partial}f_2=\lambda_2 x^J\ne 0$. Using Proposition $\ref{principle}$ we obtain the required condition. The converse is immediate.
Let ${\partial}_1$ and ${\partial}_2$ be two nonzero homogeneous LNDs on $A$ of fiber type. Suppose $e_1+e_2\notin\omega_M$. Then the Lie algebra $L$ generated by ${\partial}_1$ and ${\partial}_2$ over $\kk$ is finite dimensional and consists of locally nilpotent derivations.
We can assume without loss of generality that $\<e_1,\rho_2\>=0$ and $\<e_2,\rho_1\>=n\geqslant 0$. Then the algebra $L$ is linearly generated by the derivations ${\partial}_1,{\partial}_2$ and ${\partial}_2^{(1)},\dots,{\partial}_2^{(n)}$, where ${\partial}_2^{(i)}=\ad({\partial}_1)^i{\partial}_2$. So the dimension of $L$ equals to $n+2$. Note also that ${\partial}_2^{(i)},\ i=1,\dots,n$, are locally nilpotent derivations of degree $e_2+ie_1$ respectively, which are Demazure roots corresponding to the ray $\rho_2$. Let us assume that a derivation ${\partial}$ in $L$ is a linear combination of ${\partial}_2,{\partial}_2^{(1)},\dots,{\partial}_2^{(n)}$. Then ${\partial}$ is locally nilpotent because all these LNDs commute. If ${\partial}={\partial}_1+\lambda_0{\partial}_2+\lambda_1{\partial}_2^{(1)}+\dots+\lambda_n{\partial}_2^{(n)}$, the proof is similar to the proof of sufficiency in Theorem $\ref{main_theorem}$.
At the moment the author does not know whether the condition $\deg\,{\partial}_1+\deg\,{\partial}_2\notin\omega_M$ implies that ${\partial}_1+{\partial}_2\in\LND(A)$ for locally nilpotent derivations ${\partial}_1,\ {\partial}_2$ of horizontal type. Let us give a particular result in this direction.
Consider an effectively graded affine algebra $$A=\bigoplus_{m\in \omega_M}\widetilde{A_m}.$$ Let ${\partial}_1$ and ${\partial}_2$ be two locally nilpotent derivations on $A$ of degrees $e_1, e_2\in M$ respectively. If $\{te_1+(1-t)e_2,\ t\in\QQ\}\cap\omega_M=\varnothing$, then ${\partial}_1+{\partial}_2\in \LND(A)$.
Let us consider the grading of $A$ given by the quotient group $M'=M/(e_1-e_2)$, and denote by $\pi: M\to M'$ the factorization map. Derivations ${\partial}_1$ and ${\partial}_2$ considered as derivations on $M'$-graded algebra are homogeneous of the same degree $\pi(e_1)=\pi(e_2)$. Hence, the sum ${\partial}_1+{\partial}_2$ is $M'$-homogeneous of degree $\pi(e_1)$ as well.
The condition $\{te_1+(1-t)e_2,\ t\in\QQ\}\cap\omega_M=\varnothing$ means that the line passing through the lattice points $e_1,e_2\in M$ does not intersect the cone $\omega_M$. This implies that $\pi(e_1)\notin\pi(\omega_M)$.
Summarizing these facts, we see that ${\partial}$ is a homogeneous derivation on $M'$-graded algebra $A$ and its degree $\pi(e_1)$ does not lie in the weight cone $\pi(\omega_M)$. By Remark $\ref{deg_notin_cone}$ this is sufficient for the local nilpotency of ${\partial}$.
Unlike the case of derivations of fiber type another implication of Theorem $\ref{main_theorem}$ does not hold for derivations of horizontal type.
Consider a $\TT$-variety $\AA^2$ with the following action of a one-dimensional torus $\TT$: $t\cdot (x,y)=(tx,ty)$. The algebra of regular functions $A=\kk[x,y]$ is graded by the weight cone $\omega_M=\ZZ_{\geqslant 0}$ and $\deg\,x=\deg\,y=1$. Moreover, $\kk(X)^{\TT}=\kk\left(\dfrac xy\right)$.
Consider a homogeneous locally nilpotent derivation ${\partial}=y\dfrac{{\partial}}{{\partial}x}$ on $A$. Note that ${\partial}\left(\dfrac
xy\right)=1$, hence ${\partial}$ is of horizontal type. Let us show that two copies of the derivation ${\partial}$ give a counterexample to Theorem $\ref{main_theorem}$. Condition for commutators given in Proposition $\ref{commutator}$ for derivations of fiber type also does not hold. Indeed, note that $\deg\,{\partial}=0\in\omega_M$. Taking ${\partial}_1={\partial}_2={\partial}$, we obtain $\deg\,{\partial}_1+\deg\,{\partial}_2=0\in\omega_M$, though ${\partial}_1+{\partial}_2=2{\partial}\in \LND(A)$ and $[{\partial}_1,{\partial}_2]=[{\partial},{\partial}]=0\in \LND(A)$.
Finally, we would like to discuss torus actions and homogeneous LNDs coming from actions of reductive algebraic groups, cf. [@IA+AL].
Let $G$ be a reductive algebraic group and $\mathfrak{g}=\Lie G$. We consider a maximal torus $\TT\subset G$ and $\mathfrak{h}=\Lie \TT$. The reductive Lie algebra $\mathfrak{g}$ admits a Cartan decomposition $$\mathfrak{g}=\mathfrak{h}\oplus\bigoplus_{\alpha\in\Delta}\mathfrak{g}_{\alpha},$$ where $\Delta\subset\mathfrak{h}^*$ is the system of roots and $\mathfrak{g}_{\alpha}=\langle e_{\alpha}\rangle$.
Starting with a regular action of $G$ on an affine variety $X$ we obtain the $\TT$-action on $X$ and the corresponding grading of $A=\kk[X]$ by the lattice $M$ of characters of the torus $\TT$. For every $\alpha\in\Delta$, the nilpotent element $e_{\alpha}\in\mathfrak{g}$ defines a homogeneous locally nilpotent derivation ${\partial}_{\alpha}$ on $\kk[X]$ of degree $\alpha$.
\[alpha-beta\] Let $\alpha,\beta\in\Delta$. Then $e_{\alpha}+e_{\beta}$ is nilpotent if and only if $\alpha+\beta\ne 0$.
Assume $\beta=-\alpha$. In this case $e_{\alpha}$, $e_{-\alpha}$ and $[e_{\alpha},e_{-\alpha}]$ form an $\mathfrak{sl}_2$-triple and, thus, $e_{\alpha}+e_{-\alpha}$ is semisimple.
For $\alpha\ne -\beta$ we can consider a hyperplane in $\mathfrak{h}^*$ such that $\alpha$ and $\beta$ are contained in the same open half-space. This hyperplane determines a one-dimensional torus $T$, and the roots $\alpha$ and $\beta$ are positive with respect to $T$. Looking at the weights occuring in the decomposition of the element $\ad(e_{\alpha}+e_{\beta})^N(x)$ for some homogeneous $x\in\mathfrak{g}$ we obtain that $e_{\alpha}+e_{\beta}$ is nilpotent.
Using this lemma and the result of Theorem $\ref{main_theorem}$ we obtain the following
\[alpha-beta-cone\] Assume that $\alpha+\beta\ne 0$. If ${\partial}_{\alpha}$ and ${\partial}_{\beta}$ are of fiber type and ${\partial}_{\alpha}+{\partial}_{\beta}\in\LND(A)$, then $\alpha+\beta$ does not belong to the weight cone $\omega_M$ corresponding to the action of the maximal torus $\TT$ in $G$ on $A$.
Consider a natural $GL_n(\kk)$-action on the affine space $\AA^n$. The affine space $\AA^n$ gets a structure of a toric variety under the action of the maximal torus $$\TT^n=\{\diag(t_1,\dots,t_n)\}\subseteq GL_n(\kk).$$ The weights of $x_1,\dots,x_n$ are $\varepsilon_1,\dots,\varepsilon_n$ respectively. A root $\alpha$ determines an LND ${\partial}_{\alpha}$: $$\alpha=\varepsilon_i-\varepsilon_j, i\ne j \qquad \rightsquigarrow \qquad {\partial}_{\alpha}=x_i\frac{{\partial}}{{\partial}x_j}.$$ It is easy to see that $${\partial}_{\alpha}+{\partial}_{\beta}=x_i\frac{{\partial}}{{\partial}x_j}+x_k\frac{{\partial}}{{\partial}x_l}\in\LND(A) \quad \Leftrightarrow \quad (i,j)\ne(l,k)\quad \Leftrightarrow \quad\alpha\ne -\beta.$$ Moreover, $\alpha+\beta=\varepsilon_i-\varepsilon_j+\varepsilon_k-\varepsilon_l\notin \omega_M=\ZZ_{\geqslant 0}^n$ for $(i,j)\ne(l,k)$. This illustrates Corollary $\ref{alpha-beta-cone}$.
Consider a natural $SL_n(\kk)$-action on the affine space $\AA^n$. The maximal torus $$\TT^{n-1}=\{\diag(t_1,\dots,t_n)\ |\ \prod_i\,t_i=1\}\subseteq SL_n(\kk)$$ acts on $\AA^n$ with complexity one. The weights of $x_1,\dots,x_n$ are $\varepsilon_1,\dots,\varepsilon_{n-1}$ and $-\varepsilon_1-\dots-\varepsilon_{n-1}$ respectively, thus, $\omega_M=M$. Therefore, there exist no locally nilpotent derivations of fiber type. Taking ${\partial}_{\alpha}+{\partial}_{\beta}$ with $\alpha+\beta\ne 0$ we obtain one more counterexample to Theorem $\ref{main_theorem}$ in case of derivations of horizontal type.
Acknowledgement {#acknowledgement .unnumbered}
===============
The author is grateful to her supervisor Ivan Arzhantsev for posing the problem and permanent support in preparing the text of this paper, and to Polina Kotenkova for useful suggestions.
Ivan Arzhantsev and Alvaro Liendo. [*Polyhedral divisors and $SL_2$-actions on affine $\TT$-varieties.*]{} To appear in Michigan Math. J., see also arXiv:1105.4494v2.
Ivan Arzhantsev, Hubert Flenner, Shulim Kaliman, Frank Kutzschebauch and Mikhail Zaidenberg. [*Flexible varieties and automorphism groups.*]{} To appear in Duke Math. J., see also arXiv:1011.5375v1.
Ivan Arzhantsev, Juergen Hausen, Elaine Herppich and Alvaro Liendo. [*The automorphism group of a variety with torus action of complexity one.*]{} arXiv:1202.4568v2.
Michel Demazure. [*Sous-groupes algebriques de rang maximum du groupe de Cremona.*]{} Ann. Sci. Ecole Norm. Sup. 3 (1970), 507–588.
Miguel Ferrero, Yves Lequain and Andrzej Nowicki. [*A note on locally nilpotent derivations.*]{} J. Pure Appl. Algebra 79 (1992), 45–50.
Gene Freudenburg. [*Algebraic Theory of Locally Nilpotent Derivations.*]{} Encyclopaedia of Mathematical Sciences 136. Invariant Theory and Algebraic Transformation Groups 7. Berlin: Springer, 2006.
Polina Kotenkova. [*On restriction of roots on affine $T$-varieties.*]{} arXiv:1104.0560v3.
Alvaro Liendo. [*Affine $\TT$-varieties of complexity one and locally nilpotent derivations.*]{} Transform. Groups 15 (2010), no. 2, 389–425.
Alvaro Liendo. [*$\GG_a$-actions of fiber type on affine $\TT$-varieties.*]{} J. Algebra 324 (2010), 3653–3665.
Vladimir Popov. [*Problems for problem session.*]{} In: Affine Algebraic Geometry, Contemporary Math., Vol. 369, AMS, 2005, 12–16.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Biochemical systems that express certain chemical species of interest at the same level at any positive equilibrium are called “absolute concentration robust” (ACR). These species behave in a stable, predictable way, in the sense that their expression is robust with respect to sudden changes in the species concentration, regardless the new positive equilibrium reached by the system. Such a property has been proven to be fundamentally important in certain gene regulatory networks and signaling systems. In the present paper, we mathematically prove that a well-known class of ACR systems studied by Shinar and Feinberg in 2010 hides an internal integral structure. This structure confers these systems with a higher degree of robustness that what was previously unknown. In particular, disturbances much more general than sudden changes in the species concentrations can be rejected, and robust perfect adaptation is achieved. Significantly, we show that these properties are maintained when the system is interconnected with other chemical reaction networks. This key feature enables design of insulator devices that are able to buffer the loading effect from downstream systems - a crucial requirement for modular circuit design in synthetic biology.'
author:
- 'Daniele Cappelletti, Ankit Guptaand Mustafa Khammash'
bibliography:
- 'bib.bib'
title: A hidden integral structure endows Absolute Concentration Robust systems with resilience to dynamical concentration disturbances
---
= \[draw=olive!50!white, rectangle, very thick, fill=olive!5!white, text width=15em, text centered, minimum height=3em, rounded corners\] = \[coordinate\] = \[coordinate\]
Introduction
============
The network of chemical interactions of a biochemical system of interest can be complex and involve unknown reaction propensities. One of the main goal of reaction network theory consists in deriving dynamical properties from simpler graphical characteristic of the model, and independently on the specific value of kinetic parameters [@F:book; @toth:reaction]. The results presented in this paper follow this approach.
A qualitative property of great interest is the capability of a certain chemical species to be expressed with the same concentration at any positive steady state, independently on the initial conditions and on how many steady states are present. Namely, assume that the dynamics of the biochemical system are expressed by the $d-$dimensional ordinary differential equation $$\frac{d}{dt}x(t)=f(x(t)).$$ We say that the $i-th$ species is *absolute concentration robust* (ACR), if there exists an *ACR value* $q$ independent of the initial condition $x(0)$ such that, whenever $x(t)$ tends to a positive vector $\overline{x}$, we have $\overline{x}_i=q$. In the typical cases of interest, the positive steady state $\overline{x}$ that is reached will depend on the initial condition $x(0)$, while the entry $\overline{x}_i=q$ does not. As noted in [@AEJ:ACR], the property of absolute concentration robustness alone does not imply stability of the positive steady states: it only ensures that if a positive steady state $\overline{x}$ exists, then the value of the ACR species at $\overline{x}$ is the ACR value.
Under the assumption of stability, absolute concentration robustness provides a reliable, predictive response to environmental changes, since the species of interest reaches the equilibrium level relative to the new environmental setting, regardless the previous conditions. The existence and importance of this robustness property for various gene regulatory networks and signal transduction cascades is explored in many papers, including [@BF:robust; @BL:robustness; @BG:robustness; @MG:high; @SRA:robustness; @ASBL:robustness; @PS:Envz_book; @SRG:two; @SAF:sensitivity].
In the Control Theory community, and under the assumption of stability, the absolute concentration robustness is known as “robustness to disturbances in the initial conditions” [@XD:robust; @YHSD:robust]. To achieve robustness with respect to some disturbance, the imbalance caused by the disturbance needs to be measured first. To this aim, a quantity of interest is the *integrator*, which is a function $\phi$ of the system variables whose derivative is exactly the imbalance to be eliminated. At steady state, the derivative of $\phi$ is zero and so needs to be the imbalance. In the setting of absolute concentration robustness, one would like to find an integrator whose derivative is the difference between the ACR species and its ACR value. Unfortunately, in general this cannot be done, as shown in Section \[sec:CT\] and as discussed in [@XD:robust].
In the present paper, we systematically study for the first time the connection between ACR systems and integrators. Specifically, our first contribution is related to the existence of a linear combination of chemical species whose derivative is the difference between the ACR species and its ACR value, multiplied by a monomial. Such linear combination of species is called *constrained integrator* (CI), because it behaves similarly to an integrator given that the monomial does not vanish [@XD:robust]. We rigorously prove that such a linear CI always exists for a large class of models that strictly includes the ACR systems introduced in [@SF:ACR]. This result has some important consequences: first of all, under the assumption of stability, it implies that the expression of ACR species is not only robust to changes in the initial conditions, but also to disturbances that are applied over time. An important application in Synthetic Biology concerns the design of *insulators*. A number of biochemical systems are known to express a specific output if given a certain input. The systems can therefore be considered as modules with different functions. In cells, different modules are combined so that more complex responses to external stimuli become possible [@HHLM:modular]. In Synthetic Biology, it is desirable to combine different modules to achieve the same level of complexity [@PW:second]. However, when connected, the different modules can affect the dynamics of each other and they can lose the desired dynamical properties they had when considered in isolation [@DNS:modular]. In a simplified framework, an *upstream* module processes an input, and its output is fed to a second, *downstream* module to be further processed. Since the information is passed in form of molecules, which are then consumed or temporarily sequestrated by the downstream module, the equations governing the upstream module dynamics are perturbed and its functionality can be affected. Such effect is commonly called *loading effect* [@MRLDW:load] or *retroactivity* [@DNS:modular; @PM:retroactivity], and needs to be minimized. In other words, the upstream module needs to be *insulated* from the loading effect caused by the downstream module. We propose two ways in which the robustness of the systems studied in this paper can be used to this aim. The first solution is to simply design an upstream module which is robust to loading effects, modeled as a persistent disturbance over time. The second solution is to design an extra component, called *insulator*, which transfers the signal from the upstream module to the downstream module while at the same time shielding the dynamics of the upstream module from retroactivity effects.
We will also show how more theoretical results on reaction network models can be obtained as a consequence of our work. In Reaction Network Theory, the study of steady state invariants consitute an interesting topic of research[@SF:ACR; @KMDDG:invariants; @DDG:invariants; @F:book]. In [@SF:ACR], it has been proven that certain graphical properties of the network imply the existence of an ACR species, regardless the choice of kinetic parameters. Such sufficient conditions are generalized in the present work while they are maintained simple to check. Moreover, no way to explicitly determine the ACR value was given in [@SF:ACR], and we fill the gap by proposing a fast linear method to calculate it. Furthermore, a substantial effort in the Reaction Network community is devoted to understand under what conditions dynamical properties of single systems can be lifted to larger systems [@FCW:node; @JS:atoms; @BP:inheritance; @FW:multistationarity; @GHMS:joining]. Our contribution in this sense consist in proving that, under certain conditions, if an ACR system of the class studied in this paper is part of a larger model, the ACR species is still so in the lager system and its ACR value is maintained. Finally, it is worth mentioning that in the present work we consider the possibility of time-dependent rates for the occurrence of chemical transformations. This is more general than what is usually studied in Reaction Network Theory, with the exception of few works explicitly allowing for this scenario [@JH:solving; @BC:robust; @CNP:persistence; @CMW:parasite; @GMS:geometric; @A:time].
Examples of ACR systems
=======================
An illustrative example
-----------------------
Consider two proteins $A$ and $B$, whose interaction is described by $$\label{eq:toymodel}
\begin{split}
\schemestart
$A+B$\arrow{->[$\kappa_1$]}$2B$
\arrow(@c1.south east--.north east){0}[-90,.25]
$B$\arrow{->[$\kappa_2$]}$A$
\schemestop
\end{split}$$ where the positive constants $\kappa_1$ and $\kappa_2$ describe the propensity of a reaction to occur. If enough proteins are present and they are homogeneously spread in space, then a good model for the time evolution of the concentrations of proteins $A$ and $B$ is given by mass-action kinetics. Specifically, the concentrations of $A$ and $B$ at time $t$, denoted by $x_A(t)$ and $x_B(t)$ respectively, are assumed to solve $$\label{eq:toymodeleq}
\frac{d}{dt}\begin{pmatrix}
x_A(t)\\ x_B(t)
\end{pmatrix}
=\kappa_1 x_A(t)x_B(t)\begin{pmatrix} -1\\1\end{pmatrix}
+\kappa_2 x_B(t)\begin{pmatrix} 1\\-1\end{pmatrix}$$ It is easy to check that the steady states of are given by states $(\overline x_A, \overline x_B)$ such that either $\overline x_B=0$ or $\overline x_A=\kappa_2/\kappa_1$. Hence, $A$ is an ACR species because its expression at any positive steady state is the same. It is common during biochemical experiments to be able to control the inflow rate of some species (say $B$). Some additional chemical species may also be introduced, with the purpose of degrading some of the present components (in this case, species $C$ is introduced to faster degrade species $B$). After these modifications, (\[eq:toymodel\]) becomes $$\label{eq:toymodel_modified}
\begin{split}
\schemestart
$A+B$\arrow{->[$\kappa_1$]}$2B$
\arrow(@c1.south east--.north east){0}[-90,.25]
$0$\arrow{->[$u_1(t)$]}$B$\arrow{->[$\kappa_2$]}$A$
\arrow(@c3.south east--.north east){0}[-90,.25]
$0$\arrow{<=>[$u_2(t)$][$\kappa_3$]}$C$
\arrow(@c6.south east--.north east){0}[-90,.25]
$C+B$\arrow{->[$\kappa_4$]}$0$
\schemestop
\end{split}$$ Since we still have $$\frac{d}{dt}x_A(t)=-\kappa_1x_B(t)\left(x_A(t)-\frac{\kappa_2}{\kappa_1}\right),$$ it is still true that the value of $x_A(t)$ will converge to $\kappa_2/\kappa_1$, as long as the functions $u_1, u_2$ and the rates $\kappa_3, \kappa_4$ are such that the species $B$ is not removed from the system too fast. In this paper we will prove a general result describing when such kind of robustness to persistent perturbations is present for ACR systems.
EnvZ-OmpR osmoregulatory system
-------------------------------
Consider the system described in Figure \[fig:envz\]. The model is proposed in [@SF:ACR; @SMMA:input] as osmoregulatory system in Escherichia Coli. It is in accordance with experimental observations discussed in [@PS:Envz_book; @SRG:two; @BG:robustness]. According to the model, whose schematics is described in Figure \[fig:envz\], the activation rate of the sensor-transmitter protein EnvZ depends on the medium osmolarity. Then, an active form of EnvZ transfers its phosphoryl group to the sensory response protein OmpR, which becomes OmpR-P and promotes the production of the outer membrane porins OmpF and OmpC. Hence, it is important that the concentration of OmpR-P responds in a reliable, predictive way to changes in the medium osmolarity (which the rate constants $\kappa_i$ in the reaction network of Figure \[fig:envz\] depend upon), but not on the initial concentration of the different chemical species involved. As a matter of fact, it follows from the results developed in [@SF:ACR] that [OmpR-P]{} is an ACR species.
$$\schemestart
EnvZ-{{\color{blue}\bfseries D}}\arrow(1--2){<=>[*0{$\kappa_1$}][*0{$\kappa_2[{\text{ADP}}]$}]} EnvZ \arrow(@2--3){<=>[*0{$\kappa_3[{\text{ATP}}]$}][*0{$\kappa_4$}]} EnvZ-{{\color{brown}\bfseries T}}\arrow(@3--4){->[*0{$\kappa_5$}]} EnvZ-{{\color{red}\bfseries P}}\arrow(@1.south east--.north east){0}[-90,.35]
EnvZ-{{\color{red}\bfseries P}}\ \+ OmpR \arrow(5--6){<=>[*0{$\kappa_6$}][*0{$\kappa_7$}]} EnvZ-OmpR-{{\color{red}\bfseries P}}\arrow(@6--7){->[*0{$\kappa_8$}]} EnvZ \+ OmpR-{{\color{red}\bfseries P}}\arrow(@5.south east--.north east){0}[-90,.35]
EnvZ-{{\color{blue}\bfseries D}}\ \+ OmpR-{{\color{red}\bfseries P}}\arrow(8--9){<=>[*0{$\kappa_9$}][*0{$\kappa_{10}$}]} EnvZ-OmpR-{{\color{blue}\bfseries D}}-{{\color{red}\bfseries P}}\arrow(@9--10){->[*0{$\kappa_{11}$}]} EnvZ-{{\color{blue}\bfseries D}}\ \+ OmpR
\schemestop$$
Necessary terminology and known results
=======================================
In order to present the theory we develop, we first need to introduce some terminology. The linear combinations of chemical species appearing on either side of the chemical reactions of interest are called *complexes*, in accordance with the Reaction Network Theory literature. Be aware that the word “complex” has usually a different meaning in the Biology literature. We denote by $m$ the number of complexes present in the network, an by $d$ the number of chemical species. As an example, the complex of (\[eq:toymodel\]) are $A+B$, $2B$, $B$, and $A$. Here, $d=2$ and $m=4$. In (\[eq:toymodel\_modified\]) the complexes are $A+B$, $2B$, $0$, $B$, $A$, $C$, and $C+B$, hence $d=3$ and $m=7$. Finally, in the system depicted in Figure \[fig:envz\] $d=8$ and $m=10$. Since a complex is a linear combination of species, each complex can be regarded as a vector of length $d$. For example, for the model (\[eq:toymodel\]) we can consider $A+B$ as $(1,1)$, $2B$ as $(0,2)$, $B$ as $(0,1)$, and finally $A$ as $(1,0)$. With this in mind, we can define the *stoichiometric subspace* as $${\mathcal{S}}=\operatorname{span}_{{\mathbb{R}}} \{y_j-y_i\,:\,\text{there is a reaction from }y_i\text{ to } y_j\},$$ where $y_n$ denotes the $n$th complex, for all $1\leq n\leq m$. For example, for (\[eq:toymodel\]) we have $${\mathcal{S}}=\operatorname{span}_{{\mathbb{R}}}\left\{\begin{pmatrix}-1\\1\end{pmatrix}, \begin{pmatrix}1\\-1\end{pmatrix}\right\}=\operatorname{span}_{{\mathbb{R}}}\left\{\begin{pmatrix}-1\\1\end{pmatrix}\right\}.$$ For (\[eq:toymodel\_modified\]), we have ${\mathcal{S}}={\mathbb{R}}^3$. In the most general formulation of reaction systems, a (time-dependent) rate function $\lambda_{ij}$ is associated with the reaction from the $i$th to the $j$th complex of the network, and the concentration vector of the different chemical species is assumed to solve the differential equation $$\label{eq:Gmodel}
\frac{d}{dt}x(t)=\sum_{1\leq i,j\leq m}(y_j-y_i)\lambda_{ij}(x(t),t),$$ where if a reaction from the $i$th to the $j$th complex does not exist, then $\lambda_{ij}$ is the zero function. Note that simply sums the contributions to the dynamics given by the different chemical reactions. It is also not difficult to show that every solution to is necessarily confined within a translation of the stoichiometric subspace. If for all non-zero propensities $\lambda_{ij}$ there exists a positive constant $\kappa_{ij}$ such that $$\lambda_{ij}(x(t),t)=\kappa_{ij} \prod_{l=1}^d x_l(t)^{y_{il}},$$ then the model is a *mass-action system*. In this case, can be written as $$\frac{d}{dt}x(t)=Y A(\kappa) \Lambda(x(t)),$$ where $Y$ is a $d\times m$ matrix whose $i$th column is $y_i$, $A(\kappa)$ is a $m\times m$ matrix given by $$A(\kappa)_{ij}=\begin{cases}
\kappa_{ji}&\text{if }i\neq j\\
-\sum_{l=1}^m \kappa_{il} &\text{if }i=j
\end{cases}$$ and $\Lambda(x(t))$ is a vector of length $m$ whose $i$th entry is $\prod_{l=1}^d x_l(t)^{y_{il}}$. Examples of mass-action systems are (\[eq:toymodel\]) and the model in Figure \[fig:envz\].
A directed graph can be associated with a reaction network, where the nodes are given by the complexes and the directed edges are given by the reactions. Such a graph is called *reaction graph*. As an example, (\[eq:toymodel\]) is a reaction graph, while (\[eq:toymodel\_modified\]) is not because the complex 0 is repeated. The reaction graph corresponding to (\[eq:toymodel\_modified\]) is $$\label{eq:toymodel_modified_graph}
\begin{split}
\schemestart
$A+B$\arrow(1--2){->[*0{$\kappa_1$}]}$2B$
\arrow(@1.south east--.north east){0}[-90,.35]
$C+B$\arrow(7--3){->[*0{$\kappa_4$}]}$0$\arrow(@3--4){->[*0{$u_1(t)$}]}$B$\arrow(@4--5){->[*0{$\kappa_2$}]}$A$
\arrow(@3--6){<=>[*0{$u_2(t)$}][*0{$\kappa_3$}]}[-90]$C$
\schemestop
\end{split}$$
We denote by $\ell$ the number of connected components of the reaction graph associated with the network. For both the networks (\[eq:toymodel\]) and (\[eq:toymodel\_modified\]) $\ell=2$, as well as for the EnvZ-OmpR osmoregulatory system of Figure \[fig:envz\]. Then, we define the *deficiency* of a network as $$\delta=m-\ell-\dim {\mathcal{S}}.$$ The deficiency of a network has important geometric interpretation, and a collection of classical deficiency theory results can be found in [@F:review]. The deficiency of (\[eq:toymodel\]) is $\delta=4-2-1=1$, and the deficiency of (\[eq:toymodel\_modified\]) is $\delta=7-2-3=2$. Similarly, it can be checked that the deficiency of the EnvZ-OmpR osmoregulatory system in Figure \[fig:envz\] is 1.
Finally, we say that a complex $y$ is *terminal* if for all paths in the reaction graph leading from $y$ to another complex $y'$, there is a path leading from $y'$ to $y$. If a complex is not terminal, then it is called *non-terminal*. As an example, the only terminal complexes for (\[eq:toymodel\]) and (\[eq:toymodel\_modified\]) are $2B$ and $A$.
We recall that a species is said to be *absolute concentration robust* (ACR) if its concentration at any positive steady state of is the same. We are ready to state the following result, as presented in [@SF:ACR].
[theorem]{}[thmfeinbergweak]{}\[thm:feinberg\_weak\] Consider a mass-action system, and assume the following holds:
1. there are two non-terminal complexes $y_i$ and $y_j$ such that only one entry of $y_j-y_i$ is non-zero;
2. the deficiency is 1.
3. a positive steady state exists.
Then, the species relative to the non-zero entry of $y_j-y_i$ is ACR.
Note that a stronger version of Theorem \[thm:feinberg\_weak\] is proven in [@SF:ACR], which detects steady state invariant that are more general than the equilibrium concentration of a single species. The stronger version is stated in the Supplementary Material as Theorem \[thm:feinberg\_strong\], and an extension of it is proven in the present work.
The model (\[eq:toymodel\]) has deficiency 1, as already observed, has at least one positive equilibrium and the non-terminal complexes $A+B$ and $B$ differ only for the species $A$. Hence, Theorem \[thm:feinberg\_weak\] applies and $A$ is ACR. It is shown in [@SF:ACR] that the EnvZ-OmpR osmoregulatory system in Figure \[fig:envz\] also fulfils the hypothesis of Theorem \[thm:feinberg\_weak\], with the non-terminal complexes EnvZ-D and EnvZ-D+OmpR-P only differing for the species OmpR-P. As a consequence, OmpR-P is ACR. In Section \[sec:calculate\] we will develop a method to explicitly calculate the ACR value through symbolic linear algebra. We note that Theorem \[thm:feinberg\_weak\] cannot be applied to (\[eq:toymodel\_modified\]) for two reasons: the model is not a mass-action system and its deficiency is 2.
As noted in [@AEJ:ACR], the positive steady states of a system with an ACR species are not necessarily stable. However, as a consequence of the present work (more precisely, as a consequence of Theorem \[thm:control\] with $u$ being the zero function), we know the following: if a mass-action system as in Theorem \[thm:feinberg\_weak\] has an unstable positive steady state, then either the system oscillates around it, or some chemical species is completely consumed, or some chemical species is indefinitely produced. We give here the formal definition of “oscillation”, as intended in this paper.
We say that a function $g\colon{\mathbb{R}}_{\geq0}\to{\mathbb{R}}$ *oscillates* around a value $\overline q\in{\mathbb{R}}$ if for each $t\in{\mathbb{R}}_{\geq0}$ there exist $t_+>t$ and $t_->t$ such that $$g(t_+)>q\quad\text{and}\quad g(t_-)<q.$$
A linear constrained integrator
===============================
Control Theory background {#sec:CT}
-------------------------
In Control Theory, the focus is usually on systems of differential equations of the form $$\label{eq:system}
\frac{d}{dt} x(t)=f(x(t), u(t)),$$ where $x\colon{\mathbb{R}}_{\geq0}\to {\mathbb{R}}^{n_x}$ and $u\colon{\mathbb{R}}_{\geq0}\to{\mathbb{R}}^{n_u}$ for some $n_x,n_u\in{\mathbb{Z}}_{>0}$, and $f$ is a differentiable function. The function $u$ is called the *input of the system*. Further, a quantity of the form $z(t)=a(x(t))$ is of interest, where $a$ is a differentiable function with $a\colon{\mathbb{R}}^{n_x}\to{\mathbb{R}}^{n_z}$, for some $n_z\in{\mathbb{Z}}_{>0}$. The function $z$ is called the *output* of the system. In the usual setting, one needs to find an appropriate function $u$ such that $z$ is close to a desired level $\overline{z}\in{\mathbb{R}}^{n_z}$, either on average or for $t\to\infty$. To this aim, the existence of a function $\phi\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$ such that $$\frac{d}{dt}\phi(x(t))=z(t)-\overline{z}$$ is of high importance, an is called an *integrator*. The name derives from $\phi(x(t))$ being the integral of the error that needs to be controlled: $$\phi(x(t))=\phi(x(0))+\int_0^t (z(s)-\overline{z})ds.$$ If the function is fed back to the system and is used to tune the input, then an *integral action* or *integral feedback* is in place [@doyle:feedback; @astrom:feedback]. One of the main features of an integrator is that the derivative of $\phi(x(t))$ is zero if and only if $z(t)=\overline{z}$. If a function $\tilde\phi\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$ satisfies $$\frac{d}{dt}\tilde\phi(x(t))=r(x(t))\Big(z(t)-\overline{z}\Big)$$ for some differentiable function $r\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$, then $\tilde\phi$ is called a *constrained integrator* (CI) [@XD:robust]. The name derives from the fact that the derivative of $\tilde\phi(x(t))$ is zero if and only if $z(t)=\overline{z}$, provided that $r(x(t))\neq0$. In biology, it is common to find CIs, and the condition $r(x(t))\neq0$ is usually implied by $x(t)\neq 0$ [@XD:robust]. Note that in [@XD:robust] an explicit distinction between integrators and integral feedbacks is not made.
In the setting of systems with ACR species, the output $z$ can be considered to be the concentration of the ACR species over time, and $\overline{z}$ can be their ACR values. In (\[eq:toymodel\]), $z(t)=x_A(t)$ and $\overline{z}=\kappa_2/\kappa_1$. A CI (as noted in [@XD:robust]) is given by $\tilde\phi(x(t))=x_B(t)$, since $$\frac{d}{dt}x_B(t)=\kappa_1x_B(t)\left(x_A(t)-\frac{\kappa_2}{\kappa_1}\right).$$ The question of whether an integrator exists can be quickly answered in negative, because any point of the form $(\overline{x}_A,0)$ is a steady state. If an integrator $\phi$ existed, then by choosing $x(0)=(\overline{x}_A,0)$ we would have $$0=\frac{d}{dt}\phi(x(t))=\overline{x}_A-\frac{\kappa_2}{\kappa_1},$$ which cannot hold expect for a specific value of $\overline{x}_A$. An integrator may still exist in a weaker sense, if we restrict its domain. For example, in this case the function $\hat\phi(x(t))=\frac{1}{\kappa_1}\log x_B(t)$ would be an integrator, in the sense that if $x_B(t)>0$ then $$\frac{d}{dt}\hat\phi(x(t))=
\frac{1}{\kappa_1x_B(t)}\left(\kappa_1x_B(t)\left(x_A(t)-\frac{\kappa_2}{\kappa_1}\right)\right)=
x_A(t)-\frac{\kappa_2}{\kappa_1}$$ However, the domain of $\hat\phi$ is not the entire ${\mathbb{R}}^2$. Finally, since linear functions could always be extended continuously to the boundaries of ${\mathbb{R}}^2_{>0}$, a linear integrator cannot exist for (\[eq:toymodel\]) even if its domain is restricted.
Existence and characterization
------------------------------
We state here our result concerning linear CIs. A stronger version is proved in the Supplementary Material. The result is inspired by the analysis carried on in [@SF:ACR], which is here expanded.
For any $n\times l$ real matrix $M$ and real vector $v$ of length $n$, we denote by $(M|v)$ the $n\times l+1$ matrix obtained by adding the column $v$ at the right of the matrix $A$. Let $1\leq i,j\leq m$. To present our result, we first need to define the preimage $$\label{eq:Gamma}
\Gamma_{ij}(\kappa)=\left\{\gamma\in{\mathbb{R}}^{d+1}\,:\,
\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\gamma
=e_j\right\},$$ where $e_n$ denotes the $n$th vector in the canonical basis of ${\mathbb{R}}^m$, whose $n$th component is 1 and whose other components are 0. The role of $\Gamma_{ij}(\kappa)$ is that of providing vectors $\gamma$ satisfying $$\label{eq:explanation}
\frac{d}{dt}\langle\hat\gamma,x(t)\rangle=\hat\gamma^\top Y A(\kappa)\Lambda(x(t))=\Lambda_j(x(t))-\gamma_i\Lambda_i(x(t)),$$ where $\hat\gamma$ is the projection onto the first $d$ components of $\gamma$, and $\langle\cdot,\cdot\rangle$ is the standard scalar product. Under certain assumptions, will provide us with a CI. Then, the projection of $\Gamma_{ij}(\kappa)$ onto the first $d$ coordinates will be of interest, and we will denote it by $\hat\Gamma_{ij}(\kappa)$. The set $\Gamma_{ij}(\kappa)$ can be calculated with symbolic linear algebra. We will also prove in the Supplementary Material how $\hat\Gamma_{ij}(\kappa)$ is connected with ${\mathcal{S}}^\perp$, which is a set easily described by linear algebra and independent on the rate functions. Specifically, we will prove that if $\xi\in\hat\Gamma_{ij}(\kappa)$, then necessarily $$\label{eq:gamma_S}
\{\xi+w\,:\,w\in{\mathcal{S}}^\perp\}\subseteq\hat\Gamma_{ij}(\kappa).$$ We will also give sufficient conditions under which the inclusion in is an equality.
As an example, consider the model in Figure \[fig:envz\]. Using Matlab, we quickly obtain that a vector $\xi$ is in $\Gamma_{18}(\kappa)$, with $$\label{eq:calculation_ex}
\xi_9=\frac{\kappa_1\kappa_3\kappa_5(\kappa_{10}+\kappa_{11})[{\text{ATP}}]}{\kappa_2(\kappa_4+\kappa_5)\kappa_9\kappa_{11}[{\text{ADP}}]},$$ and it is shown in the Supplementary Material that $$\label{eq:calculation_ex_hat}
\hat\Gamma_{18}(\kappa)=\left\{\hat \xi+\begin{pmatrix}w_1\\w_1\\w_1\\w_1\\0\\w_1\\0\\w_1\end{pmatrix}+\begin{pmatrix}0\\0\\0\\0\\w_2\\w_2\\w_2\\w_2\end{pmatrix}\,:\, w_1, w_2\in{\mathbb{R}}\right\},$$ where $\hat \xi$ is the projection of $\xi$ onto its first $d=8$ coordinates.
The family of models we study in this paper concerns reaction systems with two non-terminal complexes $y_i$ and $y_j$ differing in just one entry, for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Our first result shows that such a family includes the models studied in [@SF:ACR]. The proof can be found in the Supplementary Material.
[theorem]{}[thmfeinbergimpliescond]{}\[thm:feinberg\_implies\_cond\] Consider a mass-action system, and assume the following holds:
1. there are two non-terminal complexes $y_i$ and $y_j$ such that only one entry of $y_j-y_i$ is non-zero;
2. the deficiency is 1.
3. a positive steady state exists.
Then, $\hat\Gamma_{ij}(\kappa)$ is non-empty.
As an example, we know already from direct calculation that for the EnvZ-OmpR signaling system holds, which in turn implies that the set $\hat\Gamma_{18}(\kappa)$ is non-empty. However, we could have also derived this information from Theorem \[thm:feinberg\_implies\_cond\], without explicitly calculating $\hat\Gamma_{18}(\kappa)$.
We note here that the converse of Theorem \[thm:feinberg\_implies\_cond\] does not hold. We show this with an example of a multisite phosphorylation signaling system in the Supplementary Material, which does not fall in the setting of [@SF:ACR] but for which we are able to prove absolute concentration robustness regardless the choice of rate constants, as long as a positive steady state exists. Notably, we are also able to derive information on when this occurs without working directly with the differential equation. As a consequence of this example, the family of models we analyze is proven to be strictly larger than that studied in [@SF:ACR]. The following holds.
[theorem]{}[thmlcif]{}\[thm:lcif\] Consider a mass-action system. Assume that there are two complexes $y_i$ and $y_j$ only differing in the $n$th entry, and that $\hat\Gamma_{ij}(\kappa)$ is non-empty. Let $\gamma\in\hat\Gamma_{ij}(\kappa)$, and define $$q=\gamma_{d+1}^{\frac{1}{(y_j-y_i)_n}}.$$ Then, either no positive steady state exists or the $n$th species is ACR with ACR value $q$. Moreover, $$\phi(x)=\sum_{i=1}^d \beta_i x_i$$ is a linear CI with $$\frac{d}{dt}\phi(x(t))=\Lambda_i(x(t)) \left(x_n(t)^{(y_j-y_i)_n}-q^{(y_j-y_i)_n}\right)$$ for any initial condition $x(0)$ if and only if $\beta\in\hat\Gamma_{ij}(\kappa)$.
The existence of a linear CI given by Theorem \[thm:lcif\] is essential to develop the results presented in the next sections. Before unveiling the consequences of Theorem \[thm:lcif\], however, it is important to stress that a CI does not necessarily constitute a feedback, as one may be tempted to think. Consider $$\begin{split}
\schemestart
$A+B$\arrow{->[$\kappa_1$]}$A$\arrow{<=>[$\kappa_2$][$\kappa_3$]}$0$
\arrow(@c1.south east--.north east){0}[-90,.25]
$B$\arrow{->[$\kappa_4$]}$2B$
\schemestop
\end{split}$$ with $\kappa_1\kappa_3=\kappa_2\kappa_4$. It can be shown that the system satisfies the conditions of Theorems \[thm:feinberg\_weak\] and \[thm:feinberg\_implies\_cond\], with the non-terminal complexes $A+B$ and $A$ differing only in species $A$. Hence, $A$ is ACR and the assumptions of Theorem \[thm:lcif\] hold. A linear CI as in Theorem \[thm:lcif\] is given by $\phi(x)=-x_B/\kappa_1$, since for this choice $$\frac{d}{dt}\phi(x(t))=x_B(t)\left(x_A(t)-\frac{\kappa_4}{\kappa_1}\right).$$ However, the quantity $\phi(x(t))$ does not regulate the dynamics of $A$, since $$\frac{d}{dt}x_A(t)=\kappa_2-\kappa_3x_A(t)$$ does not depend on $x_B(t)$. Since in this case the CI is not acting on the system, it is not surprising that the existence of positive steady states is lost as soon as $\kappa_1\kappa_3\neq\kappa_2\kappa_4$.
It is also worth mentioning that not all systems with ACR species have a linear CI: consider the mass-action system $$\begin{split}
\schemestart
$A+B$\arrow{->[$\kappa_1$]}$B+C$\arrow{<=>[$\kappa_2$][$\kappa_3$]}$2B$
\arrow(@c1.south east--.north east){0}[-90,.25]
$B$\arrow{<=>[$\kappa_4$][$\kappa_5$]}$2E$\arrow{->[$\kappa_6$]}$2D$
\arrow(@c4.south east--.north east){0}[-90,.25]
$C$\arrow{->[$\kappa_7$]}$A$
\arrow(@c7.south east--.north east){0}[-90,.25]
$D$\arrow{->[$\kappa_8$]}$E$
\schemestop
\end{split}$$ The model is considered in [@AC:ACR], where it is proven that the species $A$ is ACR. We show in the Supplementary Material that there exists no linear function $\phi$ whose derivative at time $t$ is of the form $r(x(t))(x_A(t)^\gamma-q)$, for some polynomial $r$ and some real numbers $\gamma, q$. Note that in this case Theorem \[thm:feinberg\_implies\_cond\] does not apply because the deficiency of the network is 2.
A method to calculate the ACR value {#sec:calculate}
===================================
The first interesting consequence of Theorem \[thm:lcif\] is that the ACR values of the mass-action systems satisfying the assumption of the theorem can be calculated by finding at least one element of the preimage $\Gamma_{ij}(\kappa)$, and this can be done via a simple symbolic linear algebra calculation. As an example, consider the EnvZ-OmpR osmoregulatory system in Figure \[fig:envz\]. Then, Theorem \[thm:lcif\] implies that the ACR value of OmpR-P is the value given in . This value is in accordance with the one found in the Supplementary Material of [@SF:ACR], however we found it by calculating a single element in the preimage of a matrix, as opposed to working with the rather complicated differential equation associated with the model. An even more involved examples is dealt with in the Supplementary Material.
Rejection of persistent disturbances
====================================
The result
----------
We state an important consequence of Theorem \[thm:lcif\], a stronger version of which is proven in the Supplementary Material:
[theorem]{}[thmcontrol]{}\[thm:control\] Consider a mass-action system, with associated differential equation $$\frac{d}{dt} x(t)=f(x(t)).$$ Assume that there are two complexes $y_i$ and $y_j$ only differing in the $n$th entry, and that $\hat\Gamma_{ij}(\kappa)$ is non-empty. Let $q$ be the ACR value of the $n$th species. Consider an arbitrary function $u$ with image in ${\mathbb{R}}^d$ such that a solution to $$\frac{d}{dt}\tilde x(t)=f(\tilde x(t))+u(\tilde x(t),t)$$ exists. Assume that there exists a $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ which is orthogonal to the vector $u(x,t)$ for any $x,t$. Then, for any initial condition $\tilde x(0)$, at least one of the following holds:
(a) \[part:infinity\_or\_zero\] the concentration of some species goes to 0 or infinity, along a sequence of times;
(b) \[part:oscillation\_MT\] $\tilde x_n(t)$ oscillates around $q$ and $\hat\gamma_k\neq0$ for some $k\neq n$;
(c) \[part:ACR\_convergence\] the integral $$\int_t^\infty |\tilde x_n(s)-q|ds$$ tends to $0$, as $t$ goes to infinity.
The result implies that if a disturbance orthogonal to a vector $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ is applied over time, then the stability of the ACR species is maintained: at most, the ACR species can be forced to oscillate around its original ACR value, but it cannot be forced to attain another equilibrium level without causing extinction or overexpression of the chemical species present. We analyze the power of Theorem \[thm:control\] by showing some examples of applications.
Consider the mass-action system (\[eq:toymodel\]), which fulfills the assumptions of Theorem \[thm:control\] as already observed. Assume the complexes are ordered as $A+B$, $2B$, $B$, and $A$, and the species are ordered alphabetically as $A$, $B$. Hence, the two non-terminal complexes differing in the ACR species $A$ are the 1st and the 3rd, and it is shown in the Supplementary Material that $$\label{eq:calculation_toy}
\hat\Gamma_{13}(\kappa)=\left\{\begin{pmatrix}
0 \\ 1
\end{pmatrix}+\begin{pmatrix}
w \\ w
\end{pmatrix}\,:\, w\in{\mathbb{R}}\right\}.$$ Hence, by choosing $w=-1$, we have that $$\label{eq:hjhjhjh}
\begin{pmatrix}
-1\\0
\end{pmatrix}\in \hat\Gamma_{13}(\kappa).$$ This vector is clearly orthogonal to any disturbance acting on the production and degradation rates of the species $B$. Hence, it follows that the stability and the ACR value of the species $A$ is maintained in (\[eq:toymodel\_modified\]), provided that no species is completely removed or indefinitely expressed. Specifically, since the entry of relative to $B$ is zero, it follows from Theorem \[thm:control\] that if all the species concentrations are bounded from below and from above by positive quantities, necessarily the concentration of the species $A$ converges to its ACR value as $t$ goes to infinity, despite the disturbances.
Consider the osmoregulatory system in Figure \[fig:envz\], whose features have already been discussed in the paper. In particular, we know the species OmpR-P is ACR with ACR value . Recall that we ordered the complexes such that the two non-terminal ones differing in OmpR-P are the 1st and the 8th. It follows from that for any chemical species, there is a vector in $\hat\Gamma_{18}(\kappa)$ with the associated entry equal to 0. It follows that even if the production and degradation of any chemical species in the model is tampered with, the stability and the ACR value of the species OmpR-P are maintained, in the sense described by Theorem \[thm:control\]. We can push the disturbances further. By appropriately choosing $w_1$ and $w_2$ in , we can see that there is vector in $\hat\Gamma_{18}(\kappa)$ whose entries relative to the species OmpR and EnvZ are both 0. Hence, it follows by Theorem \[thm:control\] that by tampering with the production and degradation rates of both these species over time, if no extinction and no overexpression occurs, then the concentration of OmpR-P still converges to the value , or oscillates around it.
As a final remark, we note that can be useful in determining whether a vector in $\Gamma_{ij}(\kappa)$ exists, with a specific component equal to 0, say the $n$th one. In fact, the existence of such a vector can be deduced without calculations, if there is a vector $w\in{\mathcal{S}}^{\perp}$ whose $n$th component is different from 0.
Insulating properties
---------------------
Here we illustrate how the theory we developed can be utilized to solve the Synthetic Biology problem of retroactivity. As explained in the Introduction, the loading effects caused by a downstream biochemical module can disrupt the functionality of upstream modules, which prevents the implementation of biochemical circuits by interconnecting biochemical modules with different functions [@DNS:modular]. A concrete example of loading effect is illustrated in Figure \[fig:application\].
Assume that a mass-action system has two complexes $y_i$ and $y_j$, that are only different in the $n$th component, which corresponds to the species $X$. Assume further that $\hat\Gamma_{ij}(\kappa)$ is non-empty and that a positive steady state exists. Hence, the species $X$ is ACR, with some ACR value $q$. It further follows from Theorem \[thm:control\] that, if there exists $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ with $\hat\gamma_n=0$, then the production and degradation rates of the species $X$ can be arbitrarily perturbed over time by an arbitrary function $u$, without compromising its robustness. Specifically, if the perturbed system is stable and no chemical species is completely consumed, then the concentration of $X$ will still converge to the same ACR value $q$ as in the original mass-action system. The key observation we make here is that the perturbation $u$ can be considered as the loading effect of a downstream module that takes the concentration of species $X$ as input. In this case, the loading effect on the original mass-action is rejected and the concentration of $X$ is maintained at a desired level $q$ at steady state. Further, the concentration of $X$ is maintained approximately constant in the transient dynamics as well, if we assume as done in [@DNS:modular] that a separation of dynamics time scale is in place. Specifically, assume $$\frac{d}{dt}\tilde x(t)=\frac{1}{\varepsilon}f(\tilde x(t))+ u(\tilde x(t), t)e_n,$$ for some small $\varepsilon>0$, with $f(\tilde x(t))$ and $u(\tilde x(t), t)$ being of the same order of magnitude. Under the assumption of stability, if $\varepsilon$ is very small then the perturbed system will quickly approach the slow manifold defined by $$0=f(\tilde x(t))+ \varepsilon u_t(\tilde x(t))e_n,$$ where $u_t$ is a function from ${\mathbb{R}}^d$ to ${\mathbb{R}}^d$ defined by $u_t(x)=u(x,t)$. By Theorem \[thm:control\] applied to the disturbance $u_t$, the species $X$ assumes its ACR value at any positive point of the slow manifold, which is exactly what we wanted.
As an example of application, consider the EnvZ-OmpR osmoregulatory system in Figure \[fig:envz\]. It follows from that there exists $\hat\gamma\in\Gamma_{18}(\kappa)$ with a zero in the entry corresponding to the ACR species OmpR-P. Hence, the production and degradation rates of OmpR-P can be arbitrarily changed over time, without altering its robustness property, in the sense described by Theorem \[thm:control\]. As observed, the statement still holds true if the perturbation is originated by a downstream module that acts on OmpR-P. Hence, the EnvZ-OmpR osmoregulatory system can be used to maintain the expression of OmpR-P at a desired level, which depends on the input rate constants, even if the species OmpR-P is used by a downstream module. Moreover, if the downstream module acts on a slower time scale, the concentration of the species OmpR-P is approximately maintained at the target level at any time point. In Figure \[fig:block\], a diagram describing this situation is proposed.
; (signaling) [EnvZ-OmpR system]{}; (plant) [Downstream module]{}; (output) ; (input) – node [$\kappa$]{} (signaling); (\[xshift=0.2cm\]signaling.south) – node [OmpR-P]{}(\[xshift=0.2cm\]plant.north); (\[xshift=-0.2cm\]plant.north) – node [OmpR-P]{} (\[xshift=-0.2cm\]signaling.south); (plant) – node [$g($OmpR-P$)$]{} (output);
; (upstream) [Upstream module]{}; ; (insulator) [Insulator\
]{}; (downstream) [Downstream module]{}; (output) ; (input) – node [$u(t)$]{} (upstream); (upstream) – node\[text width=3cm\] [$A^\star$ (not changed by the insulator)]{}(insulator); (input2) – node [$\kappa_1$, $\kappa_2$, $x_B(0)$]{} (insulator); (\[xshift=0.2cm\]insulator.south) – node [$A\approx\displaystyle\frac{\kappa_2 A^\star}{\kappa_1}$]{} (\[xshift=0.2cm\]downstream.north); (\[xshift=-0.2cm\]downstream.north) – node\[text width=3.5cm\] [$A$ (changed by downstream module)]{} (\[xshift=-0.2cm\]insulator.south); (downstream) – node [$g(A)$]{} (output);
Consider now the case were an upstream module is affected from loading effects. We show how the theory developed in this paper can be used to design an insulator. Assume that the upstream module accepts $u(t)$ as input, and modulates the concentration of the species $A^\star$ accordingly. The species $A^\star$ is then used by a downstream module, which returns a function of the concentration of $A^\star$ as output. The action of the downstream module on the species $A^\star$ causes a loading effect on the upstream module. This loading effect should be reduced. To this aim, we proposed to modify the downstream module such that it acts on a species $A$ rather than on the species $A^\star$, and to include in the system the following module, where $B$ is a species that is not used by neither the upstream nor the downstream module: $$\label{eq:insulator}
\begin{split}
\schemestart
$A + B$\arrow{->[$\kappa_1$]}[0,.8]$2B$
\arrow(@c1.south east--.north east){0}[-90,.25]
$A^\star + B$\arrow{->[$\kappa_2$]}[0,.8]$A^\star + A$
\schemestop
\end{split}$$ Assume stability is reached and that the species $B$ is not completely consumed. Then, at steady state the concentration level of $A^\star$ is fixed, and the concentration of the ACR species $A$ will converge to its ACR value $\kappa_2 x_{A^\star}/\kappa_1$ regardless any disturbance applied to the production and degradation rate of $A$. In fact, a linear CI as in the statement of Theorem \[thm:lcif\] is given by $\phi(x)=x_B/\kappa_1$, and at any time point $$\label{eq:insulator_CI}
\frac{d}{dt}\phi(x(t))=x_B(t)\left(x_A(t)-\frac{\kappa_2}{\kappa_1}x_{A^\star}(t)\right).$$ We further note that if the dynamics of (\[eq:insulator\]) occur on a faster time scale than the rest of the system, then a slow manifold is quickly approached were the concentration of the species $A$ is maintained at the level $\kappa_2 x_{A^\star}(t)/\kappa_1$ at any time point. In this case, the module (\[eq:insulator\]) approximately outputs a multiple of the concentration of $A^\star$ over the whole time line. The multiplicative constant can be tuned through the parameters $\kappa_1$ and $\kappa_2$, as well as the time scale that (\[eq:insulator\]) operates in. The time scale can be further tuned via the concentration of $B$, as it also follows from . In conclusion, the downstream module receives as input a good approximation of a multiple of the concentration of $A^\star$, and its activity does not affect the upstream module, nor (\[eq:insulator\]). Moreover, (\[eq:insulator\]) does not affect the upstream module at all, since the species $A$ appears in (\[eq:insulator\]) as a catalyst and is not changed in the catalysed reaction. The proposed insulating strategy is illustrated in Figure \[fig:insulation\], and it is applied to an example discussed in [@DNS:modular] in Figure \[fig:application\].
Isolated systems (with $x_{\overline{A}^\star}(t)=x_{A^\star}(t)$)
(inputA) ; (upstream) ; (outputA) ; (inputB) ; (downstream) ; (outputB) ; (inputA) – node [$u(t)$]{} (upstream); (upstream) – node [$x_{A^\star}(t)$]{} (outputA); (inputB) – node [$x_{\overline{A}^\star}(t)$]{} (downstream); (downstream) – node [$x_C(t)$]{} (outputB);
{width="\textwidth"}
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
Coupled systems
(inputA) ; (upstream) ; (downstream) ; (outputB) ; (inputA) – node [$u(t)$]{} (upstream); (\[xshift=0.2cm\]upstream.south) – node [$x_{A^\star}(t)$]{} (\[xshift=0.2cm\]downstream.north); (\[xshift=-0.2\]downstream.north) – node (\[xshift=-0.2\]upstream.south); (downstream) – node [$x_C(t)$]{} (outputB);
{width="\textwidth"}
{width="\textwidth"}
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
Insulated system
(inputA) ; (upstream) ; (insulator) ; (downstream) ; (outputB) ; (inputA) – node [$u(t)$]{} (upstream); (upstream) – node [$x_{A^\star}(t)$]{} (insulator); (\[xshift=0.2cm\]insulator.south) – node [$x_{A}(t)$]{} (\[xshift=0.2cm\]downstream.north); (\[xshift=-0.2\]downstream.north) – node (\[xshift=-0.2\]insulator.south); (downstream) – node [$x_C(t)$]{} (outputB);
{width="\textwidth"}
{width="\textwidth"}
Inclusion in larger systems {#sec:large}
---------------------------
In the previous section, we have seen how the stability properties of an ACR system can be transferred, in a sense, to a larger model including further chemical transformations and external inputs. As a particular case, Here, we explicitly state when the property of absolute concentration robustness can be lifted from portions of the biochemical system to the whole system. As a consequence, we further extend the set of sufficient conditions of Theorem \[thm:feinberg\_weak\] that imply the existence of an ACR species. Before stating the relevant result, which is a consequence of Theorem \[thm:lcif\], we need a definition. Given a reaction system ${\mathscr{S}}$, we say that ${\mathscr{S}}'$ is a sub-system of ${\mathscr{S}}$ if it can be obtained from ${\mathscr{S}}$ by canceling some reactions, and if the choice of rate functions for the remaining reactions is maintained. Moreover, if ${\mathscr{S}}$ and ${\mathscr{S}}'$ have $d$ and $d'$ species, respectively, we let $\pi:{\mathbb{R}}^d\to{\mathbb{R}}^{d'}$ be the projection onto the species of ${\mathscr{S}}'$. The following holds:
[corollary]{}[thmextension]{}\[thm:extension\] Consider a reaction system ${\mathscr{S}}$, and let ${\mathscr{S}}'$ be a sub-system. Assume that ${\mathscr{S}}'$ is a mass-action system with two complexes $\pi(y_i)$ and $\pi(y_j)$ only differing in the entry relative to the species $X$, and for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Moreover, assume there exists $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ such that $\pi(y_l-y_k)$ is orthogonal to $\hat\gamma$ for all $y_k\to y_l$ that are reactions of ${\mathscr{S}}$ but not reactions of ${\mathscr{S}}'$. Then, the $X$ is ACR for both ${\mathscr{S}}$ and ${\mathscr{S}}'$, with the same ACR value.
The proof of a stronger result is in the Supplementary Material. Here we illustrate how the corollary can be applied in the case of EnvZ-OmpR signaling system: consider the reaction system ${\mathscr{S}}$ described in Figure \[fig:larger\], which includes the EnvZ-OmpR osmoregulation system described in Figure \[fig:envz\] as a sub-system. We assume that a protein can misfold when the phosphoryl group is transferred from EnvZ to OmpR. Such misfold can be corrected by chaperons, which are independently produced and degraded through a mechanism which we assume unkown, but that does not involve EnvZ or OmpR proteins. We also allow for an arbitrary and persistent external control on the expression level of EnvZ sensor-transmitter protein. Finally, we consider the utilization of OmpR-P as transcription regulatory protein of the outer membrane porins OmpF and OmpC. For our purposes, we assume the details of the transcription mechanism are not known, but that only the protein Ompr-P is involved in the process. As previously done, let the complexes of the EnvZ-OmpR osmoregulation system be ordered from left to right and from top to bottom, such that EnvZ-D and EnvZ-D+OmpR-P are the 1st and the 8th complex, respectively. Also, let the species be ordered according to their appearance from left to right and from top to bottom, as EnvZ-D, EnvZ, EnvZ-T, EnvZ-P, OmpR, EnvZ-OmpR-P, OmpR-P, EnvZ-OmpR-D-P. In particular, Envz is the second species, EnvZ-OmpR-P is the sixth species, and OmprR-P is the seventh species. It follows from that, by choosing $w_1=-\xi_2$ and $w_2=-\xi_7=0$, a vector $\hat\zeta$ is in $\hat\Gamma_{18}(\kappa)$ with:
1. \[1\] $\hat\zeta_2=\hat\xi_2+w_1=0$;
2. \[2\] $\hat\zeta_6-\hat\zeta_2=\hat\xi_6+w_1+w_2-\hat\xi_2-w_1=\hat\xi_6-\hat\xi_2=0$;
3. \[3\] $\hat\zeta_7=\hat\xi_7+w_2=0$.
Denote by $E_k$ the vector of ${\mathbb{R}}^d$ with the $k$th entry equal to 1 and the other entries equal to zero. The following holds.
Misfolding of OmpR-P.
: The projection of the difference between EnvZ + OmpR$^\star$-P and EnvZ-OmpR-P onto the species of the EnvZ-OmpR signaling system is $E_2-E_6$, which is orthogonal to $\hat\zeta$ by \[2\]. The projection of the difference between OmpR-P + C and OmpR$^\star$-P + C is $E_7$, which is orthogonal to $\hat\zeta$ by \[3\].
Production and degradation of chaperons.
: By assumption, any chemical reaction $y\to y'$ involved in the production and degradation of chaperones does not consume or produce any chemical species of the EnvZ-OmpR signaling system. Hence, $\pi(y'-y)=0$, which is orthogonal to $\hat\zeta$.
External regulation of EnvZ.
: The difference between EnvZ and 0 is $\pm E_2$, which is orthogonal to $\hat\zeta$ by \[1\]
Transcription of OmpC and OmpF.
: We assume that the transcription only involves the species OmpR-P, out of all the species in the EnvZ-OmpR osmoregulation system. Hence, for all the reactions $y\to y'$ involved in the transcription, either $\pi(y'-y)=0$ or $\pi(y'-y)$ is a multiple of $E_7$. In either case, $\pi(y'-y)$ is orthogonal to $\hat\zeta$.
It follows from Corollary \[thm:extension\] that the species OmpR-P is ACR in the reaction system of Figure \[fig:larger\]. Moreover, its ACR value is still given by , as long as a positive steady state exists. Note that Corollary \[thm:extension\] could be applied even if not all chemical reactions are known, and even if the model is not mass-action. It is also worth noting that the deficiency of the model is not known, due to the lack of information on the precise reactions constituting the network, but is certainly greater than 1. Indeed, the deficiency of the sub-system constituted by the EnvZ-OmpR osmoregulation system and by the misfolding of OmpR-P is 2, and the deficiency of a system is necessarily greater than or equal to the deficiency of any sub-system [@CW:poisson Lemma 5].
EnvZ-OmpR osmoregulation system, as described in Figure \[fig:envz\]
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
Misfolding of sensory response protein OmpR-[[**P**]{}]{}, and recovery through the action of chaperones: $$\schemestart
EnvZ-OmpR-{{\color{red}\bfseries P}}\arrow{->[*0{$\kappa_{12}$}]} EnvZ \+ {\color{teal} OmpR$^\star$}-{{\color{red}\bfseries P}}\arrow(@c1.south east--.north east){0}[-90,.3]
{\color{teal} OmpR$^\star$}-{{\color{red}\bfseries P}}\+ C \arrow{->[*0{$\kappa_{13}$}]} OmpR-{{\color{red}\bfseries P}}\+ C
\schemestop$$
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
Production and degradation of chaperones:
(chap\_reg) [Gene regulatory network]{}; (C) [C]{}; (0) [0]{}; (chap\_reg) – (C); (C) – node\[above\][$u_C(t)$]{} (0);
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
External regulation of the sensor-transmitter protein EnvZ: $$\schemestart
0 \arrow{<=>[*0{$u_1(t)$}][*0{$u_2(t)$}]} EnvZ
\schemestop$$
(0,0) – (,0.5); (0-@leftlower-@boxsep,0.25)– (+@rightlower+@boxsep,0.25);
Transcription of OmpC and OmpF:
(OmpR) [OmpR-[[**P**]{}]{} binds to DNA promoter]{}; (foo) ; (OmpC) [OmpC]{}; (OmpF) [OmpF]{}; (OmpR) – (OmpC); (OmpR) – (OmpF);
Discussion
==========
We have shown in Theorem \[thm:lcif\] that a linear CI always exists for a family of ACR systems, and that this family strictly includes the models studied in [@SF:ACR]. The result, a more general version of which is proven in the Supplementary Material, has three main consequences: first, it provides an easy way to calculate the ACR value of ACR species. Secondly, as expressed in Theorem \[thm:control\], the presence of a linear CI implies that the system is robust to arbitrary disturbances that do not vanish over time, under certain conditions. This fact can be naturally exploited to design perfect insulators, which are able to reject loading effects originated from the downstream modules. Finally, as expressed in Corollary \[thm:extension\], we are able to prove that, under certain conditions, the absolute concentration robustness of a portion of a system can be lifted to the whole model, and the ACR value of the ACR species remains unchanged.
The theory we developed opens the path for future research directions. First, efficient algorithms can be designed in order to check for the existence of portions of the systems that confer absolute concentration robustness to the whole system. To this aim, the connections of $\hat\Gamma_{ij}(\kappa)$ with structural properties of the network which we show in the Supplementary Material can be useful, and theoretical results can be expanded in this direction. Second, further detailed analysis on when stability can be ensured would be welcome. Currently, in the statement of Theorem \[thm:control\] we cannot exclude the possibility that some species is completely consumed or indefinitely produced upon tampering with the model. Finding structural conditions able to eliminate this possibility would be a nice and useful contribution.
As a final remark, we think the study of stochastically modeled systems that satisfy the assumptions of Theorem \[thm:lcif\] would be interesting and fruitful. Stochastic models of reaction systems are tipycally used when few molecules of certain chemical species are available [@AK:book; @ET:book]. It is proven in [@AEJ:ACR] that systems satisfying the assumptions of Theorem \[thm:feinberg\_weak\], when stochastically modeled, undergo an extinction event almost surely. As a consequence, the desirable robustness properties of the ACR systems studied in [@SF:ACR] are completely destroyed in a low molecule copy-number regime. As an example, the model depicted in (\[eq:toymodel\]) undergoes an almost sure extinction of the chemical species $B$ when stochastically modeled, regardless the initial conditions. This is caused by the fact that all the molecules of $B$ can be consumed by the reaction $B\to A$, before the occurrence of a reaction $A+B\to 2B$. Robustness at finite time intervals of some stochastically modeled ACR systems is recovered, but only in a multiscale limit sense [@ACK:ACR]. Moreover, it is shown in [@AC:ACR] that absolute concentration robustness of the deterministic model does not necessarily imply an extinction event in the corresponding stochastic model, but the connection is still largely unexplored. The results developed in the present paper can help in this direction: consider again (\[eq:toymodel\]). The extinction of species $B$ cannot occur if production of $B$ is included in the model as in (\[eq:toymodel\_modified\]), or as in $$\label{eq:toymodel_modified_2}
\begin{split}
\schemestart
$A+B$\arrow{->[$\kappa_1$]}$2B$
\arrow(@c1.south east--.north east){0}[-90,.25]
$B$\arrow{->[$\kappa_2$]}$A$
\arrow(@c3.south east--.north east){0}[-90,.25]
$0$\arrow{<=>[$\kappa_3$][$\kappa_4$]}$B$
\schemestop
\end{split}$$ At the same time, it follows from Theorem \[thm:control\] that the stability properties of the species $A$ are maintained both in (\[eq:toymodel\_modified\]) and in (\[eq:toymodel\_modified\_2\]), when deterministically modeled. In particular, the concentration of the species $A$ still converges to the value $\kappa_2/\kappa_1$. It would be interesting to study if in this and in similar cases some form of absolute concentration robustness arise in the long-term dynamics of the stochastic models as well.
Acknowledgement {#acknowledgement .unnumbered}
===============
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme grant agreement no. 743269 (CyberGenetics project).
[**Supplementary Material**]{}
Notation {#sec:notation}
========
General notation
----------------
We will denote by ${\mathbb{R}}$, ${\mathbb{R}}_{>0}$, and ${\mathbb{R}}_{\geq0}$ the real, positive real, and non-negative real numbers, respectively. Similarly, we will denote by ${\mathbb{Z}}$, ${\mathbb{Z}}_{>0}$, and ${\mathbb{Z}}_{\geq0}$ the integer, positive integer, and non-negative integer numbers, respectively. Given a real number $a$, we will denote its absolute value by $|a|$.
We denote by $e_i$ the vector that has 1 in its $i$th entry and 0 in all other entries. The dimension of such vector will be clear from the context, and if not it will be made explicit. Given two real vectors $v$, $w$ of the same length $n$, we denote their scalar product by ${\langle v, w\rangle}$, and we use the shorthand notation $$v^w=\prod_{i=1}^n v_i^{w_i},$$ where $0^0$ is considered to be 1. We will denote the euclidean norm of $v$ by $\|v\|$.
Given two subsets $V, W\subseteq{\mathbb{R}}^n$, we denote their direct sum by $$V\oplus W=\{v+w\,:\,v\in V, w\in W\}.$$ Moreover, given a vector $v\in{\mathbb{R}}^n$, we define $$v+W=\{v+w\,:\,w\in W\}.$$
Reaction network terminology
----------------------------
### Standard notation
A reaction network ${\mathcal{G}}$ is a triple $\{{\mathcal{X}}, {\mathcal{C}},{\mathcal{R}}\}$, where
- ${\mathcal{X}}$ is a finite ordered set of $d$ different symbols, called *species*;
- ${\mathcal{C}}$ is a finite ordered set of $m$ linear combinations of species with non-negative integer coefficients, referred to as *complexes* and identified with vectors in ${\mathbb{Z}}_{\geq0}^d$;
- ${\mathcal{R}}$ is a finite ordered set of elements of ${\mathcal{C}}\times{\mathcal{C}}$, referred to as *reactions*.
We will denote by $X_i$ the $i$th species of ${\mathcal{X}}$, and by $y_j$ the $j$th complex of ${\mathcal{C}}$, for $1\leq i\leq d$ and $1\leq j\leq m$. A reaction $(y_i,y_j)$ will be denoted by $y_i\to y_j$, and for any $y_i\in{\mathcal{C}}$ we assume there is no reaction of the form $y_i\to y_i$.
As mentioned, a complex $y_i$ can be regarded as a vector of ${\mathbb{Z}}^d_{\geq0}$. Specifically, this is done by considering the $j$th entry $y_{ij}$ as the coefficient of $y_i$ relative to the $j$th species. In this regards, for any vector $v\in{\mathbb{R}}^d$ we will denote by $\operatorname{supp}v$ the subset of species such that $$X_i\in\operatorname{supp}v\text{ if and only if }v_i\neq0.$$ Similarly, for any vector $w\in{\mathbb{R}}^m$ we will denote by $\operatorname{supp}w$ the subset of complexes such that $$y_i\in\operatorname{supp}v\text{ if and only if }v_i\neq0.$$
We denote by $${\mathcal{S}}=\operatorname{span}_{{\mathbb{R}}}\{y_j-y_i\,:\,y_i\to y_j\in{\mathcal{R}}\},\qquad {{\mathcal{S}}^\perp}=\{h\in{\mathbb{R}}^d\,:\,{\langle h, y_j-y_i\rangle}=0\text{ for all }y_i\to y_j\in{\mathcal{R}}\}.$$ The subspace ${\mathcal{S}}$ is called *stoichiometric subspace*, and the elements of ${{\mathcal{S}}^\perp}$ are called *conservation laws*.
A choice of kinetics for a reaction network is a set of (time-dependent) *rate functions* $\lambda_{ij}\colon{\mathbb{R}}^{d}_{\geq0}\times{\mathbb{R}}_{\geq0}\to {\mathbb{R}}_{\geq0}$ for all $1\leq i,j\leq m$, such that $\lambda_{ij}$ is a zero function if and only if $y_i\to y_j\notin{\mathcal{R}}$. A reaction network with a choice of kinetics ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\{\lambda_{ij}\}_{1\leq i,j\leq m})$ is termed *reaction system*. A reaction system is associated with the system of differential equations $$\frac{d}{dt}x(t)=\sum_{1\leq i,j\leq m}(y_j-y_i) \lambda_{ij}(x(t),t)\quad \text{for all }t\in{\mathbb{R}}_{\geq0}.$$ We note that rate functions are commonly intended to not depend on time, but we consider this more general setting in this paper.
A reaction system is called *mass-action system*, and denoted by ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, if there is a matrix $\kappa\in{\mathbb{R}}^{m\times m}_{\geq0}$ such that $$\lambda_{ij}(x,t)=\kappa_{ij}x^{y_i}\quad\text{for all }1\leq i,j\leq m, x\in{\mathbb{R}}^{d}_{\geq0}.$$ In this case, the constants $\kappa_{ij}$ are termed *rate constants*. Define $\Lambda(x)\in{\mathbb{R}}^{d}_{\geq0}$ by $$\Lambda_i(x)=x^{y_i}\quad\text{for all }1\leq i\leq m, x\in{\mathbb{R}}^{d}_{\geq0},$$ and define the $m\times m$ matrix $A(\kappa)$ by $$A(\kappa)_{ij}=\begin{cases}
\kappa_{ji}&\text{if }i\neq j\\
-\sum_{l=1}^m \kappa_{il} &\text{if }i=j
\end{cases}$$ Finally, let $Y$ be the $d\times m$ matrix $Y$ with entries $Y_{ij}=y_{ji}$. Then, for mass-action sytems we have $$\frac{d}{dt}x(t)=\sum_{1\leq i,j\leq m}(y_j-y_i) \kappa_{ij}x(t)^{y_i}=Y A(\kappa) \Lambda(x(t))\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$
The directed graph $\{{\mathcal{C}}, {\mathcal{R}}\}$ is called *reaction graph*. Reaction systems are often presented through the reaction graph, where indication on the reaction rates are written on top of the arrows that correspond to the related reaction. Specifically, the arrow corresponding to $y_i\to y_j$ is labeled with $$\frac{\lambda_{ij}(x,t)}{x^{y}},$$ which corresponds to the rate constants for reaction rates of mass-action type. As an example, see and in the main text, or Figure \[fig:envz\]. We denote by $\ell$ the number of connected components of the reaction graph. We define the *deficiency* as the number $$\delta=m-\ell-\dim {\mathcal{S}}.$$ A *terminal component* of the reaction graph is a set of complexes $\mathcal{T}\subseteq{\mathcal{C}}$, such that
- if $y\in\mathcal{T}$ and there is a directed path in the reaction graph from $y$ to another complex $y'$, then $y'\in\mathcal{T}$;
- for any two complexes $y,y'\in\mathcal{T}$ with $y\neq y'$, there is a directed path from $y$ to $y'$, and a directed path from $y'$ to $y$.
Let $\tau$ be the number of terminal components, and denote by $\mathcal{T}^1, \mathcal{T}^2, \dots, \mathcal{T}^\tau$ the different terminal components of the network. Since each connected component contains at least one terminal component, we have $\tau\geq\ell$. We say a complex is *terminal* if it is contained in a terminal component, and we say that a complex is *non-terminal* otherwise. As an example, the terminal components in are $\{2B\}$ and $\{A\}$, hence the terminal complexes are $2B$ and $A$.
Absolute concentration robustness
---------------------------------
Consider a reaction system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}}, \{\lambda_{ij}\}_{1\leq i,j\leq m})$. A species $X_n\in{\mathcal{X}}$ is said to be *absolutely concentration robust* (ACR) if there exists $q\in{\mathbb{R}}_{>0}$ such that $c_n=q$ for all positive steady state $c$ of ${\mathscr{S}}$. In this case, $q$ is referred to as the *ACR value* of $X_n$. Finally, if an ACR species exists, then the reaction system ${\mathscr{S}}$ is said to be *ACR*.
\[rem:degenerate\_ACR\] By definition, if a reaction system has no positive steady states, or a unique one, then all the species are ACR. We can call this a *degenerate case*, since, especially if no positive steady states exist, the concept of robustness is lost.
Control Theory terminology
--------------------------
Consider a differential equation of the form $$\frac{d}{dt} x(t)=f(x(t))+g(x(t), u(t))\quad\text{for all }t\in{\mathbb{R}}_{\geq0},$$ where $x\colon{\mathbb{R}}_{\geq0}\to {\mathbb{R}}^d$ and $u\colon{\mathbb{R}}_{\geq0}\to{\mathbb{R}}^{n_u}$ for some $n_u\in{\mathbb{Z}}_{>0}$, and $f$ and $g$ are differentiable functions. The function $u$ is called the *input of the system*. Further, we define *output of the system* the quantity $z(t)=a(x(t))$, for some differentiable function $a$, with $a\colon{\mathbb{R}}^{d}\to{\mathbb{R}}^{n_z}$ and $n_z\in{\mathbb{Z}}_{>0}$. Usually, one needs to find an appropriate function $u$ such that $z$ is close to a desired level $\overline{z}\in{\mathbb{R}}^{n_z}$, either on average or for $t\to\infty$. To this aim, the existence of a function $\phi\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$ such that $$\frac{d}{dt}\phi(x(t))=z(t)-\overline{z}$$ is of high importance, an is called an *integral feedback* (IF) [@doyle:feedback; @astrom:feedback]. If a function $\tilde\phi\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$ satisfies $$\frac{d}{dt}\tilde\phi(x(t))=r(x(t))\Big(z(t)-\overline{z}\Big)$$ for some differentiable function $r\colon {\mathbb{R}}^{n_x}\to{\mathbb{R}}$, then $\tilde\phi$ is called a *constrained integral feedback* (CIF) [@XD:robust].
In the setting of ACR species, $z(t)$ is usually the concentration of the ACR species at time $t$, and $\overline{z}$ is its ACR value.
Known results
=============
The following result appears in the Appendix of [@FH:1977].
\[thm:support\] All the vectors in $\ker A(\kappa)$ have support in the terminal complexes. Specifically, there exists a basis $\{\chi^1(\kappa), \chi^2(\kappa), \dots, \chi^\tau(\kappa)\}$ of $\ker A(\kappa)$ such that $\operatorname{supp}\chi^i(\kappa)=\mathcal{T}^i$ for all $1\leq i\leq m$.
The following result can be deduced from Section 6 in [@FH:1977].
\[thm:deficiency\] We have $$\delta\geq\dim\Big(\ker Y\cap \operatorname{Im}A(\kappa)\Big)=\dim\ker YA(\kappa) - \dim\ker A(\kappa).$$
Finally, the following result is proven in the supplementary material of [@SF:ACR], where it is stated as Theorem S3.15.
\[thm:feinberg\_strong\] Consider a mass-action system $({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$. Assume the following conditions hold:
1. $y_i$ and $y_j$ are non-terminal complexes;
2. the deficiency is 1;
Then, there exists $q\in{\mathbb{R}}_{>0}$ such that $c^{y_j-y_i}=q$ for all positive steady states $c$.
For convenience, we restate here the weaker version given in the main text: Note how Theorem \[thm:feinberg\_weak\] is a straightforward consequence of Theorem \[thm:feinberg\_strong\]: if the two complexes only differ in the $n$th entry, then $$c^{y_j-y_i}=c_n^{(y_j-y_i)_n}=q$$ for all positive steady states $c$, hence $X_n$ is ACR.
In the original statement of Theorem \[thm:feinberg\_strong\], the additional hypothesis that a positive steady state exists is made. In fact, under the condition of Theorem \[thm:feinberg\_strong\], the existence of positive steady states is not guaranteed. However, if no positive steady state exists then the conclusion of the theorem holds trivially. Hence, our reformulation holds correct. See Remark \[rem:degenerate\_ACR\] for the formal relationship between ACR species and the lack of positive steady states.
Calculation and structural properties of $\Gamma_{ij}(\kappa)$
==============================================================
Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}}, \kappa)$. For any $1\leq i,j\leq m$ with $i\neq j$ define the set $$\label{eq:GammaAp}
\Gamma_{ij}(\kappa)=\left\{\gamma\in{\mathbb{R}}^{d+1}\,:\,
\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\gamma
=e_j\right\},$$ and for any real vector $v$ of length larger than $d$, let $\pi_d(v)$ be its projection onto the first $d$ components. We define $$\label{eq:gamma_hat}
\hat \Gamma_{ij}(\kappa)=\left\{\hat\gamma\in{\mathbb{R}}^d\,:\,\hat\gamma=\pi_d(\gamma)\text{ for some }\gamma\in \Gamma_{ij}(\kappa)\right\}.$$ The set $\Gamma_{ij}(\kappa)$ can be computed by first calculating a basis for $$\label{eq:Psi}
\Psi_{ij}(\kappa)=\ker \begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i \,|\, e_j
\end{pmatrix}.$$ Note that this can be easily and quickly done by using a programming language which is able to deal with symbolic linear algebra, such as Matlab. The following holds
\[prop:calculating\_Gamma\] Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, and $1\leq i,j\leq m$ with $i\neq j$. Let $\{\psi^1, \psi^2, \dots, \psi^k\}$ be a basis for $\Psi_{ij}(\kappa)$, as defined in . Then, $\Gamma_{ij}(\kappa)$ (and consequently $\hat \Gamma_{ij}(\kappa)$) is non-empty if and only if $\psi^n_{d+2}\neq 0$ for some $1\leq n\leq k$. If this is the case, then $$\begin{aligned}
\Gamma_{ij}(\kappa)&=\left\{-\frac{1}{\sum_{n=1}^k a_n\psi^n_{d+2}}\sum_{n=1}^k a_n\pi_{d+1}(\psi^n)\,:\,a_1,a_2,\dots,a_k\in{\mathbb{R}}\text{ and }\sum_{n=1}^k a_n\psi^n_{d+2}\neq 0\right\},\\
\intertext{where $\pi_{d+1}\colon{\mathbb{R}}^{d+2}\to {\mathbb{R}}^{d+1}$ is the projection onto the first $d+1$ components, and}
\hat\Gamma_{ij}(\kappa)&=\left\{-\frac{1}{\sum_{n=1}^k a_n\psi^n_{d+2}}\sum_{n=1}^k a_n\pi_{d}(\psi^n)\,:\,a_1,a_2,\dots,a_k\in{\mathbb{R}}\text{ and }\sum_{n=1}^k a_n\psi^n_{d+2}\neq 0\right\}.
\end{aligned}$$
First, note that $\gamma\in\Gamma_{ij}(\kappa)$ if and only $(\gamma, -1)\in\Psi_{ij}(\kappa)$. This can be easily deduced by the definitions and . The proof simply follows from this equivalence, and from the definition of $\hat \Gamma_{ij}(\kappa)$ given in .
In Section \[sec:models\] we will use Proposition \[prop:calculating\_Gamma\] to calculate $\hat\Gamma_{ij}(\kappa)$ for the main examples discussed in the main text. We will see that some vectors in the basis of $\Psi_{ij}(\kappa)$ do not depend on the particular choice of rate constants. As a consequence, some dynamical properties implied by the theory developed in this paper will only depend on the structure of the model rather than on a fine tuning of the parameters, which is desirable. In the following result, we explicitly derive structural properties of $\hat\Gamma_{ij}(\kappa)$.
\[prop:structure\_Gamma\] Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, and $1\leq i,j\leq m$ with $i\neq j$. Assume that $\hat\Gamma_{ij}(\kappa)$ is non-empty, and let $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$. Then, $$\label{eq:containment}
\hat\gamma+{{\mathcal{S}}^\perp}\subseteq\hat\gamma+\ker A(\kappa)^\top Y^\top\subseteq \hat\Gamma_{ij}(\kappa).$$ Moreover, if the mass-action system has a steady state $c$ with $c^{y_i}>0$, then $$\label{eq:same_q}
\gamma'_{d+1}=c^{y_j-y_i}\quad\text{for all }\gamma'\in\Gamma_{ij}(\kappa)$$ and $$\label{eq:equality1}
\hat\Gamma_{ij}(\kappa)=\hat\gamma+\ker A(\kappa)^\top Y^\top.$$ Finally, if each connected component of the reaction graph contains exactly one terminal component, then $$\label{eq:equality2}
\\ker A(\kappa)^\top Y^\top={{\mathcal{S}}^\perp}.$$
First, we have that for any $\hat \gamma\in \hat\Gamma_{ij}(\kappa)$ $$\label{eq:containment_ker}
\hat\gamma+\ker A(\kappa)^\top Y^\top\subseteq\hat\Gamma_{ij}(\kappa).$$ Indeed, for any $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$, there exists $\gamma\in\Gamma_{ij}(\kappa)$ with $\hat\gamma=\pi_d(\gamma)$. Then, for any for any $v\in\ker A(\kappa)^\top Y^\top$ we have $$\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\left(\gamma+\begin{pmatrix} v \\ 0\end{pmatrix}\right)
=\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\gamma + \begin{pmatrix}A(\kappa)^\top Y^\top v \\ 0\end{pmatrix}
=e_j.$$ Hence, $$\gamma+\begin{pmatrix} v \\ 0\end{pmatrix}\in\Gamma_{ij}(\kappa)$$ which implies that $\hat\gamma+v\in\hat\Gamma_{ij}(\kappa)$ and proves . Since for any $h\in{{\mathcal{S}}^\perp}$ and any $1\leq n\leq m$ $$\Big(A(\kappa)^\top Y^\top h\Big)_n=\sum_{l=1}^m {\langle h, y_l-y_n\rangle}\kappa_{nl}=0,$$ it follows $$\label{eq:cons_perp}
{{\mathcal{S}}^\perp}\subseteq \ker A(\kappa)^\top Y^\top.$$ follows from and .
Now assume that the mass-action system has a steady state $c$, with $c^{y_i}>0$. Then, $$0=\frac{d}{dt}{\langle \hat\gamma, x(t)\rangle}|_{x(t)=c}=\hat\gamma^\top YA(\kappa)\Lambda(c)=(e_j-\gamma_{d+1}e_i)\Lambda(c)=c^{y_j}-\gamma_{d+1}c^{y_i}.$$ Since $c^{y_i}>0$, then necessarily $\gamma_{d+1}=c^{y_j-y_i}$. For the same argument, for any other $\gamma'\in\Gamma_{ij}(\kappa)$, $\gamma'_{d+1}=\gamma_{d+1}=c^{y_j-y_i}$. Hence, holds. It follows that any $\gamma'\in\Gamma_{ij}(\kappa)$ is of the form $$\gamma'=\gamma+\begin{pmatrix} v \\ 0\end{pmatrix}$$ for some $v\in{\mathbb{R}}^{d}$. Moreover, since $\gamma'\in\Gamma_{ij}(\kappa)$ $$e_j=\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\gamma'
=\begin{pmatrix}
A(\kappa)^\top Y^\top\,|\, e_i
\end{pmatrix}\gamma + \begin{pmatrix}A(\kappa)^\top Y^\top v \\ 0\end{pmatrix}= e_j+\begin{pmatrix}A(\kappa)^\top Y^\top v \\ 0\end{pmatrix}.$$ Hence, necessarily $v\in\ker A(\kappa)^\top Y^\top$, which implies $$\hat\Gamma_{ij}(\kappa)\subseteq\hat\gamma+\ker A(\kappa)^\top Y^\top.$$ The latter, together with , implies .
To conclude the proof, we need to show that if each connected component of the reaction graph contains exactly one terminal component (i.e. if $\ell=\tau$), then holds. This follows from and $$\begin{aligned}
\dim {{\mathcal{S}}^\perp}&= d-\dim{\mathcal{S}}= d+\ell+\delta-m\\
&=d+\tau+\delta-m\geq d-m+\dim\ker YA(\kappa)\\
&=\dim \ker A(\kappa)^\top Y^\top,
\end{aligned}$$ where we utilized Theorems \[thm:support\] and \[thm:deficiency\] for the forth equality.
As a consequence, we have the following.
\[cor:structural\] Consider a mass-action system, and assume that $y_i$ and $y_j$ are two complexes for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Let $\{v_p\}_{p=1}^{H}$ be a basis for ${{\mathcal{S}}^\perp}$, where $H=d-\dim{\mathcal{S}}$. Moreover, let $X_{l_1}, X_{l_2}, \dots, X_{l_n}\in{\mathcal{X}}$ such that the rank $H\times n$ matrix $V$ has rank $n$, where $V_{pq}=v_{p l_q}$ for all $1\leq p\leq H$ and $1\leq q\leq n$. Then, there exists $\hat\gamma\in \hat\Gamma_{ij}(\kappa)$ such that ${\langle \hat\gamma, e_p\rangle}=0$ for all $1\leq p\leq n$.
Let $\hat\gamma^\star\in \hat\Gamma_{ij}(\kappa)$. It follows from Proposition \[prop:structure\_Gamma\] that for all $w\in {\mathbb{R}}^H$ $$\hat\gamma^\star+\sum_{p=1}^H w_pv_p\in \hat\Gamma_{ij}(\kappa).$$ Hence, the proof is concluded by choosing $w$ such that $$w^\top V=-(\hat\gamma^\star_{l_1},\hat\gamma^\star_{l_2},\dots,\hat\gamma^\star_{l_n}),$$ which is possible because the $H\times n$ matrix $V$ has rank $n$.
Further useful structural property of $\hat\Gamma_{ij}(\kappa)$ are following. Such properties are useful while checking whether the results of this paper can be applied, and can be used by an algorithm designed to this aim. Before stating the structural results, it is convenient to prove the following lemma.
\[lem:structural\] Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, and $1\leq i,j\leq m$ with $i\neq j$. Then, $\hat\Gamma_{ij}(\kappa)$ is non-empty if and only if $e_j-\gamma_{d+1}e_i\in(\ker YA(\kappa))^\top$ for some $\gamma_{d+1}\in{\mathbb{R}}$. Moreover, if that is the case and if the mass-action system has a steady state $c$ with $c^{y_i}>0$ and $c^{y_j}>0$, then necessarily $\gamma_{d+1}>0$.
By definition of $\hat\Gamma_{ij}(\kappa)$ given in , $\hat\Gamma_{ij}(\kappa)$ is non-empty if and only if $\Gamma_{ij}(\kappa)$ is non-empty. By , $\Gamma_{ij}(\kappa)$ is non-empty if and only if there exists $\gamma_{d+1}\in{\mathbb{R}}$ such that $e_j-\gamma_{d+1}e_i$ is in $\operatorname{Im}A(\kappa)^\top Y^\top$. By the fundamental theorem of linear algebra, the latter holds if and only if $e_j-\gamma_{d+1}e_i$ is orthogonal to $\ker YA(\kappa)$.
To conclude the proof, we note that if the mass-action system has a steady state $c$ with $c^{y_i}>0$ and $\hat\Gamma_{ij}(\kappa)$ is non-empty, then it follows by in Proposition \[prop:structure\_Gamma\] that the quantity $\gamma_{d+1}$ discussed in the first part of the proof is necessarily equal to $c^{y_j-y_i}>0$.
\[prop:empty\_gamma\_terminal\] Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, and $1\leq i,j\leq m$. Assume that one of the following holds:
- $y_i$ is non-terminal and $y_j$ is terminal;
- $y_i$ and $y_j$ are in two different terminal components.
Then, $\hat\Gamma_{ij}(\kappa)$ is empty.
By assumption, $y_j$ is terminal. By Theorem \[thm:support\], there exists a vector $$\chi\in\ker A(\kappa)\subseteq \ker YA(\kappa)$$ such that $\operatorname{supp}\chi$ is the terminal component containing $y_j$. Hence, if $y_i$ is non-terminal or if $y_i$ is in a different terminal component than $y_j$, we have that for all $\gamma_{d+1}\in{\mathbb{R}}$ $${\langle \chi, e_j-\gamma_{d+1}e_i\rangle}=\chi_j\neq0.$$ It follows from Lemma \[lem:structural\] that $\hat\Gamma_{ij}(\kappa)$ is empty, which concludes the proof.
\[prop:empty\_gamma\_symmetric\] Consider a mass-action system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}},\kappa)$, and $1\leq i,j\leq m$ with $i\neq j$. Assume that a steady state $c$ with $c^{y_i}>0$ and $c^{y_j}>0$ exists. Then, $\hat\Gamma_{ij}(\kappa)$ is empty if and only if $\hat\Gamma_{ji}(\kappa)$ is empty.
By the symmetric role of $i$ and $j$, it suffices to prove that if $\hat\Gamma_{ij}(\kappa)$ is non-empty, then necessarily $\hat\Gamma_{ij}(\kappa)$ is non-empty.
By Lemma \[lem:structural\], if $\hat\Gamma_{ij}(\kappa)$ is non-empty then there exists $\gamma_{d+1}\in{\mathbb{R}}$ with $e_j-\gamma_{d+1}e_i\in(\ker YA(\kappa))^\top$, and $\gamma_{d+1}>0$. Hence, $$e_i-\frac{1}{\gamma_{d+1}}e_j\in(\ker YA(\kappa))^\top,$$ hence by Lemma \[lem:structural\] $\hat\Gamma_{ji}(\kappa)$ is non-empty, which is what we wanted to show.
Proofs of the results stated in the main text
=============================================
In this section, we state and prove more general versions of the theorems presented in the main text.
Existence and characterization of a CIF
---------------------------------------
The following result is stated in the main text.
The following more general result holds, from which Theorem \[thm:feinberg\_implies\_cond\] can be immediately deduced.
Consider a mass-action system, and assume the following holds:
1. \[part\_non\_terminal\_complexes\] $y_i$ and $y_j$ are two distinct non-terminal complex;
2. \[part\_deficiency\_1\] the deficiency is 1;
3. \[part\_existence\_of\_pos\_eq\] a steady state $c$ with $c^{y_i}\neq0$ exists.
Then, $\hat\Gamma_{ij}(\kappa)$ is non-empty.
By condition \[part\_existence\_of\_pos\_eq\], $c$ is a steady state, hence the vector $\Lambda(c)$ is in $\ker Y A(\kappa)$. Moreover, $\Lambda_i(c)=c^{y_i}\neq0$. By condition \[part\_non\_terminal\_complexes\], $y_i$ is a non-terminal complex. Hence, by combining $\Lambda_i(c)\neq0$ with Theorem \[thm:support\], it follows that $\Lambda(c)$ is not in $\ker A(\kappa)$. Since by assumption $\delta=1$, it follows from Theorem \[thm:deficiency\] that $$\label{eq:ker_direct_sum}
\ker Y A(\kappa)=\operatorname{span}_{\mathbb{R}}\{\Lambda(c)\}\oplus\ker A(\kappa).$$ Define $$v=e_j-c^{y_j-y_i}e_i=e_j-\frac{\Lambda_j(c)}{\Lambda_i(c)}e_i.$$ Clearly, $v$ is orthogonal to $\Lambda(c)$. Moreover, by condition \[part\_non\_terminal\_complexes\], both $y_i$ and $y_j$ are non-terminal complexes. Hence, the vector $v$ is orthogonal to $\ker A(\kappa)$ by Theorem \[thm:support\]. It follows from that $v$ is orthogonal to $\ker Y A(\kappa)$, which implies that $\hat\Gamma_{ij}(\kappa)$ is non-empty by Lemma \[lem:structural\].
We now proceed to prove the following result, stated in the main text. In order to prove the result, we will show that a more general version holds, which we state here.
\[thm:lcif\_app\] Consider a mass-action system, and assume that $y_i$ and $y_j$ are two complexes for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Let $\gamma\in\hat\Gamma_{ij}(\kappa)$. Then, $$\label{eq:fixed_value}
c^{y_j-y_i}=\gamma_{d+1}$$ for all steady states $c$ satisfying $c^{y_i}\neq 0$. Moreover, $$\phi(x)=\sum_{i=1}^d \beta_i x_i$$ is a linear CIF with $$\label{eq:to_be_proved}
\frac{d}{dt}\phi(x(t))=\Lambda_i(x(t)) \Big(x(t)^{y_j-y_i}-\gamma_{d+1}\Big)$$ for any initial condition $x(0)$ if and only if $\beta\in\hat\Gamma_{ij}(\kappa)$.
follows from in Proposition \[prop:structure\_Gamma\].
Let $\beta\in\hat\Gamma_{ij}(\kappa)$. Then, by Proposition \[prop:structure\_Gamma\] we have that $\beta-\pi_d(\gamma)\in\ker A(\kappa)^\top Y^\top$. Hence, for any initial condition $x(0)\in{\mathbb{R}}^d_{\geq0}$ and any $t\in{\mathbb{R}}_{\geq0}$, $$\frac{d}{dt}{\langle \beta, x(t)\rangle}=\beta^\top Y A(\kappa)\Lambda(x(t))=\pi_d(\gamma)^\top Y A(\kappa)\Lambda(x(t))=(e_j-\gamma_{d+1} e_i)^\top\Lambda(x(t))=x(t)^{y_i}\Big(x(t)^{y_j-y_i}-\gamma_{d+1}\Big),$$ which is .
Conversely, assume that holds. Then, for all $x\in{\mathbb{R}}^d_{\geq0}$ we have $$(\beta^\top -\pi_d(\gamma))^\top Y A(\kappa)\Lambda(x)=0.$$ Since the entries of $\Lambda$ are linearly independent monomials on ${\mathbb{R}}^d$, it must be $(\beta^\top -\pi_d(\gamma)) \in\ker A(\kappa)^\top Y^\top$, which implies that $\beta\in\hat\Gamma_{ij}(\kappa)$ by Proposition \[prop:structure\_Gamma\].
Rejection of persistent disturbances
------------------------------------
We recall here the formal definition of “oscillation”, as intended in this paper.
We say that a function $g\colon{\mathbb{R}}_{\geq0}\to{\mathbb{R}}$ *oscillates* around a value $\overline q\in{\mathbb{R}}$ if for each $t\in{\mathbb{R}}_{\geq0}$ there exist $t_+>t$ and $t_->t$ such that $$g(t_+)>q\quad\text{and}\quad g(t_-)<q.$$
The following result holds.
\[thm:control\_app\] Consider a mass-action system, with associated differential equation $$\frac{d}{dt} x(t)=f(x(t)).$$ Assume that $y_i$ and $y_j$ are two complexes for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Then, there exists $\overline q\in{\mathbb{R}}_{\geq0}$ such that $c^{y_j-y_i}=\overline q$ for all steady states $c$ with $c^{y_i}\neq0$. Consider an arbitrary function $u\colon {\mathbb{R}}_{\geq0}^d\times{\mathbb{R}}_{\geq0}\to {\mathbb{R}}^d$ such that a solution to $$\frac{d}{dt}\tilde x(t)=f(\tilde x(t))+u(\tilde x(t),t)$$ exists with $\tilde x(t)\geq0$ for all $t\geq0$. Assume that there exists a $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ which is orthogonal to the vector $u(x,t)$ for any $x\in{\mathbb{R}}^d_{\geq0}$, $t\in{\mathbb{R}}_{\geq0}$. Then, for any initial condition $\tilde x(0)\in{\mathbb{R}}^d_{\geq0}$, at least one of the following statements holds:
1. \[part:infty\] There is a species $X_k\in\operatorname{supp}\hat\gamma$ such that $\limsup_{t\to\infty}\tilde x_k(t)=\infty$;
2. \[part:zero\] There is a species $X_k\in\operatorname{supp}y_i$ such that $\liminf_{t\to\infty} \tilde x_k(t)=0$;
3. \[part:oscillation\] $\tilde x(t)^{y_j-y_i}$ oscillates around $\overline q$;
4. \[part:convergence\] $\displaystyle\lim_{t\to\infty}\int_{t}^\infty \Big|\tilde x(s)^{y_j-y_i}-\overline q\Big|ds=0$.
It follows from Theorem \[thm:lcif\_app\] that there exists $\overline q\in{\mathbb{R}}_{\geq0}$ such that $c^{y_j-y_i}=\overline q$ for all steady states $c$ with $c^{y_i}\neq0$. Note that in this setting $c^{y_i}\neq0$ is equivalent to $c^{y_i}>0$, since the state space is limited to vectors of non-negative concentrations.
Fix an initial condition $\tilde x(0)\in{\mathbb{R}}^d_{\geq0}$. Assume that \[part:infty\] does not occur. Hence, there exists $M\in{\mathbb{R}}_{>0}$ such that $$|{\langle \hat\gamma, \tilde x(t)\rangle}|<M\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$ By the fact that $\hat\gamma$ is orthogonal to $u(x,t)$ for all $x\in{\mathbb{R}}^d_{\geq0}$ and all $t\geq0$, and by Theorem \[thm:lcif\_app\], we have $$\frac{d}{dt}{\langle \hat\gamma, \tilde x(t)\rangle}={\langle \hat\gamma, f(\tilde x(t))\rangle}=\tilde x(t)^{y_i}\Big(\tilde x(t)^{y_j-y_i}-\overline q\Big).$$ It follows that $$\label{eq:integral_bounded}
\left|\int_0^t \tilde x(s)^{y_i}\Big(\tilde x(s)^{y_j-y_i}-\overline q\Big)ds\right|<M-|{\langle \hat\gamma, \tilde x(0)\rangle}|\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$ If \[part:zero\] does not hold, then there exists $m\in{\mathbb{R}}_{>0}$ such that $x(t)^{y_i}>m$ for all $t\in{\mathbb{R}}_{>0}$. Together with , this would imply that there exists $M^\star\in{\mathbb{R}}_{>0}$ such that $$\label{eq:difference_bound}
\left|\int_0^t \Big(\tilde x(s)^{y_j-y_i}-\overline q\Big)ds\right|<M^\star\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$ If \[part:oscillation\] does not hold, then there exists $t^\star\in{\mathbb{R}}_{>0}$ such that $\tilde x(t)^{y_j-y_i}-\overline q$ maintains the same sign for all $t>t^\star$, which together with implies \[part:convergence\]. The proof is then concluded.
Now we prove Theorem \[thm:control\], which is stated in the main text and which we state here again for convenience. Theorem \[thm:control\] follows almost entirely from Theorem \[thm:control\_app\].
Assume \[part:infinity\_or\_zero\] does not hold. Then, there exists $\varepsilon\in{\mathbb{R}}_{>0}$ such that $$\label{eq:boxed}
\varepsilon\leq\|\tilde x(t)\|\leq \frac{1}{\varepsilon}\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$ Moreover, it follows from Theorem \[thm:control\_app\] that at least one of the following holds:
1. \[a\] $\tilde x_n(t)^{(y_j-y_i)_n}$ oscillates around $q^{(y_j-y_i)_n}$, implying that $\tilde x_n(t)$ oscillates around $q$;
2. \[b\] $\displaystyle\lim_{t\to\infty}\int_{t}^\infty \Big|\tilde x_n(s)^{(y_j-y_i)_n}- q^{(y_j-y_i)_n}\Big|ds=0$.
Since \[b\] clearly implies \[part:ACR\_convergence\], to complete the proof it suffices to show that \[a\] implies $\hat\gamma\neq \hat\gamma_n e_n$. Assume \[a\] holds. If it were $\hat\gamma=\hat\gamma_n e_n$, then by Theorem \[thm:lcif\_app\] and by the fact that $u(x,t)$ is orthogonal to $\hat\gamma$ for all $x\in{\mathbb{R}}^d_{\geq0}$ we would have $$\hat\gamma_n\frac{d}{dt} \tilde x_n(t)=\frac{d}{dt}{\langle \hat\gamma, \tilde x(t)\rangle}={\langle \hat\gamma, f(\tilde x(t))\rangle}=\tilde{x}(t)^{y_i}\Big(\tilde x_n(t)-q\Big)\quad\text{for all }t\in{\mathbb{R}}_{\geq0}.$$ By , there would be $m,M\in{\mathbb{R}}_{>0}$ such that $$m\Big(\tilde x_n(t)-q\Big)\leq \hat\gamma_n\frac{d}{dt}\Big(\tilde x_n(t)-q\Big)\leq M \Big(\tilde x_n(t)-q\Big)\quad\text{for all }t\in{\mathbb{R}}_{\geq0},$$ which would in turn imply that $\tilde x_n(t)-q$ maintains the same sign for all $t\in{\mathbb{R}}_{\geq0}$. Hence, \[a\] could not hold and the proof is concluded.
Inclusion in larger systems {#inclusion-in-larger-systems}
---------------------------
Loosely speaking, a subsystem ${\mathscr{S}}'$ of a reaction system ${\mathscr{S}}$ is simply the reaction system generated by a subset of the reactions of ${\mathscr{S}}$, with the same choice of rate functions. In order to give a formal definition, some more care is needed as the number of species involved in the subsystem may be smaller (hence the dimensions of ${\mathscr{S}}$ and ${\mathscr{S}}'$ may be different).
Let ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}}, \{\lambda_{ij}\}_{1\leq i,j\leq m})$ be a reaction system with $d$ species and $m$ complexes. A *subsystem* of ${\mathscr{S}}$ is a reaction system ${\mathscr{S}}'=({\mathcal{X}}', {\mathcal{C}}', {\mathcal{R}}', \{\lambda'_{ij}\}_{1\leq i,j\leq m'})$ with $d'$ species and $m'$ complexes such that, up to reordering of ${\mathcal{X}}$ and ${\mathcal{C}}$,
- ${\mathcal{X}}'=\{X_i\in{\mathcal{X}}\,:\,1\leq i\leq d'\}$;
- $y_{in}=0$ for all $y_i\in{\mathcal{C}}$ with $1\leq i\leq m'$ and for all $d'<n\leq d$;
- ${\mathcal{C}}'=\{y_i'\,:\,y_i'=\pi(y_i), y_i\in{\mathcal{C}}, 1\leq i\leq m'\}$, where $\pi:{\mathbb{R}}^d\to {\mathbb{R}}^{d'}$ is the projection onto the first $d'$ components;
- ${\mathcal{R}}'\subseteq\{y'_i\to y'_j\,:\, y_i\to y_j\in{\mathcal{R}}\}$;
- for all $1\leq i,j\leq m'$ and all $x\in{\mathbb{R}}^d_{\geq0}$ $$\lambda'_{ij}(\pi(x))=\begin{cases}
\lambda_{ij}(x)&\text{if }y'_i\to y'_j\in{\mathcal{R}}'\\
0&\text{otherwise.}
\end{cases}$$
With a slight abuse of notation due to the potential different lenght of the complexes of ${\mathscr{S}}$ and ${\mathscr{S}}'$, we further say that $y_k\to y_l$ is a reaction of ${\mathscr{S}}$ but is not a reaction of ${\mathscr{S}}'$ if $y_k\to y_j\in{\mathcal{R}}$ and either $\max{k,l}>m'$ or $y'_k\to y'_l\notin{\mathcal{R}}'$. In this case, we write $y_k\to y_l\in{\mathcal{R}}\setminus{\mathcal{R}}'$. The following result is stated in the main text.
Here, we prove the following, stronger version.
\[thm:extension\_app\] Consider a reaction system ${\mathscr{S}}=({\mathcal{X}}, {\mathcal{C}}, {\mathcal{R}}, \{\lambda_{ij}\}_{1\leq i,j\leq m})$, and let ${\mathscr{S}}'=({\mathcal{X}}', {\mathcal{C}}', {\mathcal{R}}', \{\lambda'_{ij}\}_{1\leq i,j\leq m'})$ be a sub-system. Assume that ${\mathscr{S}}'$ is a mass-action system with two complexes $\pi(y_i)$ and $\pi(y_j)$ for which $\hat\Gamma_{ij}(\kappa)$ is non-empty. Moreover, assume there exists $\hat\gamma\in\hat\Gamma_{ij}(\kappa)$ such that $\pi(y_l-y_k)$ is orthogonal to $\hat\gamma$ for all $y_k\to y_l\in{\mathcal{R}}\setminus{\mathcal{R}}'$. Then, there exists a value $\overline q\in{\mathbb{R}}_{\geq0}$ such that $$(c')^{\pi(y_j-y_i)}=\overline q$$ for all steady states $c'$ of ${\mathscr{S}}'$ such that $(c')^{\pi(y_i)}>0$, and $$c^{y_j-y_i}=\overline q$$ for all steady states $c$ of ${\mathscr{S}}$ such that $c^{y_i}>0$.
The existence of $\overline q\in{\mathbb{R}}_{\geq0}$ such that $$(c')^{\pi(y_j-y_i)}=\overline q$$ for all steady states $c'$ of ${\mathscr{S}}'$ with $(c')^{\pi(y_i)}>0$ follows from Theorem \[thm:lcif\_app\].
In order to prove the second part of the statement, we can write $$\frac{d}{dt} \pi(x(t))=f(\pi(x(t)))+u(x(t),t),$$ where $$\begin{aligned}
f(\pi(x(t)))&=\sum_{y'_k\to y'_l\in{\mathcal{R}}'}(y'_l-y'_k)\kappa_{ij}\pi(x(t))^{y_i}\\
u(x(t),t)&=\sum_{y_k\to y_l\in{\mathcal{R}}\setminus{\mathcal{R}}'}\pi(y_l-y_k)\lambda_{ij}(x(t),t).
\end{aligned}$$ Note that $\hat\gamma$ is orthogonal to $u(x,t)$, for all $x\in{\mathbb{R}}^d_{\geq0}$ and $t\in{\mathbb{R}}_{\geq0}$. Hence, if $c$ is a steady state of ${\mathscr{S}}$, it follows from Theorem \[thm:lcif\_app\] that $$\begin{aligned}
0&=\frac{d}{dt}{\langle \hat\gamma, \pi(c)\rangle}={\langle \hat\gamma, f(\pi(c))\rangle}\\
&=\pi(c)^{\pi(y_i)}\Big(\pi(c)^{\pi(y_j-y_i)}-\overline q\Big)\\
&=c^{y_i}\Big(c^{y_j-y_i}-\overline q\Big),
\end{aligned}$$ which concludes the proof.
Example of an ACR system with no linear CIF
===========================================
Assume that $X_i$ is an ACR species, with ACR value $q$. It is not always possible to find a *linear* combination of species whose derivative at time $t$ is of the form $r(t)(x_i(t)^\alpha-q)$, for some polynomial $r(t)$ and some real number $\alpha$. As an example, consider the following mass-action system: $$\begin{split}
\schemestart
A+B\arrow{->[$\kappa_1$]}B+C\arrow{<=>[$\kappa_2$][$\kappa_3$]}2B
\arrow(@c1.south east--.north east){0}[-90,.25]
B\arrow{<=>[$\kappa_4$][$\kappa_5$]}2E\arrow{->[$\kappa_6$]}2D
\arrow(@c4.south east--.north east){0}[-90,.25]
C\arrow{->[$\kappa_7$]}A
\arrow(@c7.south east--.north east){0}[-90,.25]
D\arrow{->[$\kappa_8$]}E
\schemestop
\end{split}$$ Let us order the species alphabetically, and the complexes as $(A+B, B+C, 2B, B, 2E, 2D, C, A, D, E)$. It is proven in [@AC:ACR] that the species $A$ is ACR, with ACR value $$q=\frac{\kappa_3\kappa_7}{\kappa_1\kappa_2}.$$ In this case there is no linear combination of species whose derivative is of the form $$\label{eq:desired_form}
r(t)\left(x_1(t)^\alpha-q^\alpha\right),$$ for some polynomial function $r(t)$ and some real number $\alpha$. Indeed, for any $\beta\in {\mathbb{R}}^{d}$, the derivative of ${\langle \beta, x(t)\rangle}$ is given by $\beta^\top Y A(\kappa) \Lambda(x(t))$, which in this case is a polynomial of the form $$w_1x_1(t)x_2(t)+w_2x_3(t)+w_3x_2(t)x_3(t)+w_4x_2(t)^2+w_5x_2(t)+w_6x_5(t)^2+w_7x_4(t),$$ for some real coefficients $w_i$. The only monomial containing $x_1(t)$ is the first one, so if we want the derivative of ${\langle \beta, x(t)\rangle}$ to be of the form , necessarily $w_1\neq0$, $\alpha=1$, and $r(t)=w_1x_2(t)$. We can further deduce that necessarily $w_2=w_3=w_4=w_6=w_7=0$, and $w_5=-w_1q.$ In matrix form, this is equivalent to $$\beta^\top Y A(\kappa)=w_1 (1,0,0,-q,0,0,0,0,0,0),$$ but this is not possible because in this case the vector on the right-hand side does not belong to the left image of $Y A(\kappa)$.
Applications to the systems introduced in the main text. {#sec:models}
========================================================
Here we use the theory developed in this work to analyze the reaction systems introduced in the main text.
Toy example
-----------
Consider the mass-action system $$ $A+B$$2B$ (@c1.south east–.north east)[0]{}\[-90,.25\] $B$$A$ $$ Order the species alphabetically, and the complexes in the appearance order from left to right and from top to bottom, as $A+B$, $2B$, $B$, and $A$. In this case, we have $$\label{eq:YAk_toyexample}
Y=\begin{pmatrix}
1 & 0 & 0 & 1\\
1 & 2 & 1 & 0
\end{pmatrix}\quad\text{and}\quad
A(\kappa)=\begin{pmatrix}
-\kappa_1 & 0 & 0 & 0\\
\kappa_1 & 0 & 0 & 0\\
0 & 0 & -\kappa_2 & 0\\
0 & 0 & \kappa_2 & 0\\
\end{pmatrix}$$ The two non-terminal complexes differing for the entry relative to $A$ are the first and the third ones. Since the model has deficiency 1, it follows from Theorem \[thm:feinberg\_implies\_cond\] that $\hat\Gamma_{13}(\kappa)$ is non-empty. We can use Matlab to explicitly calculate it. First, in order to define $Y$ and $A(\kappa)$ in Matlab, we define the symbolic variables $\kappa_1$ and $\kappa_2$ and require they are positive real numbers via
k1=sym('k1','positive');
k2=sym('k2','positive');
We then define the matrices [`Y`]{} and [`Ak`]{} as in . We can calculate $\Psi_{13}(\kappa)$ via
e=eye(4);
simplify(null([Ak'*Y' e(:,1) e(:,3)]))
which returns the following matrix, whose columns are a basis for $\Psi_{13}(\kappa)$:
$$\begin{pmatrix}
1 & -\frac{1}{\kappa_2}\\
1 & 0\\
0 & -\frac{\kappa_1}{\kappa_2}\\
0 & 1
\end{pmatrix}.$$ It then follows from Proposition \[prop:calculating\_Gamma\] that $$\Gamma_{13}(\kappa)=\left\{
\begin{pmatrix}
\frac{1}{\kappa_2}\\
0\\
\frac{\kappa_1}{\kappa_2}
\end{pmatrix}+a
\begin{pmatrix}
1 \\ 1 \\ 0
\end{pmatrix}\,:\, a\in{\mathbb{R}}\right\},$$ which in turn implies that $$\label{eq:gamma_toymodel}
\hat\Gamma_{13}(\kappa)=\left\{
\begin{pmatrix}
\frac{1}{\kappa_2}\\
0
\end{pmatrix}+a
\begin{pmatrix}
1 \\ 1
\end{pmatrix}\,:\, a\in{\mathbb{R}}\right\}$$ and together with Theorem \[thm:lcif\] that the ACR value of $A$ is $\kappa_1/\kappa_2$. Note that in this case $$\label{eq:st_toymodel}
{\mathcal{S}}=\operatorname{span}_{{\mathbb{R}}}\begin{pmatrix}
1 \\ -1
\end{pmatrix}\quad\text{and}\quad
{{\mathcal{S}}^\perp}=\operatorname{span}_{{\mathbb{R}}}\begin{pmatrix}
1 \\ 1
\end{pmatrix}.$$ Moreover, each connected component contains exactly one terminal component. Hence, Proposition \[prop:structure\_Gamma\] applies, and as a matter of fact $$\hat\Gamma_{13}(\kappa)=
\begin{pmatrix}
\frac{1}{\kappa_2}\\
0
\end{pmatrix}+{{\mathcal{S}}^\perp}.$$
Finally, while it is clear from that a vector with zero first component and a vector with zero second component are in $\hat\Gamma_{13}(\kappa)$, this could be derived from Corollary \[cor:structural\] and from without explicitly calculating $\hat\Gamma_{13}(\kappa)$. As a consequence, due to Theorem \[thm:control\], disturbances could be applied to the production and degradation rates of $A$ (or $B$) while maintaining the absolute concentration robustness of the species $A$, its ACR value, and the linear CIF described in Theorem \[thm:lcif\].
EnvZ-OmpR osmoregulatory system
-------------------------------
Consider the mass-action system $$\schemestart
EnvZ-D \arrow(1--2){<=>[*0{$\kappa_1$}][*0{$\kappa_2[{\text{ADP}}]$}]} EnvZ \arrow(@2--3){<=>[*0{$\kappa_3[{\text{ATP}}]$}][*0{$\kappa_4$}]} EnvZ-T \arrow(@3--4){->[*0{$\kappa_5$}]} EnvZ-P
\arrow(@1.south east--.north east){0}[-90,.50]
EnvZ-P \+ OmpR \arrow(5--6){<=>[*0{$\kappa_6$}][*0{$\kappa_7$}]} EnvZ-OmpR-P \arrow(@6--7){->[*0{$\kappa_8$}]} EnvZ \+ OmpR-P
\arrow(@5.south east--.north east){0}[-90,.50]
EnvZ-D \+ OmpR-P \arrow(8--9){<=>[*0{$\kappa_9$}][*0{$\kappa_{10}$}]} EnvZ-OmpR-D-P \arrow(@9--10){->[*0{$\kappa_{11}$}]} EnvZ-D \+ OmpR
\schemestop$$ Order both the species and the complexes according to their appearance order, from left to right and from top to bottom. Hence, the 8 species are ordered as EnvZ-D, EnvZ, EnvZ-T, EnvZ-P, OmpR, EnvZ-OmpR-P, OmpR-P, and EnvZ-OmpR-D-P. The 10 complexes are ordered as EnvZ-D, EnvZ, EnvZ-T, EnvZ-P, EnvZ-P+OmpR, EnvZ-OmpR-P, EnvZ+OmpR-P, EnvZ-D+OmpR-P, EnvZ-OmpR-D-P, and EnvZ-D+OmpR.
It can be checked that $\dim {\mathcal{S}}=6$ and $$\label{eq:cons_envz}
{{\mathcal{S}}^\perp}=\operatorname{span}_{\mathbb{R}}\left\{
\begin{pmatrix}
1 \\ 1 \\ 1\\ 1\\ 0 \\ 1 \\ 0\\ 1
\end{pmatrix},
\begin{pmatrix}
0 \\ 0 \\ 0\\ 0\\ 1 \\ 1\\ 1\\ 1
\end{pmatrix}
\right\}.$$ The above conservation laws correspond to the fact that the total mass of the chemical species containing some form of EnvZ is conserved, as well as the total mass of the chemical species containing some form of OmpR. We can calculate the deficiency as $$\delta=m-\ell-\dim{\mathcal{S}}=10-3-6=1.$$
The non-terminal complexes that only differ for the entry relative to OmpR-P are the first and the eighth ones. Therefore, we are interested in the study of $\hat\Gamma_{18}(\kappa)$, which by Theorem \[thm:feinberg\_implies\_cond\] is not empty. We have $$Y=\begin{pmatrix}
1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 1\\
0 & 1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0\\
0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1\\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0
\end{pmatrix}$$ and $$A(\kappa)=\begin{pmatrix}
-\kappa_{1} & \kappa_2\mathrm{[ADP]} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ \kappa_{1} & -\kappa_2\mathrm{[ADP]}-\kappa_3\mathrm{[ATP]} & \kappa_{4} & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & \kappa_3\mathrm{[ATP]} & -\kappa_{4}-\kappa_{5} & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & \kappa_{5} & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & -\kappa_{6} & \kappa_{7} & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & \kappa_{6} & -\kappa_{7}-\kappa_{8} & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & \kappa_{8} & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\kappa_{9} & \kappa_{10} & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{9} & -\kappa_{10}-\kappa_{11} & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{11} & 0
\end{pmatrix}.$$ In order to define the corresponding symbolic matrices [`Y`]{} and [`Ak`]{} in Matlab, we can first define the symbolic positive real variables
k=sym('k', [1,11], 'positive');
k(2)=k(2)*sym('ADP', 'positive');
k(3)=k(3)*sym('ATP', 'positive');
Then we calculate $\Psi_{18}(\kappa)$, as defined in , via
e=eye(10);
simplify(null([Ak'*Y' e(:,8) e(:,1)]))
The output of the last command is the following matrix, whose columns are a basis of $\Psi_{18}(\kappa)$.
$$\begin{pmatrix}
-1 & 1 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{11}\,\left(\kappa_{4}+\kappa_{5}\right)}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ -1 & 1 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{4}\,\kappa_{11}+\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{11}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{10}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{11}}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ -1 & 1 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{4}\,\kappa_{11}+\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{10}+2\,\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{11}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{10}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{11}}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ -1 & 1 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{4}\,\kappa_{10}+2\,\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{4}\,\kappa_{11}+\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{10}+2\,\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{11}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{10}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{11}}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ 1 & 0 & \frac{\mathrm{[ADP]}\,\kappa_{2}\,\left(\kappa_{4}+\kappa_{5}\right)}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}}\\ 0 & 1 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{4}\,\kappa_{11}+\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{5}\,\kappa_{11}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{10}+\mathrm{[ATP]}\,\kappa_{3}\,\kappa_{5}\,\kappa_{11}}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ 1 & 0 & 0\\ 0 & 1 & 0\\ 0 & 0 & -\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{9}\,\kappa_{11}\,\left(\kappa_{4}+\kappa_{5}\right)}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}\\ 0 & 0 & 1
\end{pmatrix}$$
For convenience, denote the last vector by $\zeta(\kappa)$. It follows from Proposition \[prop:calculating\_Gamma\] that $$\Gamma_{18}(\kappa)=
\left\{
-\pi_9(\zeta(\kappa))
+a_1
\begin{pmatrix}
-1 \\ -1 \\ -1\\ -1\\ 1 \\ 0\\ 1\\ 0\\ 0
\end{pmatrix}
+a_2
\begin{pmatrix}
1 \\ 1 \\ 1\\ 1\\ 0 \\ 1 \\ 0\\ 1\\ 0
\end{pmatrix}
\,:\, a_1,a_2\in{\mathbb{R}}\right\}.$$ It follows that $$\label{eq:hat_gamma_envz}
\hat\Gamma_{18}(\kappa)=
\left\{
-\pi_8(\zeta(\kappa))
+a_1
\begin{pmatrix}
-1 \\ -1 \\ -1\\ -1\\ 1 \\ 0\\ 1\\ 0
\end{pmatrix}
+a_2
\begin{pmatrix}
1 \\ 1 \\ 1\\ 1\\ 0 \\ 1 \\ 0\\ 1
\end{pmatrix}
\,:\, a_1,a_2\in{\mathbb{R}}\right\},$$ which corresponds to in the main text. Moreover, it follows from Theorem \[thm:lcif\] that the ACR value of the species OmpR-P is $$-\zeta_{9}(\kappa)=\frac{\mathrm{[ADP]}\,\kappa_{2}\,\kappa_{9}\,\kappa_{11}\,\left(\kappa_{4}+\kappa_{5}\right)}{\mathrm{[ATP]}\,\kappa_{1}\,\kappa_{3}\,\kappa_{5}\,\left(\kappa_{10}+\kappa_{11}\right)}$$ as reported in in the main text.
Further things can be noted about this model. First, it follows from and that $$\hat\Gamma_{18}(\kappa)=-\pi_8(\zeta(\kappa))+{{\mathcal{S}}^\perp},$$ which is in accordance with Proposition \[prop:structure\_Gamma\] because all connected components contain exactly one terminal component. Secondly, due to Corollary \[cor:structural\], we can deduce without explicitly calculating $\Gamma_{18}(\kappa)$ that there is a vector $\hat\gamma$ in $\hat\Gamma_{18}(\kappa)$ whose entries relative to species EnvZ and OmpR-P are zero (which are the second and the seventh complexes, respectively). Specifically, if we consider the basis of ${{\mathcal{S}}^\perp}$ given in and we let $X_{l_1}=X_2=$EnvZ and $X_{l_2}=X_7=$OmpR-P, we have $$V=\begin{pmatrix}
1 & 0\\
0 & 1
\end{pmatrix},$$ which has rank 2. The existence of such vector $\hat\gamma$ is used in Section \[sec:large\] and it implies by Theorem \[thm:control\] that the system is robust to persistent disturbances affecting the production and degradation rates of both EnvZ and OmpR-P.
An ACR signaling system covered by our theory and not by [@SF:ACR]
==================================================================
Consider the double-phosphorylation mass-action system in Figure \[fig:double\]. We will show that the theory developed in [@SF:ACR] stays silent on whether it is ACR. However, our theory covers this case and implies the double-phosphorylated form of the transcriptional regulatory protein is ACR, for any choice of kinetic parameters $\kappa$ such that a positive steady state exists. Note that this form of response robustness may seem a bit surprising for a multisite phosphorylation mechanism, since these are often known for their multi-stability properties, notably shown in the case of the MAPK pathway [@AFS:detection; @TTT:MAPK; @FB:MAPK; @FSWFS:MAPK].
Activation and phosphorylation of the sensor-transmitter protein: $$ $A$ (1–2)[<=>\[\*0[$\kappa_1$\[\]]{}\]\[\*0[$\kappa_2$]{}\]]{} ${{\color{brown} A^\star}}$ (@2–3)[->\[\*0[$\kappa_3$]{}\]]{} $A$-[[**P**]{}]{} $$ { \begin{tikzpicture}
\path[use as bounding box] (0,0) -- (\linewidth,0.5);
\draw[color=olive!50!white,dashed,dash phase=2pt]
(0-\kvtcb@leftlower-\kvtcb@boxsep,0.25)--
(\linewidth+\kvtcb@rightlower+\kvtcb@boxsep,0.25);
\end{tikzpicture} }
Phosphorylation of the sensory response protein:
$$ $A$-[[**P**]{}]{} $B$ (4–5)[<=>\[\*0[$\kappa_4$]{}\]\[\*0[$\kappa_5$]{}\]]{} $A$-$B$-[[**P**]{}]{}(@5–6)[->\[\*0[$\kappa_6$]{}\]]{} $A$ $B$-[[**P**]{}]{}(@4.south east–.north east)[0]{}\[-90,.50\] $A$-[[**P**]{}]{} $B$-[[**P**]{}]{}(7–8)[<=>\[\*0[$\kappa_7$]{}\]\[\*0[$\kappa_8$]{}\]]{} $A$-$B$-[[**P**]{}]{}[[**P**]{}]{}(@8–9)[->\[\*0[$\kappa_9$]{}\]]{} $A$ $B$-[[**P**]{}]{}[[**P**]{}]{}$${ \begin{tikzpicture}
\path[use as bounding box] (0,0) -- (\linewidth,0.5);
\draw[color=olive!50!white,dashed,dash phase=2pt]
(0-\kvtcb@leftlower-\kvtcb@boxsep,0.25)--
(\linewidth+\kvtcb@rightlower+\kvtcb@boxsep,0.25);
\end{tikzpicture} }
Activation and phosphorylation of the protein aggregate:
$$ $A$-$B$-[[**P**]{}]{}(10–11)[<=>\[\*0[$\kappa_{10}$\[\]]{}\]\[$\kappa_{11}$\]]{} ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{} (@10.south east–.north east)[0]{}\[-90,.50\] $A$-$B$-[[**P**]{}]{}[[**P**]{}]{}(12–13)[<=>\[\*0[$\kappa_{12}$\[\]]{}\]\[$\kappa_{13}$\]]{} ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{}[[**P**]{}]{}(@12.south east–.north east)[0]{}\[-90,.50\] ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{}(14–15)[->\[\*0[$\kappa_{14}$]{}\]]{} ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{}[[**P**]{}]{}(@15–16)[->\[\*0[$\kappa_{15}$]{}\]]{} $A$-[[**P**]{}]{} $B$-[[**P**]{}]{}[[**P**]{}]{}$${ \begin{tikzpicture}
\path[use as bounding box] (0,0) -- (\linewidth,0.5);
\draw[color=olive!50!white,dashed,dash phase=2pt]
(0-\kvtcb@leftlower-\kvtcb@boxsep,0.25)--
(\linewidth+\kvtcb@rightlower+\kvtcb@boxsep,0.25);
\end{tikzpicture} }
Dephosphorylation of the sensory response protein:
$$ $A$ $B$-[[**P**]{}]{}[[**P**]{}]{}(17–18)[<=>\[\*0[$\kappa_{16}$]{}\]\[\*0[$\kappa_{17}$]{}\]]{} $B$-[[**P**]{}]{}[[**P**]{}]{}-$A$ (@18–19)[->\[\*0[$\kappa_{18}$]{}\]]{} $A$-[[**P**]{}]{} $B$ $$
Order the species and the complexes according to their appearance order, from left to right and from top to bottom. The 11 species are then ordered as $A$, $A^\star$, $A$-P, $B$, $A$-$B$-P, $B$-P, $A$-$B$-PP, $B$-PP, $A^\star$-$B$-P, $A^\star$-$B$-PP, and $B$-PP-$A$. The 13 complexes are ordered as $A$, $A^\star$, $A$-P, $A$-P + $B$, $A$-$B$-P, $A$ + $B$-P, $A$-P + $B$-P, $A$-$B$-PP, $A$ + $B$-PP, $A^\star$-$B$-P, $A^\star$-$B$-PP, $A$-P + $B$-PP, and $B$-PP-$A$. It can be checked that $\dim {\mathcal{S}}=9$ and $$\label{eq:cons_double}
{{\mathcal{S}}^\perp}=\operatorname{span}_{\mathbb{R}}\left\{
\begin{pmatrix}
1\\ 1\\ 1\\ 0\\ 1\\ 0\\ 1\\ 0\\ 1\\ 1\\ 1
\end{pmatrix},
\begin{pmatrix}
0\\ 0\\ 0\\ 1\\ 1\\ 1\\ 1\\ 1\\ 1\\ 1\\ 1
\end{pmatrix}
\right\}.$$ Similarly as for the EnvZ-OmpR osmoregulatory system, the above conservation laws correspond express that the total mass of the chemical species containing some form of the protein $A$ is conserved, as well as the total mass of the chemical species containing some form of the protein $B$. The reaction graph is given in Figure \[fig:double\_graph\]. In particular, $\ell=2$ and the deficiency is $$\delta=m-\ell-\dim{\mathcal{S}}=13-2-9=2.$$ So, this model is not included in the class of models studied in [@SF:ACR], and Theorems \[thm:feinberg\_weak\] and \[thm:feinberg\_strong\] do not apply.
$$ $A$ (1–2)[<=>\[\*0[$\kappa_1$\[\]]{}\]\[\*0[$\kappa_2$]{}\]]{} ${{\color{brown} A^\star}}$ (@2–3)[->\[\*0[$\kappa_3$]{}\]]{} $A$-[[**P**]{}]{} (@1.south east–.north east)[0]{}\[-90,.50\] $A$-[[**P**]{}]{} $B$ (4–5)[<=>\[\*0[$\kappa_4$]{}\]\[\*0[$\kappa_5$]{}\]]{} $A$-$B$-[[**P**]{}]{}(@5–6)[->\[\*0[$\kappa_6$]{}\]]{} $A$ $B$-[[**P**]{}]{} (@5–10)[<=>\[\*0[$\kappa_{10}$\[\]]{}\]\[\*0[$\kappa_{11}$]{}\]]{}\[-90\] ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{} (@10–11)[->\[\*0[$\kappa_{14}$]{}\]]{}\[-90\] ${{\color{brown} A^\star}}$-$B$-[[**P**]{}]{}[[**P**]{}]{} (@11–8)[<=>\[\*0[$\kappa_{13}$]{}\]\[\*0[$\kappa_{12}$\[\]]{}\]]{}\[-90\]$A$-$B$-[[**P**]{}]{}[[**P**]{}]{} (@11–12)[->\[\*0[$\kappa_{15}$]{}\]]{} $A$-[[**P**]{}]{} $B$-[[**P**]{}]{}[[**P**]{}]{} (@8–9)[0]{}\[180\] $A$ $B$-[[**P**]{}]{}[[**P**]{}]{} (@9–13)[0]{}\[180\] $B$-[[**P**]{}]{}[[**P**]{}]{}-$A$ (@8–7)[<=>\[\*0[$\kappa_8$]{}\]\[\*0[$\kappa_7$]{}\]]{} $A$-[[**P**]{}]{} $B$-[[**P**]{}]{} (@9–@8)[<-\[\*0[$\kappa_9$]{}\]]{} (@13–@9)[<=>\[\*0[$\kappa_{17}$]{}\]\[\*0[$\kappa_{16}$]{}\]]{} (@13–@4)[->\[\*0[$\kappa_{18}$]{}\]]{} $$
The first and the ninth complexes are non-terminal, and they only differ for the entry relative B-PP. Hence, in order to apply Theorems \[thm:lcif\] and \[thm:control\], we are interested in studying $\hat\Gamma_{19}(\kappa)$. Since the deficiency of the model is 2, Theorem \[thm:feinberg\_implies\_cond\] does not apply, so to understand whether $\hat\Gamma_{19}(\kappa)$ is non-empty we need to explicitly calculate it. To this aim, we define in Matlab the following positive symbolic variables, which correspond to the rate constants of the model:
k=sym('k', [1,18], 'positive');
k(1)=k(1)*sym('I', 'positive');
k(10)=k(10)*sym('I', 'positive');
k(12)=k(12)*sym('I', 'positive');
We then define the symbolic matrices [`Y`]{} and [`Ak`]{} corresponding to $$Y=\left(\begin{array}{ccccccccccccc}
1 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0\\ 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 0\\ 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1\end{array}\right)$$ and to $A(\kappa)$ as described in Figure \[fig:Ak\], respectively.
$$A(\kappa)=\left(\begin{array}{ccccccccccccc}
-\kappa_{1}\mathrm{[I]} & \kappa_{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\kappa_{1}\mathrm{[I]} & -\kappa_{2}-\kappa_{3} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & \kappa_{3} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & -\kappa_{4} & \kappa_{5} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{18}\\
0 & 0 & 0 & \kappa_{4} & -\kappa_{5}-\kappa_{6}-\kappa_{10}\mathrm{[I]} & 0 & 0 & 0 & 0 & \kappa_{11} & 0 & 0 & 0\\
0 & 0 & 0 & 0 & \kappa_{6} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & -\kappa_{7} & \kappa_{8} & 0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & \kappa_{7} & -\kappa_{8}-\kappa_{9}-\kappa_{12}\mathrm{[I]} & 0 & 0 & \kappa_{13} & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{9} & -\kappa_{16} & 0 & 0 & 0 & \kappa_{17}\\
0 & 0 & 0 & 0 & \kappa_{10}\mathrm{[I]} & 0 & 0 & 0 & 0 & -\kappa_{11}-\kappa_{14} & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{12}\mathrm{[I]} & 0 & \kappa_{14} & -\kappa_{13}-\kappa_{15} & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{15} & 0 & 0\\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \kappa_{16} & 0 & 0 & 0 & -\kappa_{17}-\kappa_{18}\end{array}\right)$$
In order to calculate a basis for $\Psi_{19}(\kappa)$ as defined in , we type
e=eye(13);
simplify(null([Ak'*Y' e(:,1) e(:,9)]))
The output of the last command is shown in Figure \[fig:psi\]. While the output is rather complicated, we did not need to put much effort in its calculation, which was completed in Matlab in a matter of seconds. For convenience we denote by $\zeta(\kappa)$ the last column of the output matrix.
$$\begin{pmatrix} -1 & 1 & \frac{1}{\kappa_{16}}\\ -1 & 1 & \frac{p_1(\kappa,\mathrm{[I]})}{\kappa_{16}\kappa_{18}\left(\kappa_{2}+\kappa_{3}\right)\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ -1 & 1 & \frac{p_2(\kappa, \mathrm{[I]})}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 1 & 0 & -\frac{\left(\kappa_{17}+\kappa_{18}\right)\left(2\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+2\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+2\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+2\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{13}\kappa_{14}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\right)}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 0 & 1 & -\frac{\kappa_{17}}{\kappa_{16}\kappa_{18}}\\ 1 & 0 & -\frac{\kappa_{9}\left(\kappa_{13}+\kappa_{15}\right)\left(\kappa_{17}+\kappa_{18}\right)\left(\kappa_{6}\kappa_{11}+\kappa_{6}\kappa_{14}+\mathrm{[I]}\kappa_{10}\kappa_{14}\right)}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 0 & 1 & \frac{p_3(\kappa,\mathrm{[I]})}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 1 & 0 & 0\\ 0 & 1 & \frac{\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}-\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{17}-\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{17}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}^2\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{17}}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 0 & 1 & \frac{p_4(\kappa, \mathrm{[I]})}{\kappa_{16}\kappa_{18}\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 0 & 1 & 0\\ 0 & 0 & -\frac{\mathrm{[I]}\kappa_{1}\kappa_{3}\left(\kappa_{17}+\kappa_{18}\right)\left(\kappa_{6}\kappa_{11}+\kappa_{6}\kappa_{14}+\mathrm{[I]}\kappa_{10}\kappa_{14}\right)\left(\kappa_{9}\kappa_{13}+\kappa_{9}\kappa_{15}+\mathrm{[I]}\kappa_{12}\kappa_{15}\right)}{\kappa_{16}\kappa_{18}\left(\kappa_{2}+\kappa_{3}\right)\left(-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\mathrm{[I]}^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\mathrm{[I]}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\right)}\\ 0 & 0 & 1
\end{pmatrix}$$
where $$\begin{aligned}
p_1(\kappa, \mathrm{[I]})&=\kappa_{2}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{18}+\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{17}+2\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{18}+\kappa_{2}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{18}+\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{17}+\kappa_{2}\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}+2\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{18}\\
&\quad+\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{17}+2\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}+\kappa_{2}\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{17}+2\kappa_{3}\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{3}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{17}\\
&\quad+\mathrm{[I]}\kappa_{3}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{3}\kappa_{9}\kappa_{10}\kappa_{13}\kappa_{14}\kappa_{17}+\mathrm{[I]}\kappa_{3}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{3}\kappa_{9}\kappa_{10}\kappa_{13}\kappa_{14}\kappa_{18}-\mathrm{[I]}\kappa_{2}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{3}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}\\
&\quad+\mathrm{[I]}\kappa_{3}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{17}-\mathrm{[I]}^2\kappa_{2}\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}^2\kappa_{3}\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}\\
p_2(\kappa, \mathrm{[I]})&=\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{17}\\
&\quad+2\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}^2\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{13}\kappa_{14}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}\\
&\quad+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{13}\kappa_{14}\kappa_{18}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{17}\\
p_3(\kappa,\mathrm{[I]})&=\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}^2\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{18}\\
&\quad+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}-\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{18}\\
p_4(\kappa,\mathrm{[I]})&=\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}\kappa_{18}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{17}+2\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}^2\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}\\
&\quad+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{11}\kappa_{12}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{17}+\mathrm{[I]}\kappa_{6}\kappa_{12}\kappa_{14}\kappa_{15}\kappa_{18}+\mathrm{[I]}\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}\kappa_{17}\end{aligned}$$
It follows from Proposition \[prop:calculating\_Gamma\] that $\Gamma_{19}(\kappa)$ is non-empty, and is given by $$\label{eq:gamma_double}
\Gamma_{19}(\kappa)=\left\{-\pi_{12}(\zeta(\kappa))
+a_1
\begin{pmatrix}
-1 \\ -1 \\ -1\\ 1\\ 0 \\ 1\\ 0\\ 1\\ 0\\ 0\\ 0\\ 0
\end{pmatrix}
+a_2
\begin{pmatrix}
1 \\ 1 \\ 1\\ 0\\ 1\\ 0\\ 1\\ 0\\ 1\\ 1\\ 1\\ 0
\end{pmatrix}
\,:\, a_1,a_2\in{\mathbb{R}}\right\}.$$ It follows from and that $$\hat\Gamma_{19}(\kappa)=\left\{-\pi_{11}(\zeta(\kappa))
+a_1
\begin{pmatrix}
-1 \\ -1 \\ -1\\ 1\\ 0 \\ 1\\ 0\\ 1\\ 0\\ 0\\ 0
\end{pmatrix}
+a_2
\begin{pmatrix}
1 \\ 1 \\ 1\\ 0\\ 1\\ 0\\ 1\\ 0\\ 1\\ 1\\ 1
\end{pmatrix}
\,:\, a_1,a_2\in{\mathbb{R}}\right\}=-\pi_{11}(\zeta(\kappa))+{{\mathcal{S}}^\perp},$$ which is in accordance with Proposition \[prop:structure\_Gamma\] because every connected component of the reaction graph in Figure \[fig:double\_graph\] contains exactly one terminal component. Moreover, from and Theorem \[thm:lcif\] it follows that the ACR value of the ACR species $B$-PP is $$-\zeta_{12}(\kappa)=\frac{\kappa_1\kappa_3[{\text{I}}](\kappa_{17}+\kappa_{18})(\kappa_6\kappa_{11}+\kappa_6\kappa_{14}+\kappa_{10}\kappa_{14}[{\text{I}}])(\kappa_9\kappa_{13}+\kappa_9\kappa_{15}+\kappa_{12}\kappa_{15}[{\text{I}}])}{\kappa_{16}\kappa_{18}(\kappa_2+\kappa_3)g(\kappa, [{\text{I}}])},$$ where $$g(\kappa, [{\text{I}}])=-\kappa_{10}\kappa_{12}\kappa_{14}\kappa_{15}[{\text{I}}]^2-\kappa_{9}\kappa_{10}\kappa_{14}\kappa_{15}[{\text{I}}]+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{13}+\kappa_{6}\kappa_{9}\kappa_{11}\kappa_{15}+\kappa_{6}\kappa_{9}\kappa_{13}\kappa_{14}+\kappa_{6}\kappa_{9}\kappa_{14}\kappa_{15}.$$ This means that if a positive steady state $c$ exists, necessarily it entry relative to $B$-PP (which is the eighth species) satisfies $c_8=-\zeta_{12}(\kappa)$. Of course, this cannot occur if $-\zeta_{12}(\kappa)$ is non-positive, i.e. if $g(\kappa, [{\text{I}}])$ is non-positive, in which case no positive steady state exists. Note that we did not need to work directly with the differential equation to derive this information. We will further show that in fact a positive steady state exists if and only if $g(\kappa, [{\text{I}}])$ is positive. To this aim, note that $c$ is a steady state if and only if $\Lambda(c)\in\ker YA(\kappa)$. Then, we calculate a basis for $\ker YA(\kappa)$ by typing
simplify(null(Y*Ak))
which returns a matrix of the form $$\begin{pmatrix}
0 & 0 & 0 & b_1(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_2(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}])\\
1 & 0 & 0 & 0\\
0 & 0 & 0 & b_4(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_5(\kappa, [{\text{I}}])\\
0 & 1 & 0 & 0\\
0 & 0 & 0 & b_7(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_8(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_9(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_{10}(\kappa, [{\text{I}}])\\
0 & 0 & 0 & b_{11}(\kappa, [{\text{I}}])\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & 1
\end{pmatrix}$$ for some functions $b_i(\kappa, [{\text{I}}])$ which map positive arguments to positive real numbers. Hence, a positive steady state exists if and only if there exists a positive vector $c$ such that $$\begin{aligned}
c_1&=a_4 b_1(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}]) & c_1c_6&=a_2 & c_9&=a_4 b_{10}(\kappa, [{\text{I}}])\\
c_2&=a_4 b_2(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}]) & c_3c_6&=a_4 b_7(\kappa, [{\text{I}}]) & c_{10}&=a_4 b_{11}(\kappa, [{\text{I}}])\\
c_3&=a_1 & c_7&=a_4 b_8(\kappa, [{\text{I}}]) & c_3c_8&=a_3\\
c_3c_4&=a_4 b_4(\kappa, [{\text{I}}]) & c_1c_8&=a_4 b_9(\kappa, [{\text{I}}]) & c_{11}&=a_4\\
c_5&=a_4 b_5(\kappa, [{\text{I}}])\end{aligned}$$ for some $a_1, a_2, a_3, a_4\in{\mathbb{R}}_{>0}$. If $g(\kappa, [{\text{I}}])$ is positive, it is easy to see that such a positive vector $c$ exists. Specifically, the positive steady states are parameterized by $$\begin{aligned}
c_1&=a_4 b_1(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}]) & c_5&=a_4 b_5(\kappa, [{\text{I}}]) & c_9&=a_4 b_{10}(\kappa, [{\text{I}}])\\
c_2&=a_4 b_2(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}]) & c_6&=\frac{a_4}{a_1}b_7(\kappa, [{\text{I}}]) & c_{10}&=a_4 b_{11}(\kappa, [{\text{I}}])\\
c_3&=a_1 & c_7&=a_4 b_8(\kappa, [{\text{I}}]) & c_{11}&=a_4\\
c_4&=\frac{a_4}{a_1} b_4(\kappa, [{\text{I}}]) & c_8&=\frac{b_9(\kappa, [{\text{I}}])}{b_1(\kappa, [{\text{I}}])g(\kappa, [{\text{I}}])}\end{aligned}$$ where $a_1$ and $a_4$ vary in ${\mathbb{R}}_{>0}$. Note that the entry $c_8$ is relative to the ACR species $B$-PP and is fixed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this note we provide a concise report on the complexity of the causal ordering problem, originally introduced by Simon to reason about causal dependencies implicit in systems of mathematical equations. We show that Simon’s classical algorithm to infer causal ordering is NP-Hard—an intractability previously guessed but never proven. We present then a detailed account based on Nayak’s suggested algorithmic solution (the best available), which is dominated by computing transitive closure—bounded in time by $O(|\mathcal V|\cdot |\mathcal S|)$, where $\mathcal S(\mathcal E, \mathcal V)$ is the input system structure composed of a set $\mathcal E$ of equations over a set $\mathcal V$ of variables with number of variable appearances (density) $|\mathcal S|$. We also comment on the potential of causal ordering for emerging applications in large-scale hypothesis management and analytics.'
address:
- 'IBM Research, São Paulo, Brazil'
- 'National Laboratory for Scientific Computing (LNCC), Petrópolis, Brazil'
author:
- Bernardo Gonçalves
- Fabio Porto
bibliography:
- 'note.bib'
title: A note on the complexity of the causal ordering problem
---
$\!\!$Causal ordering ,Causal reasoning ,Structural equations ,Hypothesis management.
Introduction {#sec:intro}
============
The causal ordering problem has long been introduced by Simon as a technique to infer the causal dependencies implicit in a deterministic mathematical model [@simon1953]. For instance, let $f_1(x_1)$ and $f_2(x_1, x_2)$ be two equations defined over variables $x_1, x_2$. Then the causal ordering problem is to infer all existing causal dependencies, in this case the only one is $(x_1, x_2)$, read ‘$x_2$ causally depends on $x_1$.’ It is obtained by first matching each equation to a variable that appears in it, e.g., $f_2 \mapsto x_2$. Intuitively, this means that $f_2$ is to be assigned to compute the value of $x_2$—using the value of $x_1$, which establishes a direct causal dependency between these two variables. Indirect dependencies may then arise and can be computed, which is specially useful when the system of equations is very large.
Causal ordering inference can then support users with uncertainty management, say, towards the discovery of what is wrong with a model for enabling efficient and effective modeling intervention. If multiple values of $x_1$ are admissible for a modeler, then as a user of the causal ordering machinery she has support to track their influence on the values of $x_2$. One major application for that is probabilistic database design [@goncalves2014].
Back in the 50th’s, Simon was motivated by studies in econometrics and not very concerned with the algorithmic aspects of the Causal Ordering Problem (COP). Yet his vision on COP and its relevance turned out to be influential in the artificial intelligence literature. In a more recent study, Dash and Druzdzel revisit and motivate it in light of modern applications [@druzdzel2008]. They show that Simon’s original algorithm, henceforth the Causal Ordering Algorithm (COA), is correct in the sense that any valid causal ordering that can be extracted from a self-contained (well-posed) system of equations must be compatible with the one that is output by COA [@druzdzel2008]. Their aim has also been (sic.) to validate decades of research that has shown the causal ordering to provide a powerful tool for operating on models. In addition to the result on the correctness of COA, their note also provides a convenient survey of related work that connects to Simon’s early vision on causal reasoning.
However, Simon’s formulation of COP into COA—originally in [@simon1953], and reproduced in [@druzdzel2008], turns out to be intractable. There is a need to distinguish Simon’s COA from COP itself. The former still seems to be the main entry point to the latter in the specialized literature. In fact, there is a lack of a review on the computational properties of COA—and as we show in this note, it tries to address an NP-Hard problem as one of its steps. The interested reader who needs an efficient algorithmic approach to address COP in a real, large-scale application can only scarcely find some comments spread through Nayak [@nayak1994 p. 287-91], and then Iwasaki and Simon [@simon1994 p. 149] and Pearl [@pearl2000 p. 226] both pointing to the former. Regarding Simon’s COA itself, the classical approach to COP, it is only Nayak who suggests in words that (sic.) ‘\[it\] is a worst-case exponential time algorithm’ [@nayak1996 p. 37]. We discuss this ambiguity that exists in the most up-to-date literature shortly in §\[subsec:related-work\].
COP is significant also in view of emerging applications in large-scale hypothesis management and analytics [@goncalves2014]. The modeling of physical and socio-economical systems as a set of mathematical equations is a traditional approach in science and engineering and a very large bulk of models exist which are ever more available in machine-readable format. Simon’s early vision on the automatic extraction of the “causal mechanisms” implicit in (large-scale) models for the sake of informed intervention finds nowadays new applications in the context of open simulation laboratories [@goncalves2015cise], large-scale model management [@haas2011] and online, shared model repositories [@hunter2003; @hines2004; @chelliah2013].
In this paper we review the causal ordering problem (§\[sec:problem\]). Our core contributions are (§\[sec:coa\]) to originally show that COA aims at addressing an NP-Hard problem, confirming Nayak’s earlier intuition; and then (§\[sec:nayak\]) to organize into a concise yet complete note his hints to solve COP in polynomial time.
Informal Preliminaries {#subsec:preliminaries}
----------------------
Given a system of mathematical equations involving a set of variables, the *causal ordering problem* consists in detecting the hidden asymmetry between variables. As an intermediate step towards it, one needs to establish a one-to-one mapping between equations and variables [@simon1953].
For instance, Einstein’s famous equation $E=m\,c^2$ states the equivalence of mass and energy, summarizing (in its scalar version) a theory that can be imposed two different asymmetries for different applications. Say, given a fixed amount of mass $m=m_0$ (and recalling that $c$ is a constant), find the particle’s relativistic rest energy $E$; or rather, given the particle’s rest energy, find its mass or potential for nuclear fission. That is, the causal ordering depends on what variables are set as input and which ones are “influenced” by them. Suppose there is uncertainty, say, a user considers two values to set the mass, $\!m=m_0$ or $m= m_0^\prime$. $\!$Then this uncertainty will flow through the causal ordering and affect all variables that are dependent on it (energy $E$).
For really large systems, having structures say in the order of one million equations [@goncalves2015c], the causal ordering problem is critical to provide more specific accountability on the accuracy of the system—viz., what specific variables and subsystems account for possibly inaccurate outcomes. This is key for managing and tracking the uncertainty of alternative modeling variations systematically [@goncalves2015cise; @goncalves2015c].
Related Work {#subsec:related-work}
------------
**COA**. Dash and Druzdzel [@druzdzel2008] provide a high-level description of how Simon’s COA [@simon1953] proceeds to discover the causal dependencies implicit in a structure. It returns a ‘partial’ causal mapping: from partitions on the set of equations to same-cardinality partitions on the set of variables—a ‘total’ causal mapping would instead map every equation to exactly one variable.
They show then that any valid total causal mapping produced over a structure must be consistent with COA’s partial causal mapping. Nonetheless, no observation at all is made regarding COA’s computational properties in [@druzdzel2008], leaving in the most up-to-date literature an impression that Simon’s COA is the way to go for COP. In this note we show that Simon’s COA tries to address an NP-Hard problem in one of its steps, and then clearly recommend Nayak’s efficient approach to COP as a fix to COA.
**COP**. Inspired by Serrano and Gossard’s work on constraint modeling and reasoning [@serrano1987], Nayak describes an approach that is provably efficient to process the causal ordering: extract any valid total causal mapping and then compute the transitive closure of the direct causal dependencies, viz, the causal ordering. Nayak’s is a provably correct approach to COP, as all valid ‘total’ causal mappings must have the same causal ordering.
In this note we arrange those insights into a concise yet detailed recipe that can be followed straightforwardly to solve COP efficiently.
The Causal Ordering Problem {#sec:problem}
===========================
We start with some preliminaries on notation and basic concepts to eventually state COP formally.
For an equation $f(x_1, x_2, ..., x_\ell)=0$, we will write $Vars(f) \triangleq \{ x_1,\, x_2,\, ...,\, x_\ell \}$ to denote the set of variables that appear in it.
\[def:structure\] A **structure** is a pair $\mathcal S(\mathcal E, \mathcal V)$, where $\mathcal E$ is a set of equations over a set $\mathcal V\!$ of variables, $\mathcal V \triangleq \bigcup_{f \in\, \mathcal E} Vars(f)$, such that:
- In any subset $\mathcal E^\prime \subseteq \mathcal E$ of $k>0$ equations of the structure, at least $k$ different variables appear, i.e., $|\mathcal E^\prime| \leq |\mathcal V^\prime|$;
- In any subset of $k>0$ equations in which $r$ variables appear, $k \leq r$, if the values of any $(r - k)$ variables are chosen arbitrarily, then the values of the remaining $k$ variables can be determined uniquely—finding these unique values is a matter of solving the equations.
Note in Def. \[def:structure\] that structures are composed of equations, and variables are only part of them indirectly as part of equations. Accordingly, set inclusion and all set operations such as union, intersection and difference are computed w.r.t. the equations. That is, if $\mathcal S(\mathcal E, \mathcal V)$ and $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ are structures, then we write $\mathcal S^\prime \!\subset \mathcal S$ when $\mathcal E^\prime \!\subset \mathcal E$. An additional operation for ‘variables elimination’ shall be used. We write $\mathcal T := \mathcal S \div \mathcal S^\prime$, to denote a structure $\mathcal T$ resulting from both (i) removing equations $\mathcal E^\prime$ from $\mathcal E$, and (ii) enforcing elimination of variables $\mathcal V^\prime = \bigcup_{f \in \mathcal E^\prime} Vars(f)$ from $\mathcal E \setminus \mathcal E^\prime$.
\[def:complete\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure. We say that $\mathcal S$ is self-contained or **complete** if $|\mathcal E|=|\mathcal V|$.
In short, COP will be concerned with systems of equations that are ‘structural’ (Def. \[def:structure\]) and ‘complete’ (Def. \[def:complete\]), viz., that have as many equations as variables and no subset of equations has fewer variables than equations.[^1]
Now Def. \[def:tcm\] introduces a data structure that is an intermediate product towards addressing COP. We shall state COP formally in the sequel.
\[def:tcm\] $\!$Let $\mathcal S(\mathcal E, \mathcal V)\!$ be a complete structure. $\!\!$Then a **total causal mapping** over $\mathcal S$ is a bijection $\varphi\!: \mathcal E \to \mathcal V$ such that, for all $f \!\in \mathcal E$, if $\varphi(f)=x$ then $x \!\in Vars(f)$.
Note that such total causal mapping $\varphi$ induces a set $C_{\varphi}$ of *direct causal dependencies* (see Eq. \[eq:direct-causal-dependencies\]), which shall give us the *causal dependencies* (Def. \[def:causal-dependency\]). $$\begin{aligned}
\!\!\!C_{\varphi} \!= \{\, (x_a, x_b) \,| \;\text{there exists}\; f \!\in \mathcal E \;\text{such that}\; \varphi(f) = x_b \;\text{and}\; x_a \in Vars(f) \,\}
\label{eq:direct-causal-dependencies}\end{aligned}$$
\[def:causal-dependency\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure with variables $x_a, x_b \in \mathcal V$, and $\varphi$ a total causal mapping over $\mathcal S$ inducing set of direct causal dependencies $C_{\varphi}$ and indirectly a transitive closure $C^+_{\varphi}$. We say that $(x_a, x_b)$ is a **direct causal dependency** in $\mathcal S$ if $(x_a, x_b) \in C_{\varphi}$, and that $(x_a, x_b)$ is a **causal dependency** in $\mathcal S$ if $(x_a, x_b) \in C^+_{\varphi}$.
In other words, $(x_a, x_b)$ is in $C_{\varphi}$ iff $x_b$ direct and causally depends on $x_a$, given the causal asymmetries induced by $\varphi$. Then by transitive reasoning on $C_{\varphi}$ we shall be able to infer the transitive closure $C_{\varphi}^+$, which is the *causal ordering*. Now we can state COP more formally as Problem \[prob:cop\].
\[prob:cop\] *(COP)*. Given a complete structure $\mathcal S(\mathcal E, \mathcal V)$, find a total causal mapping $\varphi$ over $\mathcal S$ and derive a set $C_\varphi^+$ of causal dependencies induced by it.
In the sequel we shall see two different algorithmic approaches to COP (Problem \[prob:cop\]). First, the classical approach informally described by Simon in the 50th’s [@simon1953], and reproduced recently in [@druzdzel2008]; and then Nayak’s one proposed in the 90th’s [@nayak1994]. We shall present the algorithms and analyze their corresponding complexities.
Simon’s Causal Ordering Algorithm and its Complexity {#sec:coa}
====================================================
We introduce now additional concepts that are specific to Simon’s COA.
\[def:minimal\] Let $\mathcal S$ be a structure. We say that $\mathcal S$ is **minimal** if it is complete and there is no complete substructure $\mathcal S^\prime \!\subset \mathcal S$.
Consider structure $\mathcal S(\mathcal E, \mathcal V)$, where $\mathcal E \!=\! \{\, f_1(x_1),\; f_2(x_2),\; f_3(x_3),$ $f_4(x_1, x_2, x_3, x_4, x_5),\; f_5(x_1, x_3, x_4, x_5),\; f_6(x_4, x_6),\; f_7(x_5, x_7) \,\}$. Note that $\mathcal S$ is complete, as $|\mathcal E|\!=\!|\mathcal V|\!=\!7$, but not minimal. There are exactly three minimal substructures $\mathcal{S}_1, \mathcal{S}_2, \mathcal{S}_3 \subset \mathcal S$, whose sets of equations are $\mathcal{E}_1 \!=\! \{f_1(x_1)\},\, \mathcal{E}_2 \!=\! \{f_2(x_2)\},\,\mathcal{E}_3 \!=\! \{f_3(x_3)\}$. $\Box$ \[ex:structure\]
Now Lemma \[lemma:disjoint\] and Proposition \[prop:disjoint\] are stated to back up a ‘disjointness’ assumption that is made by COA upon minimal structures (Def. \[def:minimal\]). The proof we present here for Proposition \[prop:disjoint\] is a conveniently derived alternative to Simon’s own proof to his original ‘theorem 3.2’ [@simon1953 p. 59].
\[lemma:disjoint\] Let $\mathcal S_1(\mathcal E_1, \mathcal V_1)$ and $\mathcal S_2(\mathcal E_2, \mathcal V_2)$ be structures. If $\;\mathcal V_1 \cap \mathcal V_2 = \varnothing$ then $\mathcal S_1 \cap \mathcal S_2 = \varnothing$ (i.e., $\mathcal E_1 \cap \mathcal E_2 = \varnothing$). That is, disjointness of variables is strong enough to warrant disjointness of equations.
Let $\mathcal V_1 \cap \mathcal V_2 = \varnothing$. Now by contradiction assume $\mathcal S_1 \cap \mathcal S_2 \neq \varnothing$, then there must be at least one shared equation $f \in \mathcal E_1, \mathcal E_2$. Since both $\mathcal S_1, \mathcal S_2$ are structures, by Def. \[def:structure\] we know that $|Vars(f)| \geq 1$ and $Vars(f) \subseteq \mathcal V_1 \cap \mathcal V_2$. Yet $\mathcal V_1 \cap \mathcal V_2 = \varnothing$. . Therefore $\mathcal V_1 \cap \mathcal V_2 = \varnothing$ implies $\mathcal S_1 \cap \mathcal S_2 = \varnothing$. $\Box$
\[def:disjoint\] Let $\mathcal S_1(\mathcal E_1, \mathcal V_1)$ and $\mathcal S_2(\mathcal E_2, \mathcal V_2)$ be structures. Then we say that they are **disjoint** if $\;\mathcal V_1 \cap \mathcal V_2 = \varnothing$.
\[prop:disjoint\] Let $\mathcal S$ be a complete structure. If $\mathcal S_1, \mathcal S_2 \subset \mathcal S$ are any different minimal substructures of $\mathcal S$, then they are disjoint.
We show the statement by case analysis and then contradiction out of Defs. \[def:structure\]–\[def:complete\] and Defs. \[def:minimal\]–\[def:disjoint\]. See \[app:disjoint\]. $\Box$
Simon’s COA is also based on a data structure introduced in Def. \[def:matrix\].
\[def:matrix\] The **structure matrix** $A_S$ of a structure $\mathcal S(\mathcal E, \mathcal V)$, with $f_1,\, f_2, ..., f_n \in \mathcal E$ and $x_1,\, x_2, ..., x_m \in \mathcal V$, is a $|\mathcal E| \times |\mathcal V|$ matrix of 1’s and 0’s in which entry $a_{ij}$ is non-zero if variable $x_j$ appears in equation $f_i$, and zero otherwise.
Elementary row operations on the structure matrix may hinder the structure’s causal ordering and then are not valid in general [@simon1953]. This also emphasizes that the problem of causal ordering is not about solving the system of equations of a structure, but identifying its hidden asymmetries.
Simon’s Causal Ordering Algorithm {#subsec:coa}
---------------------------------
Simon has described his Causal Ordering Algorithm (COA) only informally at a high level of abstraction [@simon1953]. It is given a complete structure $\mathcal S(\mathcal E, \mathcal V)$ and computes a causal mapping $\varphi$. The causal ordering itself is to be obtained as a post-processing (transitive closure) out of the causal mapping $\varphi$ and its induced set $C_\varphi$ of direct causal dependencies. Example \[ex:coa\] (continued) warms up for Simon’s algorithm.
(continued). Fig. \[fig:coa-a\] shows the matrix of the structure $\mathcal S$ given above in this example. By eliminating the variables identified with the minimal substructures $\mathcal{S}_1, \mathcal{S}_2, \mathcal{S}_3 \subset \mathcal S$, a smaller structure $\mathcal T$ is derived to be input at the next recursive step (see Fig. \[fig:coa-b\]). This is the main insight of Simon’s to arrive at his recursive causal ordering algorithm, as described next. $\Box$ \[ex:struct-matrix\]
[1.2]{}
[0.29]{}
\(A) \[matrix of math nodes, nodes = [node style ge]{},column sep=0.6mm\] [ & (x1) [x\_1]{}; & (x2) [x\_2]{}; & (x3) [x\_3]{}; & (x4) [x\_4]{}; & (x5) [x\_5]{}; & (x6) [x\_6]{}; & (x7) [x\_7]{};\
(f1) [f\_1]{}; & (a11) [1]{}; & (a12) [0]{}; & (a13) [0]{}; & (a14) [0]{}; & (a15) [0]{}; & (a16) [0]{}; & (a17) [0]{};\
(f2) [f\_2]{}; & (a21) [0]{}; & (a22) [1]{}; & (a23) [0]{}; & (a24) [0]{}; & (a25) [0]{}; & (a26) [0]{}; & (a27) [0]{};\
(f3) [f\_3]{}; & (a31) [0]{}; & (a32) [0]{}; & (a33) [1]{}; & (a34) [0]{}; & (a35) [0]{}; & (a36) [0]{}; & (a37) [0]{};\
(f4) [f\_4]{}; & (a41) [1]{}; & (a42) [1]{}; & (a43) [1]{}; & (a44) [1]{}; & (a45) [1]{}; & (a46) [0]{}; & (a47) [0]{};\
(f5) [f\_5]{}; & (a51) [1]{}; & (a52) [0]{}; & (a53) [1]{}; & (a54) [1]{}; & (a55) [1]{}; & (a56) [0]{}; & (a57) [0]{};\
(f6) [f\_6]{}; & (a61) [0]{}; & (a62) [0]{}; & (a63) [0]{}; & (a64) [1]{}; & (a65) [0]{}; & (a66) [1]{}; & (a67) [0]{};\
(f7) [f\_7]{}; & (a71) [0]{}; & (a72) [0]{}; & (a73) [0]{}; & (a74) [0]{}; & (a75) [1]{}; & (a76) [0]{}; & (a77) [1]{};\
]{};
=\[rectangle, fill=gray!0, inner sep=0.025cm, rounded corners=1mm\] =\[rectangle, fill=gray!30, inner sep=0.025cm, rounded corners=1mm\] =\[rectangle, fill=gray!70, inner sep=0.025cm, rounded corners=1mm\]
[0.33]{}
\(A) \[matrix of math nodes, nodes = [node style ge]{},column sep=0.6 mm\] [ & (x1) [x\_1]{}; & (x2) [x\_2]{}; & (x3) [x\_3]{}; & (x4) [x\_4]{}; & (x5) [x\_5]{}; & (x6) [x\_6]{}; & (x7) [x\_7]{};\
(f1) [f\_1]{}; & (a11) [1]{}; & (a12) [0]{}; & (a13) [0]{}; & (a14) [0]{}; & (a15) [0]{}; & (a16) [0]{}; & (a17) [0]{};\
(f2) [f\_2]{}; & (a21) [0]{}; & (a22) [1]{}; & (a23) [0]{}; & (a24) [0]{}; & (a25) [0]{}; & (a26) [0]{}; & (a27) [0]{};\
(f3) [f\_3]{}; & (a31) [0]{}; & (a32) [0]{}; & (a33) [1]{}; & (a34) [0]{}; & (a35) [0]{}; & (a36) [0]{}; & (a37) [0]{};\
(f4) [f\_4]{}; & (a41) [1]{}; & (a42) [1]{}; & (a43) [1]{}; & (a44) [1]{}; & (a45) [1]{}; & (a46) [0]{}; & (a47) [0]{};\
(f5) [f\_5]{}; & (a51) [1]{}; & (a52) [0]{}; & (a53) [1]{}; & (a54) [1]{}; & (a55) [1]{}; & (a56) [0]{}; & (a57) [0]{};\
(f6) [f\_6]{}; & (a61) [0]{}; & (a62) [0]{}; & (a63) [0]{}; & (a64) [1]{}; & (a65) [0]{}; & (a66) [1]{}; & (a67) [0]{};\
(f7) [f\_7]{}; & (a71) [0]{}; & (a72) [0]{}; & (a73) [0]{}; & (a74) [0]{}; & (a75) [1]{}; & (a76) [0]{}; & (a77) [1]{};\
]{}; (a11.north west) to (a11.south east); (a22.north west) to (a22.south east); (a33.north west) to (a33.south east); (a44.north west) to (a55.south east); (a66.north west) to (a66.south east); (a77.north west) to (a77.south east);
[background]{} ; ; ;
[0.33]{}
=\[circle,fill=black!22,minimum size=16pt,inner sep=0pt\] = \[vertex, fill=red!50\] = \[draw,thick,->,bend left\] = \[font=\] = \[draw,line width=5pt,-,red!50\] = \[draw,line width=5pt,-,black!20\]
/in [[(-0.5,2)/x\_1]{}, [(2,2)/x\_2]{}, [(4.5,2)/x\_3]{}, [(0.7,0)/x\_4]{}, [(3.3,0)/x\_5]{}, [(1,-2)/x\_6]{}, [(3,-2)/x\_7]{}]{} () at ; (x\_1) to (x\_4); (x\_1) to (x\_5); (x\_2) to (x\_4); (x\_3) to (x\_4); (x\_3) to (x\_5); (x\_4) to\[out=10,in=200\] (x\_5); (x\_5) to\[out=170,in=-20\] (x\_4); (x\_4) to (x\_6); (x\_5) to (x\_7);
[1.1]{}
$\mathcal S$ given is complete, i.e., $|\mathcal E|=|\mathcal V|$ Returns $C_\varphi^+$, the causal ordering of $\mathcal S$ $\varphi \gets \textsf{RTCM}(\mathcal S)$ $C_\varphi \gets \varnothing$ $C_\varphi \gets C_\varphi \cup \{(x_a, x)\}$ $\textsf{TC}(C_\varphi)$
------------------------------------------------------------------------
Structure $\mathcal S$ given is complete, i.e., $|\mathcal E|=|\mathcal V|$ Returns total causal mapping $\varphi: \mathcal E \to \mathcal V$ $\varphi \gets \varnothing$, $\mathcal S^\star \gets \varnothing$, $D \gets \varnothing$ identify all minimal substructures $\mathcal S^\prime \subseteq \mathcal S$ $\mathcal S^\star \gets \mathcal S^\star \cup \mathcal S^\prime$ $x \gets \text{any} \;x_a$ such that $x_a \in Vars(f)$ and $x_a \notin D$ $\varphi \gets \varphi \cup \langle f,\, x \rangle$ $D \gets D \cup \{x\}$ $\mathcal T \gets \mathcal S \div \mathcal S^\star$ $\varphi \;\cup\;$$\textsf{RTCM}$$(\mathcal T)$ $\varphi$
Algorithm \[alg:coat\] describes the variant of Simon’s original description that returns a ‘total’ causal mapping (satisfies Def. \[def:tcm\]).[^2] We refer to its core procedure as RTCM (recursive total causal mapping). It comprises the actual source of intractability in Simon’s original description, while lines 3-7 of the COA procedure were not described by himself but only considered as matter of a post-processing. We illustrate RTCM through Example \[ex:coa\] and Fig. \[fig:coa\].
(continued). Let $\mathcal T = \mathcal S \div (\mathcal S_1 \cup \mathcal S_2 \cup \mathcal S_3)$ be the structure returned by COA’s first recursive step $k=0\,$ for this example. $\!$Then a valid total causal mapping that can be returned at $k=1$ (see Fig.\[fig:coa-b\]) is COA$(\mathcal T) = \{ \langle f_4, x_4 \rangle,\, \langle f_5, x_5 \rangle\}$. Since $x_4$ and $x_5$ are strongly coupled, COA maps them arbitrarily (e.g., it could be $f_4 \mapsto x_5,\, f_5 \mapsto x_4$ instead). Such total causal mapping $\varphi$ renders a cycle in the directed causal graph $G_{\varphi}$ induced by $\varphi$ (see Fig.\[fig:causal-graph\]), which might not be desirable for some applications. $\Box$ \[ex:coa\]
Hardness of Simon’s Recursive COA {#subsec:hard}
---------------------------------
First of all, we state a decision problem associated with finding the minimal structures in a given structure (line 3 of Simon’s RTCM procedure in Algorithm \[alg:coat\]). For short, we shall refer to this problem as the Complete Substructure Decision Problem (CSDP).
(CSDP). Given a complete structure $\mathcal S(\mathcal E, \mathcal V)$ with $|\mathcal E|=|\mathcal V|=m$ and an integer $1 \leq \ell < m$, does $\mathcal S$ have a complete substructure $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ with $|\mathcal E^\prime|=|\mathcal V^\prime|=\ell$?
In this section we carry out an original study on CSDP and show that it is NP-Complete. We consider a basic observation by Nayak [@nayak1994] apud. [@serrano1987], that there is a correspondence between Simon’s structures and bipartite graphs. A graph is said *bipartite* if its vertices can be divided into two disjoint sets $V_1$ and $V_2$ and every edge connects a vertex in $V_1$ to one in $V_2$ [@bondy1976]. Moreover it is said to be $\ell$-*balanced* if $|V_1|=|V_2|=\ell$, and is said to be *connected* if $deg(w) \geq 1$ for all $w \in V_1 \cup V_2$. Def. \[def:bipartite\] introduces the mentioned correspondence and provides us some shorthand notation.
\[def:bipartite\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure, and $G=(V_1 \cup V_2, E)$ be a bipartite graph where $V_1 \mapsto \mathcal E$ and $V_2 \mapsto \mathcal V$, and $E \mapsto \mathcal S$ so that an edge $(f, x) \in E$ if and only if we have $x \in Vars(f)$. We say that $G$ is the bipartite graph that **corresponds to** structure $\mathcal S$, and for short write $G \sim \mathcal S$.
Fig. \[fig:bipartite\] shows the bipartite graph $G \sim \mathcal S$, where $\mathcal S$ is the initial structure given in Example \[ex:struct-matrix\].
Def. \[def:structural\] introduces a bipartite graph property of our interest, and then Lemma \[lemma:bipartite\] originally establishes an equivalence of two problems: searching for complete substructures $\mathcal S^\prime \subset \mathcal S$ and searching for specific bipartite subgraphs $G^\prime \subset G$.
\[def:structural\] Let $G(V_1 \cup V_2, E)$ be a bipartite graph. We say that $G$ is **structural** if, for every $V_1^\prime \subseteq V_1$, there is a connected bipartite subgraph $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$ with $|V_1^\prime| \leq |V_2^\prime|$. (Note resemblance with Def. \[def:structure\]).
=\[rectangle, thick, minimum size=0.3cm, draw=black\] =\[circle, thick, minimum size=0.3cm, draw=black\] =\[circle,fill=black!10,minimum size=20pt,inner sep=0pt\] = \[vertex, fill=red!24\] = \[draw,thick,->,bend left\] = \[font=\] = \[draw,line width=5pt,-,red!50\] = \[draw,line width=5pt,-,black!20\]
/in [[(0,10)/f\_1]{}, [(6,10)/x\_1]{}, [(0,9)/f\_2]{}, [(6,9)/x\_2]{}, [(0,8)/f\_3]{}, [(6,8)/x\_3]{}, [(0,7)/f\_4]{}, [(6,7)/x\_4]{}, [(0,6)/f\_5]{}, [(6,6)/x\_5]{}, [(0,5)/f\_6]{}, [(6,5)/x\_6]{}, [(0,4)/f\_7]{}, [(6,4)/x\_7]{}]{} () at ; (f\_1) to (x\_1); (f\_2) to (x\_2); (f\_3) to (x\_3); (f\_4) to (x\_1); (f\_4) to (x\_2); (f\_4) to (x\_3); (f\_4) to (x\_4); (f\_4) to (x\_5); (f\_5) to (x\_1); (f\_5) to (x\_3); (f\_5) to (x\_4); (f\_5) to (x\_5); (f\_6) to (x\_4); (f\_6) to (x\_6); (f\_7) to (x\_5); (f\_7) to (x\_7);
\[lemma:bipartite\] Let $\mathcal S(\mathcal V, \mathcal E)$ be a complete structure with $|\mathcal E|=|\mathcal V|=m$ and $1 \leq \ell < m$ provide an instance of CSDP. Let also $G(V_1 \cup V_2, E)$ be a bipartite graph $G \sim \mathcal S$. Then $\mathcal S$ has a substructure $\mathcal S^\prime$ that gives a yes answer to CSDP if and only if $G$ has a bipartite subgraph $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$ such that $G^\prime \sim \mathcal S^\prime$ and all of these conditions hold:
- Bipartite subgraph $G^\prime$ is structural;
- For every $f \in V_1^\prime$, there is an edge $(f, x) \in E$ only if $x \in V_2^\prime$;
- Bipartite subgraph $G^\prime$ is $\ell$-balanced, that is, $|V_1^\prime| = |V_2^\prime|=\ell$;
We establish conditions (i-iii) as the bipartite subgraph properties that correspond to a yes answer to CSDP. See \[app:coa-bipartite\]. $\Box$
We now reach the key property in our argument to show COA’s hardness. A *biclique* (or complete bipartite graph) is a bipartite graph $G=(V_1 \cup V_2,\, E)$ such that for every two vertices $u \in V_1$, $v \in V_2$, we have $(u,\, v) \in E$ [@even2011]. Thus the number of edges in a biclique is $|E| = |V_1| \cdot |V_2|$. A biclique with partitions of size $|V_1|=m\,$ and $|V_2|=n$ is denoted $K_{m, n}$. For instance, the bipartite graph $G$ shown in Fig. \[fig:bipartite\] has a $K_{2, 2}$ biclique, viz., $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$, where $V_1^\prime = \{f_4, f_5\}$, $V_2^\prime = \{x_4, x_5\}$ and $E^\prime \!=\! \{(f_4, x_4), (f_4, x_5), (f_5, x_4), (f_5, x_5)\}$. Let us now consider Example \[ex:hardness\].
[1.2]{}
=\[rectangle, fill=gray!17, inner sep=0.025cm, rounded corners=1mm\] =\[rectangle, fill=gray!65, inner sep=0.025cm, rounded corners=1mm\]
[0.33]{}
\(A) \[matrix of math nodes, nodes = [node style ge]{},column sep=1.0 mm\] [ & (x1) [x\_1]{}; & (x2) [x\_2]{}; & (x3) [x\_3]{}; & (x4) [x\_4]{};\
(f1) [f\_1]{}; & (a11) [1]{}; & (a12) [0]{}; & (a13) [1]{}; & (a14) [0]{};\
(f2) [f\_2]{}; & (a21) [1]{}; & (a22) [1]{}; & (a23) [0]{}; & (a24) [0]{};\
(f3) [f\_3]{}; & (a31) [0]{}; & (a32) [1]{}; & (a33) [1]{}; & (a34) [0]{};\
(f4) [f\_4]{}; & (a41) [1]{}; & (a42) [1]{}; & (a43) [1]{}; & (a44) [1]{};\
]{};
[background]{} ; ;
[0.32]{} =\[circle, thick, minimum size=0.2cm, draw=black\] =\[circle,fill=black!10,minimum size=18pt,inner sep=0pt\] = \[vertex, fill=black!30\] = \[draw,thick,->,bend left\] = \[font=\] = \[draw,line width=5pt,-,red!50\] = \[draw,line width=5pt,-,black!20\]
/in [[(-0.5,10)/f\_1]{}, [(3.0,10)/x\_1]{}, [(-0.5,9)/f\_2]{}, [(3.0,9)/x\_2]{}, [(-0.5,8)/f\_3]{}, [(3.0,8)/x\_3]{}]{} () at ; (f\_4) at (-0.5,7) [$f_4$]{}; (x\_4) at (3.0,7) [$x_4$]{};
(f\_1) to (x\_1); (f\_1) to (x\_3); (f\_2) to (x\_1); (f\_2) to (x\_2); (f\_3) to (x\_2); (f\_3) to (x\_3); (f\_4) to (x\_1); (f\_4) to (x\_2); (f\_4) to (x\_3); (f\_4) to (x\_4);
[0.35]{} =\[circle, thick, minimum size=0.2cm, draw=black\] =\[circle,fill=black!10,minimum size=18pt,inner sep=0pt\] = \[vertex, fill=black!30\] = \[draw,thick,->,bend left\] = \[font=\] = \[draw,line width=5pt,-,red!50\] = \[draw,line width=5pt,-,black!20\]
/in [[(3.0,10)/x\_1]{}, [(3.0,9)/x\_2]{}, [(3.0,8)/x\_3]{}]{} () at ; (f\_1) at (-0.5,10) [$f_1$]{}; (f\_2) at (-0.5,9) [$f_2$]{}; (f\_3) at (-0.5,8) [$f_3$]{}; (f\_4) at (-0.5,7) [$f_4$]{}; (x\_4) at (3.0,7) [$x_4$]{};
(f\_1) to (x\_2); (f\_1) to (x\_4); (f\_2) to (x\_3); (f\_2) to (x\_4); (f\_3) to (x\_1); (f\_3) to (x\_4);
\[ex:hardness\] We introduce another structure $\mathcal S$, whose structure matrix is shown in Fig. \[fig:hard-matrix\] together with the bipartite graph $G \sim \mathcal S$ in Fig. \[fig:hard-graph\]. Let us consider subgraph $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$ in $G$ that has $V_1^\prime=\{f_1, f_2, f_3\}$ and $V_2^\prime=\{x_1, x_2, x_3\}$. Observe that we have $G^\prime \sim \mathcal S^\prime$, where $\mathcal S^\prime \subset \mathcal S$ is the complete substructure represented by the shaded $3 \times 3$ matrix in Fig. \[fig:hard-matrix\].
Note also that such bipartite subgraph $G^\prime$ satisfies the conditions (i-iii) of Lemma \[lemma:bipartite\] and in fact $\mathcal S^\prime$ is a complete substructure in $\mathcal S$. Clearly, $G^\prime \sim \mathcal S^\prime$ is not a biclique, as it is not the case that $deg(w) = 3$ for all $w \in V_1^\prime \cup V_2^\prime$. So there is no obvious connection between identifying complete substructures in a structure and bicliques in a bipartite graph. $\Box$
The key insight to COA’s hardness comes as follows—consider Example \[ex:hardness\] and Fig. \[fig:hardness\] for illustration. Recall from Lemma \[lemma:bipartite\](ii) that, if we had an edge, say, connecting $(f_1, x_4) \in E$, then by Def. \[def:structure\] the substructure $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ with $\mathcal E^\prime = \{f_1, f_2, f_3\}$ would have $\mathcal V^\prime = \bigcup_{f \in \mathcal E^\prime} Vars(f) = \{x_1, x_2, x_3, x_4\}$ instead. That is, it would no more be a complete substructure. In fact, verifying such a negative property (Lemma \[lemma:bipartite\].ii) in structural bipartite graphs translates onto a neat positive property (biclique) in the bipartite complement of bipartite graph $G$.
The *bipartite complement* of a bipartite graph $G(V_1 \cup V_2 , E)$ is a bipartite graph $G^c(V_1 \cup V_2, E^c)$ where an edge $(u, v) \in E^c$ iff $(u, v) \notin E$ for every $u \in V_1$ and $v \in V_2$. Given a bipartite graph $G(V_1 \cup V_2, E)$, it is easy to see that we can render $G^c(V_1 \cup V_2, E^c)$ in polynomial time—consider, e.g., the biadjacency matrix of $G$ (viz., the structure matrix in Fig. \[fig:hard-matrix\]), and run a full scan on it to switch the boolean value of each entry in time $O(|V_1|\cdot |V_2|)$. Moreover, this operation is clearly invertible, as there is a one-to-one correspondence between $G$ and $G^c$.
Fig. \[fig:hard-cgraph\] shows the bipartite complement graph $G^c$ of the bipartite graph $G$ from Fig. \[fig:hard-graph\]. Note that $G^c$ has a biclique $K_{3, 1}$ with its vertices shaded in dark grey. To emphasize the point, if we had an edge $(f_1, x_4) \in E$ (see Fig. \[fig:hard-graph\]), then such a biclique $K_{3, 1}$ would not exist in $G^c$ (see Fig. \[fig:hard-cgraph\]). We would have a $K_{2,\, 1}$ biclique instead with all edges in $\{f_2, f_3\} \times \{x_4\}$, but note that $2+1=3$ does not sum up to $|V_1|=|V_2|=m=4$.
We can now establish the result we seek. We introduce below the Exact Node Cardinality Decision Problem (ENCD), which is a variant of biclique problem in bipartite graphs that is known to be NP-Complete [@dawande2001 p. 393]. Theorem \[thm:biclique\] establishes its connection with CSDP.
$\!$(ENCD). Given a bipartite graph $G=(V_1 \cup V_2,\, E)$ and two positive integers $a, b$, does $G$ have a biclique $K_{a,\, b}$?
\[thm:biclique\] CSDP is NP-Complete.
We shall construct an instance of ENCD and describe its polynomial-time reduction to an instance of CSDP. We refer to Def. \[def:structural\] and Lemma \[lemma:bipartite\] and present the argument in detail in \[app:coa-npcomplete\]. $\Box$
Finally, we formulate an optimization problem associated with CSDP. We refer to it as the Minimal Substructures Problem (MSP). Corollary \[cor:coa-hardness\] then finally establishes the hardness of Simon’s COA based on RTCM.
(MSP). Given a complete structure $\mathcal S(\mathcal E, \mathcal V)$ with $|\mathcal E|=|\mathcal V|=m$, list all its complete substructures $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ with $|\mathcal E^\prime|=|\mathcal V^\prime|=\ell$ where $1 \leq \ell < m$ is minimal.
\[cor:coa-hardness\] Let $\mathcal S$ be a complete structure. The extraction of its causal ordering by Simon’s $COA(\mathcal S)$ through its RTCM procedure requires solving MSP, which is NP-Hard.
Clearly, MSP is the optimization problem that needs to be solved at each recursive step $k$ of Simon’s RTCM procedure — Algorithm \[alg:coat\], line 3, “find all minimal substructures $\mathcal S^\prime \subseteq \mathcal S$.” But MSP is clearly an optimization problem that subsumes CSDP, which we know from Theorem \[thm:biclique\] that is NP-Complete by a reduction from ENCD.
In fact, an instance of ENCD$^{\,\prime}$ (as an optimization variant of ENCD) that can be reduced to MSP is as follows: given a bipartite graph $G(V_1 \cup V_2, E)$ that bears the complement structural property (cf. Theorem \[thm:biclique\]) and has $|V_1|=|V_2|=m$, list all bicliques $K_{\ell,\, m-\ell}$ contained in $G$ where $1 \leq \ell < m$ is minimal. In worst-case scenario, it requires searching for all bicliques $K_{\ell,\, m-\ell}$ for each of the $m-1$ possible values of $\,\ell$.
ENCD is NP-Complete, therefore ENCD$^{\,\prime}$ is NP-Hard. Accordingly, CSDP is NP-Complete (cf. Theorem \[thm:biclique\]) therefore MSP is NP-Hard. $\Box$
COP (Problem \[prob:cop\]), nonetheless, can be solved efficiently by means of a different approach due to Nayak [@nayak1994], which we describe in next section.
Nayak’s Efficient Algorithm to COP {#sec:nayak}
==================================
The first part of COP requires finding a total causal mapping $\varphi\!:\, \mathcal E \to \mathcal V$ over a given complete structure $\mathcal S$. While Simon’s RTCM goes into an intractable step, inspired by Serrano and Gossard’s work [@serrano1987] on constraint modeling and reasoning Nayak has found a polynomial-time approach to that task. We cover it next in all of its steps and see their complexity in some detail.
Total Causal Mappings {#sec:tcm}
---------------------
Given a structure $\mathcal S$, there may be more than one total causal mappings over $\mathcal S$ (recall Example \[ex:struct-matrix\]). So a question that arises is whether the transitive closure $C^+_{\varphi}$ is the same for any total causal mapping $\varphi$ over $\mathcal S$; that is, whether the causal ordering of $\mathcal S$ is unique. Proposition \[prop:causal-ordering\], from Nayak [@nayak1994], ensures that is the case.
Before proceeding, we introduce Def. \[def:strongly-coupled\] in order to detach the notion of ‘strongly coupled’ variables from ‘minimal structures’ (Def. \[def:minimal\]) and connect it to the concept ‘causal dependency’ (Def. \[def:causal-dependency\]).
\[def:strongly-coupled\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure with variables $x_a, x_b \in \mathcal V$, and $C^+_{\varphi}$ be the set of causal dependencies induced by total causal mapping $\varphi$ over $\mathcal S$. We say that $x_a$ and $x_b$ are **strongly coupled** if we have both $(x_a, x_b), (x_b, x_a) \in C_{\varphi}^+$.
Recall from Example \[ex:coa\] the strongly coupled variables, $x_4$ and $x_5$. Now we can see it only in terms of causal dependencies.
\[prop:causal-ordering\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure, and $\varphi_1\!:\, \mathcal E \to \mathcal V$ and $\varphi_2\!:\, \mathcal E \to \mathcal V$ be any two total causal mappings over $\mathcal S$. Then $C^+_{\varphi_1}$ = $C^+_{\varphi_2}$.
The proof is based on an argument from Nayak [@nayak1994], which we present in a bit more of detail (see \[coa-efficient\]). Intuitively, it shows that if $\varphi_1$ and $\varphi_2$ differ in the variable an equation $f$ is mapped to, then such variables, viz., $\varphi_1(f)=x_a$ and $\varphi_2(f)=x_b$, must be causally dependent on each other (strongly coupled). $\Box$
Another issue is concerned with the precise conditions under which total causal mappings exist (i.e., whether or not all variables in the equations can be causally determined). In fact, by Proposition \[prop:mapping-existence\], based on Nayak [@nayak1994] apud. Hall [@even2011 p. 135-7], we know that the existence condition holds if and only if the given structure is complete. We refer to Even [@even2011] to briefly introduce the additional graph-theoretic concepts that are necessary here:
- A *matching* in a graph is a subset of edges such that no two edges in the matching share a common node.
- A matching is said *maximum* if no edge can be added to the matching (without hindering the matching property).
- Finally, a matching in a graph is said ‘perfect’ if every vertex is an end-point of some edge in the matching — in a bipartite graph, a perfect matching is said to be a *complete* matching.
\[prop:mapping-existence\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure. Then a total causal mapping $\varphi\!:\, \mathcal E \to \mathcal V$ over $\mathcal S$ exists if and only if $\mathcal S$ is complete.
We observe that a total causal mapping $\varphi\!:\, \mathcal E \to \mathcal V$ over $\mathcal S$ corresponds exactly to a complete matching $M$ in a bipartite graph $B = (V_1 \cup V_2, E)$, where $V_1 \mapsto \mathcal E$, $V_2 \mapsto \mathcal V$, and $E \mapsto \mathcal S$. In fact, by Even apud. Hall’s theorem (cf. [@even2011 135-7]), we know that $B$ has a complete matching iff (a) for every subset of vertices $F \subseteq V_1$, we have $|F| \leq |E(F)|$, where $E(F)$ is the set of all vertices connected to the vertices in $F$ by edges in $E$; and (b) $|V_1|=|V_2|$. By Def. \[def:structure\] (no subset of equations has fewer variables than equations), and Def. \[def:complete\] (number of equations is the same as number of variables), it is easy to see that conditions (a) and (b) above hold iff $\mathcal S$ is a complete structure. $\Box$
The problem of finding a maximum matching is a well-studied algorithmic problem. The Hopcroft-Karp algorithm is a classical solution [@karp1973], bounded in polynomial time by $O(\sqrt{|V_1|+|V_2|}\,|E|)$. It solves maximum matching in a bipartite graph efficiently as a problem of maximum flow in a network (cf. [@even2011 p. 135-7], or [@cormen2009 p. 763]). That is, we can handle the problem of finding a total causal mapping $\varphi$ over a structure $\mathcal S$ (see Alg. \[alg:tcm\]) by first translating it to the problem of maximum matching in a bipartite graph in time $O(|\mathcal S|)$. Then we can just apply the Hopcroft-Karp algorithm to get the matching and finally translate it back to the total causal mapping $\varphi$. This procedure has been suggested by Nayak in connection with his Proposition \[prop:mapping-existence\] and its respective proof [@nayak1994].
Fig. \[fig:matching\] shows the complete matching found by the Hopcroft-Karp algorithm for the structure given in Example \[ex:struct-matrix\].
=\[rectangle, thick, minimum size=0.3cm, draw=black\] =\[circle, thick, minimum size=0.3cm, draw=black\] =\[circle,fill=black!10,minimum size=20pt,inner sep=0pt\] = \[vertex, fill=red!24\] = \[draw,thick,->,bend left\] = \[font=\] = \[draw,line width=5pt,-,red!50\] = \[draw,line width=5pt,-,black!20\]
/in [[(0,12)/f\_1]{}, [(6,12)/x\_1]{}, [(0,11)/f\_2]{}, [(6,11)/x\_2]{}, [(0,10)/f\_3]{}, [(6,10)/x\_3]{}, [(0,9)/f\_4]{}, [(6,9)/x\_4]{}, [(0,8)/f\_5]{}, [(6,8)/x\_5]{}, [(0,7)/f\_6]{}, [(6,7)/x\_6]{}, [(0,6)/f\_7]{}, [(6,6)/x\_7]{}]{} () at ; (f\_1) to (x\_1); (f\_2) to (x\_2); (f\_3) to (x\_3); (f\_4) to (x\_4); (f\_5) to (x\_5); (f\_6) to (x\_6); (f\_7) to (x\_7);
[1.1]{}
$\mathcal S$ given is a complete structure, i.e., $|\mathcal E|=|\mathcal V|$ Returns a total causal mapping $\varphi$ $B(V_1 \cup V_2, E) \gets \varnothing$ $\varphi \gets \varnothing$ $V_1 \gets V_1 \cup \{f\}$ $V_2 \gets V_2 \cup \{x\}$ $E \gets E \cup \{(f, x)\}$ $M \gets \textsf{Hopcroft-Karp}(B)$ $\varphi \gets \varphi \cup \{\langle f, x\rangle\}$ $\varphi$
Corollary \[cor:tcm\] and Remark \[rmk:correct-mapping\] summarize the results presented in this note.
\[cor:tcm\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a complete structure. Then a total causal mapping $\varphi\!:\, \mathcal E \to \mathcal V$ over $\mathcal S$ can be found by (Alg. \[alg:tcm\]) TCM in time that is bounded by $O(\sqrt{|\mathcal V|} \cdot |\mathcal S|)$.
Let $B = (V_1 \cup V_2, E)$ be the bipartite graph corresponding to complete structure $\mathcal S$ given to TCM, where $V_1 \mapsto \mathcal E$, $V_2 \mapsto \mathcal V$, and $E \mapsto \mathcal S$. The translation of $\mathcal S$ into $B$ is done by a scan over it. This scan is of length $|\mathcal S| = |E|$. Note that number $|E|$ of edges rendered is precisely the length $|\mathcal S|$ of structure, where the denser the structure, the greater $|\mathcal S|$ is. The re-translation of the matching computed by internal procedure Hopcroft-Karp, in turn, is done at expense of $|\mathcal E| = |\mathcal V| \leq |\mathcal S|$. Thus, it is easy to see that TCM is dominated by the maximum matching algorithm Hopcroft-Karp, which is known to be $O(\sqrt{|V_1|+|V_2|}\cdot |E|)$, i.e., $O(\sqrt{|\mathcal E|+|\mathcal V|} \cdot |\mathcal S|)$. Since $\mathcal S$ is assumed complete, we have $|\mathcal E| \!=\! |\mathcal V|$ then $\sqrt{|\mathcal V|+|\mathcal V|} = \sqrt{2}\,\sqrt{|\mathcal V|}$. Therefore, must have running time at most $O(\sqrt{|\mathcal V|}\cdot |\mathcal S|)$. $\Box$
Computing Transitive Closure {#subsec:closure}
----------------------------
Finally, recall that the set $C_{\varphi}$ of direct causal dependencies induced by a total causal mapping $\varphi$ over a given structure $\mathcal S(\mathcal E, \mathcal V)$ produces to the so-called ‘causal graph,’ i.e., a directed graph (digraph) $G(V, E)$ where $V \mapsto \mathcal V$ and $E \mapsto C_{\varphi}$. So, computing set $C_{\varphi}^+$ of causal dependencies given $C_{\varphi}$ corresponds to computing transitive closure (reachability links) on $G$. Note that $|V|=|\mathcal V|$, and also note that $|E|=|C_{\varphi}|=|\mathcal S|\!-\!|\mathcal V|=O(|\mathcal S|)$.
Classical algorithms for such task (e.g., Floyd-Warshall’s) are bounded in time $O(|\mathcal V|^3)$ [@cormen2009 p. 697]. Another way to do it is by discovering reachability links using either one of the popular graph traversal algorithms, breadth-first search or depth-first search (DFS) [@cormen2009 p. 603]. Algorithm \[alg:tc\] describes DFS-based transitive closure over digraph $G(V, E)$. It runs in time $O(|V|\cdot |E|)$, which means $O(|\mathcal V|\cdot |\mathcal S|)$ for a complete structure $\mathcal S(\mathcal E, \mathcal V)$.
[1.1]{}
$E^+ \gets \varnothing$ $D \gets \varnothing$ $\text{DFS}(G,v, D)$ $D \gets D \setminus \{v\}$ $E^+ \gets \bigcup_{u \in D} \{(u, v)\} \cup E^+$ $G^+(V, E^+)$
------------------------------------------------------------------------
$D \gets D \cup \{v\}$ $\text{DFS}(G,u, D)$
\[rmk:correct-mapping\] Let $\mathcal S(\mathcal E, \mathcal V)$ be a complete structure. Then we know (cf. Proposition \[prop:mapping-existence\]) that a total causal mapping over $\mathcal S$ exists. Let it be defined $\varphi \triangleq TCM(\mathcal S)$, which can be done in $O(\sqrt{|\mathcal V|} \cdot |\mathcal S|)$. Then the causal ordering implicit in $\mathcal S$ can be correctly extracted (cf. Proposition \[prop:causal-ordering\]) by computing $C^+_\varphi$, the set of causal dependencies induced by $\varphi$, in terms of graph transitive closure (TC). The latter is bounded in time by $O(|\mathcal V|\cdot |\mathcal S|)$, that is, the complexity of COP is dominated by TC.
In other words, the complete recipe to solve COP consists in replacing Simon’s RTCM by Nayak’s TCM in COA (Algorithm \[alg:coat\]). Transitive closure (TC) in turn is computed as described in Algorithm \[alg:tc\]. $\Box$
Conclusions
===========
Causal ordering inference is a classical problem in the AI literature, and still relevant in light of modern applications [@druzdzel2008], e.g., large-scale hypothesis management and analytics [@goncalves2015cise]. In this note we have shown that Simon’s classical algorithm (COA) tries to address an NP-Hard problem; and then we have given a detailed account on the state-of-the-art algorithms for the causal ordering problem (COP, stated as Problem \[prob:cop\]). The key points are:
- By Theorem \[thm:biclique\] and Corollary \[cor:coa-hardness\], we know (an original hardness result) that Simon’s approach to COP requires solving an NP-Hard problem;
- From the seminal work of Simon [@simon1953] (cf. §\[sec:problem\]) and Nayak [@nayak1994] (cf. §\[sec:nayak\], and Propositions \[prop:causal-ordering\] and \[prop:mapping-existence\]), an approach is conveyed to solve COP efficiently;
- By Corollary \[cor:tcm\], we know how to process a complete structure into a total causal mapping in time that is bounded by $O(\sqrt{|\mathcal V|}\cdot |\mathcal S|)$. This is a core step to solve COP, which Simon’s COA in turn makes intractable.
- By Remark \[rmk:correct-mapping\], we know how to extract the *causal ordering* of a complete structure in time $O(|\mathcal V|\cdot |\mathcal S|)$, that is, in sub-quadratic time on the structure density (number of variable appearances). The machinery of causal ordering is then suitable for processing very large structures.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank three anonymous reviewers for their careful reading and sharp criticism on a previous version of this manuscript. This work has been supported by the Brazilian funding agencies CNPq (grants n$^o\!$ 141838/2011-6, 309494/2012-5) and FAPERJ (grants INCT-MACC E-26/170.030/2008, ‘Nota $\!$10’ $\!$E-26/100.286/2013). We thank IBM for a Ph.D. Fellowship award.
Proof of Proposition \[prop:disjoint\] {#app:disjoint}
======================================
*Let $\mathcal S$ be a complete structure. If $\mathcal S_1, \mathcal S_2 \subset \mathcal S$ are any different minimal substructures of $\mathcal S$, then they are disjoint.*
We show the statement by case analysis and then contradiction out of Defs. \[def:structure\]–\[def:minimal\]. By assumption both $\mathcal S_1, \mathcal S_2$ are minimal (hence complete). Let their size be $|\mathcal V_1|=|\mathcal E_1|=\ell$ and $|\mathcal V_2|=|\mathcal E_2|=m$. Let also $\ell \leq m$. The argument is analogous otherwise but it shall be convenient to keep a placeholder for the size of the smaller substructure (with no loss of generality).
By Def. \[def:minimal\] (minimal structures), we know that $\mathcal S_1 \not\subseteq \mathcal S_2$ and $\mathcal S_1 \not\supseteq \mathcal S_2$. Now suppose $\mathcal S_1, \mathcal S_2$ are not disjoint. Then by Def. \[def:disjoint\] there must be at least one shared variable $x \in \mathcal V_1, \mathcal V_2$, and thus we must have $|\mathcal V_1 \cup \mathcal V_2| \leq \ell+m-1$.
We can then proceed through case analysis by inquiring how many equations are shared by $\mathcal S_1, \mathcal S_2$. Since $\mathcal S_1$ is minimal with $|\mathcal E_1|=|\mathcal V_1|=\ell$ for $1 \leq \ell \leq m$, the number of equations that are shared with $\mathcal S_2$ could be any $0 \leq k < \ell$. (Note that the case of $k=\ell$ shared equations would lead to the more obvious contradiction that $\mathcal S_1 \subseteq \mathcal S_2$, even though $\mathcal S_2$ is minimal).
Let us start with the case $\mathcal E_1 \cap \mathcal E_2 = \varnothing$ to illustrate the rationale in its simplest form. In this case, no equations are shared yet at least one variable is. Then we have $|\mathcal E_1 \cup \mathcal E_2| = \ell+m$, but $|\mathcal V_1 \cup \mathcal V_2| \leq \ell+m-1$. Since we have both $\mathcal S_1, \mathcal S_2 \subset \mathcal S$, in fact we have their sets of equations $\mathcal E_1, \mathcal E_2 \subset \mathcal E$ as well and then $\mathcal E_1 \cup \mathcal E_2 \subseteq \mathcal E$. Now, by Def. \[def:structure\] (valid structure), we know that in any subset of $k>0$ equations of $\mathcal S$, at least $k$ different variables must appear. But rather we have $|\mathcal E_1 \cup \mathcal E_2| = \ell+m$ and yet $|\mathcal V_1 \cup \mathcal V_2| \leq \ell+m-1$. That is, we reach a contradiction to Def. \[def:structure\], viz., $|\mathcal E_1 \cup \mathcal E_2| > |\mathcal V_1 \cup \mathcal V_2|$. .
The next case is when one equation is shared ($|\mathcal E_1 \,\cap\, \mathcal E_2| = 1$). Note that, if we had $|\mathcal E_1|=|\mathcal V_1|=\ell=1$ in particular then the only equation $f \in \mathcal E_1$ would have $|Vars(f)|=1$ and be shared with $\mathcal E_2$, making $\mathcal S_1 \subseteq \mathcal S_2$ even though $\mathcal S_2$ is assumed minimal. . We rather know that $|\mathcal E_1|=\ell \geq 2$. Also, note that we must have $|Vars(f)| \geq 2$ for all $f \in \mathcal E_1$, otherwise there would be some $g \in \mathcal E_1$ with $|Vars(g)|=1$ even though $|\mathcal E_1| \geq 2$. That is, we would have a minimal substructure within $\mathcal S_1$, although it is minimal.
So, since one equation is shared and for all $f \in \mathcal E_1$ we have $|Vars(f)| \geq 2$, then at least two variables must be shared. Observe then that $|\mathcal E_1 \cup \mathcal E_2| = \ell+m-1$ (since exactly one equation is shared) and $|\mathcal V_1 \cup \mathcal V_2| \leq \ell+m-2$ (at least two variables are shared). Again, we see the same contradiction in face of Def. \[def:structure\], viz., $|\mathcal E_1 \cup \mathcal E_2| > |\mathcal V_1 \cup \mathcal V_2|$. .
Now we complete the case analysis by making the argument abstract for any number of shared equations, $0 \leq k < \ell$ (an inductive step, n.b., is not required because $k \in \mathbb N$ is bounded. Note that, for any such number $0 \leq k < \ell$, we must have at least $k+1$ shared variables, otherwise the shared substructure having $k$ equations, formed out of $\mathcal E_1 \cap \mathcal E_2$, would be minimal as well even though $\mathcal E_1 \cap \mathcal E_2 \subseteq \mathcal E_1, \mathcal E_2$ (that is, rendering both $\mathcal S_1, \mathcal S_2$ non-minimal. ). However, once more we see that this contradicts Def. \[def:structure\]. . $\Box$
Proof of Lemma \[lemma:bipartite\] {#app:coa-bipartite}
==================================
First, we consider the ‘if’ statement—that is, all conditions (i-iii) together are sufficient. Let $G^\prime \subset G$ be a bipartite subgraph $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$ that satisfies all conditions (i-iii), and $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ be a substructure of $\mathcal S$ with $G^\prime \sim \mathcal S^\prime$. We shall see that such $\mathcal S^\prime$ does give a yes answer to CSDP, that is, it is a complete substructure with $|\mathcal E^\prime|=|\mathcal V^\prime|=\ell$.
From condition (i) we know that $G^\prime$ is structural (Def. \[def:structural\]). That is, for every $V_1^{\prime\prime} \subseteq V_1^\prime$, there is a connected bipartite subgraph $G^{\prime\prime}(V_1^{\prime\prime} \cup V_2^{\prime\prime}, E^{\prime\prime})$ with $|V_1^{\prime\prime}| \leq |V_2^{\prime\prime}|$. Since $V_1^\prime \mapsto \mathcal E^\prime$, $V_2^\prime \mapsto \mathcal V^\prime$ and $E^\prime \mapsto \mathcal S^\prime$, such property bears obvious resemblance with Def. \[def:structure\] (structure). That is, the ‘connected’ bipartite subgraph aspect implies that, for any subset of $|\mathcal E^{\prime\prime}|$ equations in $\mathcal E^\prime$, at least $|\mathcal V^{\prime\prime}| \geq |\mathcal E^{\prime\prime}|$ variables appear and each such variable $x \in \mathcal V^{\prime\prime}$ appears in some $f \in \mathcal E^{\prime\prime}$, otherwise $x \in V_2^{\prime\prime}$ would be disconnected in $G^{\prime\prime}(V_1^{\prime\prime} \cup V_2^{\prime\prime}, E^{\prime\prime})$.
Condition (ii) ensures in addition that $\bigcup_{f \in \mathcal E^\prime} Vars(f) = \mathcal V^\prime$. That is, the variables in $\mathcal V^\prime$ are exhaustive w.r.t. $\mathcal E^\prime$. Thus, structure $\mathcal S^\prime$ satisfies Def. \[def:structure\]. Finally, condition (iii) ensures that $\mathcal S^\prime$ is complete with $|\mathcal E^\prime|=|\mathcal V^\prime|=\ell$.
We consider now the ‘only if’ statement—i.e., every condition (i-iii) is necessary. We assume a bipartite graph $G^\prime \sim \mathcal S^\prime$ and show that lacking any one such condition implies that $\mathcal S^\prime$ cannot be complete or cannot be a structure at all. First, it is easy to see that when condition (iii) does not hold for $G^\prime$ then a structure $\mathcal S^\prime$ with $G^\prime \sim \mathcal S^\prime$ cannot be complete.
Now suppose condition (ii) does not for $G^\prime$. That is, there is some $f \in V_1^\prime$ that has incidence with some $x \in V_2 \setminus V_2^\prime$. Thus we have $V_1^\prime \mapsto \mathcal E^\prime$ and $V_2^\prime \mapsto \mathcal V^\prime$ but $\bigcup_{f \in \mathcal E^\prime} Vars(f) \neq \mathcal V^\prime$. Therefore either $\mathcal S^\prime$ does not satisfy Def. \[def:structure\] or we cannot actually have $G^\prime \sim \mathcal S^\prime$. .
Finally, consider that $G^\prime$ is not structural (Def. \[def:structural\]). That is, there is some $V_1^{\prime\prime} \subseteq V_1^\prime$ such that no connected bipartite subgraph $G^{\prime\prime}(V_1^{\prime\prime} \cup V_2^{\prime\prime}, E^{\prime\prime})$ exists in $G^\prime$ with $|V_1^{\prime\prime}| \leq |V_2^{\prime\prime}|$. Considering $G^\prime \sim \mathcal S^\prime$, that would imply for $\mathcal S^\prime(\mathcal E^\prime, \mathcal V^\prime)$ either an equation $f \in \mathcal E^\prime$ with no variables (a disconnected vertex $x \in V_1^\prime$), or a redundant subset of equations—number of equations is larger than number of variables appearing in it. Either conditions violate Def. \[def:structure\], so $\mathcal S^\prime$ cannot be a structure even though $G^\prime \sim \mathcal S^\prime$. . $\Box$
Proof of Theorem \[thm:biclique\] {#app:coa-npcomplete}
=================================
*CSDP is NP-Complete.*
We shall construct an instance of ENCD and describe its polynomial-time reduction to an instance of CSDP by using Lemma \[lemma:bipartite\].
**Constructing an instance of ENCD**. Let $G(V_1 \cup V_2, E)$ be a bipartite graph such that, for every $V_1^\prime \subseteq V_1$, there is a bipartite subgraph $G^\prime(V_1^\prime \cup V_2^\prime, E^\prime)$ with $|V_1^\prime| \leq |V_2^\prime|$ and $deg(f) < |V_2^\prime|$ for all $f \in V_1^\prime$. Note that this is the complement property of the structural bipartite graph property (see Def. \[def:structural\]). It is meant to ensure that the bipartite complement graph $G^c(V_1 \cup V_2, E^c)$ of $G$ is structural—satisfies Def. \[def:structural\]. That is, when we produce $G^c$, we know that it can possibly correspond to a structure $\mathcal S$ such that $G^c \sim \mathcal S$. Let also $G$ have $|V_1|=|V_2|=m$ in order to ensure that such structure $\mathcal S$ will be complete as well—recall that $\mathcal S$ given in CSDP is assumed complete indeed.
Now let $G$ and an integer $1 \leq \ell < m$ provide an instance of ENCD for integers $a = \ell$ and $b = m - \ell$. That is, our problem is to decide whether $G$ has a biclique $K_{\ell,\, m-\ell}$. Imposing both of the above properties on $G$, n.b., incurs in no loss of generality w.r.t. ENCD as it does not open a pruning opportunity in searching for a biclique $K_{\ell,\, m-\ell}$ in powerset $\mathcal P(V_1 \times V_2)$. Such a biclique $K_{\ell,\, m-\ell}$, if existing in $G$, can be denoted $C(V_1^\prime \cup V_2^\star, K)$, where $|V_1^\prime|=\ell$ and $|V_2^\star|=m-\ell$, and $K$ is a complete set of edges between $V_1^\prime$ and $V_2^\star$. Note also that $V_1^\prime \subset V_1$ and $V_2^\star \subset V_2$.
**Production of an instance of CSDP from the ENCD one**. Let $G^c(V_1 \cup V_2, E^c)$ be the bipartite complement graph of $G$, where an edge $(f, x) \in E^c$ if and only if $(f, x) \notin E$ for $f \in V_1$ and $x \in V_2$. Clearly, bipartite graph $G^c$ can be produced in polynomial time from $G$—as mentioned in §\[subsec:hard\], consider the ‘structure matrix’ (biadjacency matrix) of $G$ and run a full scan on it to switch the boolean value of each entry in time $O(|V_1|\cdot |V_2|)$ and then get $G^c$. **Decision problem correspondence**. Now we show that a biclique $K_{\ell,\, m-\ell}$ in $G$, if existing, corresponds to a bipartite subgraph $G^{c\,^\prime}(V_1^\prime \cup V_2^\prime, E^{c\,\prime})$ in $G^c$ that satisfies the conditions (i-iii) of Lemma \[lemma:bipartite\]. That is, we show that a yes answer to ENCD implies a yes answer to CSDP.
In fact, as $G^c$ is the bipartite complement graph of $G$, then the biclique $C(V_1^\prime \cup V_2^\star, K)$ in $G$ becomes a bipartite subgraph $C^c(V_1^\prime \cup V_2^\star, \varnothing)$ in $G^c$. Now let $G^{c\,\prime}(V_1^\prime \cup V_2^\prime, E^{c\,\prime})$ be such that $V_2^\prime = V_2 \setminus V_2^\star$. We observe that:
- The presence of biclique $C(V_1^\prime \cup V_2^\star, K)$ in $G$ indicates that $V_2^\star$ could not have contributed to satisfy the complement structural property for $V_1^\prime$, only $V_2^\prime = V_2 \setminus V_2^\star$ could. But such property turns into the structural property in $G^c$, thus $G^{c\,\prime}(V_1^\prime \cup V_2^\prime, E^{c\,\prime})$ must be structural indeed. That is, condition (i) of Lemma \[lemma:bipartite\] is ensured.
- By the fact that we have $C^c(V_1^\prime \cup V_2^\star, \varnothing)$ in $G^c$ we know that, for all $f \in V_1^\prime$, there can only be an edge $(f, x) \in E^c$ if $x \in V_2^\prime$ indeed. That is, condition (ii) of Lemma \[lemma:bipartite\] is ensured.
- The presence of biclique $C(V_1^\prime \cup V_2^\star, K)$ of form $K_{\ell,\, m-\ell}$ in $G$ implies that $V_1^\prime$ has size $|V_1^\prime|=\ell$. Besides, $V_2^\prime$ will have size $|V_2^\prime|=|V_2| - |V_2^\star| = m - (m-\ell) = \ell$. That is, we must have $|V_1^\prime| = |V_2^\prime| = \ell$ and then condition (iii) of Lemma \[lemma:bipartite\] is ensured as well.
We have then established that the existence of a biclique $C \subset G$ of form $K_{\ell,\, m-\ell}$ implies the existence of a bipartite subgraph $G^{c\,\prime} \subset G^c$, where $G^{c\,\prime}$ satisfies the conditions (i-iii) of Lemma \[lemma:bipartite\]. That is, we get a yes answer to CSDP if we find one to ENCD. It remains to show the ‘only if’ part of the correspondence.
In fact, suppose no biclique $C(V_1^\prime \cup V_2^\star, K)$ of form $K_{\ell,\, m-\ell}$ exists in $G(V_1 \cup V_2, E)$. Clearly, it means that for any $V_1^\prime \subset V_1$ where $|V_1^\prime|=\ell$, there is at least one $f \in V_1^\prime$ such that an edge $(f, x)$ with $x \in V_2^\star$ is missing from $E$. Accordingly, in $G^c(V_1 \cup V_2, E^c)$, we cannot have $G^{c\,\prime} \subset G^c$ with condition (ii) of Lemma \[lemma:bipartite\] satisfied.
ENCD is NP-Complete. Thus CSDP must be NP-Complete as well. $\Box$
Proof of Proposition \[prop:causal-ordering\] {#coa-efficient}
=============================================
*Let $\mathcal S(\mathcal E, \mathcal V)$ be a structure, and $\varphi_1\!:\, \mathcal E \to \mathcal V$ and $\varphi_2\!:\, \mathcal E \to \mathcal V$ be any two total causal mappings over $\mathcal S$. Then $C^+_1$ = $C^+_2$.*
The proof is based on an argument from Nayak [@nayak1994], which we reproduce here in a bit more of detail. Intuitively, it shows that if $\varphi_1$ and $\varphi_2$ differ in the variable an equation $f$ is mapped to, then such variables, viz., $\varphi_1(f)$ and $\varphi_2(f)$, must be causally dependent on each other (strongly coupled).
To show $C^+_1$ = $C^+_2$ reduces to show both $C^+_1 \subseteq C^+_2$ and $C^+_2 \subseteq C^+_1$. We show the first containment, and the second is understood as following by symmetry. Closure operators are extensive, $X \subseteq cl(X)$, and idempotent, $cl(cl(X)) = cl(X)$. That is, if we have $C_1 \subseteq C_2^+$, then we shall have $C_1^+ \subseteq (C_2^+)^+$ and, by idempotence, $C_1^+ \subseteq C_2^+$.
Then it suffices to show that $C_1 \subseteq C_2^+$, i.e., for any $(x^\prime,\, x) \in C_1$, we must show that $(x^\prime,\, x) \in C_2^+$ as well. Observe by Def. \[def:tcm\] that both $\varphi_1$ and $\varphi_2$ are bijections, then, invertible functions. If $\varphi_1^{-1}(x) = \varphi_2^{-1}(x)$, then we have $(x^\prime,\, x) \in C_2$ and thus, trivially, $(x^\prime,\, x) \in C_2^+$. Else, $\varphi_1$ and $\varphi_2$ disagree in which equations they map onto $x$. But we show next, in any case, that we shall have $(x^\prime,\, x) \in C_2^+$.
Take all equations $g \in \mathcal E^\prime \subseteq \mathcal E$ such that $\varphi_1(g) \neq \varphi_2(g)$, and let $n \leq |\mathcal E|$ be the number of such ‘disagreed’ equations. Now, let $f \in \mathcal E^\prime$ be such that its mapped variable is $x = \varphi_1(f)$. Construct a sequence of length $2n$ such that, $s_0 = \varphi_1(f) = x$ and, for $1 \leq i \leq 2n$, element $s_i$ is defined $s_i = \varphi_2(\varphi_1^{-1}(s_{i-1}))$. That is, we are defining the sequence such that, for each equation $g \in \mathcal E^\prime$, its disagreed mappings $\varphi_1(g)=x_a$ and $\varphi_2(g)=x_b$ are such that $\varphi_1(g)$ is immediately followed by $\varphi_2(g)$. As $x_a,\, x_b \in Vars(g)$, we have $(x_a,\, x_b) \in C_2$ and, symmetrically, $(x_b,\, x_a) \in C_1$. The sequence is of form $s=\langle \underbrace{x,\, x_f}_{f}, \hdots, \underbrace{x_a,\, x_b}_{g}, \hdots, \underbrace{x_{2n-1},\, x_{2n}}_{h} \rangle$.
Since $x$ must be in the codomain of $\varphi_2$, we must have a repetition of $x$ at some point $2 \leq k \leq 2n$ in the sequence index, with $s_k=x$ and $s_{k-1}=x^{\prime\prime}$ such that $(x^{\prime\prime},\, x) \in C_2$. If $x^{\prime\prime}=x^\prime$, then $(x^\prime,\, x) \in C_2$ and obviously $(x^\prime,\, x) \in C_2^+$. Else, note that $x_f$ must also be in the codomain of $\varphi_1$, while $x^{\prime\prime}$ in the codomain of $\varphi_2$. Let $\ell$ be the point in the sequence, $3 \leq \ell \leq 2n\!-\!1$, at which $s_\ell=x_f=x_a$ and $s_{\ell+1}=x_b$ for some $x_b$ such that $(x_f,\, x_b) \in C_2$. It is easy to see that, either we have $x_b=x^{\prime\prime}$ or $x_b \neq x^{\prime\prime}$ but $(x_b,\, x^{\prime\prime}) \in C_2^+$. Thus, by transitivity on such a causal chain, we must have $(x_f,\, x^{\prime\prime}) \in C_2^+$ and eventually $(x_f,\, x) \in C_2^+$. Finally, since $x^\prime \in Vars(f)$ and $\varphi_2(f)=x_f$, we have $(x^\prime,\, x_f) \in C_2$ and, by transitivity, $(x^\prime,\, x) \in C_2^+$. $\Box$
[^1]: Also, for inferring causal ordering the systems of equations given as input is expected to be ‘independent,’ i.e., can only have non-redundant equations.
[^2]: This slight variation takes place in lines 7–10 of RTCM in Algorithm \[alg:coat\], and is irrelevant to its intractability—which we shall see is due to line 3. Besides, ‘total’ and ‘partial’ causal mappings are interchangeable [@druzdzel2008]. In particular, recovering the latter from the former is straightforward: just merge ‘strongly coupled’ variables in a cluster. Intuitively, these are variables whose values can only be determined simultaneously. To be precise, let $x_1, x_2 \in \mathcal V$ be variables in a structure $\mathcal S(\mathcal E, \mathcal V)$. We say $x_1, x_2$ are *strongly coupled* if $\mathcal S$ is minimal.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the flow equation of the O($N$) $\varphi^4$ model in $d$ dimensions at the next-to-leading order (NLO) in the $1/N$ expansion. Using the Schwinger-Dyson equation, we derive 2-pt and 4-pt functions of flowed fields. As the first application of the NLO calculations, we study the running coupling defined from the connected 4-pt function of flowed fields in the $d+1$ dimensional theory. We show in particular that this running coupling has not only the UV fixed point but also an IR fixed point (Wilson-Fisher fixed point) in the 3 dimensional massless scalar theory. As the second application, we calculate the NLO correction to the induced metric in $d+1$ dimensions with $d=3$ in the massless limit. While the induced metric describes a 4-dimensional Euclidean Anti-de-Sitter (AdS) space at the leading order as shown in the previous paper, the NLO corrections make the space asymptotically AdS only in UV and IR limits. Remarkably, while the AdS radius does not receive a NLO correction in the UV limit, the AdS radius decreases at the NLO in the IR limit, which corresponds to the Wilson-Fisher fixed point in the original scalar model in 3 dimensions.'
author:
- Sinya Aoki
- Janos Balog
- Tetsuya Onogi
- Peter Weisz
title: Flow equation for the scalar model in the large $N$ expansion and its applications
---
Introduction {#sec:intro}
============
In the previous paper[@Aoki:2016ohw], the present authors studied the proposal[@Aoki:2015dla] that a $d+1$ dimensional induced metric can be constructed from a $d$ dimensional field theory using gradient flow[@Narayanan:2006rf; @Luscher:2010iy; @Luscher:2009eq; @Luscher:2013vga], applying the method to the O($N$) $\varphi^4$ model. We have shown that in the large $N$ limit the induced metric becomes classical and describes Euclidean Anti-de-Sitter (AdS) space in both ultra-violet (UV) and infra-red (IR) limits of the flow direction. The method proposed in Ref. [@Aoki:2015dla] may provide an alternative way to understand the AdS/CFT (or more generally Gravity/Gauge theory) correspondence[@Maldacena:1997re], and the result in Ref. [@Aoki:2016ohw] might be related to the correspondence between O($N$) vector models in $d$-dimensions and (generalized) gravity theories in $d+1$ dimensions[@Klebanov:2002ja].
To further investigate a possible connection between Ref. [@Aoki:2016ohw] and Ref. [@Klebanov:2002ja] at the quantum level, one must calculate, for example, the anomalous dimension of the O($N$) invariant operator $\phi^2(x)$, which requires the next-to-leading order (NLO) of the $1/N$ expansion for the flow equation to evaluate necessary quantum corrections. Since the method employed in Refs. [@Aoki:2015dla; @Aoki:2016ohw] is a specific one adopted for the large $N$ limit, some systematic way to solve the flow equation in the $1/N$ expansion is needed.
In this paper, we employ the Schwinger-Dyson equation (SDE) to solve the flow equation in the $1/N$ expansion for the O($N$) invariant $\varphi^4$ model in $d$ dimensions. Using this method we explicitly calculate the 2-pt and 4-pt functions at the NLO.
As the first application of the NLO calculations, we define a running coupling from the connected 4-pt function of flowed fields, which runs with the flow time $t$ such that $t=0$ corresponds to the UV limit while $t=\infty$ is the IR limit. This property establishes that the flow equation can be interpreted as a renormalization group transformation. In particular at $d=3$, we show that the running coupling so defined has not only the asymptotic free UV fixed point but also a Wilson-Fisher IR fixed point for the massless case.
As the second application, we investigate the NLO correction to the induced metric in $3+1$ dimensions from the massless scalar model in 3 dimensions. In the massless limit, the whole 4-dimensional space becomes AdS at the leading order, as shown in Ref. [@Aoki:2016ohw]. The NLO corrections give a small perturbation to the metric, which makes the space asymptotically AdS in UV ($t=0$) and IR ($t=\infty$) limits only. A remarkable thing is that, while the NLO corrections do not change the AdS radius in the UV limit, the AdS radius is reduced by the NLO correction in the IR limit, which corresponds to the Wilson-Fisher IR fixed point of the original theory. In other words, a nontrivial fixed point in the field theory leads to a change of the AdS radius in the geometry at the NLO. The induced metric at NLO describes a 4-dimensional space connecting one asymptotically AdS space at UV to an other asymptotically AdS space at IR, which have different radii.
This paper is organized as follows. In Sec. \[sec:flow\], we first introduce the O($N$) invariant $\varphi^4$ model in $d$ dimensions. We then formulate the Schwinger-Dyson equation (SDE) for the flowed fields, and solve it to derive 2-pt and 4-pt functions of flowed fields at the NLO. In Sec. \[sec:coupling\], we define a running coupling from the connected 4-pt function of flowed fields and investigate its behavior as a function of the flow time $t$. In Sec. \[sec:metric\], we study the induced metric from the 3 dimensional massless scalar model at the NLO. We finally give a summary of this paper in Sec. \[sec:summary\]. We collect all technical details in appendices. In appendix \[app:d-dim\], using the SDE, we present results at the NLO in the $1/N$ expansion of the $d$ dimensional theory necessary for the main text. We also perform the renormalization of the $d$ dimensional theory at the NLO, and explicitly calculate renormalization constants for various $d$. In appendix \[app:SDE\_flow\], we give detailed derivations of solutions to the SDE for the flow fields at the NLO. We explicitly evaluate 2-pt and 4-pt functions of the flowed field in appendix \[app:massless\] while we derive the induced metric in appendix \[app:metric\], for the massless scalar theory in 3 dimensions.
$1/N$ expansion of the flow equation in $d+1$ dimensions {#sec:flow}
========================================================
Model in $d$ dimensions
-----------------------
In this paper, we consider the $N$ component scalar $\varphi^4$ model in $d$ dimensions, defined by the action $$\begin{aligned}
S(\mu^2,u) &=& N \int {{\rm d}}^d x\,\left[ \frac{1}{2}\partial^k \varphi (x)\cdot \partial_k \varphi (x) + \frac{\mu^2}{2}\varphi^2(x) +\frac{u}{4!} \left(\varphi^2(x)\right)^2
\right] ,
\label{eq:action}\end{aligned}$$ where $\varphi^a(x)$ is an $N$ component scalar field, $(\ \cdot \ )$ indicates an inner product of $N$ component vectors such that $\varphi^2(x) \equiv \varphi(x)\cdot\varphi(x) =\sum_{a=1}^N \varphi^a(x)\varphi^a(x)$, $\mu^2$ is the bare scalar mass parameter, and $u$ is the coupling constant of the $\varphi^4$ interaction, whose canonical dimension is $4-d$. While it is consistent to take $u$ as $N$ independent, as will be seen later, the mass parameter $\mu^2$ is expanded as $$\begin{aligned}
\mu^2 &=& \mu_0^2 + \frac{1}{N}\mu_1^2 + \cdots,\end{aligned}$$ where $\mu_i^2$ is cut-off dependent in order to make the physical mass finite order by order in the $1/N$ expansion.
This model describes the free massive scalar at $u=0$, while it is equivalent to the non-linear $\sigma$ model (NLSM) in the $u\rightarrow \infty$ limit, whose action is obtained from eq. (\[eq:action\]) as $$\begin{aligned}
S(\lambda) &=& \frac{ N}{2\lambda} \int {{\rm d}}^d x\,\partial^k \sigma (x)\cdot \partial_k \sigma (x) , \qquad
\sigma^2(x) = 1,
\label{eq:action_NLSM}\end{aligned}$$ with the replacement $$\begin{aligned}
\sigma^a(x)&=&\sqrt{\lambda} \varphi^a(x), \qquad \lambda
= \lim_{u\rightarrow\infty} -\frac{u}{6\mu^2} . \end{aligned}$$
Some regularization which preserves O($N$) symmetry is assumed in this paper, so that we can always make formal manipulations without worrying about divergences.[^1] Calculations of 2-pt and 4-pt functions at the next-to-leading order (NLO) of the $1/N$ expansion in $d$ dimensions will be given in appendix \[app:d-dim\].
Flow equation in the $1/N$ expansion
-------------------------------------
In this paper, we consider the flow equation, given by $$\begin{aligned}
\frac{\partial}{\partial t} \phi^a(t,x) &=& -\left. \frac{1}{N} \frac{\delta S(\mu_f^2,u_f)}{\delta \varphi^a(x)}\right\vert_{\varphi\rightarrow\phi}
= \left(\Box -\mu_f^2\right)\phi^a(t,x) -\frac{u_f}{6}\phi^a(t,x)
\phi^2(t,x),
\label{eq:flow} \\
\phi^a(0,x) &=&\varphi^a(x), \nonumber\end{aligned}$$ where $\mu_f^2$ and $u_f$ can be different from $\mu^2$ and $u$ in the original $d$ dimensional theory. As in the case of $d$ dimensions, $u_f$ is kept fixed and $N$ independent, whereas $\mu^2_f$ is adjusted as $$\mu_f^2 = m_f^2 - \frac{u_f}{6} Z(m_f), \qquad Z(m_f) \equiv \int {{\rm D}}q \frac{1}{q^2+m_f^2}, \quad {{\rm D}}q \equiv \frac{{{\rm d}}^d q}{(2\pi)^d},
\label{eq:Zm}$$ where $m_f$ is a renormalized mass. The flow with $\mu_f=\mu$ and $u_f=u$ is called gradient flow, as it is given by the gradient of the original action.
In the case of the free flow ($u_f=0$), the solution is easily given by $$\begin{aligned}
\phi^a(t,x) &=& {{\rm e}}^{t(\Box -\mu_f^2)} \varphi^a(x).\end{aligned}$$ We therefore consider the interacting flow ($u_f\not=0$) hereafter unless otherwise stated.
The above flow equation leads to the Schwinger-Dyson equation (SDE)[@Aoki:2014dxa] as $$\begin{aligned}
\langle D^f_z \phi^a(z) {\cal O} \rangle &=& -\frac{u_f}{6} \langle \phi^a(z) \phi^2(z) {\cal O}\rangle,
\quad D_z^f \equiv \frac{\partial}{\partial t} -(\Box -\mu_f^2), \end{aligned}$$ where $z=(t,x)$, ${\cal O}$ is an arbitrary operator and the expectation value $\langle{\cal O} \rangle$ should be calculated in $d$ dimensions as $$\begin{aligned}
\langle {\cal O}(\varphi)\rangle &\equiv & \frac{1}{Z} \int \left[{\cal D}\varphi\right] {\cal O}(\varphi){{\rm e}}^{-S(\mu^2,u)}, \quad
Z = \int \left[{\cal D}\varphi\right] {{\rm e}}^{-S(\mu^2,u)} .
\label{eq:vev_d}\end{aligned}$$
If we take ${\cal O} =\displaystyle\prod_{i=1}^{2n-1} \phi^{a_i}(z_i)$ the SDE becomes $$\begin{aligned}
D_z^f \, \Gamma_{2n}^{aa_1\cdots a_{2n-1}}(z,z_1,\cdots,z_{2n-1}) &=&
-\frac{u_f}{6 N^2}\sum_b \Gamma_{2n+2}^{abba_1\cdots a_{2n-1}}(z,z,z,z_1,\cdots,z_{2n-1}),~~\end{aligned}$$ where $\Gamma_n$ is the $n$-point function, defined by $$\begin{aligned}
\Gamma_n^{a_1\cdots a_n}(z_1,\cdots,z_n) = N^{n-1}\langle \prod_{i=1}^n \phi^{a_i}(z_i)\rangle
\equiv \Gamma_n[12\cdots n] ,\end{aligned}$$ which is analogous to the $d$ dimensional counterpart in eq. (\[eq:G2n\_d\]). We consider only the symmetric phase in this paper, where $\Gamma_{2n-1} = 0$ for all positive integers $n$.
We consider the next-to-leading order of the $1/N$ expansion, so that the following two SDE’s need to be considered. $$\begin{aligned}
D_1^f\, \Gamma_2[12] &=& -\frac{u_f}{6N^2} \sum_b \Gamma_4[1bb2], \\
D_1^f\, \Gamma_4[1234] &=&
-\frac{u_f}{6N^2} \sum_b \Gamma_6[1bb234], \end{aligned}$$ where $z_b=z_1$, so that the sum over $b$ runs over the O($N$) index only.
The connected part of 4- and 6- pt functions are introduced as $$\begin{aligned}
\Gamma_4[1234] &=& K_4[1234] + N\left\{\Gamma_2[12] \Gamma_2[34] + \Gamma_2[13] \Gamma_2[24] + \Gamma_2[14] \Gamma_2[23]\right\}, \\
\Gamma_6[123456] &=& K_6[123456] + N\left\{\Gamma_2[12] K_4[3456] + \mbox{ 14 perms.} \right\} \nonumber \\
&+& N^2\left\{\Gamma_2[12] \Gamma_2[34]\Gamma_2]56] + \mbox{ 14 perms.} \right\} .\end{aligned}$$ Furthermore decompositions in O($N$) indices are given by $$\begin{aligned}
\Gamma_2[12] &=& \delta^{a_1a_2}\Gamma(z_1,z_2), \\
K_4[1234] &=& \delta^{a_1a_2}\delta^{a_3a_4} K(z_1,z_2;z_3,z_4) +\mbox{ 2 perms.}, \\
K_6[123456] &=& \delta^{a_1a_2}\delta^{a_3a_4}\delta^{a_5a_6} H(z_1,z_2;z_3,z_4;z_5,z_6) +
\mbox{ 14 perms.} ,\end{aligned}$$ where $\Gamma(z_1,z_2)$, $K(z_1,z_2;z_3,z_4)$ and $H(z_1,z_2;z_3,z_4;z_5,z_6)$ are invariant under the exchange of arguments such that $z_{2i-1}\leftrightarrow z_{2i}$ or $(z_{2i-1},z_{2i}) \leftrightarrow (z_{2j-1},z_{2j})$.
By expanding $\Gamma$, $K$ and $H$ as $$\Gamma =\sum_{i=0}^\infty \frac{\Gamma_i}{N^i}, \qquad
K=\sum_{i=0}^\infty \frac{K_i}{N^i}, \qquad
H=\sum_{i=0}^\infty \frac{H_i}{N^i},$$ the above two SDE are reduced to $$\begin{aligned}
D_1^f \Gamma_0(12) &=& -\frac{u_f}{6} \Gamma_0(12)\Gamma_0(11)
\label{eq:SDE_LO}\end{aligned}$$ at the LO of the $1/N$ expansion, and $$\begin{aligned}
D_1^f \Gamma_1(12) &=& -\frac{u_f}{6}\left[K_0(12;11) +\Gamma_0(12)\Gamma_1(11)+
\Gamma_1(12)\Gamma_0(11) +2\Gamma_0(12)\Gamma_0(11)\right], ~~~~~
\label{eq:SDE_G1}\\
D_1^f K_0(12;34) &=& -\frac{u_f}{6}\left[\Gamma_0(12) K_0(11;34) +\Gamma_0(11) K_0(12;34)
+2\Gamma_0(12)\Gamma_0(13) \Gamma_0(14)\right] ~~~~~~~~
\label{eq:SDE_K0}\end{aligned}$$ at the NLO.
Solutions to the flowed SDE at NLO
----------------------------------
The solutions to the SDE at NLO are summarized below. Details of calculations can be found in appendix \[app:SDE\_flow\].
At the NLO, the 2-pt function is given by $$\begin{aligned}
\langle \phi^{a_1}(z_1) \phi^{a_2}(z_2) \rangle &=&\frac{\delta_{a_1a_2}}{N}
\frac{Z(m_f)}{\sqrt{\zeta(t_1)\zeta(t_2)}}
\int {{\rm D}}p \frac{{{\rm e}}^{-p^2(t_1+t_2)}{{\rm e}}^{i p(x_1-x_2)}}{p^2+m^2}
\left[ 1+\frac{1}{N} G_1(t_1,t_2\vert p) \right], ~~~~~\end{aligned}$$ where $\zeta(t)$ is defined in eq. (\[eq:zeta\]), and the NLO contribution $G_1(t_1,t_2\vert p)$ is given in appendix \[app:SDE\_flow\_G1\]. In the continuum limit, $\zeta(t)$ approaches to $\zeta_0(t)$ and is finite as long as $t> 0$, where $$\begin{aligned}
\zeta_0(t) &\equiv& \int {{\rm D}}p\frac{{{\rm e}}^{-2p^2t}}{p^2+ m^2}=\frac{{{\rm e}}^{2tm^2} m^{d-2}}{(4\pi)^{d/2}}\Gamma(1-d/2,2tm^2)
\label{eq:zeta0}\end{aligned}$$ with the incomplete gamma function $\Gamma(a,x)$, while $Z(m_f)$ diverges at $d > 1$.
The leading contribution of the connected 4-pt function appearing at the NLO of the $1/N$ expansion can be obtained as $$\begin{aligned}
\langle \phi^{a_1}(z_1) \phi^{a_2}(z_2) \phi^{a_3}(z_3)\phi^{a_4}(z_4)\rangle_c
&=& \frac{1}{N^3}\left[\delta_{a_1a_2}\delta_{a_3a_4}K_0(12;34) + \mbox{ 2 permutations} \right] ,
\label{eq:4pt_c}\end{aligned}$$ where $$\begin{aligned}
K_0(12;34) &=& \int {{\rm d}}P_4\ g(12;34\vert 12;34),
\quad {{\rm d}}P_4 \equiv \prod_{j=1}^4 {{\rm D}}p_j \sqrt{\frac{Z(m_f)}{\zeta(t_j)}} \frac{{{\rm e}}^{ip_j x_j} {{\rm e}}^{-p_j^2 t_j}}{p_j^2+m^2},
\label{eq:K0g}
\\
g(12;34\vert 12;34) &=& X(23\vert 12;34)+X(13\vert 21;34)+X(24\vert 12;43)+X(14\vert 21;43)\nonumber \\
&+&Y(2\vert 12;34) + Y(1\vert 21;34) +Y(3\vert 43;12) +Y(4\vert 34;12) \nonumber \\
&+& Z(\vert 12;34) .\end{aligned}$$ Here the variables to the left of the vertical line refer to flow times and those to the right refer to momenta. Explicitly we have in the continuum or NLSM limits $$\begin{aligned}
X(t_1,t_2\vert 12;34) &=& \hat\delta (p_2^2+m^2)(p_3^2+m^2)\int_0^{t_1} {{\rm d}}s_1\,
\int_0^{t_2} {{\rm d}}s_2\, {{\rm e}}^{s_1(p_2^2-p_1^2)} {{\rm e}}^{s_2(p_3-p_4^2)} \omega(s_1,s_2\vert p_{34}),~~~~~~ \\
Y(t\vert 21;34) &=& \hat\delta (p_1^2+m^2)\int_0^t {{\rm d}}s\, {{\rm e}}^{s(p_1^2-p_2^2)}\psi(s\vert 34), \\
Z(\vert 12;34) &=& -\hat\delta \frac{2}{6/u+ B(0\vert p_{34})}, \end{aligned}$$ where $\hat\delta \equiv (2\pi)^d \delta(p_1+p_2+p_3+p_4)$, $p_{34} =p_3+p_4$, $$\begin{aligned}
B(t\vert Q) &=& \int {{\rm D}}q_1 {{\rm D}}q_2 \frac{{{\rm e}}^{-t(q_1^2+q_2^2)}}{(q_1^2+m^2)(q_2^2+m^2)} (2\pi)^d\delta (q_{12}-Q), \quad q_{12}=q_1+q_2,
\label{eq:Bt}\end{aligned}$$ and thus $B(0\vert Q) = B(Q^2)$, defined in appendix \[app:d-dim\]. Here $\psi$ and $\omega$ satisfy $$\begin{aligned}
\rho(t\vert 34) +\int_0^t {{\rm d}}s\, K(t,s\vert p_{34}) \psi(s\vert 34) &=& 0, \label{eq:psi} \\
\rho(t_1,t_2\vert Q) - 2 \int_0^{t_1} {{\rm d}}s_1\, K(t_1,s_1\vert Q) \int_0^{t_2} {{\rm d}}s_2\, K(t_2,s_2\vert Q)\, \omega(s_1,s_2\vert Q)
&=& 0,
\label{eq:omega}\end{aligned}$$ where $$\begin{aligned}
K(t,s\vert Q) &=& \int {{\rm D}}q_1 {{\rm D}}q_2\, (2\pi)^d \delta(q_{12}-Q) \frac{{{\rm e}}^{-(t+s)q_1^2-(t-s)q_2^2}}{q_1^2+m^2} ,
\label{eq:Kts}\\
\rho(t\vert 34) &=&{{\rm e}}^{- t(p_3^2+p_4^2)} -\frac{B(t\vert p_{34})}{6/u + B(0\vert p_{34})} ,
\label{eq:rho_t}\\
\rho(t_1,t_2\vert Q) &=& B(t_1+t_2\vert Q) -\frac{B(t_1\vert Q) B(t_2\vert Q)}{6/u +B(0\vert Q)} .
\label{eq:rho_ts}\end{aligned}$$ The derivation of these results is given in appendix \[app:SDE\_flow\].
Running coupling from flowed fields {#sec:coupling}
===================================
Definitions
-----------
Using the connected 4-pt functions $g \equiv \hat \delta \hat g$ for the flow fields given in eq. (\[eq:4pt\_c\]), we define the $t$-dependent dimensionless coupling as $$\begin{aligned}
g(t) = -3 \hat g(t,t; t,t\vert \{ p\}_{\rm sym}) t^{2-d/2}, \end{aligned}$$ where $\{ p\}_{\rm sym}$ is given by $p_i^2 t = 3\Delta/4$ ($i=1\sim 4$) and $p_{12}^2 t =p_{34}^2 t=\Delta$ ($p_{ij}=p_i + p_j$), which is the symmetric point for $d > 2$, and $t^{2-d/2}$ is introduced to make the coupling dimensionless. Here $\Delta$ is an arbitrary dimensionless constant but we can take $\Delta = 1$ without loss of generality by the rescaling $t\rightarrow \Delta t$. Explicitly we have $$\begin{aligned}
\hat g(t,t; t,t\vert \{ p\}_{\rm sym}) &=& 4 \hat X(t,t\vert \{ p\}_{\rm sym}) + 4 \hat Y(t\vert \{ p\}_{\rm sym}) + \hat Z(\vert \{ p\}_{\rm sym}),\end{aligned}$$ where we remove $\hat\delta$ by defining ${\cal O} = \hat\delta \hat {\cal O}$ for ${\cal O} = g, X, Y, Z$, and $$\begin{aligned}
\hat X(t_1,t_2\vert 12;34) &=& (p_2^2+m^2)(p_3^2+m^2)\int_0^{t_1} {{\rm d}}s_1\int_0^{t_2} {{\rm d}}s_2\,
{{\rm e}}^{s_1(p_2^2-p_1^2)}{{\rm e}}^{s_2(p_3^2-p_4^2)}\omega(s_1,s_2\vert p_{34}),~~~~~~ \\
\hat Y(t\vert 12;34) &=& (p_2^2+m^2) \int_0^t {{\rm d}}s\, {{\rm e}}^{s(p_2^2-p_1^2)}\psi(s\vert 34), \\
\hat Z(\vert 12;34) &=& -\frac{1}{3}\frac{u}{1+\frac{u}{6} B(0\vert p_{34})} .\end{aligned}$$
Free flow
---------
For simplicity, we first consider the free flow, where $\hat g(t,t;t,t\vert \{p\}_{\rm sym}) = \hat Z( \vert \{ p\}_{\rm sym})$. Taking $\Delta=1$, the running coupling is given by $$\begin{aligned}
g(t) &=& \frac{u t^{2-d/2}}{1+\dfrac{u}{6} B\left(1/t\right)},\end{aligned}$$ where $B(p^2) = B(0\vert p)$.
### $d=2$
In 2-dimensions, we obtain $$\begin{aligned}
g(t)&=& \frac{ut}{1+\dfrac{ut}{6\pi\sqrt{1+4m^2t}}\tanh^{-1}\left(\dfrac{1}{\sqrt{1+4m^2t}}\right)},\end{aligned}$$ which behaves in the UV limit ($t\rightarrow 0$) and IR limit ($t\rightarrow\infty$) as $$\begin{aligned}
g(t) &\simeq &\left\{
\begin{array}{ccc}
\dfrac{ ut}{1- u t \log (m^2 t) /(12\pi)} & \rightarrow 0, & t=0 \\
\\
\dfrac{u t}{1+u/(24\pi m^2)} &\rightarrow \infty , & t=\infty \\
\end{array}
\right. .\end{aligned}$$ In the massless limit $m^2\rightarrow 0$, we have $$\begin{aligned}
g(t) &\simeq &- \frac{12\pi}{\log (m^2 t)}\rightarrow 0.\end{aligned}$$
### $d=3$
At $d=3$, the running coupling is given by $$\begin{aligned}
g(t) &=&\dfrac{ u\sqrt{t}}{1+\dfrac{u\sqrt{t}}{24\pi}\arctan\left(\dfrac{1}{\sqrt{4m^2 t}}\right)},\end{aligned}$$ which behaves as $$\begin{aligned}
g(t) &\simeq &\left\{
\begin{array}{ccc}
\dfrac{ u\sqrt{t}}{1+ u \sqrt{t}/48} & \rightarrow 0, & t=0 \\
\\
\dfrac{u\sqrt{t}}{1+u /(48\pi m)} &\rightarrow \infty , & t=\infty \\
\end{array}
\right. .\end{aligned}$$
In the massless limit, we have $$\begin{aligned}
g(t) &=& \dfrac{ u\sqrt{t}}{1+ u \sqrt{t} /48} =
\left\{
\begin{array}{cc}
\rightarrow 0, & t\rightarrow 0 \\
\rightarrow 48, & t\rightarrow\infty \\
\end{array}
\right. ,\end{aligned}$$ which correspond to the asymptotic free UV fixed point and the Wilson-Fisher IR fixed point, respectively.
### $d\ge 4$
Since $B(Q^2)$ diverges as $\Lambda^{d-4}$ ($\log \Lambda$ at $d=4$) at $d\ge 4$, the running coupling vanishes as the cut-off is removed ($\Lambda\rightarrow\infty$). Thus the theory is trivial in the continuum limit at $d\ge 4$.
Interacting flow in the massless limit at $d=3$
-----------------------------------------------
### Massless limit
We next consider the interacting flow case, where we need to evaluate $\hat X$ and $\hat Y$, which are difficult to calculate in general. We therefore consider the massless limit.[^2] In this limit, the kernel function is reduced to $$\begin{aligned}
K(t,s\vert \{ p \}_{\rm sym.}) &=& D^{d/2-1} k_0(D t, D s),\end{aligned}$$ where $$\begin{aligned}
k_0(w,v) &=& \frac{{{\rm e}}^{v-w} w^{1-d/2}}{2^{d-1}(2\pi)^{d/2}}\int_0^1 {{\rm d}}z\, z^{d/2-2}\exp\left[\dfrac{(w-v)^2 z}{2w}\right],\end{aligned}$$ and we regard $D\equiv Q^2=\Delta/t$ as an independent variable. Here the $z$ integral is convergent for $d>2$ while the bubble integral $B(0\vert Q)$ is finite for $d<4$. We thus concentrate on the $d=3$ case hereafter.
In this limit, we obtain (see appendix \[app:massless\] for details) $$\begin{aligned}
\hat Z(\vert \{ p \}_{\rm sym.}) &=& -16\sqrt{D}\frac{\bar u(D) }{1 + \bar u (D)}, \qquad \bar u(D) \equiv \frac{u}{48\sqrt{D}}, \label{eq:hatZ}\\
\hat Y(t\vert \{ p \}_{\rm sym.}) &=& \frac{3 }{4}\sqrt{D} \left\{\xi_0^{(1)}(\Delta) - 8\xi_0^{(2)}(\Delta)
\frac{\bar u(D)}{1+\bar u(D)}\right\}, \label{eq:hatY}\\
\hat X(t,t\vert \{ p \}_{\rm sym.}) &=& \frac{9}{16}\sqrt{D}\left\{ \Xi_0(\Delta) -4\{\xi_0^{(2)}(\Delta)\}^2\frac{\bar u(D)}{1+\bar u(D)}\right\}, \label{eq:hatX}\end{aligned}$$ where $$\begin{aligned}
\xi_0^{(i)}(\Delta) &=& \int_0^\Delta {{\rm d}}w\, \phi_0^{(i)}(w), \quad i=1,2, \\
\Xi_0(\Delta) &=&\int_0^\Delta {{\rm d}}w\, \int_0^\Delta {{\rm d}}v\, \Omega_0(w,v) ,\end{aligned}$$ and $\phi_0^{(i)}$ and $\Omega_0$ are solutions to the integral equations $$\begin{aligned}
{{\rm e}}^{-3w/2} + \int_0^w {{\rm d}}v\, k_0(w,v)\, \phi_0^{(1)}(v) &=& 0, \label{eq:phi0_1}\\
b_0(w) +\int_0^w {{\rm d}}v\, k_0(w,v)\, \phi_0^{(2)}(v) &=& 0 , \label{eq:phi0_2}\\
b_0(w+v) -2 \int_0^w {{\rm d}}x\, k_0(w,x) \int_0^v {{\rm d}}y\, k_0(v,y)\, \Omega_0 (x,y) &=& 0,
\label{eq:Omega0}\end{aligned}$$ where $b_0(w)$ is the massless bubble integral given by eq. (\[eq:b0\]). These equations can be solved numerically, and at $\Delta =1$, for example, we have $\xi_0^{(1)}(1) = -14.8440(1)$, $\xi_0^{(2)}(1) = -1.60557(1)$ and $\Xi_0(1) = 16.6753(1)$.
### Running coupling and $\beta$ function
Using the above results, the running coupling at $d=3$ is given by $$\begin{aligned}
g_0(\mu) &=& G_1 + G_2 \frac{\bar u(\Delta) \sqrt{t}}{1+ \bar u(\Delta) \sqrt{t}}, \qquad \bar u(\Delta)= \frac{ u}{48\sqrt{\Delta}} ,\end{aligned}$$ where $\mu =1/\sqrt{t}$ and $$\begin{aligned}
G_1&=& -9\sqrt{\Delta} \left[ \xi_0^{(1)}(\Delta) +\frac{3}{4}\Xi_0(\Delta) \right], \quad
G_2 = 48\sqrt{\Delta} \left[ 1 +\frac{3}{4}\xi_0^{(2)}(\Delta)\right]^2 \ge\ 0 . \end{aligned}$$ With the numerical values given above we obtain $G_1 = 21.0378(1)$ and $G_2 = 2.00105(1)$ at $\Delta=1$. [^3]
We then calculate the $\beta$ function for $ g_0( \mu )$ as $$\begin{aligned}
\beta(g_0)&\equiv&\mu \frac{\partial }{\partial \mu} g_0(\mu) = \frac{(g_0(\mu) - G_1-G_2)(g_0(\mu) - G_1)}{G_2}, \end{aligned}$$ which becomes zero at $g_0(\mu) =G_1$ and $g_0(\mu) =G_1+G_2$. The coupling $g_0(\mu)$ near $G_1$ behaves as $$\begin{aligned}
g_0(\mu) - G_1 \simeq C_{UV} \frac{u}{\mu} \rightarrow 0 , \qquad \mu\rightarrow \infty, \qquad
C_{UV} = \left[ 1 +\frac{3}{4}\xi_0^{(2)}(\Delta)\right]^2,\end{aligned}$$ approaching to the UV fixed point from above, while near $G_1+G_2$ we have the IR fixed point as $$\begin{aligned}
g_0(\mu) - G_1-G_2 \simeq - C_{IR} \frac{\mu}{u} \rightarrow 0 , \qquad \mu\rightarrow 0, \qquad
C_{IR} =\left\{48\sqrt{\Delta} \left[ 1 +\frac{3}{4}\xi_0^{(2)}(\Delta)\right]\right\}^2,\end{aligned}$$ where the coupling approaches from below to the Wilson-Fisher fixed point in the 3 dimensional scalar theory. Note that the derivative of the $\beta$ function with respect to $g_0$ at the fixed point becomes $$\begin{aligned}
\beta^\prime(g_0) \equiv \frac{d\beta(g_0)}{d g_0} &=&
\left\{
\begin{array}{ll}
-1, & g_0=G_1 \\
1, & g_0 = G_1+G_2 \\
\end{array}
\right. ,\end{aligned}$$ which should be compared with the same quantities calculated for the standard running coupling in the 3 dimensional massless theory in Ref. [@Aoki:2014yra], where $\beta^\prime (0) = -1$ (UV) and $\beta^\prime (48) = 1$ (IR). The derivative of the $\beta$ function at the fixed point gives the anomalous dimension of the operator conjugate to the coupling in the conformal theory at the fixed point, and thus is universal. Our flow coupling indeed satisfies this condition and the derivatives at the two fixed points agree with those for the conventional definition of the coupling. This establishes that our flow coupling gives a good definition of the running coupling of the theory. The scaling dimension $\gamma$ of the operator conjugate to the running coupling $g_0$ is given by $\gamma=d+\beta^\prime(g_0)$, so that $\gamma_{\rm UV}= 2$ and $\gamma_{IR}=4$ in this model. Interestingly $\gamma_{\rm UV}= 2$ corresponds to the canonical dimension of the $\varphi^4$ operator in 3 dimensions, which is the interaction term in our theory.
By the redefinition of the coupling as $g(\mu) \equiv ( g_0(\mu) - G_1 )/G_2$, the corresponding $\beta$ function is simplified as $$\begin{aligned}
\beta (g) &\equiv&\mu \frac{\partial }{\partial \mu} g(\mu) = g(\mu) (g(\mu) - 1) .\end{aligned}$$
NLO corrections to the induced metric {#sec:metric}
=====================================
In Ref. [@Aoki:2016ohw], the induced metric has been calculated from the flowed scalar field in the large $N$ limit. It has been shown that the metric from the massive scalar field describes a space which becomes the Euclidean AdS space asymptotically both in UV and IR limits, where the radius $R_{\rm IR}$ in the IR is larger than the radius $R_{\rm UV}$ in UV as $$\begin{aligned}
R_{\rm UV}^2 &=&\frac{d-2}{2} R_{\rm IR}^2 < R_{\rm IR}^2,\end{aligned}$$ while the metric describes the whole AdS space in the massless limit with the radius $R_{\rm UV}$. In this section, we consider the NLO correction to the induced metric in the $1/N$ expansion as another application of the NLO calculation, in particular, in the massless case at $d=3$, in order to see whether the space remains AdS or not and how the radius changes at the NLO.
Induced metric at NLO
---------------------
The VEV of the induced metric is defined from the normalized flowed field as[@Aoki:2016ohw] $$\begin{aligned}
g_{\mu\nu} (z) &=& R_0^2 \langle \partial_\mu \sigma^a (z) \partial_\nu \sigma^a (z) \rangle \end{aligned}$$ with some length scale $R_0$, where $z=(\tau=2\sqrt{t}, x)$ and $\mu,\nu =0,1,\cdots, d$. Here $\sigma^a(z)$ is the normalized flowed field such that $\langle \sigma^2(z) \rangle =1$, and the corresponding 2-point function is explicitly given at NLO as $$\begin{aligned}
\langle \sigma^{a_1}(z_1) \sigma^{a_2}(z_2) \rangle &=& \frac{\delta^{a_1a_2}}{N}\frac{1}{\sqrt{\zeta_0(t_1)\zeta_0(t_2)}}\left( 1-\dfrac{\zeta_1(t_1)+\zeta_1(t_2)}{2N}\right) \nonumber \\
&\times&
\int {{\rm D}}p \frac{{{\rm e}}^{-p^2(t_1+t_2)}{{\rm e}}^{i p(x_1-x_2)}}{p^2+m^2} \left[1 +\frac{G_1(t_1,t_2\vert p)}{N} \right],\end{aligned}$$ where $$\begin{aligned}
\zeta_1(t) &=& \frac{1}{\zeta_0(t)} {\cal H}\left[ G_1(t,t\vert p) \right] , \qquad {\cal H}\left[ f(t\vert p) \right]\equiv\int {{\rm D}}p \frac{{{\rm e}}^{-2p^2t} }{p^2+m^2} f(t\vert p) .
\label{eq:def_H}\end{aligned}$$
After some algebra (see appendix \[app:metric\]), we obtain $$\begin{aligned}
g_{ij}(\tau) &=& \delta_{ij}\frac{ R_0^2}{d} A(t), \quad (i,j=1, 2, \cdots, d), \qquad g_{00}(\tau) = - \frac{ R_0^2\, t}{2}\partial_t A(t), \end{aligned}$$ where $$\begin{aligned}
A(t) &=& -\frac{1}{2} \frac{\partial_t\zeta_0(t)}{\zeta_0(t)} +\frac{1}{N}A_1(t), \end{aligned}$$ and $A_1(t)$ in general is a very complicated function given in appendix \[app:metric\].
Induced metric in the massless limit at $d=3$
---------------------------------------------
In the massless limit at $d=3$, the metric at the LO is given by $$\begin{aligned}
g_{ij}(\tau) &=&\delta_{ij} \frac{R_0^2}{3\tau^2}, \qquad g_{00}(\tau) = \frac{R_0^2}{2\tau^2},\end{aligned}$$ which describes the AdS space for all $\tau$.
At the NLO, $A_1(t)$ is given by $$\begin{aligned}
A_1(t) &=& \frac{1}{2\sqrt{t}}\int {{\rm D}}Q\, h_{\rm total}(Q^2)\frac{\bar u(Q^2)}{(1+\bar u(Q^2)\sqrt{t})^2},
\quad \bar u(Q^2) = \frac{u}{48\sqrt{Q^2}}, \label{eq:A1} \\
\partial_t A_1(t) &=& -\frac{1}{4\sqrt{t^3}}\int {{\rm D}}Q\, h_{\rm total}(Q^2)\frac{\bar u(Q^2)(1+3\bar u(Q^2)\sqrt{t})}{(1+\bar u(Q^2)\sqrt{t})^3}, \label{eq:pA1} \end{aligned}$$ which leads to $$\begin{aligned}
g_{ij}(\tau) &=&\delta_{ij}\frac{R_0^2}{3 \tau^2} \left[1 +\frac{\tau}{N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\frac{\bar u(Q^2)}{(1+\bar u(Q^2)\tau/2)^2}\right],
\label{eq:metric_space}\\
g_{00}(\tau) &=& \frac{R_0^2}{2\tau^2} \left[1 +\frac{\tau}{2N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\frac{\bar u(Q^2)(1+3\bar u(Q^2)\tau/2)}{(1+\bar u(Q^2)\tau/2)^3} \right] ,
\label{eq:metric_time}\end{aligned}$$ where $h_{\rm total}(Q^2)$ is a function given in appendix \[app:metric\].
UV and IR limits
----------------
The above expression in the UV limit ($\tau\rightarrow 0$) leads to $$\begin{aligned}
g_{ij}(\tau) &\simeq&\delta_{ij}\frac{R_0^2}{3\tau^2} \left[1 +\frac{\tau}{N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\bar u(Q^2)\right], \quad \tau\rightarrow 0, \\
g_{00}(\tau) &\simeq& \frac{R_0^2}{2 \tau^2} \left[1 +\frac{\tau}{2N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\bar u(Q^2) \right] , \quad \tau\rightarrow 0,\end{aligned}$$ which shows that the NLO correction is less singular than the LO contribution. Therefore the space becomes asymptotically AdS in the UV limit at the NLO whose AdS radius is equal to that at the LO.
We cannot naively take the $\tau\rightarrow\infty$ limit in eqs. (\[eq:metric\_space\]) and (\[eq:metric\_time\]), on the other hand, due to the enhancement of the UV contribution of the $Q$ integrals. Careful evaluations of these $Q$ integrals in appendix \[app:metric\] give $$\begin{aligned}
g_{ij}(\tau) &\simeq&\delta_{ij}\frac{R_0^2}{3\tau^2} \left[1 +\frac{r}{N}\right], \quad
g_{00}(\tau) \simeq \frac{R_0^2}{2 \tau^2} \left[1 +\frac{r}{N}\right], \quad \tau\rightarrow \infty, \end{aligned}$$ where $r= -0.41869(1)$.[^4] Therefore, the space becomes asymptotically AdS again in the IR limit, whose radius, however, is smaller than that in the UV limit.[^5] The induced metric at the NLO describes a 4 dimensional space which is asymptotically AdS in both UV and IR regions with different radii but non-AdS in-between.
It is clear that the NLO correction to the AdS radius in the IR limit is related to the Wilson-Fisher fixed point in the original 3 dimensional scalar theory, since the eqs. (\[eq:metric\_space\]) and (\[eq:metric\_time\]) can be written as $$\begin{aligned}
g_{ij}(\tau) &=& \delta_{ij}\frac{R_0^2}{3 \tau^2} \left[1 -\frac{1}{24N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\beta ( g ( 48 \mu\sqrt{Q^2}))\right], \\
g_{00}(\tau) &=& \frac{R_0^2}{2\tau^2} \left[1 -\frac{1}{24N}\int {{\rm D}}Q\, h_{\rm total}(Q^2)
\left\{ 1 + \frac{\mu}{2}\frac{ \partial}{\partial \mu}\right\} \beta ( g ( 48\mu\sqrt{Q^2}))
\right] ,\end{aligned}$$ where $\mu =1/\sqrt{t} =2/\tau$, and $\beta(g(x))$ is the $\beta$ function for the running coupling $g(x)$ from the free flowed field defined in the previous section with $\Delta=1$ as $$\begin{aligned}
\beta(g) &=& \frac{g(g-48)}{48}, \qquad
g(x) = 48 \frac{ u }{x+ u }.\end{aligned}$$
Summary {#sec:summary}
=======
In this paper, we studied the flow equation of the O($N$) $\varphi^4$ model in $d$ dimensions at the NLO in the $1/N$ expansion, employing the Schwinger-Dyson equation. We calculated the 2-pt and 4-pt functions at the NLO.
As an application of the NLO calculation, we investigated the running coupling defined from the connected 4-pt function of flowed fields. In particular at $d=3$ in the massless limit, we showed that the running coupling has two fixed points, the asymptotic free one in the UV region and the Wilson-Fisher one in the IR region. We also derived the corresponding $\beta$ function. Our study suggests that the flow equation can be interpreted as a renormalization group transformation.
We also calculated the NLO correction to the $d+1$ dimensional metric induced from the massless scalar field theory at $d=3$. In the massless limit, the whole 4-dimensional space becomes AdS at the LO of the $1/N$ expansion[@Aoki:2016ohw]. We found that the NLO corrections give small perturbations to the metric, which make the space only asymptotically AdS in both UV ($t=0$) and IR ($t=\infty$) limits. In addition, while the NLO corrections do not change the AdS radius at the LO in the UV limit, the AdS radius is reduced by the NLO correction in the IR limit, which corresponds to the Wilson-Fisher IR fixed point of the original theory. The nontrivial fixed point in the field theory appears as a change of the AdS radius at the NLO. The induced metric at NLO describes a 4-dimensional space which connects one asymptotically AdS space at UV to the other asymptotically AdS space at IR.
This paper contains two important messages. One is that the flow equation can provide an alternative method to define a renormalization group transformation. The other is that the massless scalar field in $d$ dimensions plus the extra dimension from the RG scale not only generates a $d+1$ dimensional AdS space at LO[@Aoki:2016ohw] but also gives a NLO correction, which makes the $d+1$ dimensional space asymptotically AdS only in UV and IR limits at $d=3$. In particular, the AdS radius in the IR limit, which corresponds to the Wilson-Fisher fixed point, becomes smaller than that in the UV limit, which is equal to the radius at the LO. Although the relation found in this paper between the massless scalar field theory and the induced geometry is very suggestive, further studies will be needed to establish an alternative interpretation of AdS/CFT correspondences proposed in Ref. [@Aoki:2015dla] in terms of field theories.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to thank Satoshi Yamaguchi for very useful comments and discussions. S. A. is supported in part by the Grant-in-Aid of the Japanese Ministry of Education, Sciences and Technology, Sports and Culture (MEXT) for Scientific Research (No. JP16H03978), by a priority issue (Elucidation of the fundamental laws and evolution of the universe) to be tackled by using Post “K" Computer, and by Joint Institute for Computational Fundamental Science (JICFuS). This investigation has also been supported in part by the Hungarian National Science Fund OTKA (under K116505). S. A. and J. B. would like to thank the Max-Planck-Institut für Physik for its kind hospitality during their stay for this research project. T.O. is supported in part by the Grant-in-Aid of the Japanese Ministry of Education, Sciences and Technology, Sports and Culture (MEXT) for Scientific Research (No. 26400248).
[99]{} S. Aoki, J. Balog, T. Onogi and P. Weisz, PTEP [**2016**]{} (2016) no.8, 083B04 doi:10.1093/ptep/ptw106 arXiv:1605.02413 \[hep-th\].
S. Aoki, K. Kikuchi and T. Onogi, PTEP [**2015**]{} (2015) no.10, 101B01 doi:10.1093/ptep/ptv131 \[arXiv:1505.00131 \[hep-th\]\].
R. Narayanan and H. Neuberger, JHEP [**0603**]{}, 064 (2006) \[hep-th/0601210\].
M. Lüscher, JHEP [**1008**]{}, 071 (2010) \[JHEP [**1403**]{}, 092 (2014)\] \[arXiv:1006.4518 \[hep-lat\]\].
M. Lüscher, Commun. Math. Phys. [**293**]{}, 899 (2010) \[arXiv:0907.5491 \[hep-lat\]\].
M. Lüscher, PoS LATTICE [**2013**]{}, 016 (2014) \[arXiv:1308.5598 \[hep-lat\]\].
J. M. Maldacena, Int. J. Theor. Phys. [**38**]{}, 1113 (1999) \[Adv. Theor. Math. Phys. [**2**]{}, 231 (1998) \] \[hep-th/9711200\].
I. R. Klebanov and A. M. Polyakov, Phys. Lett. B [**550**]{} (2002) 213 doi:10.1016/S0370-2693(02)02980-5 \[hep-th/0210114\].
S. Aoki, K. Kikuchi and T. Onogi, JHEP [**1504**]{}, 156 (2015) \[arXiv:1412.8249 \[hep-th\]\].
0 S. Giombi and X. Yin, J. Phys. A [**46**]{} (2013) 214003 doi:10.1088/1751-8113/46/21/214003 \[arXiv:1208.4036 \[hep-th\]\].
M. Lüscher and P. Weisz, JHEP [**1102**]{}, 051 (2011) \[arXiv:1101.0963 \[hep-th\]\].
H. Makino and H. Suzuki, PTEP [**2015**]{}, no. 3, 033B08 \[arXiv:1410.7538 \[hep-lat\]\].
K. Hieda, H. Makino, H. Suzuki, \[arXiv:1604.06200 \[hep-lat\]\].
F. Capponi, L. Del Debbio, S. Ehret, R. Pellegrini and A. Rago, arXiv:1512.02851 \[hep-lat\].
K. Fujikawa, JHEP [**1603**]{} (2016) 021 doi:10.1007/JHEP03(2016)021 \[arXiv:1601.01578 \[hep-lat\]\]. H. Suzuki, private communication.
M. Moshe, J. Zinn-Justin, Phys.Rept. 385 (2003) 69-228 S. Aoki, K. Kikuchi and T. Onogi, JHEP [**1504**]{}, 156 (2015) \[arXiv:1412.8249 \[hep-th\]\].
H. Makino, F. Sugino and H. Suzuki, arXiv:1412.8218 \[hep-lat\].
S. Aoki, J. Balog and P. Weisz, JHEP [**1409**]{} (2014) 167 Erratum: \[JHEP [**1507**]{} (2015) 037\] doi:10.1007/JHEP07(2015)037, 10.1007/JHEP09(2014)167 \[arXiv:1407.7079 \[hep-lat\]\].
S. S. Pufu, arXiv:1608.02960 \[hep-th\].
The $1/N$ expansion in the $d$ dimensional theory {#app:d-dim}
=================================================
In this appendix, we consider the $1/N$ expansion in the $d$ dimensional theory.
Schwinger-Dyson equation(SDE)
-----------------------------
In order to perform the $1/N$ expansion, we consider the SDE of this model, which can be written compactly as $$\begin{aligned}
\langle \delta^a_x X[\varphi] \rangle &=& \langle X[\varphi] \delta^a_x S(\mu^2,u) \rangle,\end{aligned}$$ where $ \delta^a_x \varphi^b(y) = \delta^{ab}\delta^{(d)}(x-y) \epsilon$ with a small parameter $\epsilon$, so that $$\begin{aligned}
\delta^a_x S(\mu^2,u) &=& N \epsilon\left[ (- \Box +\mu^2)\varphi^a(x) + \frac{u}{3!}\varphi^a(x)\varphi^2(x)\right] .\end{aligned}$$ Here the vacuum expectation value of an operator ${\cal O}$ is defined in eq. (\[eq:vev\_d\]).
We define $2n$-point functions $\Gamma_{2n}$[^6] as $$\begin{aligned}
\Gamma^{a_1a_2\cdots a_{2n}}(x_1,x_2,\cdots,x_{2n}) &=&
N^{2n-1}\left\langle \prod_{i=1}^{2n} \varphi^{a_i}(x_i) \right\rangle \equiv
\Gamma_{2n}[12\cdots (2n) ]
\label{eq:G2n_d}\end{aligned}$$ which can be written in terms of their connected parts $K_{2n}$ as $$\begin{aligned}
\Gamma_4[1234] &=& K_4[1234] + N\left\{\Gamma_2[12]\Gamma_2[34] + \Gamma_2[13]\Gamma_2[24]
+\Gamma_2[14]\Gamma_2[23] \right\}, \\
\Gamma_6[123456] &=& K_6[123456] + N\left\{\Gamma_2[12]K_4[3456] + \mbox{ 14 perms. } \right\}\nonumber \\
&+& N^2\left\{\Gamma_2[12] \Gamma_2[34]\Gamma_2[56]+ \mbox{ 14 perms. } \right\}\end{aligned}$$ and so on. As mentioned in the main text, we assume we are working in a phase where O$(N)$ symmetry is not broken. We therefore do not add the external source term $h \varphi(x)$ to the action, so that the action has the symmetry under $\varphi \rightarrow -\varphi$, which implies $\Gamma_{2n-1} = 0$ for all positive integers $n$.
In terms of these, the SDE for $X(\varphi) = \varphi^{a_2}(x_2)$ becomes $$\begin{aligned}
\delta_{12} &=& (- \Box +\mu^2)_{x_1} \Gamma_2[12] +\frac{u}{3! N^2}\sum_b \left(K_4[bb12]
+N\left\{ \Gamma_2[bb]\Gamma_2[12] + 2\Gamma_2[b1]\Gamma_2[b2]\right\} \right)~~~
\label{eq:SDE_2pt}\end{aligned}$$ where $\delta_{12}\equiv\delta^{a_1a_2}\delta^{(d)}(x_1-x_2)$ and $x_b=x_1$, so that $b$ in the summation runs over the O($N$) indices only.
For $X(\varphi) = \varphi^{a_2}(x_2) \varphi^{a_3}(x_3)\varphi^{a_4}(x_4)$, on the other hand, we have $$\begin{aligned}
\delta_{12}\Gamma_2[34] +
\mbox{ 2 perms.}
&=& (- \Box +\mu^2)_{x_1} \frac{1}{N}\left( K_4[1234]+ N\left\{\Gamma_2[12]\Gamma_2[34] + \mbox{ 2 perms.}\right\}\right) \nonumber \\
&+&
\frac{u}{3! N^3}\sum_b \left(K_6[bb1234] + N\left\{\Gamma_2[bb]K_4[1234] + \mbox{ 14 perms.} \right\}\right.
\nonumber \\
&+&\left. N^2\left\{\Gamma_2[bb]\Gamma_2[12]\Gamma_2[34]+\mbox{ 14 perms.} \right\}\right),\end{aligned}$$ which can be simplified by using eq. (\[eq:SDE\_2pt\]) as $$\begin{aligned}
0&=& (- \Box +\mu^2)_{x_1} K_4[1234] + \frac{u}{3! N^2}\sum_b \Bigl( K_6[bb1234]
+ N\Gamma_2[bb]K_4[1234] \nonumber \\
&+& 2N\left\{\Gamma_2[b1]K_4[b234] + \Gamma_2[b2]K_4[1b34]+\Gamma_2[b3]K_4[12b4]+\Gamma_2[b4]K_4[123b]
\right\} \nonumber \\
&+&N\left\{ \Gamma_2[12]K_4[bb34] + \Gamma_2[13]K_4[b2b4]+\Gamma_2[14]K_4[b23b]
\right\} \nonumber \\
&+& 2N^2\left\{\Gamma_2[b2][\Gamma_2[b3]\Gamma_2[14] + \Gamma_2[b2][\Gamma_2[b4]\Gamma_2[13]
+\Gamma_2[b3][\Gamma_2[b4]\Gamma_2[12]
\right\} \Bigr) .
\label{eq:SDE_4pt}\end{aligned}$$
Using the O($N$) symmetry and assuming translational invariance (e.g. infinite volume or periodic boundary condition), we can write $$\begin{aligned}
\Gamma_2[12]&\equiv& \delta^{a_1a_2}\Gamma(x_{12}),\qquad x_{12}\equiv x_1-x_2 \\
K_4[1234]&\equiv& \delta^{a_1a_2} \delta^{a_3a_4} K(x_1,x_2;x_3,x_4) + \mbox{ 2 perms.} , \\
K_6[123456] &\equiv& \delta^{a_1a_2} \delta^{a_3a_4} \delta^{a_5a_6} H(x_1,x_2;x_3,x_4;x_5,x_6)
+ \mbox{ 14 perms.},\end{aligned}$$ where $K(x_1,x_2;x_3,x_4)$ is invariant under $1\leftrightarrow 2$ or $3\leftrightarrow 4$ as well as $(12)\leftrightarrow (34)$, and similar invariances hold for $H(x_1,x_2;x_3,x_4;x_5,x_6)$.
We finally obtain $$\begin{aligned}
\delta^{(d)}(x_1-x_2) &=& \left[ (- \Box +\mu^2)_{x_1} +\frac{u}{3!} \Gamma(0)\right] \Gamma(x_{12})\nonumber \\
&+& \frac{u}{3! N}\left[\left(1+\frac{2}{N}\right) K(x_1,x_1;x_1,x_2) + 2\Gamma(0)\Gamma(x_{12})\right] ,
\label{eq:SDE_2ptA}\end{aligned}$$ and $$\begin{aligned}
0&=& \left[ (- \Box +\mu^2)_{x_1} +\frac{u}{3!}\left(1+\frac{2}{N}\right)\Gamma(0)\right] K(x_1,x_2;x_3,x_4)
\nonumber \\
&+&
\frac{ u}{3! }\Gamma(x_{12}) \left[
2 \Gamma(x_{13})\Gamma(x_{14}) +\left(1+\frac{2}{N}\right) K(x_1,x_1;x_3,x_4)
+\frac{2}{N}K(x_1,x_3;x_1,x_4)\right]
\nonumber \\
&+& \frac{u}{3! N}\left[\left(1+\frac{2}{N}\right) H(x_1,x_1;x_1,x_2;x_3,x_4) + \frac{2}{N}H(x_1,x_2;x_1,x_3;x_1,x_4)\right]\nonumber\\
&+& \frac{2u}{3! N}
\left[ \Gamma(x_{13})K(x_1,x_2;x_1,x_4)+ \Gamma(x_{14})K(x_1,x_2;x_3,x_1) \right] .~~~~~
\label{eq:SDE_4ptA}\end{aligned}$$
The leading order in the $1/N$ expansion
----------------------------------------
We introduce the $1/N$ expansion as $$\begin{aligned}
\Gamma(x_{12}) &=&\sum_{i=0}^\infty N^{-i} \Gamma_i(x_{12}), \qquad
K(x_1,x_2;x_3,x_4) = \sum_{i=0}^\infty N^{-i} K_i(x_1,x_2;x_3,x_4), \end{aligned}$$ and so on, together with $\displaystyle \mu^2 =\sum_{i=0}^\infty N^{-i} \mu^2_i $.
At the leading order (LO) of the $1/N$ expansion, the eq. (\[eq:SDE\_2ptA\]) in momentum space becomes $$\begin{aligned}
1 &=& \left( p^2+\mu^2_0 +\frac{u}{6} \int {{\rm D}}q\, \widetilde{\Gamma}_0(q) \right) \widetilde{\Gamma}_0(p) ,
\qquad \Gamma_0(x) =\int Dp\, \widetilde{\Gamma}_0(p)\, {{\rm e}}^{ipx}, \end{aligned}$$ which can easily be solved as $$\begin{aligned}
\widetilde{\Gamma}_0(p) &=&\frac{1}{p^2+m^2}, \qquad m^2 = \mu^2_0 + \frac{u}{6} Z(m),
\label{eq:sol_2pt}\end{aligned}$$ where $m \ge 0$ is the renormalized mass and $Z(m)$ is given in eq. (\[eq:Zm\]). Thus the 2-pt function at the LO becomes $$\begin{aligned}
\langle \varphi^a(x) \varphi^b(y)\rangle &=& \frac{\delta^{ab}}{N}\int {{\rm D}}p \frac{{{\rm e}}^{ip(x-y)}}{p^2+m^2} .\end{aligned}$$
Eq. (\[eq:SDE\_4ptA\]) at the LO leads to $$\begin{aligned}
(- \Box +m^2)_{x_1} K_0(x_1,x_2;x_3,x_4)
&+&\frac{ u}{3! } \Gamma_0(x_{12})K_0(x_1,x_1;x_3,x_4)\nonumber \\
&=& -\frac{ 2 u}{3! }
\Gamma_0(x_{12})\Gamma_0(x_{13})\Gamma_0(x_{14}) .\end{aligned}$$ Introducing a function $G_0(p_1,p_2,p_3,p_4)$ to rewrite $K_0(x_1,x_2,x_3,x_4)$ as $$\begin{aligned}
K_0(x_1,x_2;x_3,x_4) &=& \left\{\prod_{i=1}^4 \int {{\rm D}}p_i \frac{{{\rm e}}^{i p_i x_i}}{p_i^2+m^2}\right\}
G_0(p_1,p_2,p_3,p_4)(2\pi)^d \delta(p_1+p_2+p_3+p_4),~~~~~~~~\end{aligned}$$ we obtain $$\begin{aligned}
G_0(p_1,p_2,p_3,p_4) &=& G_0(p_1+p_2) = -\frac{2 u}{6+ u B(p_{12}^2)},
\label{eq:4pt_LO}\end{aligned}$$ where $p_{12}=p_1+p_2$, and $$\begin{aligned}
B(Q^2) &=& \int {{{\rm D}}}q_1 {{{\rm D}}}q_2 \frac{(2\pi)^d\delta(q_1+q_2-Q)}{(q_1^2+m^2)(q_2^2+m^2)}
= \int_0^1 {{\rm d}}x\, \int {{{\rm D}}}q_1 \frac{ \theta(\Lambda^2-q_1^2)}{(q_1^2+m^2+ x(1-x) Q^2)^2} . ~~~~~~~\end{aligned}$$ This agrees with the previous result obtained by a different method[@Aoki:2014yra]. We here specify the way we introduce the cut-off $\Lambda$ for the case where $B(Q^2)$ diverges.
NLO correction to the 2-pt functions
------------------------------------
Let us consider the next-to-leading order (NLO) correction to the 2-pt function $\Gamma_2$. At the NLO, eq. (\[eq:SDE\_2ptA\]) leads to $$\begin{aligned}
0&=& (-\Box +m^2)\Gamma_1(x_{12}) +\left\{\frac{u}{6}(2 Z(m) + \gamma_1) +\mu^2_1\right\} \Gamma_0(x_{12}) +\frac{u}{6}K_0(x_1,x_1;x_1,x_2), ~~~~ \\
\gamma_1 &=& \int {{\rm D}}q\, \widetilde{\Gamma}_1(q), \end{aligned}$$ which can be solved in momentum space as $$\begin{aligned}
\widetilde{\Gamma}_1(p) &=& -\frac{1}{(p^2+m^2)^2}\left( \mu^2_1+\frac{u}{6}\gamma_1 + \frac{u}{3} S(p^2)\right),
\label{eq:Gamma_1}\end{aligned}$$ where $$\begin{aligned}
S(p^2) =\int \frac{ {{\rm D}}Q }{(p-Q)^2+m^2}\frac{6}{6+u B(Q^2)} ,\end{aligned}$$ and the condition for $\gamma_1$ is solved as $$\begin{aligned}
\gamma_1 &=& - \frac{\mu_1^2 B(0) +C_2}{1+\frac{u}{6} B(0)}, \qquad
C_2\equiv - \int \frac{ {{\rm D}}Q }{\frac{6}{u}+ B(Q^2)}\frac{d }{d m^2} B(Q^2) .
\label{eq:gamma_1}\end{aligned}$$ Substituting eq. (\[eq:gamma\_1\]) into eq. (\[eq:Gamma\_1\]), we finally obtain $$\begin{aligned}
\widetilde{\Gamma}_1(p) &=& - \frac{1}{(p^2+m^2)^2}\left\{ g(p^2) + \tilde C \right\},\end{aligned}$$ where $$\begin{aligned}
g(p^2) &=& \int \frac{{{\rm D}}Q}{\frac{6}{u} +B(Q^2)}\left\{ \frac{1}{(Q+p)^2+m^2} + \frac{1}{(Q-p)^2+m^2} - \frac{2}{Q^2+m^2} \right\} , \\
\tilde C &=& C_1 +\frac{\mu_1^2}{1+\frac{u}{6}B(0)}
-\frac{C_2}{\frac{6}{u}+B(0)} , \quad
C_1 = \int \frac{{{\rm D}}Q}{\frac{6}{u} +B(Q^2)} \frac{2}{(Q^2+m^2)}, \end{aligned}$$ and $g(p^2)$ can be expanded as $$\begin{aligned}
g(p^2) &=& Z_1 p^2 + \tilde g(p^2), \quad \tilde g(p^2) = O(p^4),\end{aligned}$$ where $$\begin{aligned}
Z_1 &=&\frac{2}{d} \int \frac{{{\rm D}}Q}{6/u + B(Q^2)}\left[\frac{4-d}{(Q^2+m^2)^2} -\frac{4 m^2}{ (Q^2+m^2)^3}
\right] .
\label{eq:Z1}\end{aligned}$$
Renormalization
---------------
Let us now consider the renormalization of the theory.
Our renormalization condition for the renormalized 2-pt function $\Gamma_R$ is given in momentum space as $$\begin{aligned}
\widetilde{\Gamma}_R^{-1}(p) &\simeq & p^2+ m^2, \qquad p^2\simeq 0,\end{aligned}$$ where $m$ is interpreted as the renormalized mass, which is independent of both $N$ and the cut-off. Relating the bare field to the renormalized field by the renormalization constant $Z_R$ as $Z_R^{1/2} \varphi_R = \varphi$, we explicitly obtain $$\begin{aligned}
Z_R \widetilde{\Gamma}_R(p) &=& \widetilde{\Gamma}(p) = \frac{1}{p^2+m^2 +\dfrac{1}{N}\Sigma_1(p^2)} + O\left(\frac{1}{N^2}\right) ,\end{aligned}$$ where $$\begin{aligned}
\Sigma_1(p^2) &=& Z_1 p^2 + \tilde C +\tilde g(p^2) .\end{aligned}$$
At the LO of the $1/N$ expansion, the above condition implies $$\begin{aligned}
\mu_0^2 &=& m^2 -\frac{u}{6} Z(m) , \qquad Z_R = 1, \end{aligned}$$ where $Z(m)$ is potentially divergent at $d > 1$. We therefore introduce the momentum cut-off $\Lambda$ to regulate the integral, and $\mu^2_0$ is tuned to cancel the effect of $Z(m)$ including such divergences, in order to keep the renormalized mass $m$ finite and constant. The lattice regularization or dimensional regularization is more consistent than the momentum cut-off, but calculations become much more complicated in the lattice regularization or power divergences are difficult to deal with in the dimensional regularization. Since the momentum cut-off is enough to see the leading divergences, we adopt it in this paper.
At the NLO, the renormalization condition implies $$\begin{aligned}
Z_R &=& 1 -\frac{Z_1}{N}, \qquad
\mu_1^2 = \left(1+\frac{u}{6}B(0) \right) Z_1 m^2 + \frac{u}{6} C ,
\label{eq:Z1_mu1}\end{aligned}$$ where $$\begin{aligned}
C
&=&
-\int \frac{{{\rm D}}Q}{\frac{6}{u}+B(Q^2)}\left[\frac{d B(Q^2)}{d m^2} +2\frac{\frac{6}{u} + B(0)}{Q^2+m^2}\right].
\label{eq:C}\end{aligned}$$
The renormalization condition for the coupling, which first appears at the NLO of the $1/N$ expansion, is given by $G_0(Q^2=s) = -u_r(s)/3$, so that $u_r(s)$ is regarded as the renormalized coupling at the scale $s$. Eq. (\[eq:4pt\_LO\]) thus leads to $$\begin{aligned}
u_r(s) &=& \frac{u}{1+\frac{u}{6} B(s)},\end{aligned}$$ where $B(Q^2)$ is divergent at $d \ge 4$. Therefore the renormalized coupling goes to zero as $$\begin{aligned}
u_r(s) \simeq \frac{6}{B(s)} \rightarrow 0, \quad \Lambda \rightarrow \infty \end{aligned}$$ at $d\ge 4$. This indicates the triviality of the $\varphi^4$ theory at $d\ge 4$.
Renormalization constants
-------------------------
We here explicitly evaluate the renormalization constants.
### $d=1$
At $d=1$, $\mu_0^2$ is finite as $$\begin{aligned}
Z(m) &=& \frac{1}{\pi m} \arctan \left(\frac{\Lambda}{m} \right) \end{aligned}$$ is finite, and the coupling is also finite and nonzero since $$\begin{aligned}
B(Q^2) &=& \frac{1}{ m (Q^2+4m^2)} \simeq \frac{1}{m Q^2}+\cdots,\qquad Q^2\to\infty,\end{aligned}$$ has a finite limit as $\Lambda\to\infty$.
The most divergent part of $Z_1$ is given by $$\begin{aligned}
Z_1 &\simeq&\left\{
\begin{array}{ll}
\displaystyle \int {{\rm D}}Q \frac{u}{(Q^2+m^2)^2}, &u\not=\infty \\
\\
\displaystyle \int \frac{{{\rm D}}Q}{B(Q^2)}\frac{6}{(Q^2+m^2)^2} , &u=\infty \\
\end{array}
\right. ,\end{aligned}$$ which shows that $Z_1$ is finite for all $u$ including $u=\infty$. Eqs. (\[eq:Z1\_mu1\]) and (\[eq:C\]) thus tell us that $\mu_1^2$ is also finite for all $u$ including $u=\infty$, and therefore, there is no divergence at $d=1$ up to the NLO.
### $d=2$
At $d=2$, $\mu_0^2$ is logarithmically divergent as $$\begin{aligned}
\mu_0^2 &= & m^2-\frac{u}{6} Z(m), \qquad Z(m)\simeq\frac{1}{ 4\pi} \log \left(\frac{\Lambda^2+m^2}{m^2} \right).\end{aligned}$$
On the other hand, $B(Q^2)$ is finite as $$\begin{aligned}
B(Q^2) &=& \frac{\tanh^{-1}\left(\sqrt{\frac{Q^2}{Q^2+4m^2}}\right)}{\pi \sqrt{Q^2(Q^2+4m^2)}}
\simeq\frac{1}{2\pi Q^2}\log\frac{Q^2}{m^2} -\frac{m^2}{\pi (Q^2)^2}\left(\log \frac{Q^2}{m^2}-1\right)+\cdots,~~
\\
\frac{d B(Q^2)}{d m^2} &\simeq& -\frac{2 B(0)}{Q^2 +4m^2}\left[1+\frac{2m^2}{Q^2} \log\frac{Q^2}{m^2}+\cdots\right],
\quad B(0) =\frac{1}{4\pi m^2},\end{aligned}$$ so that the renormalized coupling becomes $$\begin{aligned}
u_r(s) &=& \frac{6u}{\displaystyle 6 +u \frac{\tanh^{-1}\left(\sqrt{\frac{s}{s+4m^2}}\right)}{\pi \sqrt{s(s+4m^2)}} } \simeq \frac{12\pi u s}{12\pi s +u \log (s/m^2)}, \qquad s\rightarrow\infty .\end{aligned}$$
The most singular term of $Z_1$ for $u\not=\infty$ becomes $$\begin{aligned}
Z_1 &\simeq&\frac{u}{6} \int {{{\rm D}}} Q \frac{2}{(Q^2+m^2)^2},\end{aligned}$$ which is manifestly finite, while at $u=\infty$, we have $$\begin{aligned}
Z_1 &=& \int \frac{{{\rm D}}Q}{B(Q^2)}\left[\frac{2}{(Q^2+m^2)^2} -\frac{4 m^2}{ (Q^2+m^2)^3}\right],\end{aligned}$$ which diverges as $Z_1 \simeq \log\left(\log \Lambda^2\right)$.
The most divergent part of $\mu_1^2$ is given by $$\begin{aligned}
\mu^2_1 &\simeq& \left\{
\begin{array}{ll}
-\dfrac{u}{3} Z(m)\delta_1, \quad (\delta_1=1), & u \not= \infty \\
\\
\dfrac{u}{12\pi}\log \left(\dfrac{\Lambda^2+4m^2}{4m^2}\right) , & u\to\infty \\
\end{array}
\right. .\end{aligned}$$
### $d=3$
At $d=3$, $\mu_0^2$ is linearly divergent as $$\begin{aligned}
\mu_0^2 &= & m^2-\frac{u}{6} Z(m), \qquad Z(m)\simeq\frac{1}{ 2\pi^2}\left[\Lambda - m \arctan \left(\frac{\Lambda}{m} \right) \right],\end{aligned}$$ while $B(Q^2)$ is finite as $$\begin{aligned}
B(Q^2) &=& \frac{1}{4\pi\sqrt{Q^2} } \arctan \left(\sqrt{\frac{Q^2}{4 m^2}}\right)
\simeq\ \frac{1}{8\vert Q\vert} -\frac{ m }{2\pi Q^2} +\frac{2 m^3}{3\pi (Q^2)^2}+\cdots, ~~
\\
\frac{d B(Q^2)}{d m^2} &=& -\frac{2 B(0)}{Q^2+4m^2},
\quad B(0) =\frac{1}{8\pi m},\end{aligned}$$ and the renormalized coupling becomes $$\begin{aligned}
u_r(s) &=& \frac{6u}{\displaystyle 6 + \frac{u}{4\pi\sqrt{s} } \arctan \left(\sqrt{\frac{s}{4 m^2}}\right)}
\simeq \frac{ u }{ \displaystyle 1 +\frac{u}{48 \sqrt{s}} }, \qquad s\rightarrow\infty .\end{aligned}$$
The most singular term of $Z_1$ for $u\not=\infty$ becomes $$\begin{aligned}
Z_1 &\simeq&\frac{u}{ 9} \int {{\rm D}}Q\, \frac{1}{(Q^2+m^2)^2},
\label{eq:Z_1}\end{aligned}$$ which is manifestly finite at $d=3$. On the other hand, at $u=\infty$, we have $$\begin{aligned}
Z_1 &=&\frac{2}{3} \int \frac{{{\rm D}}Q}{B(Q^2)}\left[\frac{1}{(Q^2+m^2)^2} -\frac{4 m^2}{ (Q^2+m^2)^3}\right],\end{aligned}$$ whose divergent part becomes $$\begin{aligned}
Z_1 &\simeq &
\dfrac{4}{3\pi^2}\log \Lambda^2 .\end{aligned}$$
The most divergent part of $\mu_1^2$ becomes $$\begin{aligned}
\mu^2_1 &\simeq&\left\{
\begin{array}{ll}
\displaystyle -\frac{u}{3}Z(m)\delta_1, \quad (\delta_1=1), & u\not=\infty \\
\\
\displaystyle - m\frac{2 u }{9\pi^3} \log \Lambda^2 , & u\rightarrow\infty
\end{array} \right. .\end{aligned}$$
### $d=4$
At $d=4$, $\mu_0^2$ is quadratically divergent as $$\begin{aligned}
\mu_0^2 &= & m^2-\frac{u}{6} Z(m), \qquad Z(m)\simeq \frac{1}{ 16\pi^2}\left[\Lambda^2- m^2\log \left(\frac{\Lambda^2+m^2}{m^2} \right)\right].\end{aligned}$$
On the other hand, at $d=4$, we have $$\begin{aligned}
B(Q^2)&=& \frac{1}{(4\pi)^2}\left[ \log\left(\frac{\Lambda_m^2}{m^2}\right)
+2\frac{Q^2+4\Lambda_m^2-2\Lambda^2}{\sqrt{Q^2(Q^2+4\Lambda_m^2)}}\tanh^{-1}
\sqrt{\frac{Q^2}{Q^2+4\Lambda_m^2}}\right. \nonumber \\
&& -\left. 2\sqrt{\frac{Q^2+4m^2}{Q^2}}\tanh^{-1}\sqrt{\frac{Q^2}{Q^2+4m^2}}\right] , \\
B(0) &=& \frac{1}{(4\pi)^2}\left[\log \frac{\Lambda_m^2}{m^2}-\frac{\Lambda^2}{\Lambda_m^2}\right],
\qquad \Lambda_m^2\equiv\Lambda^2+m^2, \end{aligned}$$ which diverge logarithmically, so that $u_r(s) =0$ as $\Lambda\rightarrow\infty$.
Since $ \tanh^{-1} (x) \simeq_{x\to 1} -\frac{1}{2}\log\left(\frac{1-x}{2}\right)$, we have $$\begin{aligned}
B(Q^2) +\frac{6}{u}&=&\hat B\left(q^2, \alpha^2\right), \quad q^2=\frac{Q^2}{\Lambda^2}, \quad \alpha =\frac{m}{\Lambda}, \\
\hat B(q^2, 0) &=& -c_0 \log q^2 + \frac{6}{u}+ c_0 F(q^2), \quad c_0 = \frac{1}{(4\pi)^2},\end{aligned}$$ where $$\begin{aligned}
F(q^2) &=& \frac{2(q^2+2)}{\sqrt{q^2(q^2+4)}}\tanh^{-1}\sqrt{\frac{q^2}{q^2+4}} .\end{aligned}$$
Let us now consider the continuum limit of $Z_1$. By rescaling the momentum, we have $$\begin{aligned}
Z_1 &= & -\frac{\alpha^2}{8\pi^2}\int_0^1 \frac{t {{\rm d}}t}{\hat B(t,\alpha^2) (t+\alpha^2)^3} .
\label{eq:Z1_4d}\end{aligned}$$ As $\alpha^2\rightarrow 0$ in the $\Lambda\rightarrow\infty$ limit, we have $$\begin{aligned}
\int_0^1 \frac{ t {{\rm d}}t}{\hat B(t,\alpha^2) (t+\alpha^2)^3} &\simeq& \int_0^1 \frac{ t {{\rm d}}t}{\hat B(t,0) (t+\alpha^2)^3}\nonumber \\
&=&\int_0^{\frac12} \frac{ t {{\rm d}}t}{\hat B(t,0) (t+\alpha^2)^3} + \int_{\frac12}^{1} \frac{ t {{\rm d}}t}{\hat B(t,0) (t+\alpha^2)^3},\end{aligned}$$ where the second term is finite in this limit, while the first term is bounded from above $$\begin{aligned}
\int_0^{\frac12} \frac{ t {{\rm d}}t}{\hat B(t,0) (t+\alpha^2)^3} &\le &
-\frac{1}{c_0} \int_0^{\frac12}\frac{t {{\rm d}}t}{(t+\alpha^2)^3 \log (t+\alpha^2)} \nonumber \\
&=&
\frac{1}{c_0} \left[ \log \vert \log \alpha^2 \vert +\sum_{r=1}^\infty \frac{ (-\log\alpha^2)^r}{r\, r!}
+\mbox{(finite terms)} \right],\end{aligned}$$ so that $Z_1$ in eq. (\[eq:Z1\_4d\]) vanishes as $\alpha^2\rightarrow 0$.
The most divergent part of $\mu_1^2$ becomes $$\begin{aligned}
\mu^2_1 &\simeq& -\frac{u}{3} \frac{\Lambda^2}{16\pi^2} \delta_1 , \qquad \delta_1 = \int_0^1 { {{\rm d}}q^2}\left(\frac{c_0 T(q^2)-6/u}{c_0\{\log q^2 -F(q^2) \} - 6/u}\right),\\
T(q^2)&\equiv& \log q^2 +1 -\frac{q^2}{q^2+4}\left( 1+\frac{q^2+6}{q^2+2} F(q^2) \right) , \nonumber\end{aligned}$$ where $\delta_1$ is finite, but is not universal as it depends on how we regulate the integral.
### $d > 4$
At $d> 4$, $\mu_0^2$ is $O(\Lambda^{d-2})$ as $$\begin{aligned}
\mu_0^2 &=& m^2-\frac{u}{6} Z(m), \qquad Z(m)\simeq
\frac{d}{ (4\pi)^{d/2}(d-2) \Gamma(1+d/2)}\Lambda^{d-2} .\end{aligned}$$
We also write $$\begin{aligned}
B(Q^2) &=& \frac{d}{(4\pi)^{d/2}\Gamma(1+d/2)}\int_0^1 {{\rm d}}x\,\int_0^\Lambda \frac{ p^{d-1} {{\rm d}}p}{\left[p^2+m^2 + Q^2 x(1-x)\right]^2} ,\end{aligned}$$ from which we obtain $$\begin{aligned}
B(Q^2)&=& \Lambda^{d-4} \hat B\left(\frac{Q^2}{\Lambda^2}, \frac{m^2}{\Lambda^2}\right), \qquad
\hat B(0,0) =\frac{d}{(d-4)} \frac{1}{(4\pi)^{d/2}\Gamma(1+d/2)},\\
\label{eq:BQ_d5}
\frac{d B(Q^2)}{d m^2} &=& -2 \Lambda^{d-6} \hat B_m\left(\frac{Q^2}{\Lambda^2}, \frac{m^2}{\Lambda^2}\right),\end{aligned}$$ where $$\begin{aligned}
\hat B(q^2,\alpha^2) &=& \frac{d}{(4\pi)^{d/2}\Gamma(1+d/2)}\int_0^1 {{\rm d}}x\,\int_0^1 \frac{ y^{d-1} {{\rm d}}y}{\left[y^2+\alpha^2 + q^2 x(1-x)\right]^2} , \\
\hat B_m(q^2,\alpha^2) &=& \frac{d}{(4\pi)^{d/2}\Gamma(1+d/2)}\int_0^1 {{\rm d}}x\,\int_0^1 \frac{ y^{d-1} {{\rm d}}y}{\left[y^2+\alpha^2 + q^2 x(1-x)\right]^3} \end{aligned}$$ so that $B(Q^2) =O(\Lambda^{d-4})$. As in the case at $d=4$, $u_r(s)=0$ in the limit that $\Lambda\rightarrow\infty$.
By the change of variable $Q=\Lambda q$ in eq. (\[eq:Z1\]) and then taking the limit $\Lambda\rightarrow \infty$, we obtain $$\begin{aligned}
Z_1 &= & \frac{2(4-d)}{d} \int_{q^2<1} \frac{{{\rm D}}q}{\hat B(q^2,0)}\frac{1}{(q^2)^2}.\end{aligned}$$ The fact that $\hat B(0,0)\not=0$ establishes that $Z_1$ is finite at $d > 4$.
The most divergent part of $\mu_1^2$ is given by $$\begin{aligned}
\mu_1^2 &\simeq& -\frac{u}{3} Z(m) \delta_1,\end{aligned}$$ where $$\begin{aligned}
\delta_1 &=& (d-2) \int_0^1 \frac{q^{d-1} {{\rm d}}q}{\hat B(q^2,0)} \left( \frac{\hat B(0,0)}{q^2} -\hat B_m(q^2,0) \right)\end{aligned}$$ with the change of variables as $q^2 = Q^2/\Lambda^2$. It is easy to show that $\delta_1$ is finite.
Solving the SED for the flow equation {#app:SDE_flow}
=====================================
In this appendix we explicitly solve the SDE in $d+1$ dimensions, in order to obtain the 2-pt and 4-pt functions for the flow fields at the NLO.
Solution for $\Gamma_0$
-----------------------
We first solve the equation at the LO for $\Gamma_0$. If we introduce one unknown function $F(t,p)$ as $$\begin{aligned}
\Gamma_0(12) &=& \int {{\rm D}}p \frac{F(t_1,p) F(t_2, p)}{p^2+m^2} {{\rm e}}^{-(p^2+\mu_f^2)(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)}\end{aligned}$$ with the initial condition $F(0,p) =1$, we have $$\begin{aligned}
D_1^f \, \Gamma_0(12) &=& \int {{\rm D}}p \frac{\dot F(t_1,p) F(t_2, p)}{p^2+m^2} {{\rm e}}^{-(p^2+\mu_f^2)(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)} \\
-\frac{u_f}{6}\Gamma_0(12)\Gamma_0(11) &=& -\frac{u_f}{6} \int {{\rm D}}p \frac{F(t_1,p) F(t_2, p)}{p^2+m^2} {{\rm e}}^{-(p^2+\mu_f^2)(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)}\Gamma_0(t_1), \\
\Gamma_0(t_1) &=& \Gamma_0(11)= \int {{\rm D}}p \frac{F^2(t_1,p)}{p^2+m^2} {{\rm e}}^{-2(p^2+\mu_f^2)t_1} ,\end{aligned}$$ where $\dot F$ means a $t$-derivative of $F$. Then, the SDE (\[eq:SDE\_LO\]) becomes $$\begin{aligned}
\frac{\dot F(t,p)}{F(t,p)} &=& -\frac{u_f}{6} \Gamma_0(t) ,\end{aligned}$$ which tells us that $F(t,p)$ is independent of $p$, so we put $F(t,p)=F(t)$. The above equation is thus reduced to $$\begin{aligned}
\dot F(t) &=& - \frac{u_f}{6}F^3(t) {{\rm e}}^{-2\mu_f^2 t} \zeta_0(t), \end{aligned}$$ where $\zeta_0(t)$ is defined in eq. (\[eq:zeta0\]), whose solution is given by $$\begin{aligned}
F^{-2}(t) = 1+ \frac{u_f}{3}\int_0^t {{\rm d}}s \zeta_0(s) {{\rm e}}^{-2\mu_f^2 s} \equiv {{\rm e}}^{-2\mu_f^2 t} \frac{\zeta(t)}{Z(m_f)}, \qquad \zeta(t) =\zeta_0(t) +\Delta(t)
\label{eq:zeta}\end{aligned}$$ where $m_f$ is defined in eq. (\[eq:Zm\]) and $$\begin{aligned}
\Delta(t) &=& {{\rm e}}^{2t \mu_f^2} \left( Z(m_f) - Z(m) \right) +\int {{\rm D}}p \left(\frac{p^2+m_f^2}{p^2+m^2}\right)\frac{{{\rm e}}^{2t\mu_f^2} -{{\rm e}}^{-2t p^2}}{p^2+\mu_f^2} .\end{aligned}$$
In the case of the interacting flow with $u_f > 0$, $\mu_f^2$ negatively diverges as $Z(m_f)\rightarrow+\infty $ in the continuum limit at $d > 1$ or as $u_f\rightarrow +\infty$ in the NLSM limit. In these limits, $\Delta(t)$ vanishes as $$\begin{aligned}
\lim_{\mu_f^2\rightarrow -\infty} \Delta(t) \simeq -\frac{m_f^2 \zeta_0(t) -\dot \zeta_0(t)/2}{\mu_f^2}+O\left(1/\mu_f^4\right) \end{aligned}$$ for $t > 0$. In the case of free flow ($u_f=0$), we simply have $F(t)=1$.
We then obtain $$\begin{aligned}
\Gamma_0(12) &=& \left\{
\begin{array}{ll}
\displaystyle \dfrac{Z(m_f)}{\sqrt{\zeta(t_1)\zeta(t_2)} } \int {{\rm D}}p \dfrac{{{\rm e}}^{-p^2(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)}} {p^2+m^2}, & u_f \not= 0 \\
\\
\displaystyle \int {{\rm D}}p \dfrac{{{\rm e}}^{-(p^2+\mu_f^2)(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)}} {p^2+m^2}, & u_f=0 \\
\end{array}
\right. .\end{aligned}$$
Solution for $K_0$
------------------
We consider $K_0$, which appears at the NLO. The equation for $K_0$ in eq. (\[eq:SDE\_K0\]) is closed, once $\Gamma_0$ is obtained. Using eq. (\[eq:K0g\]), we have $$\begin{aligned}
D_1^f K_0(12;34) &=& \int {{\rm d}}P_4 \left[ \frac{\dot F(t_1)}{F(t_1)} +\partial_{t_1}\right] g(12;34\vert 12;34)\nonumber \\
&=& \int {{\rm d}}P_4 \left[ -\frac{u_f}{6} F^2(t_1){{\rm e}}^{-2\mu_f^2 t_1}\zeta_0(t_1) +\partial_{t_1}\right] g(12;34\vert 12;34), \\
\Gamma_0(12) \Gamma_0(13) \Gamma_0(14) &=& \int {{\rm d}}P_4 \hat\delta\, (p_1^2+m^2) F^2(t_1) {{\rm e}}^{-2\mu_f^2 t_1} {{\rm e}}^{(p_1^2-p_2^2-p_3^2-p_4^2)t_1}, \\
\Gamma_0(11) K_0(12;34) &=& F^2(t_1) {{\rm e}}^{-2\mu_f^2 t_1} \zeta_0(t_1)\int {{\rm d}}P_4\, g(12;34\vert 12;34),\\
\Gamma_0(12) K_0(11;34) &=& F^2(t_1) {{\rm e}}^{-2\mu_f^2 t_1} \int {{\rm d}}P_4\hat\delta\, (p_1^2+m^2) {{\rm e}}^{t_1(p_1^2-p_2^2)} \nonumber \\
&\times& \int {{\rm D}}q_1{{\rm D}}q_2\, \frac{{{\rm e}}^{-t_1(q_1^2+q_2^2)}}{(q_1^2+m^2)(q_2^2+m^2)}\, g(11;34\vert q_1q_2;34), \end{aligned}$$ so that the SDE leads to $$\begin{aligned}
\partial_{t_1} g(12;34\vert 12;34) &=& -\frac{u_f}{6}F(t_1)^2 {{\rm e}}^{-2\mu_f^2 t_1} (p_1^2+m^2){{\rm e}}^{t_1(p_1^2-p_2^2)} \hat\delta \Bigl[ 2{{\rm e}}^{-t_1(p_3^2+p_4^2)}\nonumber \\
&+& \int {{\rm D}}q_1{{\rm D}}q_2\, \frac{{{\rm e}}^{-t_1(q_1^2+q_2^2)}}{(q_1^2+m^2)(q_2^2+m^2)}\, g(11;34\vert q_1q_2;34) \Bigr].
\label{eq:SDE_g}\end{aligned}$$ From eq. (\[eq:SDE\_g\]), one can easily see $\partial_{t_2}\partial_{t_1}g(12;34\vert 12;34) =0$, which implies $$\begin{aligned}
g(12;34\vert 12;34) &=&X(23\vert 12;34) + X(13\vert 21;34)+X(24\vert 12;43)+X(14\vert 21;43)\nonumber \\
&+& Y(2\vert 12;34)+ Y(1\vert 21;34)+ Y(3\vert 43;12)+ Y(4\vert 34;12) \nonumber \\
&+& Z(\vert 12;34),\end{aligned}$$ where we require that $X$ and $Y$ satisfy $$\begin{aligned}
X(\tau,\tau^\prime \vert 12;34) &=& X(\tau^\prime,\tau \vert 43;21), \qquad
X(\tau,0 \vert 12;34) =0, \\
Y(\tau\vert 12;34) &=& Y(\tau\vert 12;43), \qquad Y(0 \vert 12;34) = 0.
\label{eq:Y}\end{aligned}$$ Since $g(12;34\vert 12;34)$ agrees with the amputated connected 4-pt function in the $d$ dimensional theory at $ \tau_i = 0$ ($i=1,2,3,4$), we obtain $$\begin{aligned}
Z(\vert p_1,p_2,p_3,p_4) &=& -\hat\delta \frac{2}{6/u+ B(0 \vert p_{34})},\end{aligned}$$ where $B(t \vert Q)$ is defined in eq. (\[eq:Bt\]). Then one can easily check that $g$ satisfies the required symmetries $$\begin{aligned}
g(12;34\vert 12;34) &=& g(21;34\vert 21;34) =g(12;43\vert 12;43)=g(34;12\vert 34;12) .\end{aligned}$$
### Solution for $Y$
Terms which depend only on $t_1$ in eq. (\[eq:SDE\_g\]) can be written as $$\begin{aligned}
\partial_t Y(t\vert 21;34) &=& -\frac{u_f}{3} F^2(t) {{\rm e}}^{-2t\mu_f^2}(p_1^2+m^2) {{\rm e}}^{t(p_1^2-p_2^2)}
\hat\delta \nonumber \\
&\times& \left[ \rho(t\vert 34) +\int {{\rm D}}q_1{{\rm D}}q_2 \frac{{{\rm e}}^{-t(q_1^2+q_2^2)} Y(t\vert q_1,q_2;34)}{(q_1^2+m^2)(q_2^2+m^2)}\right],
\label{eq:SDE_Y}\end{aligned}$$ where $\rho(t\vert34)$ is defined in eq. (\[eq:rho\_t\]). To solve this equation, we set $$\begin{aligned}
Y(t\vert 21;34) &=& \hat\delta (p_1^2+m^2)\int_0^t {{\rm d}}s\, {{\rm e}}^{s(p_1^2-p_2^2)}\psi(s\vert 34),\end{aligned}$$ satisfying eq. (\[eq:Y\]). Eq. (\[eq:SDE\_Y\]) is reduced to $$\begin{aligned}
\psi(t\vert 34) &=& -\frac{u_f}{3} F^2(t){{\rm e}}^{-2t \mu_f^2} \left[\rho(t\vert 34) +\int_0^t {{\rm d}}s\, K(t,s\vert p_{34}) \psi(s\vert 34) \right] ,\end{aligned}$$ which shows $\psi$ does not depend on $p_1,p_2$, where $K$ is defined in eq. (\[eq:Kts\]). Since $u_f F^2(t) {{\rm e}}^{-2t\mu_f^2} = u_f Z(m_f)/\zeta(t)$ goes to infinity in the continuum limit at $t>0$ and $d>1$ or in the NLSM limit $u_f\rightarrow\infty$, eq. (\[eq:psi\]) must hold in either of the two limits.
### Solution for $X$
We next consider the solution for $X$. Terms depending on both $t_1$ and $t_3$ in eq. (\[eq:SDE\_g\]), and thereafter replacing $t_3$ by $t_2$ and interchanging $p_1\leftrightarrow p_2$, gives $$\begin{aligned}
\partial_{t_1} X(t_1,t_2\vert 12;34) &=& -\frac{u_f}{6}F^2(t_1) {{\rm e}}^{-2t_1\mu_f^2}(p_2^2+m^2) {{\rm e}}^{t_1(p_2^2-p_1^2)} \hat\delta \int {{\rm D}}q_1{{\rm D}}q_2
\nonumber \\
&\times& \frac{{{\rm e}}^{-t_1(q_1^2+q_2^2)}}{(q_1^2+m^2)(q_2^2+m^2)} \left\{ 2 X(t_1,t_2\vert q_1,q_2;34) +Y(t_2\vert 43;q_1,q_2) \right\},~~~\end{aligned}$$ where $$\begin{aligned}
Y(t\vert 43;q_1,q_2) &=& (2\pi)^d \delta(p_{34}+q_{12}) (p_3^2+m^2) \int_0^t {{\rm d}}s\, {{\rm e}}^{s(p_3^2-p_4^3)}\psi(s\vert q_1,q_2) .\end{aligned}$$ We define $$\begin{aligned}
\partial_{t_2}\partial_{t_1} X(t_1,t_2\vert 12;34) &=& \hat\delta (p_2^2+m^2) (p_3^2+m^2) {{\rm e}}^{t_1(p_2^2-p_1^2)} {{\rm e}}^{t_2(p_3^2-p_4^2)} \beta(t_1,t_2\vert 12;34),\end{aligned}$$ where properties of $X$ imply $\beta (t_1,t_2\vert 12;34) =\beta(t_2,t_1\vert 43;21)$ and $\beta(t,0\vert 12;34) =\beta(0,t\vert 12;34)=0$. Then the above equation becomes $$\begin{aligned}
\beta(t_1,t_2\vert 12;34) &=& -\frac{u_f}{6}F^2(t_1) {{\rm e}}^{-2t_1 \mu_f^2}\Bigl[g(t_1,t_2\vert p_{34}) + 2
\int_0^{t_1} {{\rm d}}s_1 \int {{\rm D}}q_1 {{\rm D}}q_2 (2\pi)^d\delta(q_{12}+p_{34})\nonumber \\
&\times& \frac{{{\rm e}}^{-(t_1+s_1)q_1^2 -(t_1-s_1)q_2^2}}{q_1^2+m^2}\beta(s_1,t_2\vert q_1,q_2;34)\Bigr],\end{aligned}$$ where $$\begin{aligned}
g(t_1,t_2 \vert Q) = \int {{\rm D}}q_1{{\rm D}}q_2(2\pi)^d\delta(q_{12}+Q)\frac{{{\rm e}}^{-t_1(q_1^2+q_2^2)}}{(q_1^2+m^2)(q_2^2+m^2)} \psi(t_2\vert q_1,q_2) .\end{aligned}$$ Since the above expression tells us that $\beta$ depends only on $p_{34}$, we can write $$\begin{aligned}
\beta(t_1,t_2\vert 12;34) &=& \omega(t_1,t_2\vert p_{34})=\omega(t_1,t_2\vert -p_{34}),\end{aligned}$$ so that we have $$\begin{aligned}
\omega(t_1,t_2\vert p_{34}) &=& -\frac{u_f}{6}F^2(t_1){{\rm e}}^{-2t_1 \mu_f^2}\Bigl[g(t_1,t_2\vert p_{34}) + 2\int_0^{t_1} {{\rm d}}s_1 K(t_1,s_1\vert p_{34}) \omega(s_1,t_2\vert p_{34})\Bigr], ~~~~~~~~\end{aligned}$$ which is reduced to $$\begin{aligned}
g(t_1,t_2\vert Q) + 2\int_0^{t_1} {{\rm d}}s_1 K(t_1,s_1\vert Q) \omega(s_1,t_2\vert Q)
= 0
\label{eq:g-omega}\end{aligned}$$ in the continuum limit or NLSM limit. Eq. (\[eq:g-omega\]) leads to eq. (\[eq:omega\]) in the main text, since $$\begin{aligned}
\int_0^{t_2} ds_2\, K(t_2,s_2\vert Q) g(t_1,s_2\vert Q) &=& - \rho(t_1,t_2\vert Q).\end{aligned}$$
Solution for $\Gamma_1$
-----------------------
### SDE at NLO
The SDE for $\Gamma_1$ is a little modified as $$\begin{aligned}
D_1^f \Gamma_1(12) +\mu_{1,f}^2 \Gamma_0(12)&=& -\frac{u_f}{6}\Bigl[ K_0(12;11)+\Gamma_0(12)\Gamma_1(11) +\Gamma_1(12)\Gamma_0(11) + 2\Gamma_0(12)\Gamma_0(11)\Bigr], \nonumber \\\end{aligned}$$ where we replace $\mu_f^2$ by $\mu_f^2+\dfrac{\mu_{1,f}^2}{N}$, so that $D_1^f \rightarrow D_1^f +\dfrac{1}{N} \mu_{1,f}^2 $. Here $u_{1,f}^2$ is given by eq. (\[eq:Z1\_mu1\]) with the replacement $u,m \rightarrow u_f, m_f$.
We parametrize $\Gamma_1$ as $$\begin{aligned}
\Gamma_1(12) &=& F(t_1)F(t_2)\int {{\rm D}}p\frac{{{\rm e}}^{-(p^2+\mu_f^2)(t_1+t_2)}{{\rm e}}^{ip(x_1-x_2)}}{p^2+m^2} G_1(t_1,t_2\vert p) \end{aligned}$$ with the boundary condition $$\begin{aligned}
G_1(0,0\vert p) &\equiv& b(p) = -\frac{\Sigma_1(p)}{p^2+m^2} , \end{aligned}$$ where $\Sigma_1(p)$ is the self-energy at the NLO in the $d$ dimensional theory.
The NLO SDE becomes $$\begin{aligned}
\partial_{t_1} G_1(t_1,t_2\vert p_1) + \mu_{1,f}^2&=& -\frac{u_f}{6}F^2(t_1){{\rm e}}^{-2t_1\mu_f^2}
{\cal H}\left[G_1(t_1,t_1|p)\right] + \lambda(t_1,t_2|p_1)\,,
\label{eq:G1}\end{aligned}$$ where ${\cal H}$ is defined in eq. (\[eq:def\_H\]) and $$\begin{aligned}
\lambda(t_1,t_2\vert p_1) &\equiv&
-\frac{u_f}{6}F^2(t_1) {{\rm e}}^{-2t_1\mu_f^2}\Delta(t_1,t_2\vert p_1), \\
\Delta(t_1,t_2\vert p_1) &=& 2 \zeta_0(t_1)
+{{\rm e}}^{t_1p_1^2}\int \prod_{i=2}^4
\frac{{{\rm D}}p_i {{\rm e}}^{-t_1 p_i^2}}{p_i^2+m^2}\Bigl\{ Z(\vert 21;34) +2Y(1\vert 34;21) \nonumber \\
&+&2X(11\vert 12;34) +Y(1\vert12;34)+2X(21\vert 21;34) + Y(2\vert 21;34)\Bigr\} .\end{aligned}$$
Using solutions $X$ and $Y$, we have in the continuum limit $$\begin{aligned}
\lambda(t_1,t_2\vert p_1)
&=& \int {{\rm d}}p_2\, \frac{{{\rm e}}^{t_1(p_1^2-p_2^2)}}{p_2^2+m^2}\Biggl[\psi(t_1\vert 12)+(p_2^2+m^2)\int_0^{t_1} {{\rm d}}s\, {{\rm e}}^{s(p_2^2-p_1^2)}\omega(t_1,s\vert p_{12}) \nonumber \\
&+&(p_1^2+m^2)\int_0^{t_2} {{\rm d}}s\, {{\rm e}}^{s(p_1^2-p_2^2)}\omega(t_1,s\vert p_{12})\Biggr] .
\label{eq:lambda}\end{aligned}$$ Since the right-hand side of eq. (\[eq:lambda\]) is finite, $\Delta(t_1,t_2\vert p)\rightarrow 0$ in the continuum limit.
### Solution to the SDE {#app:SDE_flow_G1}
Let us define $$\begin{aligned}
G_1(t_1,t_2\vert p) &\equiv & b(p) + \kappa(t_1,t_2\vert p) + H(t_1) + H(t_2) \end{aligned}$$ with $\kappa(t_1,t_2\vert p) = \kappa(t_2,t_1\vert p)$ and $\kappa(0,0\vert p)=H(0)= 0$, where $$\begin{aligned}
\partial_{t_1}\kappa(t_1,t_2\vert p) &=&\lambda(t_1,t_2\vert p),
\label{eq:kappa}\\
\frac{d H(t)}{d t} &=& -\frac{u_f}{6}F^2(t){{\rm e}}^{-2t\mu_f^2}\left[
{\cal H}\left[G_1(t,t\vert p)\right]
-2\zeta(t)\delta_1\right].
\label{eq:second}\end{aligned}$$ The second equation (\[eq:second\]) can be rewritten as $$\begin{aligned}
\frac{d H(t)}{d t} &=& -\frac{u_f}{6}F^2(t){{\rm e}}^{-2t\mu_f^2}\left[2\zeta_0(t) H(t) +b_0(t) +\kappa_0(t) -2\zeta(t) \delta_1\right],\end{aligned}$$ so that we have in the continuum limit $$\begin{aligned}
H(t) &=& -\frac{b_0(t)+\kappa_0(t)}{2\zeta_0(t)} +\delta_1,\end{aligned}$$ where we define $b_0(t) = {\cal H}[b(p)]$ and $ \kappa_0(t) ={\cal H}[\kappa(t,t\vert p)]$.
The first equation (\[eq:kappa\]) can be solved as $$\begin{aligned}
\kappa(t_1,t_2\vert p) &=& k_2(t_1,t_2\vert p) + k_1(t_1\vert p)+ k_1(t_2\vert p),\end{aligned}$$ where $$\begin{aligned}
k_1(t\vert p) &=& \int_0^t {{\rm d}}s\, \lambda_1(s\vert p),
\label{eq:k1}\\
\lambda_1(t\vert p) &=&
\int {{\rm D}}q \frac{{{\rm e}}^{(p^2-q^2)t}}{q^2+m^2}\psi(t\vert p,q) +
\int_0^t {{\rm d}}s \int {{\rm D}}q {{\rm e}}^{(p^2-q^2)(t-s)} \omega(t,s\vert Q),
\label{eq:lambda1} \\
k_2(t_1,t_2\vert p) &=& \int_0^{t_1} {{\rm d}}s_1\int_0^{t_2} {{\rm d}}s_2 \int {{\rm D}}q \frac{p^2+m^2}{q^2+m^2} {{\rm e}}^{(p^2-q^2)(s_1+s_2)}\omega(s_1,s_2\vert Q)
\label{eq:k2}\end{aligned}$$ with $Q=p+q$.
Calculations in the massless limit at $d=3$ {#app:massless}
===========================================
It can be shown that the flow bubble integral can be represented as $$\begin{aligned}
B(t\vert \{ p\}_{\rm sym.}) &=& -2\int_0^{ t} {{\rm d}}s\, K (s,0 \vert \{ p\}_{\rm sym.}) + B(0\vert \{ p\}_{\rm sym.}), \quad B(0\vert \{ p\}_{\rm sym.}) =\frac{1}{8\sqrt{D}},~~~~~~~\end{aligned}$$ which can be rescaled as $$\begin{aligned}
B(t\vert \{ p\}_{\rm sym.}) &=& \frac{1}{\sqrt{D}}b_0(Dt),\end{aligned}$$ where $$\begin{aligned}
b_0(w) &=&\frac{1}{8} -\frac{\sqrt{w}}{2(2\pi)^{3/2}}\int_0^1\frac{ {{\rm d}}x}{\sqrt{x}}{{\rm e}}^{-wx}\int_0^1\frac{ {{\rm d}}z}{\sqrt{z}}{{\rm e}}^{wzx/2}.
\label{eq:b0}\end{aligned}$$
Rescaling $$\begin{aligned}
\rho(t\vert \{ p\}_{\rm sym.}) &=& R_0(Dt, D),
\quad
\psi(t \vert \{ p\}_{\rm sym.}) = \sqrt{D} \phi_0(Dt,D),\end{aligned}$$ the integral equation for $\psi$ in the massless limit is written as $$\begin{aligned}
R_0(w, D) +\int_0^w {{\rm d}}v\, k_0(w,v) \phi_0(v,D) &=& 0 ,\end{aligned}$$ where $$\begin{aligned}
R_0(w,D) &=& {{\rm e}}^{-3w/2} -8b_0(w)\dfrac{\bar u(D)}{1+ \bar u(D)},
\qquad \bar u(D)= \dfrac{u}{48\sqrt{D}}.\end{aligned}$$ Since the problem is linear, we can write $$\begin{aligned}
\phi_0(w,D) &=& \phi_0^{(1)}(w) - 8 \phi_0^{(2)}(w) \frac{\bar u(D)}{1+\bar u(D)} ,\end{aligned}$$ where $\phi_0^{(i)}$, $i=1,2$ solve the momentum-independent equations (\[eq:phi0\_1\]) and (\[eq:phi0\_2\]). We thus finally obtain eq. (\[eq:hatY\]).
As the source term can be rescaled as $$\begin{aligned}
\rho(t,s\vert \{ p\}_{\rm sym.}) &=&\frac{1}{\sqrt{D}}\left[ b_0(D(t+s)) -8 b_0(Dt) b_0(Ds)\frac{\bar u(D)}{1+\bar u(D)} \right], \end{aligned}$$ the equation for $\omega$ in the massless limit is written for $\omega(t,s\vert \{ p\}_{\rm sym.})= \sqrt{D}W_0(Dt,Ds,D)$ as $$\begin{aligned}
&&b_0(D(t+s)) -8 b_0(Dt) b_0(Ds)\frac{\bar u(D)}{1+\bar u(D)} \nonumber \\&=&
2\int_0^{Dt} {{\rm d}}u\, k_0(Dt,u) \int_0^{Ds} {{\rm d}}v\, k_0(Ds,v) W_0( u,v,D),\end{aligned}$$ which can be solved as $$\begin{aligned}
W_0(w,v , D) &=& \Omega_0(w,v) - 4 \phi_0^{(2)}(w) \phi_0^{(2)}(v) \frac{\bar u(D)} {1+\bar u(D)},\end{aligned}$$ where $\Omega_0$ solves the momentum ($D$) independent equation (\[eq:Omega0\]). We thus obtain eq. (\[eq:hatX\]).
Induced metric in the massless limit at $d=3$ {#app:metric}
=============================================
Induced metric
--------------
The space component of the induced metric is given by $$\begin{aligned}
g_{ij}(z) &=& \delta_{ij}\frac{R_0^2}{d\zeta_0(t)} \left(1-\frac{\zeta_1(t)}{N}\right) {\cal H}\left[ p^2\left(1 +\frac{G_1(t,t\vert p)}{N}\right)\right] .\end{aligned}$$ We then evaluate $$\begin{aligned}
\zeta_1(t) &=& \frac{1}{\zeta_0(t)} {\cal H}[ G_1(t,t\vert p)] = 2 \delta_1, \quad
{\cal H}[1] = \zeta_0(t), \qquad {\cal H}[p^2] = -\frac{\partial_t\zeta_0(t)}{2}, \\
{\cal H}[ p^2 G_1(t,t\vert p)] &=& {\cal H}[\lambda(t,t\vert p)] +\zeta_0(t) \partial_t H(t) -\partial_t\zeta_0(t)\delta_1
=\zeta_0(t) \partial_t H(t) -\partial_t\zeta_0(t)\delta_1, ~~~~~\end{aligned}$$ where in the last equation we use $ {\cal H}[\lambda(t,t\vert p)] = 0$. Altogether we obtain $$\begin{aligned}
g_{ij}(z) &=& \delta_{ij} R_0^2\left[ g^{(0)}(t) + \frac{1}{N} g^{(1)}(t) \right], \quad
g^{(0)}(t) =-\frac{\partial_t\zeta_0(t)}{2d\zeta_0(t)}, \ g^{(1)}(t) =\frac{\partial_t H(t)}{d}.~~\end{aligned}$$
The time component is evaluated as $$\begin{aligned}
g_{00}(t)&=&t\partial_{t_1}\partial_{t_2}\left[
\frac{ R_0^2}{\sqrt{\zeta_0(t_1)\zeta_0(t_2)}}\int {{\rm D}}p\frac{{\mathrm e}^{-p^2(t_1+t_2)}}{p^2+m^2}
\left(1+\frac{\widetilde G_1(t_1,t_2|p)}{N}\right)\right]_{t_1=t_2=t}\\
&=& R_0^2\left\{ g_{00}^{(0)}(t) +\frac{1}{N}g_{00}^{(1)}(t)\right\},\end{aligned}$$ where $$\widetilde G_1(t_1,t_2|p)=-2\delta_1+G_1(t_1,t_2|p).$$ The leading term is $$g_{00}^{(0)}(t)=\frac{t}{4}\partial_t^2\left[\log\zeta_0(t)\right]$$ and for the NLO term we have $$\frac{1}{t}\,g^{(1)}_{00}(t)
=\partial_{t_1}\partial_{t_2}\frac{I(t_1,t_2)}{\sqrt{\zeta_0(t_1)\zeta_0(t_2)}}
\Bigg\vert_{t_1=t_2=t},$$ where $$I(t_1,t_2)=
\int {{\rm D}}p\,\frac{{\mathrm e}^{-p^2(t_1+t_2)}}{p^2+m^2}\widetilde G_1(t_1,t_2|p).$$ With this notation $$\frac{1}{t}\,g^{(1)}_{00}(t)=
\frac{1}{4}\,\frac{(\partial_t\zeta_0(t))^2}{\zeta_0^3(t)}\,I(t,t)
-\frac{1}{2}\,\frac{\partial_t\zeta_0(t)}{\zeta_0^2(t)}\,\partial_t I(t,t)
+\frac{1}{\zeta_0(t)}\partial_{t_1}\partial_{t_2} I(t_1,t_2)
\big\vert_{t_1=t_2=t}.$$ Since $$I(t,t)={\cal H}[\widetilde G_1(t,t|p)]=0,$$ the first two terms vanish. Further, $$\partial_{t_1}\partial_{t_2} I(t_1,t_2)\big\vert_{t_1=t_2=t}=
{\cal H}\big[(p^2)^2 \widetilde G_1(t,t|p)-2p^2\lambda(t,t|p)
+\partial_{t_2}\lambda(t,t_2|p)\big\vert_{t_2=t}\big]
+\partial_t H(t)\partial_t\zeta_0(t).$$ Using the identities $${\cal H}[\lambda(t,t|p)]=0;\qquad
{\cal H}[p^2\widetilde G_1(t,t|p)]=\zeta_0(t)\,\partial_t H(t)$$ and their derivatives this can be further simplified: $$\partial_{t_1}\partial_{t_2} I(t_1,t_2)\big\vert_{t_1=t_2=t}=
-\frac{1}{2}\,\zeta_0(t)\partial_t^2 H(t)+
{\cal H}\big[\partial_{t_2}\lambda(t,t_2|p)\big\vert_{t_2=t}
-\,\partial_t\lambda(t,t|p)/2].$$ Here the second term vanishes and we finally obtain $$g^{(1)}_{00}(t)=-\frac{t}{2}\,\partial_t^2 H(t).$$
Calculation of $H(t)$ in the massless limit
-------------------------------------------
We recall the definition of $H(t)$ as $$\begin{aligned}
H(t) &=& -\frac{b_0(t) +\kappa_0(t)}{2\zeta_0(t)} +\delta_1\end{aligned}$$ where $$\begin{aligned}
b_0(t)&=&{\cal H}[b(p)], \qquad \kappa_0(t) ={\cal H}[\kappa(t,t\vert p)],\end{aligned}$$ with $$\begin{aligned}
b(p) &=& -\frac{\Sigma_1(p)}{p^2+m^2}, \quad
\kappa(t,t\vert p) = k_2(t,t \vert p) + 2 k_1(t\vert p) .\end{aligned}$$ Here $k_1$ and $k_2$ are given in eqs. (\[eq:k1\]), (\[eq:lambda1\]) and (\[eq:k2\]).
Hereafter we consider the massless limit at $d=3$, where we have $\zeta_0(t)^{-1} = 2(2\pi)^{3/2} \sqrt{t}$.
### Calculation of $b_0(t)$
We first calculate $b_0(t)$. In the massless limit, we have $$\begin{aligned}
H_b(t) &\equiv& -\frac{b_0(t)}{2\zeta_0(t)} =\frac{1}{2\zeta_0(t)}\int {{\rm D}}p \frac{{{\rm e}}^{-2p^2 t}} {(p^2)^2} g(p^2) \end{aligned}$$ since $\tilde C = Z_1 m^2 =0$ and $$\begin{aligned}
g(p^2) &=&\frac{u}{3} \int \frac{{{\rm D}}Q}{1+\bar u(Q^2)}\left\{\frac{1}{(Q+p)^2} -\frac{1}{Q^2} \right\}.\end{aligned}$$ After rescaling, we obtain $$\begin{aligned}
H_b(t) &=& \int {{\rm D}}Q\, h_b(Q^2) \frac{\bar u(Q^2) \sqrt{t}}{1+\bar u(Q^2) \sqrt{t}},\end{aligned}$$ where $$\begin{aligned}
h_b(Q^2) &=& 32\sqrt{2} \sqrt{\pi^3}\sqrt{Q^2} \int {{\rm D}}p\, \frac{{{\rm e}}^{-2p^2}}{(p^2)^2} \left\{ \frac{1}{(Q+p)^2}-\frac{1}{Q^2}\right\} .\end{aligned}$$
### Calculation of $\kappa_0(t)$
For this we need $\psi$ and $\omega$ in the massless limit, which can be obtained as $$\begin{aligned}
\psi_0(t\vert p,q) &=& \sqrt{Q^2}\left[ \varphi_0(Q^2t, z) - 8\phi_0^{(2)}(Q^2 t) \frac{\bar u(Q^2)}{1+\bar u(Q^2)} \right], \\
\omega_0(t,s \vert Q) &=& \sqrt{Q^2}\left[ \Omega_0(Q^2t,Q^2s) - 4\phi_0^{(2)}( Q^2t) \phi_0^{(2)}( Q^2s)
\frac{ \bar u(Q^2)}{1+\bar u(Q^2)} \right] \end{aligned}$$ with $z=(p^2+q^2)/Q^2$, where $\phi_0^{(2)}$ and $\Omega_0$ are already obtained in section \[sec:coupling\], while $\varphi_0$ satisfies $$\begin{aligned}
{{\rm e}}^{-z w} +\int_0^{w} {{\rm d}}x\, k_0(w,x) \varphi_0(x,z) &=& 0,\end{aligned}$$ instead of eq. (\[eq:phi0\_1\]) and thus $\varphi_0(x,3/2) =\phi_0^{(1)}(x) $.
Using these, we first calculate $$\begin{aligned}
H_{\kappa}^{(1)}(t) &\equiv & -\frac{1}{\zeta_0(t)} \int {{\rm D}}p \frac{{{\rm e}}^{-2p^2t}}{p^2} \int_0^t {{\rm d}}s\, \int {{\rm D}}q
\frac{{{\rm e}}^{(p^2-q^2)s}}{q^2} \psi_0(s\vert p,q) \nonumber \\
&=& H_\kappa^{(1)}(0) +\int {{\rm D}}Q\, \int_0^1 {{\rm d}}x\, \phi_0^{(2)}(Q^2 x) h_{11}(x,Q^2) \frac{ \bar u(Q^2)\sqrt{t}}{1+\bar u(Q^2) \sqrt{t}},\end{aligned}$$ where $H_\kappa^{(1)}(0)$ is some constant and $$\begin{aligned}
h_{11}(x,Q^2) &=& 32\sqrt{2}\sqrt{\pi^3}\sqrt{Q^2}\int{{\rm D}}p\, {{\rm D}}q\, (2\pi)^3\delta(q+p-Q) \frac{{{\rm e}}^{-(2-x)p^2-xq^2}}{p^2 q^2} .\end{aligned}$$
Similarly we have $$\begin{aligned}
H_{\kappa}^{(2)}(t) &\equiv & -\frac{1}{\zeta_0(t)} \int {{\rm D}}p \frac{{{\rm e}}^{-2p^2t}}{p^2} \int_0^t {{\rm d}}s\, \int {{\rm D}}q\,
{{\rm e}}^{(p^2-q^2)s} \int_0^s {{\rm d}}r\, {{\rm e}}^{(q^2-p^2)r}\omega_0(s,r\vert Q) \nonumber \\
&=& H_{\kappa}^{(2)}(0) + 2 \int {{\rm D}}Q\, \int_0^1 {{\rm d}}x\, \phi_0^{(2)}(Q^2x) \int_0^x {{\rm d}}y\, \phi_0^{(2)}(Q^2y) \,
h_{10}(x-y,Q^2) \nonumber \\
&\times & \frac{\bar u(Q^2)\sqrt{t}}{1+ \bar u(Q^2)\sqrt{t}},\end{aligned}$$ where $$\begin{aligned}
h_{10}(z,Q^2) &=& 8\sqrt{2}\sqrt{\pi^3}\sqrt{Q^2}\int {{\rm D}}p\, {{\rm D}}q\, (2\pi)^{3}\delta(q+p-Q) \frac{{{\rm e}}^{-(2-z)p^2-z q^2}}{p^2} .~~~~\end{aligned}$$
The last contribution becomes $$\begin{aligned}
H_{\kappa}^{(3)}(t) &\equiv & -\frac{1}{2\zeta_0(t)} \int {{\rm D}}p\, {{\rm e}}^{-2p^2t} \int_0^t {{\rm d}}s\, \int {{\rm D}}q\,
\frac{{{\rm e}}^{(p^2-q^2)s}}{q^2} \int_0^t {{\rm d}}r\, {{\rm e}}^{(p^2-q^2)r}\omega_0(s,r\vert Q)\nonumber \\
&=& H_{\kappa}^{(3)}(0) + \int {{\rm D}}Q\, \int_0^1 {{\rm d}}x\, \phi_0^{(2)}(Q^2x) \int_0^1 {{\rm d}}y\, \phi_0^{(2)}(Q^2y) \, h_{10}(2-x-y,Q^2)
\nonumber \\
&\times & \frac{\bar u(Q^2)\sqrt{t}}{1+ \bar u(Q^2)\sqrt{t}}.\end{aligned}$$
Total contributions
-------------------
We thus obtain the $H(t)$ as[^7] $$\begin{aligned}
H(t) &=& H(0) +\int {{\rm D}}Q\ h_{\rm total}(Q^2) \frac{\bar u(Q^2)\sqrt{t}}{1+ \bar u(Q^2)\sqrt{t}},\end{aligned}$$ where $$\begin{aligned}
H(0) &=& H_\kappa^{(1)}(0) +H_\kappa^{(2)}(0) + H_\kappa^{(3)}(0)+\delta_1 \\
h_{\rm total}(Q^2) &=& h_b(Q^2) + \int_0^1 {{\rm d}}x\, \phi_0^{(2)}(Q^2x) \left\{h_{11}(x,Q^2)
+2\int_0^x {{\rm d}}y\, \phi_0^{(2)}(Q^2y) h_{10}(x-y,Q^2) \right. \nonumber \\
&+&\left.
\int_0^1 {{\rm d}}y\, \phi_0^{(2)}(Q^2y)h_{10}(2-x-y,Q^2) \right\} ,\end{aligned}$$ which leads to eqs. (\[eq:A1\]) and (\[eq:pA1\]) by $A_1(t) \equiv \partial_t H(t)$ and $ \partial_t A_1(t) \equiv \partial^2_t H(t)$.
IR behaviors
------------
### Some definitions
We write the NLO induced metric as $$\begin{aligned}
g_{ij}(\tau) &=&\delta_{ij} \left\{ \frac{R_0^2}{12 t} \left[ 1+\frac{R(t)}{N} \right]\right\}, \qquad
g_{00}(\tau) = -t \partial_t \left\{ \frac{R_0^2}{8 t} \left[ 1+\frac{R(t)}{N} \right]\right\},\end{aligned}$$ where the relative correction is a sum of four contributions, $$R(t)=R_b(t)+\sum_{i=1}^3R^{(i)}_\kappa(t),\quad R_b(t) \equiv 4t\partial_t H_b(t), \ R_\kappa^{(i)}(t) \equiv 4t\partial_t H_\kappa^{(i)}(t).$$ We also introduce $G(v)$ by $$\phi^{(2)}_{ 0}(v)=-\frac{(2\pi)^{3/2}}{\sqrt{v}}G(v),\qquad G(0)=1/8, \quad G(v)\sim \exp(-v/2),\ v\rightarrow\infty$$ and use the time variable $T = u\sqrt{t}/48$.
In the following we will use the fact that a double 3-dimensional integral of any function depending only on the absolute values $p$, $q$ and $|Q|$, where $Q=p+q$, can be written $$\int {{\rm D}}p\int {{\rm D}}q \, f(p,q,Q^2)=\frac{1}{(2\pi)^4}\int_0^\infty p {{\rm d}}p\, \int_0^\infty
q{{\rm d}}q\, \int_{(q-p)^2}^{(q+p)^2}{{\rm d}}Q^2\, f(p,q,Q^2) .$$
### The $R_b$ contribution
Here we can do the angular part of the $Q^2$ integral analytically and find $$R_b(t)=\frac{32T}{\sqrt{2\pi}^5}\int_0^\infty\frac{q{{\rm d}}q}{(q+T)^2}\, \rho_b(q),$$ where $$\rho_b(q)=q^2\int_0^\infty\frac{{{\rm d}}p}{p^3}\, {{\rm e}}^{-2p^2}\left\{
\ln\frac{(p+q)^2}{(p-q)^2}-\frac{4p}{q}\right\},$$ which behaves as $\rho_b(q)= O(q) $ for small $q$, while $$\rho_b(q)\sim \frac{\sqrt{2\pi}}{3q},$$ for large $q$. Thus we can establish that $R_b(t)=O(T)$ for small $t$, while for large $t$ $$r_b\equiv R_b(\infty)=\frac{8}{3\pi^2}=0.27019.$$
### The $R_\kappa^{(1)}$ contribution
We have $$R^{(1)}_\kappa(t)=-32(2\pi)^3\int {{\rm D}}p\int {{\rm D}}q\int_0^1\frac{{{\rm d}}x}{\sqrt{x}}
\frac{{{\rm e}}^{-p^2(2-x)-q^2x}}{p^2q^2}\frac{|Q|T}{(T+|Q|)^2}G(Q^2x).$$ Doing the $q^2$ integral first and introducing $x=y^2$ we can rewrite it as $$-\frac{32}{\pi}\int_0^\infty\frac{{{\rm d}}p}{p}\, {{\rm e}}^{-2p^2}\int_0^\infty
\frac{Q^2T}{(T+Q)^2}{{\rm d}}Q\int_0^1{{\rm d}}y\, G(Q^2y^2)
\int_{(Q-p)^2}^{(Q+p)^2}\frac{{{\rm e}}^{(p^2-q^2)y^2}}{q^2}{{\rm d}}q^2.$$ After some further rescaling we get $$R^{(1)}_\kappa(t)=-\frac{64}{\pi}\int_0^\infty{{\rm d}}Q\, \frac{QT}{(T+Q)^2}\,
\rho^{(1)}_\kappa(Q),$$ where $$\rho^{(1)}_\kappa(Q)=\int_0^\infty\frac{{{\rm d}}p}{p}{{\rm e}}^{-2p^2}
\int_0^Q{{\rm d}}z\, G(z^2)Y(\frac{p}{Q},z),$$ $$Y(\varepsilon,z)=\int_{|1-\varepsilon|}^{1+\varepsilon}\frac{{{\rm d}}\xi}{\xi}\,
{{\rm e}}^{(\varepsilon^2-\xi^2)z^2}=2\varepsilon{{\rm e}}^{-z^2}+O(\varepsilon^2).$$ From this we see that $\rho_\kappa^{(1)}(Q) = O(Q)$ for small $Q$, while $$\rho^{(1)}_\kappa(Q)\sim\frac{2}{Q}
\int_0^\infty{{\rm d}}p\, {{\rm e}}^{-2p^2}\int_0^\infty
{{\rm d}}z\, G(z^2)
{{\rm e}}^{-z^2}=\frac{1}{Q}\sqrt{\frac{\pi}{2}}\int_0^\infty{{\rm d}}z\,
G(z^2){{\rm e}}^{-z^2}$$ for large $Q$, so that we numerically obtain $$r^{(1)}_\kappa \equiv R^{(1)}_\kappa(\infty)=-\frac{64}{\sqrt{2\pi}}\int_0^\infty{{\rm d}}z\, G(z^2)
{{\rm e}}^{-z^2}=-1.14734.$$
### The $R_\kappa^{(2)}$ contribution
Similarly $$R^{(2)}_\kappa(t)=16\sqrt{2\pi}^9\int {{\rm D}}p\int {{\rm D}}q\int_0^1
\frac{{{\rm d}}x}{\sqrt{x}}\int_0^x\frac{{{\rm d}}y}{\sqrt{y}}
\frac{T}{(T+|Q|)^2}G(Q^2x)G(Q^2y)
\frac{{{\rm e}}^{-2p^2}}{p^2}
{{\rm e}}^{(p^2-q^2)(x-y)}.$$ Doing the $q^2$ integrations first, we have $$\begin{split}
R^{(2)}_\kappa(t)=64\sqrt{2\pi}&\int_0^\infty{{\rm d}}Q \frac{QT}{(T+Q)^2}
\int_0^\infty\frac{{{\rm d}}p}{p}{{\rm e}}^{-2p^2}\\
&\times \int_0^1{{\rm d}}x\int_0^x{{\rm d}}y \, G(Q^2x^2) \, G(Q^2y^2)
\int_{(Q-p)^2}^{(Q+p)^2} {{\rm e}}^{(p^2-q^2)(x^2-y^2)}{{\rm d}}q^2.
\end{split}$$ The $q^2$ integral can be done analytically and we find $$R^{(2)}_\kappa(t)=128\sqrt{2\pi}\int_0^\infty{{\rm d}}Q \frac{QT}{(T+Q)^2}
\rho^{(2)}_\kappa(Q),$$ where $$\rho^{(2)}_\kappa(Q)=
\int_0^\infty\frac{{{\rm d}}p}{p}{{\rm e}}^{-2p^2}
\int_0^Q{{\rm d}}z\int_0^z{{\rm d}}w \, G(z^2) \, G(w^2)
\frac{{{\rm e}}^{w^2-z^2}}{z^2-w^2}\sinh\frac{2p}{Q}(z^2-w^2).$$ Thus $\rho_\kappa^{(2)} = O(Q)$ for small $Q$, while $$\rho^{(2)}_\kappa(Q)\sim\frac{1}{Q}
\int_{-\infty}^\infty{{\rm d}}p\, {{\rm e}}^{-2p^2}
\int_0^\infty{{\rm d}}z\int_0^z{{\rm d}}w \, G(z^2) \, G(w^2)
{{\rm e}}^{w^2-z^2}$$ for large $Q$, and $$r^{(2)}_\kappa \equiv R_\kappa^{(2)}(\infty) =
128\pi
\int_0^\infty{{\rm d}}z\int_0^z{{\rm d}}w \, G(z^2)\, G(w^2)
{{\rm e}}^{w^2-z^2}=0.45846.$$
### The $R_\kappa^{(3)}$ contribution
For $R_\kappa^{(3)}$ we find $$R_\kappa^{(3)}(t)=32\sqrt{2\pi}\int_0^\infty{{\rm d}}Q\frac{QT}{(T+Q)^2}
\rho^{(3)}_\kappa(Q)$$ with $$\rho^{(3)}_\kappa(Q)=\int_0^1{{\rm d}}x \int_0^1{{\rm d}}y
\int_0^\infty p{{\rm d}}p\, {{\rm e}}^{-2p^2+p^2(x^2+y^2)}G(Q^2x^2)G(Q^2y^2)
\int_{(Q-p)^2}^{(Q+p)^2}\frac{{{\rm e}}^{-q^2(x^2+y^2)}}{q^2} {{\rm d}}q^2.$$ After rescaling $$\rho^{(3)}_\kappa(Q)=\frac{1}{Q^2}\int_0^Q{{\rm d}}z \int_0^Q{{\rm d}}w\,
G(z^2)\, G(w^2)\int_0^\infty p{{\rm d}}p\, {{\rm e}}^{-2p^2}
Z\left(\frac{p}{Q},z^2+w^2\right),$$ where $$Z(\varepsilon,A)=2{{\rm e}}^{A\varepsilon^2}\int_{|1-\varepsilon|}^{1+\varepsilon}
\frac{{{\rm e}}^{-A\xi^2}}{\xi}{{\rm d}}\xi
\approx 4\varepsilon{{\rm e}}^{-A}, \quad \varepsilon\rightarrow 0.$$ Thus $\rho_\kappa^{(3)}(Q) = O(Q)$ for small $Q$, while $$\begin{split}
\rho^{(3)}(Q)\sim\frac{4}{Q^3}\int_0^\infty{{\rm d}}z &\int_0^\infty{{\rm d}}w\,
G(z^2)\, G(w^2)\int_0^\infty p^2{{\rm d}}p\, {{\rm e}}^{-2p^2-z^2-w^2}\\
&=\sqrt{\frac{\pi}{8}}\left(\int_0^\infty{{\rm d}}z\, G(z^2){{\rm e}}^{-z^2}\right)^2
\frac{1}{Q^3},
\end{split}$$ for large $Q$, which leads to $$r^{(3)}_\kappa \equiv R^{(3)}_\kappa(\infty)=0 .$$
Thus the total relative correction is negative: $$r=r_b+r^{(1)}_\kappa+r^{(2)}_\kappa+r^{(3)}_\kappa=
-0.41869.$$
[^1]: We will call the infinite cutoff ($\Lambda\to\infty$) limit the ’continuum limit’.
[^2]: We will indicate the massless limit by a subscript $_0$.
[^3]: It turns out that $G_2(\Delta)$ has only one zero at $\Delta=0.36228(1)$.
[^4]: This is independent of $u_f\not= 0$ (the interacting flow). In the case of free flow ($u_f=0$), however, $r=\dfrac{8}{3\pi^2}\simeq0.27019$.
[^5]: It is interesting and also suggestive to see that the F-coefficient of the 3 dimensional $O(N)$ scalar model is given by $F_{\rm IR} = F_{UV} - \zeta(3)/(8\pi^2) + O(1/N)$, where $F_{\rm UV} = N F_S$ with $F_S\simeq 0.0638$ as an example of a conjecture, the so-called “the F-theorem”, which claims that the F-coefficient monotonically decreases along a RG trajectory connecting two 3 dimensional CFTs. Furthermore, in the holographic dual picture, the F-coefficient is proportional to the AdS radius squared. (See Ref. [@Pufu:2016zxm] and references therein.)
[^6]: Note that we use the same notation $\Gamma_{2n}$ for the 2$n$-point functions in both $d$ and $d+1$ dimensions, since no confusion may occur.
[^7]: Here $H(0)$ is potentially divergent but it does not contribute to the metric.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We revisit the minimal supersymmetric left–right model with $B-L=2$ triplet Higgs fields and show that a self–consistent picture emerges with automatic $R$–parity conservation even in the absence of higher dimensional operators. By computing the effective potential for the Higgs system including heavy Majorana neutrino Yukawa couplings we show that the global minimum of the model can lie in the charge and $R$–parity conserving domain. The model provides natural solutions to the SUSY phase problem and the strong CP problem and makes several interesting predictions. Quark mixing angles arise only after radiative corrections from the lepton sector are taken into account. A pair of doubly charged Higgs fields remain light below TeV with one field acquiring its mass entirely via renormalization group corrections. We find this mass to be not much above the Bino mass. In the supergravity framework for SUSY breaking, we also find similar upper limits on the stau masses. Natural solutions to the $\mu$ problem and the SUSY CP problem entails light $SU(2)_L$ triplet Higgs fields, leading to rich collider phenomenology.'
---
OSU–HEP–08-05\
UMD-PP-08-011
[**Minimal Supersymmetric Left–Right Model\
with Automatic $R$–Parity**]{}\
0.5cm [*${}^1\,$Department of Physics, Oklahoma State University,\
Stillwater, OK 74078, USA\
\[0.1truecm\] ${}^2\,$Maryland Center for Fundamental Physics and Department of Physics,\
University of Maryland, College Park, MD 20742, USA\
*]{}
.5cm
\#1 \#1
Introduction
============
Left-right symmetric extensions of the Standard Model (SM) based on the gauge group [@patimohap] $SU(3)_C\times SU(2)_L\times
SU(2)_R\times U(1)_{B-L}$ have many attractive features. These include an understanding of the origin of parity violation, and a compelling rationale for small neutrino masses via the seesaw mechanism. The enlarged gauge symmetry allows for parity to be defined as an exact symmetry, which is broken only spontaneously. Right–handed neutrino is required to exist in order to complete the $SU(2)_R$ multiplet, and so neutrino mass is natural. In the domain of flavor physics, the supersymmetric version of this theory resolves several problems of the popular minimal supersymmetric standard model (MSSM): [**(i)**]{} $R$–parity emerges as an exact symmetry of MSSM, preventing rapid proton decay and providing a naturally stable dark matter candidate [@rnm]. This is possible if the $SU(2)_R\times U(1)_{B-L}$ gauge symmetry is broken down to $U(1)_Y$ by Higgs triplet fields carrying $B-L=\pm 2$. $R$–parity, which is part of the original $B-L$ symmetry, will remain unbroken even after symmetry breaking in this case. [**(ii)**]{} It solves the SUSY CP problem [@rasin; @babu] because of parity invariance. Parity makes the Yukawa couplings and the corresponding SUSY breaking $A$ terms hermitian, and the gluino mass and the $\mu$ term real. The electric dipole moments of fermions will then vanish at the scale of parity restoration. [**(iii)**]{} Finally, it has all the ingredients necessary to solve the strong CP problem without the need for an axion, again by virtue of parity symmetry [@rasin; @babu]. This is achieved by ensuring that the quark mass matrix has a real determinant, which is possible since the Yukawa couplings are hermitian.
Previous studies of this model focussed on two versions: (i) A TeV scale version where $R$–parity is spontaneously broken by the vacuum expectation value (VEV) of the right–handed sneutrino [@kuchi], or alternatively (ii) an $R$–parity conserving version [@goran; @chacko] where non-renormalizable (NR) higher dimensional operators were included and played an essential role. The reason for considering only these two versions is that in the absence of the above features, i.e., $\left\langle \tilde{\nu}^c
\right\rangle \neq 0$, or the presence of NR operators, the global minimum of the theory that is both $R$–parity conserving and parity violating, breaks electric charge and is therefore unacceptable. In the first version with $\left\langle
\tilde{\nu}^c \right\rangle \neq 0$, the $W_R$ scale must necessarily be in the TeV range [@kuchi], whereas in the second one, it is necessarily above $10^{11}$ GeV. In the first version, SUSY dark matter candidate is lost. In the second version, the possibility of solving strong CP problem via parity symmetry is eliminated due to the essential presence of higher dimensional operators which makes $\overline{\theta}$ large. It is also difficult to solve the SUSY phase problem, since these higher dimensional operators typically generate parity violating effects in the fermion mass matrices. Extensions of the minimal model which use additional Higgs multiplets have been proposed. Ref. [@bdm] introduces Higgs doublets in addition to triplets, but in such models $R$–parity conservation is exterior to parity symmetry. In Ref. [@benakli] $B-L=0$ Higgs triplets are introduced in addition to the $B-L = \pm 2$ triplets, which is clearly non–minimal.
In this note we revisit the minimal SUSYLR model with $B-L=2$ Higgs triplets. We assume that the higher dimensional operators are absent or small, so that the solutions to the strong CP and the SUSY phase problems are still intact. The global minimum of the tree–level Higgs potential is either charge violating, or $R$–parity violating, as noted. However, we find that inclusion of the heavy Majorana neutrino Yukawa couplings in the effective potential automatically cures this problem. The vacuum that preserves both electric charge and $R$–parity can naturally be the global minimum of the full potential. We study the consequences of such a setup.[^1]
The main results of our investigation can be summarized as follows: [**(i)**]{} In this general class of models, there are two doubly charged Higgs and Higgsino fields with masses below a TeV. One combination of these doubly charged Higgs boson fields has a vanishing mass at the scale of $SU(2)_R \times U(1)_{B-L}$ breaking (denoted as $v_R$). So its mass is calculable, arising through renormalization group effects between $v_R$ and the weak scale. We find its squared mass to be positive with the Higgs boson having a mass close to the Bino mass. [**(ii)**]{} There exist two pairs of Higgs doublets in the low energy, although one pair is unlikely to be observed directly at the LHC. This naturally leads to calculable flavor violation, which are within experimental limits. [**(iii)**]{} Renormalization group evolution plays a crucial role in the generation of quark mixing angles. In fact, an asymmetry in the $\mu$ terms of the Higgs doublets generated by the leptonic Yukawa couplings is what induces CKM mixings. [**(iv)**]{} In the version that solves the SUSY phase and the strong CP problems and which provides an understanding of the $\mu$ problem, there are also light $SU(2)_L$ triplet superfields with TeV to sub-TeV scale masses with interesting collider signature [@huitu; @han]. These fields couple to left–handed leptons with the couplings proportional to the heavy Majorana neutrino masses.
The basic structure of the model
================================
Quarks and leptons in the model have the following left-right symmetric assignment under the $SU(3)_C \times SU(2)_L \times
SU(2)_R \times (1)_{B-L}$ gauge group. $$\begin{aligned}
Q(3,2,1,{1 \over 3}) &=& \left(\matrix{u \cr d}\right);~~~~
Q^c(3^*,1,2,-{1 \over 3}) = \left(\matrix{d^c \cr -u^c}\right)
\nonumber \\
L(1,2,1,-1) &=& \left(\matrix{\nu_e \cr e}\right);~~~~
L^c(1,1,2,1) = \left(\matrix{e^c \cr -\nu_e^c}\right)~.\end{aligned}$$ The minimal Higgs sector consists of the following superfields: $$\begin{aligned}
\Delta(1,3,1,2) &=&\left(\matrix{{\delta^{+} \over \sqrt{2}} & \delta^{++} \cr \delta^{0} & -{\delta^+ \over \sqrt{2}}}\right);~~
\overline{\Delta}(1,3,1,-2) = \left(\matrix{{\overline{\delta}^{-} \over \sqrt{2}} & \overline{\delta}^{0} \cr \overline{\delta}^{--} & -{\overline{\delta}^{-} \over \sqrt{2}}}\right);~~ \nonumber \\
\Delta^c(1,1,3,-2) &=& \left(\matrix{{\delta^{c^-} \over \sqrt{2}} & \delta^{c^0} \cr \delta^{c^{--}} & -{\delta^{c^-} \over \sqrt{2}}}\right);~~
\overline{\Delta^c}(1,1,3,2) = \left(\matrix{{\overline{\delta}^{c^+} \over \sqrt{2}} & \overline{\delta}^{c^{++}} \cr \overline{\delta}^{c^0} & -{\overline{\delta}^{c^+} \over \sqrt{2}}}\right);~~ \nonumber \\
\Phi_a(1,2,2,0) &=& \left(\matrix{\phi_1^+ & \phi_2^0 \cr \phi_1^0
& \phi_2^-}\right)_a~~(a=1-2);~~~S(1,1,1,0)~~.\end{aligned}$$ This is the minimal Higgs system in the following sense. The $(\Delta^c + \overline{\Delta^c})$ fields are needed for $SU(2)_R \times U(1)_{B-L}$ symmetry breaking without inducing $R$–parity violating couplings. The $(\Delta + \overline{\Delta})$ fields are their left–handed partners needed for parity invariance. Two bidoublet fields $\Phi_a$ are needed in order to generate quark and lepton masses and CKM mixings. The singlet field $S$ is introduced so that $SU(2)_R \times U(1)_{B-L}$ symmetry breaking occurs in the supersymmetric limit.
The superpotential of the model is given by $$\begin{aligned}
W &=& Y_u Q^T \tau_2 \Phi_1 \tau_2 Q^c + Y_d Q^T \tau_2 \Phi_2 \tau_2
Q^c + Y_\nu L^T \tau_2 \Phi_1 \tau_2 L^c + Y_\ell L^T \tau_2
\Phi_2 \tau_2
L^c \nonumber \\
&+& i \left(f^* L^T \tau_2 \Delta L + f L^{cT} \tau_2 \Delta^c
L^c
\right) \nonumber \\
&+& S \left[ {\rm Tr} \left(\lambda^* \Delta \bar \Delta +\lambda \Delta^c
\bar \Delta^c
\right) +\lambda'_{ab} {\rm Tr} \left( \Phi_a^T \tau_2 \Phi_b \tau_2
\right)- {\cal M}_R^2 \right] + W'\end{aligned}$$ where $$W' = \left[M_\Delta {\rm Tr}(\Delta \bar \Delta) +
M_\Delta^* {\rm Tr}(\Delta^c \bar \Delta^c)\right] + \mu_{ab} {\rm Tr} \left( \Phi_a^T \tau_2 \Phi_b \tau_2 \right) + {\cal M}_S S^2
+ \lambda_S S^3~.$$ $Y_{u,d}$ and $Y_{\nu,\ell}$ in Eq. (3) are quark and lepton Yukawa coupling matrices, while $f$ is the Majorana neutrino Yukawa coupling matrix. The $W'$ term listed in Eq. (4) is optional, in fact when terms in $W'$ are set to zero, the theory has an enhanced $R$ symmetry. Under this $R$–symmetry, $\{Q,
~Q^c,~ L,~L^c\}$ fields have charge $+1$, $S$ has charge $+2$, and all other fields have charge zero with $W$ carrying charge $+2$. While the general setup of the minimal model includes $W'$, the special case of $W'=0$ is interesting, as it leads to an understanding of the $\mu$ term. In the supersymmetric limit, the VEV of the singlet $S$ is zero, but after SUSY breaking, $\left\langle S \right\rangle \sim m_{\rm SUSY}$. Thus the $\mu$ term for the bidoublet $\Phi$ will arise from the coupling $\lambda'_{ab}$, with a magnitude of order $m_{\rm SUSY}$ [@shafi]. It is also in the limit where $W'=0$ that the SUSY CP problem and the strong CP problem can be explained naturally. The main difference between the cases $W' \neq 0$ and $W'=0$ from the low energy perspective is that in the latter case the left–handed triplet superfields $(\Delta + \overline{\Delta})$ will remain light, also with masses of order $m_{\rm SUSY}$.
The superpotential of Eq. (3) is invariant under the parity transformation under which $\Phi\to \Phi^\dagger$, $\Delta\to\Delta^{c*}$, $\overline{\Delta} \to
\overline{\Delta}^{c*}$, $S \to S^*$, $Q\to Q^{c*}$, $L \to
L^{c*}$, $\theta\to \bar{\theta}$, etc. Parity invariance implies that the Yukawa coupling matrices $Y_{u,d},~ Y_{\nu,\ell}$ are hermitian, i.e. $Y_u=Y_u^\dagger$, etc. Additionally, $\lambda'_{ab}$ are real, as is ${\cal M}_R^2$. This means that the effective $\mu$ terms of the bidoublet will be real, provided that $\left\langle S \right \rangle$ is real. If the $\Phi$ VEVs are also real, this setup will provide a solution to the SUSY CP problem and the strong CP problem [@rasin; @babu]. Below we will study under what conditions this is achieved and what the implications of this theory are.
We will work in the ground state corresponding to the following charge preserving VEV pattern for the triplet fields. $$\begin{aligned}
\label{VEV}
\langle\Delta^c \rangle = \left( \begin{array}{cc}
0 & v_R \\
0 & 0
\end{array} \right), \
\langle\bar \Delta^c\rangle = \left( \begin{array}{cc}
0 & 0 \\
\overline{v}_R & 0
\end{array} \right)~.\end{aligned}$$ The VEVs of the left–handed triplet fields $(\Delta +
\overline{\Delta})$ are assumed to be zero since no interaction in the model induces such VEVs. There are two important implications of this setup: [**(i)**]{} Above the parity breaking scale ${\cal
M}_R$, this model has an enhanced global $U(3,c)$ (complexified $U(3)$) symmetry which is broken by the above VEVs to $U(2,c)$. This leads to five massless superfields. Three of these superfields are absorbed by the gauge fields via the super–Higgs mechanism. There remains two light superfields, which are the doubly charged Higgs and Higgsino fields $\delta^{c^{--}}$ and $\overline{\delta}^{c^{++}}$. These fields will consistently acquire masses of order TeV or less, as we shall explicitly show in the next section. Even after soft SUSY breaking terms are turned on, there is a $U(3)$ symmetry in the potential, which leads to one massless doubly charged Higgs boson (and its conjugate). This field will acquire positive squared mass from the renormalization group evolution below $v_R$ proportional to the Bino mass $M_1$. [**(ii)**]{} The bi-doublet fields, when expressed in terms of the components $H_{u,a}, H_{d,a}$ ($a=1,2$), have a symmetric mass matrix in $W$ due to parity symmetry which requires $\mu_{12}=\mu_{21}$. (When $W'=0$, $\mu_{ij} =
\lambda'_{ij} \left\langle S \right\rangle$.) Therefore, if we make one pair of doublets light at the scale $v_R$, it would lead to vanishing CKM mixing angle. This happens in spite of having two Yukawa coupling matrices. Consistency then requires that both pairs of doublets be light below $v_R$. In this case RGE extrapolation brings in an asymmetry between $\mu_{12}$ and $\mu_{21}$. Thus, not only are the potential problems solved by RGE extrapolation, but the resulting scenario becomes very predictive.
Symmetry breaking and the mass of the doubly charged Higgs boson
================================================================
To be specific, we will analyze the model with $W'=0$ of Eq. (3). In the SUSY limit we have from the vanishing of $D$ and $F$ terms, $$|v_R| = |\overline{v}_R|,~~\lambda v_R \overline{v}_R = {\cal M}_R^2,~~ \left\langle S \right\rangle = 0~.$$ It is easy to determine the VEV of $S$ field that is generated after SUSY breaking. Only linear terms in SUSY breaking are relevant for this purpose. We have $$V_{\rm soft} = A_\lambda \lambda S {\rm Tr}(\Delta^c
\overline{\Delta}) - C_\lambda {\cal M}_R^2 S + h.c.$$ Minimization of the resulting potential yields $$\left\langle S^* \right \rangle = {1 \over 2|\lambda|} (C_\lambda - A_\lambda)~.$$ Note that this is of order $m_{\rm SUSY}$. If the coupling $|\lambda|$ is somewhat small, then $\left\langle S
\right\rangle$ can be above the SUSY breaking scale. This feature can be used to make one pair of Higgs doublet superfields somewhat heavier than the SUSY breaking scale. However, the masses of doubly charged fermionic fields, which are equal to $|\lambda|\left\langle S \right \rangle$ must remain below a TeV. Phenomenology of doubly charged Higgsino has been studied in Ref. [@huitu; @han; @gunion; @other].
Parity symmetry requires ${\cal M}_R^2$ and $C_\lambda$ be real. If the trilinear soft breaking terms are proportional to the corresponding Yukawa coupling matrices, then we have $A_{\lambda}$ real as well. Proportionality will require $A_\lambda \lambda = A_0 \lambda$, with the universal $A_0$ being real. Since the trilinear $A$ terms in the quark sector must be hermitian by parity, and since the Yukawa coupling matrices are hermitian, $A_0$ must be real. This condition is realized in many models of SUSY breaking such as Poloni type supergravity breaking, gauge mediated SUSY breaking, anomaly mediated SUSY breaking, etc. We shall adopt this proportionality relation for all the $A$ terms. We see that $\left\langle S \right \rangle$ is then real. The resulting $\mu$ terms will also be real. This helps solve the strong CP problem and the SUSY phase problem.
The full potential of the model relevant for symmetry breaking has $F$ term, $D$ term and soft SUSY breaking contributions. They are given by $$\begin{aligned}
V_F &=& \left|\lambda {\rm Tr}(\Delta^c \overline{\Delta}^c +
\lambda'_{ab} {\rm Tr} \left( \Phi_a^T \tau_2 \Phi_b \tau_2
\right)-{\cal M}_R^2\right|^2 + |\lambda|^2 \left|{\rm
Tr}(\Delta^c \Delta^{c \dagger}) + {\rm
Tr}(\overline{\Delta}^c~\overline{\Delta}^{c \dagger}) \right|
\nonumber \\
V_{\rm soft} &=& M_1^2 {\rm Tr} (\Delta^{c \dagger} \Delta^c) +
M_2^2 {\rm Tr} (\overline{\Delta}^{c \dagger} \overline{\Delta}^c) +M_S^2 |S|^2\nonumber \\
&+& \{A_\lambda \lambda S {\rm Tr}(\Delta^c \Delta^{c \dagger}) -
C_\lambda {\cal M}_R^2 S + h.c.\} \nonumber \\
V_D &=& {g_R^2 \over
8}\sum_{a}\left|{\rm Tr}(2 \Delta^{c \dagger} \tau_a \Delta^c +
2 \overline{\Delta}^{c \dagger} \tau_a \overline{\Delta}^c + \Phi_a \tau_a^T \Phi_a^\dagger)\right|^2 \nonumber \\
&+& {g'^2 \over 8} \left|{\rm Tr}(2 \Delta^{c \dagger} \Delta^c +
2 \overline{\Delta}^{c \dagger} \overline{\Delta}^c )\right|^2~.\end{aligned}$$
Minimizing the potential yields the following two complex conditions. $$\begin{aligned}
&~& v_R^*\left[|\lambda|^2|S|^2+M_1^2+g_R^2(|v_R^2
-|\overline{v}_R|^2 + {X \over 2})+g'^2 (|v_R|^2-
|\overline{v}_R|^2)\right]\nonumber \\
&+& \overline{v}_R\left[\lambda A_\lambda S
+ |\lambda|^2(v_R \overline{v}_R-{{\cal M}_R^2
\over \lambda})^*\right] = 0, \nonumber \\
&~& \overline{v}_R^*\left[|\lambda|^2|S|^2+M_2^2-g_R^2(|v_R^2
-|\overline{v}_R|^2 + {X \over 2})-g'^2 (|v_R|^2-
|\overline{v}_R|^2)\right]\nonumber \\
&+& {v}_R\left[\lambda A_\lambda S
+ |\lambda|^2(v_R \overline{v}_R-{{\cal M}_R^2
\over \lambda})^*\right] = 0,\end{aligned}$$ where we defined $X= \sum_{a=1}^2\left\langle |\phi_1^0|^2 -
|\phi_2^0|^2 \right\rangle_a$. Applying these conditions, we obtain the following mass squared matrix for the doubly charged Higgs bosons $(\delta^{c^{-- *}}, \overline{\delta}^{c^{++}})$. $$\begin{aligned}
{\cal M}^2_{\delta^{++}} = \left(\matrix{-2 g_R^2(|v_R|^2-|\overline{v}_R|^2+{X \over 2})-{\overline{v}_R
\over v_R^*} Y & Y^* \cr Y & 2 g_R^2(|v_R|^2 - |\overline{v}_R|^2+{X \over 2}) - {v_R \over \overline{v}_R^*}Y
}\right)\end{aligned}$$ where $Y = \lambda A_\lambda S +|\lambda|^2(v_R \overline{v}_R -
{{\cal M}_R^2 \over \lambda})^*$. It is clear that as the $D$ term is set to zero, there is one massless mode in this sector. Actually, if $v_R$ is much larger than the SUSY breaking terms, turning on the $D$ term makes one of the masses negative. This is the pseudo–Goldstone boson of the model. There is no inconsistency, as this zero squared-mass will turn positive via RGE evolution.
Below the scale $v_R$, the mass matrix of the doubly charged Higgs boson fields has the form $$\begin{aligned}
{\cal M}^2_{\delta^{++}} = \left(\matrix{M_{++}^2 + \mu_\delta^2 +
\delta_1 & (B\mu)_{\delta} + \delta_{12} \cr (B\mu)_{\delta}^*
+\delta_{12}^* & M^2_{--} + \mu_\delta^2 + \delta_2}\right)\end{aligned}$$ where $\mu_{\delta} \delta^{++} \delta^{--}$ is the effective superpotential mass term, $M_{++}^2$ and $M_{--}^2$ are the soft mass parameters, and $\delta_i$ denote RGE correction factors corresponding to running from $v_R$ down to the SUSY breaking scale. Eq. (12) should match Eq. (11) at $v_R$, which implies that $M_{++}^2 \simeq M_{--}^2$, $|(B \mu)_\delta| \simeq M_{++}^2
+ \mu_\delta^2$ at $v_R$. In the large $v_R$ limit, the light Higgs resulting from Eq. (11) is $(\delta^{*--} -
\delta^{++})/\sqrt{2}$, so the squared mass of this state, including RGE corrections is $[\delta_1 + \delta_2-2 {\rm
Re}(\delta_{12})]/2$. There is an upper limit on this mass, which can be derived as follows. Let us ignore the off–diagonal entry for the moment. The renormalizaion group equation for $M_{++}^2$ has the form [@martin] $${d M_{++}^2 \over dt} = -{c \over 16 \pi^2} g_1^2 M_1^2 + ...$$ where $c = (96/5)$. Here we have displayed only the positive contributions to the mass-squared, which would be relevant for determining the upper limit. Along with $${d g_1 \over dt} = {b_1 \over 16 \pi^2} g_1^3,~~~{dM_1 \over dt} =
{2 b_1 \over 16 \pi^2} g_1^2 M_1^2,$$ we can solve for $M_{++}^2$. In the present model $b_1 = (78/5)$ when the $(\Delta + \overline{\Delta})$ are light, and $b_1 =
12$ when these fields are heavy. We find $$M_{++}^2(m_Z) < {24 \over 5 b_1} M_1^2(m_Z)\left[{\alpha_1^2(v_R)
\over \alpha_1^2(m_Z)}-1\right]~.$$ The gauge couplings in this model will remain perturbative up to about $10^{12}$ GeV when $b_1=(78/5)$ and up to about $10^{14}$ GeV when $b_1= 12$. If we choose $\alpha_1(v_R) = 0.1$, we find the upper limit on $M_{++} < 3.7 M_1$ (for the case where $b_1=12$). The running of the $(B\mu)_{\delta}$ will also contribute to the mass of this state, but this evolution depends on other SUSY breaking parameters. We expect the entire contribution to be of order few times $M_1$.
Effective potential and the global minimum of the theory
========================================================
Let us now turn attention to the electric charge and/or $R$–parity breaking global minimum of the model and see how this problem is cured by taking loop corrections induced by the heavy Majorana neutrino Yukawa couplings into account. We will show that the results of Ref. [@kuchi] gets significantly modified, allowing for the desired charge conserving minimum to be the global minimum for some domain of the parameters.
It is easy to see why the tree–level potential has a deeper minimum that violates electric charge and/or $R$–parity. In the desired minimum which preserves these quantum numbers, the VEVs of the triplet fields are as shown in Eq. (5). Consider the following alternative VEV configuration. $$\begin{aligned}
\langle\Delta^c \rangle = {1 \over \sqrt{2}}\left( \begin{array}{cc}
0 & {v_R} \\
{v_R} & 0
\end{array} \right), \
\langle\bar \Delta^c\rangle = {1 \over \sqrt{2}} \left( \begin{array}{cc}
0 & {\overline{v}_R} \\
{\overline{v}_R} & 0
\end{array} \right)~.\end{aligned}$$ This pattern of course breaks electric charge. All terms in the scalar potential are exactly the same for this configuration of VEVs and that of Eq. (5), except in the $SU(2)_R$ $D$–terms. Since the VEVs of Eq. (16) are along $\tau_1$, the $D$–terms vanish for this configuration, while it is nonzero and positive for the desired configuration. This proves that the desired VEV pattern does not correspond to the global minimum of the potential.
We proceed to compute the Coleman–Weinberg potential of the model by keeping one family of neutrino Yukawa couplings to the $\Delta^c$ field, as shown by the $f$ coupling in Eq. (3). To be able to compare different minima, we use a the general background with the full $\Delta^c$ and $\overline{\Delta}^c$ fields. The field–dependent masses of the $(e^c,~\nu^c$) fermionic and scalar fields can be expressed in terms of the invariant combinations $$\begin{aligned}
D^2_{1,2} = {1 \over 2} \left[{\rm Tr}(\Delta^{c \dagger} \Delta^c) \pm \sqrt{ {\rm Tr}(\Delta^{c \dagger} \Delta^c)^2
-{\rm Tr}(\Delta^c \Delta^c) {\rm Tr} (\Delta^{c \dagger} \Delta^{c \dagger})}\right]~.\end{aligned}$$ We also define $\overline{D}_{1,2}^2$ in an analogous way, with the replacement of $\Delta^c$ by $\overline{\Delta}^c$ in Eq. (17). Including the soft SUSY breaking contributions, the $F$–term contributions, and the $D$–term contributions, the field–dependent masses of the sleptons $(\tilde{e^c},~\tilde{\nu^c})$, and the corresponding fermions are found to be $$\begin{aligned}
m_{1,2}^2 &=& |f|^2 D_1^2 + m_{L^c}^2 + {g_R^2 \over 2}[(D_2^2 -
\overline{D}_2^2) - (D_1^2 - \overline{D}_1^2)] -{{g'^2} \over
2}[(D_1^2 - \overline{D}_1^2) +
(D_2^2 - \overline{D}_2^2)], \nonumber \\
&\pm& \left|A_f f D_1 + \lambda^* S^* f \overline{D}_1 \right|^2
\nonumber \\
m_{3,4}^2 &=& |f|^2 D_2^2 + m_{L^c}^2 + {g_R^2 \over 2}[(D_1^2 -
\overline{D}_1^2) - (D_2^2 - \overline{D}_2^2)] -{{g'^2} \over
2}[(D_1^2 - \overline{D}_1^2) +
(D_2^2 - \overline{D}_2^2)], \nonumber \\
&\pm& \left|A_f f D_2 + \lambda^* S^* f \overline{D}_2 \right|^2 \nonumber \\
m_{F_1}^2 &=& |f D_{1}|^2,~\nonumber\\
m_{F_2}^2 &=& |f D_{2}|^2.\end{aligned}$$ Here the $m_{1-4}$ correspond to the masses of the four real scalar states, while $m_{F_{1,2}}$ are the masses of the two fermionic states.
With these mass eigenvalues, one can compute the effective potential in the Landau gauge in the $\overline{DR}$ scheme from the expression $$V_{\rm eff}^{\rm 1-loop} = {1 \over 64 \pi^2}\sum_i(-1)^{2s}
(2s+1) M_i^4 \left[{\rm Log}({M^2_i \over \mu^2}) - {3 \over
2}\right] ~.$$ We expand this potential in the limit where SUSY breaking parameters are small compared to the VEVs of the $(\Delta^c, \overline{\Delta}^c)$ fields. In the SUSY limit, vanishing of the $D$–terms require $D_1^2 = \overline{D}_1^2,~D_2^2 = \overline{D}_2^2$. So we use the expansion $$\overline{D}_1^2 - D_1^2 = a_1 m_{L^c}^2,~~~~~\overline{D}_2^2 -
D_2^2 = a_2 m_{L^c}^2$$ where $m_{L^c}^2$ denotes the soft SUSY breaking mass of the slepton doublet. Defining $$x = {{{\rm Tr}(\Delta^c \Delta^c){\rm Tr}(\Delta}^{c \dagger}
\Delta^{c \dagger}) \over [{\rm Tr}(\Delta^{c \dagger}
\Delta^c)]^2}$$ we find the leading contribution to $V_{\rm eff}$ to be [$$\begin{aligned}
\hspace*{-0.2in} V_{\rm eff}^{\rm 1-loop} &=&
-{|f|^2 m_{L^c}^2 {\rm
Tr}(\Delta^c {\Delta}^{c \dagger}) \over 64 \pi^2} \left
[(4+2~{\rm ln}2)+2 (a_1-a_2)g_R^2\sqrt{1-x}+ 2(a_1+a_2)g'^2+
\right.
\nonumber \\
&-& \left\{2+(a_2-a_1) g_R^2 +
(a_2+a_1)g'^2\right\}\left(1-\sqrt{1-x}\right){\rm ln}
\left({|f|^2 {\rm Tr}(\Delta^c {\Delta}^{c \dagger}) \over 2
\mu^2}\left(1-\sqrt{1-x}\right)\right) \nonumber \\
&+& \left\{\left((a_2-a_1) g_R^2 -
(a_2+a_1)g'^2\right)\left(1+\sqrt{1-x}\right)-2\sqrt{1-x}\right\}{\rm
ln} \left({|f|^2 {\rm Tr}(\Delta^c {\Delta}^{c \dagger})
\over 2 \mu^2}\left(1+\sqrt{1-x}\right)\right) \nonumber \\
&-&\left. 2~{\rm ln}\left({|f|^2 {\rm Tr}(\Delta^c {\Delta}^{c
\dagger}) \over \mu^2}\left(1+\sqrt{1-x}\right)\right) \right]\end{aligned}$$]{}
--------------------------------------------------------------------------------------------------------- --
{width="0.5\linewidth"}
--------------------------------------------------------------------------------------------------------- --
Clearly, these loop contributions vanish in the SUSY limit. The non–vanishing terms arise because the cancelation between the first two diagrams of Fig. 1 is no longer exact, once SUSY breaking is turned on. And diagram (c) has no fermionic counterpart. The most interesting aspect of the one–loop effective potential is the appearance of the structure ${\rm
Tr}(\Delta^c \Delta^c){\rm Tr}({\Delta}^{c \dagger} {\Delta}^{c
\dagger})$, which was absent in the tree–level potential. If we make a further expansion in small $x$, Eq. (22) will result in the following quartic coupling: $$\begin{aligned}
V^{\rm quartic} &=& -{|f|^2 m_{L^c}^2 {\rm Tr}(\Delta^c
\Delta^c){\rm Tr}({\Delta}^{c \dagger} {\Delta}^{c \dagger}) \over
128 \pi^2|v_R|^2} \left [\{2-\{a_1-a_2)g_R^2-(a_1+a_2)g'^2\}(1+
2~{\rm ln}2) \right.
\nonumber \\
&+& \left. (a_1-a_2)g_R^2 ~{\rm ln}{|fv_R|^2 \over \mu^2} -
\{2-(a_1-a_2)g_R^2+(a_1+a_2)g'^2\}~{\rm ln}x\right] + ...\end{aligned}$$ where the ... indicates higher order terms in $x$ and $x$–independent terms. In the desired vacuum we have $D_2=\overline{D}_2 =0$, so that the coefficient $a_2$ is zero. Minimization conditions (Eq. (10)) determine $a_1$ as $$(g_R^2 + g'^2) a_1 m_{L^c}^2 \simeq {1 \over 2} (M_1^2 - M_2^2 +g_R^2
X)~,$$ where $M_{1,2}^2$ are the soft mass squared of the $(\Delta^c,~
\overline{\Delta}^c$) fields. In supergravity type SUSY breaking, one would expect $M_1^2 \leq M_2^2$, as $\Delta^c$ has the Majorana Yukawa coupling which would lower its mass from the universal mass, while $\overline{\Delta}^c$ does not. Using this we find that for the charge conserving vacuum to be lower than the charge breaking vacuum, we would need $m_{L^c}^2$ to be negative. In such a situation, we can derive upper limits on the stau masses. (We assume that the third family fermions have the largest Majorana Yukawa coupling $f$.) Note that the positive contributions to the masses of $\tilde{\tau}_R$ and $\tilde{\tau}_L$ arise from the gaugino masses $M_1$ and $M_2$ [@martin]. $$\begin{aligned}
16 \pi^2 {d m^2_{\tilde{\tau}_R}\over dt} &=& -{24 \over 5} g_1^2
M_1^2 + ...\nonumber
\\
16 \pi^2 {d m^2_{\tilde{\tau}_L}\over dt} &=& -{6 \over 5} g_1^2
M_1^2 - 6 g_2^2 M_2^2 + ...\end{aligned}$$ where the ... denote terms that would decrease the scalar mass in the evolution from $v_R$ to $m_Z$. We have $b_2= 6$ when $(\Delta
+ \overline{\Delta})$ fields are light, and $b_2=2$ when they are heavy. The upper limits on the stau masses are found to be $$\begin{aligned}
M_{\tilde{\tau_R}}^2(m_Z) &<& {6 \over 5 b_1}
M_1^2(m_Z)\left[{\alpha_1^2(v_R) \over
\alpha_1^2(m_Z)}-1\right]~,\nonumber \\
M_{\tilde{\tau_L}}^2(m_Z) &<& {3 \over 10 b_1}
M_1^2(m_Z)\left[{\alpha_1^2(v_R) \over \alpha_1^2(m_Z)}-1\right]
+{3 \over 2 b_2} M_2^2(m_Z)\left[{\alpha_2^2(v_R) \over
\alpha_2^2(m_Z)}-1\right] .\end{aligned}$$ Both these limits are in th acceptable range. For $\alpha_1(v_R) =
0.1$, we find the right–handed stau mass to be bounded by about 1.9 $M_1$ (for $b_1=12$), with the left–handed stau roughly two times heavier.
CKM angles out of radiative corrections
=======================================
As noted earlier, our model predicts that the CKM angles vanish at the tree–level due to left–right symmetry. The reason for this is that the $2\times 2$ ($H_u,~H_d$) Higgsino mass matrix is symmetric. When one pair of light MSSM Higgs superfields is extracted from such a symmetric matrix, it follows that the up and down quark Yukawa coupling matrices to these light doublets will be the same. This is assuming that only one pair of doublets survives below $v_R$. Therefore once electroweak symmetry breaks, we have $M_u~=~\xi M_d$ and hence $V_{CKM}~=~1$. Consistency with CKM mixings then requires that both pairs of Higgs doublets remain light below $v_R$. In that case, below $v_R$, the bidoiblet mass terms $\mu_{ab}$ will receive asymmetric radiative RGE corrections, in the momentum range $v_R$ to $\mu_\Phi$, because parity is violated in this regime. (We denote the scale of the heavy doublet mass as $\mu_\Phi$.) To leading order the quark Yukawa couplings do not induce an asymmetry in $\mu_{ab}$. However, since the right–handed neutrinos decouple below $v_R$, the lepton sector induces an asymmetry. Only the charged lepton Yukawa couplings contribute to the evolution of $\mu_{ab}$, making the RGE contribution to $\mu_{12}$ different from that of $\mu_{21}$. As a result, when the $H_u,H_d$ mass matrix is diagonalized at a scale $\mu_\Phi$ below $v_R$, so that only one pair of Higgs superfields remain light, the resulting light Higgs doublets couple to up and down quarks with different Yukawa coupling matrices.
The RGE for the asymmetry between $\mu_{12}$ and $\mu_{21}$ (to leading order) is $$\begin{aligned}
{d \over dt}(\mu_{12} - \mu_{21}) = {\mu_{12} + \mu_{21} \over 32
\pi^2}{\rm Tr}(Y_\nu^\dagger Y_\nu - Y_\ell^\dagger Y_\ell),\end{aligned}$$ which can be solved to determine the asymmetry in $\mu_{ij}$. We obtain $(\mu_{12}-\mu_{21})/(\mu_{12}+\mu_{21}) \simeq 1/(16 \pi^2){\rm
Tr}(Y_\nu^\dagger Y_\nu - Y_\ell^\dagger Y_\ell){\rm
ln}(v_R/\mu_\Phi)$, where $\mu_\Phi$ is the mass of the heavy bidoublet. The suppression factor that apperas in the CKM angles is about 0.1 when one of the leptonic Yukawa coupling entries is of order one. This can lead to reasonable values for the CKM angles.
FCNC, the strong CP and the SUSY CP problems
============================================
The presence of a second pair of Higgs doublets coupling to fermions implies that there will be tree–level flavor changing neutral currents mediated by the Higgs. Experimental constraints will require that one pair of Higgs doublets be heavy, with mass of the order of few to 50 TeV [@ji; @pospelov]. This can be seen from the mass matrices of the quarks, $$\begin{aligned}
M_u &=& Y_u \kappa_u + Y_d \kappa_u' \nonumber \\
M_d &=& Y_u \kappa_d' + Y_d \kappa_d\end{aligned}$$ where $\kappa_i$ are the VEVs of the neutral components. These equations can be used to solve for the Yukawa coupling matrices. For example, $Y_d = (\kappa_u M_u - \kappa_d' M_d)/(\kappa_u
\kappa_d - \kappa_u' \kappa_d')$. In a basis where $M_d$ is diagonal, $M_u = \hat{V}^T D_u \hat{V}^*$, where $\hat{V} =
P.V.Q$, with $V$ being the CKM matrix in the standard parametrization, and $P,Q$ being phase matrices. $D_u$ is the diagonal up–quark mass matrix. Flavor changing Higgs couplings can be then readily derived: $${\cal L}^{FCNC} = \left({\kappa_u \over \kappa_u \kappa_d -
\kappa_u' \kappa_d'}\right)Q_i Q_j^* (D_u)_kV_{ki} V_{kj}^* H^0 +
h.c.$$ Due to the hermiticity of this matrix, the unknown phase matrix $Q$ disappears from processes such as $\epsilon_K$. We find stringent limit on the mass of $H_0$, $m_{H^0} \geq (30-50)$ TeV, if there is no cancelation between the Higgs exchange and the SUSY squark–gluino exchange box diagram. If such cancelations are allowed, the limit on $H^0$ mass is considerably reduced [@ji]. As noted after Eq. (8), the model allows for one pair of Higgs doublets to be naturally heavier than the SUSY breaking scale, thus satisfying the FCNC constraint.
Since two pairs of Higgs doublets must survive below $v_R$, there are calculable FCNC via SUSY diagrams. The most significant ones are the gluino box diagram for $K^0 - \overline{K}^0$ mixing. We find that these constraints are met in the model.
The basic idea behind parity as a solution for the strong CP problems is that left–right symmetry leads to hermitian Yukawa couplings [@senj]. If the VEVs of bi–doublet Higgs fields are real, this would lead to a solution to the strong CP problem. The reality of the VEVs is not guaranteed by parity and always involves additional assumptions. Supersymmetry provides this extra symmetry in minimal left–right models without any singlet fields as shown in [@rasin1]. When there are gauge singlet fields in the theory, this needs to be reinvestigated. As we noted in the symmetry breaking discussion, if $W'=0$, which can be enforced by an $R$–symmetry, one can have a scenario where the singlet VEV is real. In such a setup not only is the strong CP problem solved, but the weak SUSY CP problem is also solved. The EDMs of the electron and the neutron will be vanishing due to parity at the scale $v_R$. Renormalization group evolution does induce small EDMs, but well within experimental limits.
Conclusion
==========
In conclusion, we have pointed out that the minimal renormalizable supersymmetric left-right model is completely consistent phenomenologically without any need for higher dimensional operators or spontaneous $R$–parity violation. The scale of left–right symmetry can now be higher than TeV. The model can solve the strong CP problem without fear of large contributions to $\overline{\theta}$ from non-renormalizable terms (since they are now not needed). The model also provides a simple solution based on parity symmetry for the SUSY CP problem. The effective potential of the theory, which has important contributions from heavy Majorana Yukawa couplings, allows for the charge conserving and $R$–parity conserving minimum to be the global minimum. The model predicts light (sub–TeV) doubly charged Higgs bosons and their superpartners.
Acknowldgements {#acknowldgements .unnumbered}
===============
We wish to thank Zurab Tavartkiladze for discussions. KSB is supported in part by DOE grants DE-FG02-04ER41306 and DE-FG02-ER46140. RNM is supported by NSF grant No. PHY-0652363.
[90]{}
R. N. Mohapatra and J. C. Pati, Phys. Rev. D [**11**]{}, 2558 (1975); G. Senjanovic and R. N. Mohapatra, Phys. Rev. D [**12**]{}, 1502 (1975).
R. N. Mohapatra and A. Rasin, Phys. Rev. D [**54**]{}, 5835 (1996).
K. S. Babu, B. Dutta and R. N. Mohapatra, Phys. Rev. D [**65**]{}, 016005 (2002).
R. N. Mohapatra, Phys. Rev. [**D 34**]{}, 3457 (1986); A. Font, L. E. Ibanez and F. Quevedo, Phys. Lett. B [**228**]{}, 79 (1989); S. P. Martin, Phys. Rev. D [**46**]{}, 2769 (1992).
R. Kuchimanchi and R. N. Mohapatra, Phys. Rev. D [**48**]{}, 4352 (1993); R. Kuchimanchi and R. N. Mohapatra, Phys. Rev. Lett. [**75**]{}, 3989 (1995).
C. S. Aulakh, A. Melfo and G. Senjanovic, Phys. Rev. D [**57**]{}, 4174 (1998).
Z. Chacko and R. N. Mohapatra, Phys. Rev. D [**58**]{}, 015003 (1998).
K. S. Babu, B. Dutta and R. N. Mohapatra, Phys. Rev. Lett. [**85**]{}, 5064 (2000).
C. S. Aulakh, K. Benakli and G. Senjanovic, Phys. Rev. Lett. [**79**]{}, 2188 (1997).
K. Huitu, J. Maalampi, A. Pietila and M. Raidal, Nucl. Phys. B [**487**]{}, 27 (1997).
T. Han, B. Mukhopadhyaya, Z. Si and K. Wang, Phys. Rev. D [**76**]{}, 075013 (2007); P. Fileviez Perez, T. Han, G. y. Huang, T. Li and K. Wang, arXiv:0805.3536 \[hep-ph\]; D. A. Demir, M. Frank, K. Huitu, S. K. Rai and I. Turan, arXiv:0805.4202 \[hep-ph\].
G. R. Dvali, G. Lazarides and Q. Shafi, Phys. Lett. B [**424**]{}, 259 (1998); S. F. King and Q. Shafi, Phys. Lett. B [**422**]{}, 135 (1998); K. S. Babu, B. Dutta and R. N. Mohapatra, Phys. Rev. D [**65**]{}, 016005 (2002); R. Kitano and N. Okada, Prog. Theor. Phys. [**106**]{}, 1239 (2001); L. J. Hall, Y. Nomura and A. Pierce, Phys. Lett. B [**538**]{}, 359 (2002).
J. F. Gunion, C. Loomis and K. T. Pitts, [*In the Proceedings of 1996 DPF / DPB Summer Study on New Directions for High-Energy Physics (Snowmass 96), Snowmass, Colorado, 25 Jun - 12 Jul 1996, pp LTH096*]{} \[arXiv:hep-ph/9610237\].
G. Azuelos, K. Benslama and J. Ferland, J. Phys. G [**32**]{}, 73 (2006); A. G. Akeroyd and M. Aoki, Phys. Rev. D [**72**]{}, 035011 (2005); M. Muhlleitner and M. Spira, Phys. Rev. D [**68**]{}, 117701 (2003); M. Kuze and Y. Sirois, Prog. Part. Nucl. Phys. [**50**]{}, 1 (2003); J. Maalampi and N. Romanenko, Phys. Lett. B [**532**]{}, 202 (2002); M. Lusignoli and S. Petrarca, Phys. Lett. B [**226**]{} 397 (1989).
See for eg., S. Martin and M. Vaughn, Phys. Rev. [**D 50**]{}, 2282 (1994).
Y. Zhang, H. An and X. d. Ji, arXiv:0710.1454 \[hep-ph\].
G. Ecker and W. Grimus, Nucl. Phys. B [**258**]{}, 328 (1985); M. E. Pospelov, Phys. Rev. D [**56**]{}, 259 (1997).
M. A. B. Beg and H. S. B. Tsao, Phys. Rev. Lett. [**41**]{}, 278 (1978); R. N. Mohapatra and G. Senjanovic, Phys. Lett. B [**79**]{}, 283 (1978).
R. N. Mohapatra and A. Rasin, Phys. Rev. Lett. [**76**]{}, 3490 (1996); R. Kuchimanchi, Phys. Rev. Lett. [**76**]{}, 3486 (1996).
[^1]: It is not strictly required that the vacuum we live in correspond to the global minimum of the potential. Metastable vacua are acceptable, provided that the tunnelling rate from that vacuum to the true vacuum is sufficiently slow in comparison to the age of the Universe.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'While the Bayesian Information Criterion ([<span style="font-variant:small-caps;">bic</span>]{}) and Akaike Information Criterion ([<span style="font-variant:small-caps;">aic</span>]{}) are powerful tools for model selection in linear regression, they are built on different prior assumptions and thereby apply to different data generation scenarios. We show that in the finite dimensional case their respective assumptions can be unified within an augmented model-plus-noise space and construct a prior in this space which inherits the beneficial properties of both [<span style="font-variant:small-caps;">aic</span>]{} and [<span style="font-variant:small-caps;">bic</span>]{}. This allows us to adapt the [<span style="font-variant:small-caps;">bic</span>]{} to be robust against misspecified models where the signal to noise ratio is low.'
bibliography:
- 'dekockeggers-submitv2.bib'
---
**Bayesian Model Selection for Misspecified Models in Linear Regression**\
M.B. de Kock${}^1$ and H.C. Eggers${}^{1,2}$\
${}^1$*Institute of Theoretical Physics and Department of Physics, Stellenbosch University, P/Bag X1,*\
*7602 Matieland, Stellenbosch, South Africa*\
${}^2$*National Institute for Theoretical Physics,*\
*P/Bag X1, 7602 Matieland, South Africa*
Introduction {#sec:intro}
============
The selection of a model between multiple competing models is a well established tool of data analysis in a wide spectrum of fields ranging from ecology to psychology[@burnham2003model; @vrieze2012model; @aho2014model]. If the true model is one of the candidates then Bayesian methods are consistent in that they will select the true model with probability one as the sample size increases[@Nishii]. On the other hand if the underlying data generating process is nonparametric then to estimate the underlying regression function we require a minimax-rate optimal rule[@shao]. This dichotomy is closely related to the competition between [<span style="font-variant:small-caps;">aic</span>]{} and [<span style="font-variant:small-caps;">bic</span>]{} , where [<span style="font-variant:small-caps;">bic</span>]{} represents the Bayesian methods and [<span style="font-variant:small-caps;">aic</span>]{} the loss-optimal rules. In general, it is though to be impossible to combine the properties of both, see [@yang2005can].
Our goal is not to address the difference between the parametric and nonparametric cases but to give a Bayesian construction for linear regression that is robust to model misspecification. It is similar to [@muller2013], but does not introduce an artificial posterior but augments the likelihood with a larger parameter space. This extends the idea of Akaike[@akaike1978bayesian] which appeared after Akaike [@akaike1974new] and Schwartz introduced,[@schwarz1978estimating], their information criterion, [<span style="font-variant:small-caps;">aic</span>]{} and [<span style="font-variant:small-caps;">bic</span>]{}, respectively. Akaike, in the limited context of linear regression, considers the model selection problem in a parameter space expanded from the $K$-dimensional model space into a larger one and then shows how different Bayesian priors in this space give arise to the two different information criterion. This reveals that the BIC implicitly includes a prior which fixes the extra non-model “noise parameters” to be exactly zero, while the [<span style="font-variant:small-caps;">aic</span>]{} allows them to vary to some degree around zero. This confirms our own experience in that as the noise-to-signal ratios is lowered to unity the [<span style="font-variant:small-caps;">bic</span>]{} fails completely. Statistically speaking, the [<span style="font-variant:small-caps;">bic</span>]{} assumes one of the candidate models is the true model while [<span style="font-variant:small-caps;">aic</span>]{} does not. It is this property that we wish to extend to the [<span style="font-variant:small-caps;">bic</span>]{} case.
Armed with this insight, we shall construct priors not only for the $K$ model parameters, but for an additional set of $L = N-K$ noise parameters in the spirit of Akaike. Unlike their [<span style="font-variant:small-caps;">bic</span>]{} and [<span style="font-variant:small-caps;">aic</span>]{} predecessors, however, these new priors will take into account the crucial information that the modes of signal parameter priors should be located some distance away from the origin of parameter space. This allows us to adapt the Bayesian Information Criterion to not assume that one of the candidate models are true.
The paper is organised as follows. In Section \[sec:blr\], we review the framework of Bayesian model comparison and linear regression, augmenting in Section \[sec:nsp\] the model space by a noise space in preparation for the new priors. Reconsideration in Section \[sec:prcn\] of the priors underlying the [<span style="font-variant:small-caps;">bic</span>]{} and [<span style="font-variant:small-caps;">aic</span>]{} forms the basis and motivation for the construction of spherically symmetric priors for both model and noise parameters in Section \[sec:nic\]. The results are tested numerically and compared to the traditional information criteria in Section \[sec:smm\], followed by a brief summary and discussion in Section \[sec:dsc\].
Bayesian linear regression {#sec:blr}
==========================
Bayesian model selection {#sec:bmc}
------------------------
Given data $\mathcal{D}$, Bayesian model selection is based on the evidence or marginal likelihood for the model ${{\mathcal{H}}}_{{\scriptscriptstyle K}}$ which has $K$ parameters ${{\bm{\alpha}}}_{{\scriptscriptstyle K}}= (\alpha_1,\ldots,\alpha_{{\scriptscriptstyle K}})$. The evidence is an average over the likelihood $p(\mathcal{D}{\,|\,}{{\bm{\alpha}}}_{{\scriptscriptstyle K}},{{\mathcal{H}}}_{{\scriptscriptstyle K}})$ weighted by the parameter prior $p({{\bm{\alpha}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})$, $$\begin{aligned}
\label{bsb}
p(\mathcal{D} {\,|\,}{{\mathcal{H}}}_{{\scriptscriptstyle K}})
&= \int p(\mathcal{D}{\,|\,}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}, {{\mathcal{H}}}_{{\scriptscriptstyle K}})
\,p({{\bm{\alpha}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})\, {\mathrm{d}}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}.\end{aligned}$$ where $p({{\bm{\alpha}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})$ may contain hyperparameters as necessary. Bayes’ theorem used twice for competing models ${{\mathcal{H}}}_{{\scriptscriptstyle K}}$, ${{\mathcal{H}}}_{{{\scriptscriptstyle K}}'}$ relates the ratio of model posteriors $p({{\mathcal{H}}}_{{\scriptscriptstyle K}}{\,|\,}\mathcal{D})$ to the corresponding model evidences by $$\begin{aligned}
\label{bsc}
\frac{p({{\mathcal{H}}}_{{\scriptscriptstyle K}}{\,|\,}\mathcal{D})}{p({{\mathcal{H}}}_{{{\scriptscriptstyle K}}'}{\,|\,}\mathcal{D})}
&=
\frac{p({{\mathcal{H}}}_{{\scriptscriptstyle K}})}{p({{\mathcal{H}}}_{{\scriptscriptstyle K}}')}\,
\frac{p(\mathcal{D}{\,|\,}{{\mathcal{H}}}_{{\scriptscriptstyle K}})}{p(\mathcal{D}{\,|\,}{{\mathcal{H}}}_{{{\scriptscriptstyle K}}'})}.\end{aligned}$$ Barring good reasons to deviate from the Principle of Indifference, model priors would normally be set equal, $p({{\mathcal{H}}}_{{\scriptscriptstyle K}}) = p({{\mathcal{H}}}_{{\scriptscriptstyle K}}') = \tfrac{1}{2}$, in which case the posterior odds equals the Bayes Factor [@kass1995bayes] $$\begin{aligned}
\label{bsd}
\frac{p({{\mathcal{H}}}_{{\scriptscriptstyle K}}{\,|\,}\mathcal{D})}{p({{\mathcal{H}}}_{{{\scriptscriptstyle K}}'}{\,|\,}\mathcal{D})}
&= \frac{p(\mathcal{D}{\,|\,}{{\mathcal{H}}}_{{\scriptscriptstyle K}})}{p(\mathcal{D}{\,|\,}{{\mathcal{H}}}_{{{\scriptscriptstyle K}}'})}
\equiv {\mathrm{BF}}[K,K'].\end{aligned}$$ When more than two models are to be compared, it is convenient to define a reference model against which all others are measured. In this paper, we use as reference model ${{\mathcal{H}}}_{{\scriptscriptstyle N}}$, the case where $N$ data points are modelled by $K=N$ free parameters, implying of course an exact fit and no noise. Among the $N$ competing models with $K=1,2,\ldots N$ parameters, the best model is the one with maximal Bayes Factor or equivalently minimum information criterion $\mathrm{IC} = -2\log{\mathrm{BF}}$.
Linear Regression {#sec:lrg}
-----------------
We briefly review the canonical formalism for linear regression and introduce the language and notation to be used in later sections. By assumption the data $\mathcal{D}$ comes in the form of $N$ data points $y_n \in {{\bm{y}}}= (y_1,\ldots,y_{{\scriptscriptstyle N}})$ measured at locations $x_n \in {{\bm{x}}}= (x_1,\ldots,x_{{\scriptscriptstyle N}})$ with fixed experimental uncertainties $\sigma_n \in {{\bm{\sigma}}}$. The immediate aim is to find joint distributions (posterior, evidence etc) of the $K$ coefficients $\alpha_k \in {{\bm{\alpha}}}_{{\scriptscriptstyle K}}= (\alpha_1,\ldots,\alpha_{{\scriptscriptstyle K}})$ of a linear model function $$\begin{aligned}
\label{lry}
y(x{\,|\,}{{\bm{\alpha}}}_{{\scriptscriptstyle K}},K) &= \sum_{k=1}^K f_k(x)\,\alpha_k,\end{aligned}$$ where the choice of basis functions $f_k(x), k=1,\ldots, K$ forms part of the model specification and we subscript model-dependent quantities by $K$ in preparation for the extensions of Section \[sec:nsp\]. By assumption, the differences $\varepsilon_n \equiv y_n - y(x_n{\,|\,}{{\bm{\alpha}}}_{{\scriptscriptstyle K}},K)$ between each data point and the corresponding model point are normally distributed, $$\begin{aligned}
\label{bmep}
p({{\bm{\varepsilon}}}|{{\bm{\sigma}}})
&= \prod_{n=1}^N \frac{e^{-\varepsilon_n^2/2\sigma_n^2} }{\sigma_n\sqrt{2\pi}},\end{aligned}$$ resulting in a joint likelihood $$\begin{aligned}
\label{likz}
L\left[{{\bm{\alpha}}}_{{\scriptscriptstyle K}}\right]
= p({{\bm{y}}}{\,|\,}{{\bm{\alpha}}}_{{\scriptscriptstyle K}},K)
&= \prod_{n=1}^N \frac{1}{\sigma_n\sqrt{2\pi}}
\exp\left[-\frac{1}{2}\left(
\frac{y_n}{\sigma_n} - \sum_{k=1}^K \frac{f_k(x_n)}{\sigma_n}\alpha_k\right)^2\,\right].\end{aligned}$$ The $K$-dimensional model space ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$ is spanned by $N$-dimensional basis vectors $$\begin{aligned}
{{\bm{v}}}_k^{{\scriptscriptstyle\sf T}}&= [f_k(x_1)/\sigma_1,\ldots,f_k(x_{{\scriptscriptstyle N}})/\sigma_{{\scriptscriptstyle N}}]\qquad k = 1,\ldots,K,\end{aligned}$$ which together constitute the $(N{\times}K)$-dimensioned design matrix ${\mathbb{X}}_{{\scriptscriptstyle K}}= [{{\bm{v}}}_1,\ldots,{{\bm{v}}}_{{\scriptscriptstyle K}}]$. In terms of the standardised data vector ${{\bm{z}}}=[y_1/\sigma_1,\ldots,y_{{\scriptscriptstyle N}}/\sigma_{{\scriptscriptstyle N}}]$ and collecting constants into $C = (2\pi)^{-N/2}[\textstyle\prod_n\sigma_n]^{-1}$, the likelihood can be written in three ways, $$\begin{aligned}
\label{lika}
L\left[{{\bm{\alpha}}}_{{\scriptscriptstyle K}}\right]
&= C \exp\left[-\tfrac{1}{2}
\left({{\bm{z}}}{-}{\mathbb{X}}_{{\scriptscriptstyle K}}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}\right)^{{\scriptscriptstyle\sf T}}\!
\left({{\bm{z}}}{-}{\mathbb{X}}_{{\scriptscriptstyle K}}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}\right)\right]
\\
&= C \exp\left[-\tfrac{1}{2} ({{\bm{z}}}{-} {{\bm{f}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{z}}}{-} {{\bm{f}}}_{{\scriptscriptstyle K}}) \right]
\ =\ C \exp\left[-\tfrac{1}{2} {{\bm{\chi}}}^{{\scriptscriptstyle\sf T}}{{\bm{\chi}}}\right]\end{aligned}$$ where ${{\bm{f}}}_{{\scriptscriptstyle K}}= {\mathbb{X}}_{{\scriptscriptstyle K}}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}= \sum_{k=1}^K {{\bm{v}}}_k\alpha_k$ is the model-dependent vector aspiring to approximate the data vector ${{\bm{z}}}$ and ${{\bm{\chi}}}= {{\bm{z}}}- {{\bm{f}}}_{{\scriptscriptstyle K}}$ is the discrepancy between data and model. Description of the data ${{\bm{z}}}$ is thereby decomposed into a “noise” component ${{\bm{\chi}}}$ and a “signal” component ${{\bm{f}}}_{{\scriptscriptstyle K}}$. As illustrated in Figure \[fig:dspace\], ${{\bm{\chi}}}$ and ${{\bm{f}}}_{{\scriptscriptstyle K}}$ are in general not orthogonal. The length of the minimum-chisquared vector $\hat{{{\bm{\chi}}}}$ represents the minimum distance between the data vector ${{\bm{z}}}$ and model space ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$, so that it is orthogonal to model space ${{\bm{f}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{\chi}}}} = 0$ for all ${{\bm{f}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}$. The resulting maximum-signal vector can be found directly from $$\begin{aligned}
\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}&= {\mathbb{X}}_{{\scriptscriptstyle K}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}= \textstyle\sum_{k=1}^K {{\bm{v}}}_k{\bm{\hat{\alpha}}}_k\end{aligned}$$ where the maximum-likelihood parameter vector is determined by the usual Moore-Penrose inverse $$\begin{aligned}
\label{ptg}
{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}&= {\mathbb{H}}_{{\scriptscriptstyle K}}^{-1}{\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{z}}},\end{aligned}$$ with ${\mathbb{H}}_{{\scriptscriptstyle K}}= {\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\mathbb{X}}_{{\scriptscriptstyle K}}$ the Hessian with elements $({\mathbb{H}})_{k k'} = {{\bm{v}}}_k^{{\scriptscriptstyle\sf T}}{{\bm{v}}}_{k'} = \sum_n
f_k(x_n)f_{k'}(x_n)/\sigma_n^2$. Since ${{\bm{z}}}= \hat{{{\bm{\chi}}}} + \hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}$, the squared data vector $z^2 = {{\bm{z}}}^{{\scriptscriptstyle\sf T}}{{\bm{z}}}$ can hence be written as the Pythagorean sum of the usual minimum chisquared $\chi^2 = \hat{{{\bm{\chi}}}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{\chi}}}}$ and the squared signal vector $F_{{\scriptscriptstyle K}}^2 = \hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}$, $$\begin{aligned}
\label{pth}
z^2 &= \chi^2 + F_{{\scriptscriptstyle K}}^2.\end{aligned}$$ Following the usual diagonalisation by orthonormal eigenvector matrix ${\mathbb{S}}_{{\scriptscriptstyle K}}$ and rescaling by diagonal eigenvalue matrix ${\mathbb{L}}_{{\scriptscriptstyle K}}$ and transforming to hyperspherical parameters ${{\bm{\beta}}}_{{\scriptscriptstyle K}}= {\mathbb{L}}_{{\scriptscriptstyle K}}^{\!\!1/2}{\mathbb{S}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\alpha}}}_{{\scriptscriptstyle K}}$, and corresponding modes ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}= {\mathbb{L}}_{{\scriptscriptstyle K}}^{\!\!1/2}{\mathbb{S}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}$, the likelihood becomes $$\begin{aligned}
\label{likb}
L\left[{{\bm{\alpha}}}_{{\scriptscriptstyle K}}\right]
&= C \exp\left[-\tfrac{1}{2}\chi^2 -\tfrac{1}{2}
({{\bm{\alpha}}}_{{\scriptscriptstyle K}}{-}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}{\mathbb{H}}_{{\scriptscriptstyle K}}({{\bm{\alpha}}}_{{\scriptscriptstyle K}}{-}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}})\right],
\\
\label{likc}
L\left[{{\bm{\beta}}}_{{\scriptscriptstyle K}}\right]
&= C \exp\left[-\tfrac{1}{2}\chi^2
-\tfrac{1}{2}
({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})\right],\end{aligned}$$ and the squared signal vector transforms to $$\begin{aligned}
\label{rsq}
F_{{\scriptscriptstyle K}}^2
&
= \hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}= {\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\mathbb{H}}_{{\scriptscriptstyle K}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}= {{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}= \sum_{k=1}^K \hat{\beta}_k^2.\end{aligned}$$ All of the above is the standard fare of linear regression.
(0,6) node (yaxis) \[above\] [$e_2$]{} |- (10,0) node (xaxis) \[right\] [$e_1$]{}; (0,0)–(-2,2) node (zaxis) \[above\] [$e_3$]{}; (0,0) – (4,5) node \[pos=1,above\] [$\bm{z} = \mbox{data}$]{}; (5.2,2.6) – (4,5) node \[pos=0.6,right\] [$\hat{{{\bm{\chi}}}}$]{}; (0,0) – (5.2,2.6) node \[pos=0.95,below\] [$\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}$]{}; (3.5,1.75) – (4,5) node \[pos=0.5,left\] [${{\bm{\chi}}}$]{}; (0,0)–(3.5,1.75) node \[pos=0.65,above\] [${{\bm{f}}}_{{\scriptscriptstyle K}}$]{}; (3.55,2.025) arc (90:190:0.35) node\[pos=0.6,above\] [$\theta$]{}; (5,2.5) – (4.9,2.7); (4.9,2.7) – (5.1,2.8); in [1,...,8]{} (-0.866,/2+0.5) – (+0.866,/2-0.5); (0,0) – (8,4); (1-0.866,1) – (8-0.866,4.5); (1+0.866,0) – (8+0.866,3.5) node \[pos=0.5,below\] [${{\mathcal{A}}}_{{\scriptscriptstyle K}}$]{}; (9.2,1.6) – (6.8,6.4) node \[pos=1.0,above\] [${{\mathcal{A}}}_{{\scriptscriptstyle L}}$]{}; (7.9,3.7)–(7.7,4.1); (7.7,4.1)–(7.9,4.2); (7.9,3.7)–(8.1,3.8);
Model space and noise space {#sec:nsp}
===========================
The expanded model-noise space introduced in this Section is best understood in the context of Akaike’s rederivation in a Bayesian framework of the [<span style="font-variant:small-caps;">bic</span>]{} and [<span style="font-variant:small-caps;">aic</span>]{} in [@akaike1978bayesian]. His central message was that both could be understood by introducing, over and above the $K$ parameters ${{\bm{\beta}}}_{{\scriptscriptstyle K}}= (\beta_1,\ldots,\beta_{{\scriptscriptstyle K}})$ making up the model, an additional set of parameters ${{\bm{\beta}}}_{{\scriptscriptstyle L}}= (\beta_{{{\scriptscriptstyle K}}+1},\ldots,\beta_{{\scriptscriptstyle M}})$ with $K < M \leq N$, for which particular choices of priors yield the [<span style="font-variant:small-caps;">bic</span>]{} and [<span style="font-variant:small-caps;">aic</span>]{}. Details of the derivation are postponed to Section \[sec:prcn\].
The introduction of additional parameters in Akaike’s derivations allows the method to account for misspecified model functions in that if there is some signal left the noise parameters would be able to fit the shift in the residuals. The difference between a model ${{\mathcal{H}}}_{{\scriptscriptstyle K}}$ with $K$ parameters and another model ${{\mathcal{H}}}_{{{\scriptscriptstyle K}}+1}$ with $K{+}1$ parameters must then be found not in the existence or nonexistence of additional parameter $\beta_{{{\scriptscriptstyle K}}+1}$ but in different priors $p(\beta_{{{\scriptscriptstyle K}}+1}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})\neq p(\beta_{{{\scriptscriptstyle K}}+1}|{{\mathcal{H}}}_{{{\scriptscriptstyle K}}+1})$. In this view, all *model parameters* ${{\bm{\beta}}}_{{\scriptscriptstyle K}}$ should be assigned priors which allow them to exhibit large deviations from zero, while the additional *noise parameters* ${{\bm{\beta}}}_{{\scriptscriptstyle L}}$ should be assigned priors which are not exactly zero but restricted to small intervals around the origin.
Taking this line of thought to its logical conclusion, we let $M\equiv N$ and introduce $L=N-K$ additional noise parameters $\beta_\ell \in {{\bm{\beta}}}_{{\scriptscriptstyle L}}= (\beta_{{{\scriptscriptstyle K}}+1},\ldots,\beta_{{\scriptscriptstyle N}})$ along with $L$ additional basis functions $\{f_\ell(x)\}_{\ell=K{+}1}^N$ spanning what we shall call the *noise space* ${{\mathcal{A}}}_{{\scriptscriptstyle L}}$. While the mathematics does not preclude overlap, it seems natural to demand that model space (also called signal space) ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$ and noise space ${{\mathcal{A}}}_{{\scriptscriptstyle L}}$ partition the data space, $$\begin{aligned}
{{\mathcal{A}}}_{{\scriptscriptstyle N}}&= {{\mathcal{A}}}_{{\scriptscriptstyle K}}\cup{{\mathcal{A}}}_{{\scriptscriptstyle L}},
\qquad {{\mathcal{A}}}_{{\scriptscriptstyle K}}\cap{{\mathcal{A}}}_{{\scriptscriptstyle L}}= \emptyset.\end{aligned}$$ In this view, model construction is seen as a successive decomposition of the data space ${{\mathcal{A}}}_{{\scriptscriptstyle N}}$ into sequences of partitions $\{{{\mathcal{A}}}_{{\scriptscriptstyle K}},{{\mathcal{A}}}_{{\scriptscriptstyle L}}\}_{K=1}^N$ with progressively increasing $K$ and decreasing $L$, with model selection based on the maximum evidence or Bayes Factor as a function of $K$. The partitioning property can be enforced by constructing, if necessary by a Gram-Schmidt procedure, a set of *noise functions* $f_\ell(x)$ which are orthogonal to all model functions $f_k(x)$, $$\begin{aligned}
\sum_{n=1}^N f_k(x_n)\,f_\ell(x_n) &= 0
\qquad \forall\; k=1,\ldots,K \text{ and } \ell=K{+}1,\ldots,N,\end{aligned}$$ thereby ensuring[^1] that the Hessian of the complete basis set $\{f_k\}_{k=1}^N$ is block-diagonal, ${\mathbb{H}}_{{\scriptscriptstyle N}}= {\mathbb{H}}_{{\scriptscriptstyle K}}{\oplus}{\mathbb{H}}_{{\scriptscriptstyle L}}$. We note that the basis functions $f_\ell(x)$ may have to be adapted as $K$ changes to safeguard block-diagonality. The resulting sequence of models is therefore not nested in the strict sense.
As already mentioned in Section \[sec:lrg\], ${{\bm{f}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{\chi}}}} = 0 $ for all ${{\bm{f}}}_{{\scriptscriptstyle K}}$ because, with the help of Eq. (\[ptg\]), ${{\bm{\alpha}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}({{\bm{z}}}- {\mathbb{X}}_{{\scriptscriptstyle K}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}) =
{{\bm{\alpha}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}({{\bm{z}}}-
{\mathbb{X}}_{{\scriptscriptstyle K}}{\mathbb{H}}_{{\scriptscriptstyle K}}^{-1}{\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{z}}}) = 0$. Together with ${{\mathcal{A}}}_K\cap{{\mathcal{A}}}_L=\emptyset$, this means that that $\hat{{{\bm{\chi}}}}$ is a vector in ${{\mathcal{A}}}_{{\scriptscriptstyle L}}$ and hence has a representation in the noise-space basis ${{\bm{v}}}_\ell = [f_\ell(x_1)/\sigma_1,\ldots,f_\ell(x_{{\scriptscriptstyle N}})/\sigma_{{\scriptscriptstyle N}}]$ with coefficients ${\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}= (\hat{\alpha}_{{{\scriptscriptstyle K}}+1},\ldots,\hat{\alpha}_{{\scriptscriptstyle N}})$ or equivalently in terms of the noise-space design matrix ${\mathbb{X}}_{{\scriptscriptstyle L}}= [{{\bm{v}}}_{{{\scriptscriptstyle K}}{+}1},\ldots,{{\bm{v}}}_{{\scriptscriptstyle N}}]$, $$\begin{aligned}
\hat{{{\bm{\chi}}}} &= \sum_{\ell={{\scriptscriptstyle K}}+1}^N {{\bm{v}}}_\ell\,\hat{\alpha}_\ell
= {\mathbb{X}}_{{\scriptscriptstyle L}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}.\end{aligned}$$ Diagonalisation by ${\mathbb{S}}_{{\scriptscriptstyle L}}$ and rescaling by ${\mathbb{L}}_{{\scriptscriptstyle L}}$ in noise space results in $\hat{{{\bm{\chi}}}} = {\mathbb{X}}_{{\scriptscriptstyle L}}{\mathbb{S}}_{{\scriptscriptstyle L}}{\mathbb{L}}_{{\scriptscriptstyle L}}^{-1/2}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}$ with ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}= {\mathbb{L}}_{{\scriptscriptstyle L}}^{1/2}{\mathbb{S}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}$ just as in model space, so that, in close analogy with Eq. (\[rsq\]), $$\begin{aligned}
\label{qkb}
\chi^2
&
= \hat{{{\bm{\chi}}}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{\chi}}}}
= {\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{\mathbb{H}}_{{\scriptscriptstyle L}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}= {{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}= \sum_{\ell=K+1}^{N} \hat{\beta}^2_\ell.\end{aligned}$$ Apart from the requirement that the basis functions $f_\ell(x)$ must be orthogonal to those in model space, their specific choice is arbitrary; correspondingly, individual coefficients ${{\bm{\alpha}}}_{{\scriptscriptstyle L}}$, ${{\bm{\beta}}}_{{\scriptscriptstyle L}}$ and their maximum-likelihood cases ${\bm{\hat{\alpha}}}_{{\scriptscriptstyle L}}$ and ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}$ are not fixed by the model. All that matters is that $\chi^2$ can be written as a sum of $L=N{-}K$ squared components $\hat{\beta}_\ell$.
The extension from ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$ to the larger space ${{\mathcal{A}}}_{{\scriptscriptstyle K}}\cup{{\mathcal{A}}}_{{\scriptscriptstyle L}}$ has an important consequence for the likelihood and evidence. Rather than using the conventional $K$-dimensional version which would result from Eq. (\[likb\]), $$\begin{aligned}
\label{likd}
L[{{\bm{\beta}}}_{{\scriptscriptstyle K}}]
&= p({{\bm{y}}}{\,|\,}{{\bm{\beta}}}_{{\scriptscriptstyle K}},{{\mathcal{H}}})
= C \exp\left[-\tfrac{1}{2}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}-\tfrac{1}{2}
({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})\right],\end{aligned}$$ the limited model $\sum_{k=1}^K f_k(x)\alpha_k$ in the likelihood of Eq. (\[likz\]) is replaced with the full set, $$\begin{aligned}
\label{likf}
L\left[{{\bm{\alpha}}}_{{\scriptscriptstyle K}},{{\bm{\alpha}}}_{{\scriptscriptstyle L}}\right]
&= C \exp\left[-\frac{1}{2}\sum_{n=1}^N\left(
\frac{y_n}{\sigma_n}
- \sum_{k=1}^K \frac{f_k(x_n)}{\sigma_n}\alpha_k
- \sum_{\ell=K+1}^N \frac{f_\ell(x_n)}{\sigma_n}\alpha_\ell\right)^2
\,\right],\end{aligned}$$ which for block-diagonal ${\mathbb{H}}_{{\scriptscriptstyle N}}={\mathbb{H}}_{{\scriptscriptstyle K}}\oplus{\mathbb{H}}_{{\scriptscriptstyle L}}$ takes the form $$\begin{aligned}
\label{likg}
p({{\bm{y}}}{\,|\,}{{\bm{\beta}}}_{{\scriptscriptstyle K}},{{\bm{\beta}}}_{{\scriptscriptstyle L}})
=
L[{{\bm{\beta}}}_{{\scriptscriptstyle K}},{{\bm{\beta}}}_{{\scriptscriptstyle L}}]
&= C \exp\left[
-\tfrac{1}{2}
({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})
-\tfrac{1}{2}
({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})\right],\end{aligned}$$ for which the evidence factorises into noise and signal parts, $$\begin{aligned}
\label{evxt}
p({{\bm{y}}}{\,|\,}K,{{\mathcal{H}}})
&= \int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}\,{\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle L}}\, L[{{\bm{\beta}}}_{{\scriptscriptstyle K}},{{\bm{\beta}}}_{{{\scriptscriptstyle L}}}]\,
p({{\bm{\beta}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})\,p({{\bm{\beta}}}_{{{\scriptscriptstyle L}}}|{{\mathcal{H}}}_{{\scriptscriptstyle L}})
\nonumber\\
&= C
\left\{\int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle L}}\,
\exp\left[-\tfrac{1}{2} ({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})\right]
\,p({{\bm{\beta}}}_{{\scriptscriptstyle L}}|{{\mathcal{H}}}_{{\scriptscriptstyle L}}) \right\}
\nonumber\\
&\quad\times
\left\{\int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}\,
\exp\left[-\tfrac{1}{2}
({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})\right]
\,p({{\bm{\beta}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})
\right\}.\end{aligned}$$ We briefly consider the case $K{=}N, L{=}0$ as this constitutes our reference model for Bayes Factors. The design matrix ${\mathbb{X}}_{{\scriptscriptstyle N}}$ is invertible so that ${\bm{\hat{\alpha}}}_{{\scriptscriptstyle N}}= {\mathbb{X}}_{{\scriptscriptstyle N}}^{-1}{{\bm{z}}}$, all the data becomes signal (${{\bm{z}}}= \hat{{{\bm{f}}}}_{\scriptscriptstyle K=N}$, $z^2 = F_{{\scriptscriptstyle N}}^2$), there is no noise ($\hat{{{\bm{\chi}}}}_{\scriptscriptstyle{L} =0}=0$), and the model amounts to a change of basis for ${{\mathcal{A}}}_{{\scriptscriptstyle N}}$ from ${{\bm{v}}}_{{\scriptscriptstyle N}}$ to ${{\bm{\alpha}}}_{{\scriptscriptstyle N}}$.
Akaike’s BIC/AIC priors in model and noise space {#sec:prcn}
================================================
We now rederive Akaike’s central insight in the language of model and noise space. The [<span style="font-variant:small-caps;">bic</span>]{} results when the priors for the model parameters $\beta_k,k=1,\ldots,K$ are normally distributed with variance $N$, while the additional parameters $\beta_\ell, \ell = K{+}1,\ldots,N$ are set to zero exactly by means of Dirac delta functions,[^2] $$\begin{aligned}
\label{ccb}
& p(\beta_k|{{{\mathcal{H}}}_{\textsc{bic}}}) = \frac{1}{\sqrt{2\pi N}} e^{-\beta_k^2/2N},
&& p(\beta_\ell|{{{\mathcal{H}}}_{\textsc{bic}}}) = \delta(\beta_{\ell}).\end{aligned}$$ The evidence (\[evxt\]) and corresponding $K{=}N$ reference evidence are then $$\begin{aligned}
\label{ccc}
p({{\bm{y}}}{\,|\,}K,{{{\mathcal{H}}}_{\textsc{bic}}})
&= C (N{+}1)^{-K/2}
\exp\left[-\frac{\chi^2}{2} - \frac{F_{{\scriptscriptstyle K}}^2}{2(N{+}1)} \right],
\\
\label{cce}
p({{\bm{y}}}{\,|\,}N,{{{\mathcal{H}}}_{\textsc{bic}}})
&= C (N{+}1)^{-N/2}
\exp\left[- \frac{\chi^2 + F_{{\scriptscriptstyle K}}^2}{2(N{+}1)} \right],\end{aligned}$$ yielding a Bayes Factor $$\begin{aligned}
\label{ccf}
{\mathrm{BF}}[K,N]
&= (N{+}1)^{(N-K)/2} \exp\left[-\frac{N \chi^2}{2(N{+}1)} \right].\end{aligned}$$ Dropping $K$-independent constants and assuming $N\gg 1$, we recover the BIC from the logarithm $$\begin{aligned}
\label{ccg}
-2\log {\mathrm{BF}}[K,N] &\ \simeq\ \chi^2 + K\log N \ =\ \mbox{{\textsc{bic}}}\end{aligned}$$ since $\chi^2 \propto -2\log$(maximum likelihood). In rederiving the [<span style="font-variant:small-caps;">aic</span>]{}, [@akaike1978bayesian] similarly suggested that model and noise parameters be treated on the same basis but with different scales for their priors, $$\begin{aligned}
\label{cch}
& p(\beta_k|{{{\mathcal{H}}}_{\textsc{aic}}}) = \frac{1}{\sqrt{2\pi \Delta^2}} e^{-\beta_k^2/2\Delta^2},
&& p(\beta_\ell|{{{\mathcal{H}}}_{\textsc{aic}}}) = \frac{1}{\sqrt{2\pi \delta^2}} e^{-\beta_\ell^2/2 \delta^2}\end{aligned}$$ resulting in evidence, reference evidence and Bayes Factor $$\begin{aligned}
\label{cci}
p({{\bm{y}}}{\,|\,}K,{{{\mathcal{H}}}_{\textsc{aic}}})
&= \frac{C}
{\left[1{+}\Delta^2\right]^{K/2} \left[1{+}\delta^2\right]^{L/2}}
\exp\left[-\frac{\chi^2}{2(1{+}\delta^2)} - \frac{F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)} \right],
\\
\label{ccj}
p({{\bm{y}}}{\,|\,}N,{{{\mathcal{H}}}_{\textsc{aic}}})
&= \frac{C}
{\left[1{+}\Delta^2\right]^{N/2} }
\exp\left[-\frac{\chi^2 {+} F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)} \right],
\\
BF[K,N]
&= \left[\frac{1{+}\Delta^2}{1{+}\delta^2}\right]^{L/2}
\exp\left[-\left(\frac{1}{1{+}\delta^2} -\frac{1}{1{+}\Delta^2} \right)\frac{\chi^2}{2} \right],\end{aligned}$$ and information criterion $$\begin{aligned}
\label{cck}
-2 \log BF[K,N]
&= \left(\frac{1}{1{+}\delta^2} -\frac{1}{1{+}\Delta^2} \right)
\chi^2 - \left(N{-}K\right) \log \frac{1{+}\Delta^2}{1{+}\delta^2}.\end{aligned}$$ At this point, Akaike argued that $\delta$ and $\Delta$ should approach 1 from above and below, where 1 represents the situation of equal signal and noise magnitude. This is the critical case where it is difficult to distinguish between model and noise and the Bayes Factor will tend to zero. To find the next to leading order behaviour of the Bayes Factor we as Akaike take the limit $\delta\rightarrow 1^{-}$ and $\Delta\rightarrow 1^{+}$, and recover the [<span style="font-variant:small-caps;">aic</span>]{}, $$\begin{aligned}
\label{ccl}
\lim_{\substack{\delta\rightarrow 1\\\Delta\rightarrow 1}}
\left(\frac{1}{1+\delta^2} -\frac{1}{1+\Delta^2} \right)^{\!\!-1}
\!\!\left( -2\log BF[K,N]\right)
&\ =\ \chi^2 + 2K \ =\ \mathrm{{\textsc{aic}}}.\end{aligned}$$
Figure \[fig:schema\] summarises schematically the generic form of the data and model and noise parameter priors for the various information criteria. If the true behaviour of the system resulted from some true model with ${{S}}$ parameters plus noise, the data would correspond to ${{S}}$ non-zero parameters $\hat{\beta}_k$ and $N{-}{{S}}$ near-null parameters, both within some uncertainty range; this is sketched in the two leftmost columns. The third and fourth columns in Fig. \[fig:schema\] remind us that [<span style="font-variant:small-caps;">bic</span>]{} and [<span style="font-variant:small-caps;">aic</span>]{} set priors for both model and noise parameters centered around zero, differing only in the scale of the variation around zero for the noise parameters. The [<span style="font-variant:small-caps;">bic</span>]{} conflates probabilistic intervals for model parameters with point probabilities for noise parameters, a strong assumption which reduces its effectiveness for weak-signal cases. While consistently using intervals for all parameters, the [<span style="font-variant:small-caps;">aic</span>]{}fails to take account of the fact that model parameters will usually not be centered around zero. Consistent with the generic data behaviour, the robust version of [<span style="font-variant:small-caps;">bic</span>]{} displayed in the last two columns explicitly shifts model parameter priors away from zero with the help of a hyperparameter $\gamma$ and consistently uses intervals for all.
Noncentral radial priors and information criterion {#sec:nic}
==================================================
The lesson of Fig. \[fig:schema\] is that properties and performance of information criteria depend crucially both on their treatment of noise parameters and the location of the model parameters priors’ modes. Seen from this perspective, a generic weakness of the [<span style="font-variant:small-caps;">aic</span>]{} and [<span style="font-variant:small-caps;">bic</span>]{} is self-evident: their model parameter priors are maximal near zero, while the likelihood will be maximal for nonzero values of ${{\bm{\beta}}}_{{\scriptscriptstyle K}}$, resulting in poor overlap of prior and likelihood. This is exactly the problem which the Empirical Bayes criterion tries to correct, albeit in a nonrigorous way [@george2000calibration].
There is a good reason, of course, to maximise these priors around the origin. By definition, the state of knowledge embodied in a prior excludes the location of the data-dependent maximum ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}$, so that, barring supplementary prior information, the origin becomes the preferred mode for $p({{\bm{\beta}}}_{{\scriptscriptstyle K}}|{{\mathcal{H}}}_{{\scriptscriptstyle K}})$. However, while ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}$ itself may not be used, we can and should take into consideration the generic fact that for a good model the parameters ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}$ will be nonzero. We do not know *where* in ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$ the ${{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}$ is located, but we do know that the *posterior model radius* $({{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{})^{1/2}$ is significantly nonzero for any and all data, which implies that the prior for what we call the *prior model radius* ${\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle K}}{\parallel} =
({{\bm{\beta}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}_{{\scriptscriptstyle K}})^{1/2}$ should be chosen to be significantly nonzero. Moreover, the identity $\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{} = {{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{}$ in Eq. (\[rsq\]) and generally ${{\bm{f}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{f}}}_{{\scriptscriptstyle K}}^{} =
{{\bm{\beta}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}^{}$ imply that *the same model radius* ${\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle K}}{\parallel}$ sets the scale both in model space ${{\mathcal{A}}}_{{\scriptscriptstyle K}}$ and in the corresponding parameter space ${{\mathcal{A}}}_{\beta,{{\scriptscriptstyle K}}}$. Likewise, we have generic knowledge that a good model will be characterised by a near-zero *posterior noise radius* ${\parallel}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}{\parallel} = ({{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})^{1/2}$ for which a near-zero *prior noise radius* ${\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle L}}{\parallel} =
({{\bm{\beta}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}_{{\scriptscriptstyle L}})^{1/2}$ is of course appropriate, and *the same noise radius* $q = {\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle L}}{\parallel}$ sets the scale in both noise space ${{\mathcal{A}}}_{{\scriptscriptstyle L}}$ and its parameter space ${{\mathcal{A}}}_{\beta,{{\scriptscriptstyle L}}}$. These considerations are compactly summarised in the last column of Fig. \[fig:schema\] and in the equation set
----------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Model Noise
Prior ${\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle K}}{\parallel} ${\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle L}}{\parallel}
= \sqrt{{{\bm{\beta}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}^{}{\mbox{\rule[0pt]{0pt}{9pt}}}} = \sqrt{{{\bm{f}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{f}}}_{{\scriptscriptstyle K}}^{}{\mbox{\rule[0pt]{0pt}{9pt}}}} > 0 $ = \sqrt{{{\bm{\beta}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}_{{\scriptscriptstyle L}}^{}{\mbox{\rule[0pt]{0pt}{9pt}}}} = \sqrt{{{\bm{\chi}}}^{{\scriptscriptstyle\sf T}}{{\bm{\chi}}}^{}{\mbox{\rule[0pt]{0pt}{9pt}}}} \simeq 0$
Posterior ${\parallel}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}{\parallel} ${\parallel}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}{\parallel}
= \sqrt{{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}^{}} = \sqrt{\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{f}}}}_{{\scriptscriptstyle K}}^{}} > 0$ = \sqrt{{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}^{}} = \sqrt{\hat{{{\bm{\chi}}}}^{{\scriptscriptstyle\sf T}}\hat{{{\bm{\chi}}}}^{}{\mbox{\rule[0pt]{0pt}{9pt}}}} \simeq 0$
----------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
All of this constitutes prior knowledge without reference to the particulars of the data. Crucially, this knowledge pertains to the *radii*. The model evidence is therefore expanded in terms of two radial parameters $q$ and $r$ and two radial priors, $$\begin{aligned}
\label{ncc}
p({{\bm{y}}}{\,|\,}{{\mathcal{H}}}_{{\scriptscriptstyle K}}) &= \int_0^\infty \!{\mathrm{d}}r\,{\mathrm{d}}q \, p(r{\,|\,}{{\mathcal{H}}}_r)\,p(q{\,|\,}{{\mathcal{H}}}_q) \,p({{\bm{y}}}{\,|\,}r,q,{{\mathcal{H}}}_{{\scriptscriptstyle K}}),\end{aligned}$$ where the evidence conditioned on $r$ and $q$ is $$\begin{aligned}
\label{ncd}
p({{\bm{y}}}{\,|\,}r,q,{{\mathcal{H}}}_{{\scriptscriptstyle K}}) &= \int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}\,{\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle L}}\,
p({{\bm{y}}}{\,|\,}{{\bm{\beta}}}_{{\scriptscriptstyle K}},{{\bm{\beta}}}_{{\scriptscriptstyle L}})
\,p({{\bm{\beta}}}_{{\scriptscriptstyle K}}| r)
\,p({{\bm{\beta}}}_{{\scriptscriptstyle L}}| q).\end{aligned}$$ Given the factorised likelihood (\[likg\]), the conditioned evidence also factorises, $$\begin{aligned}
\label{nce}
p({{\bm{y}}}{\,|\,}r,q,{{\mathcal{H}}}_{{\scriptscriptstyle K}})
&= p({{\bm{y}}}{\,|\,}q,{{\mathcal{H}}}_{{\scriptscriptstyle K}}) \, p({{\bm{y}}}{\,|\,}r,{{\mathcal{H}}}_{{\scriptscriptstyle K}})
\\
&= \int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle L}}\, p({{\bm{\beta}}}_{{\scriptscriptstyle L}}| q)\,
e^{-({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle L}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}})/2}
\int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}\, p({{\bm{\beta}}}_{{\scriptscriptstyle K}}| r)\,
e^{-({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})^{{\scriptscriptstyle\sf T}}({{\bm{\beta}}}_{{\scriptscriptstyle K}}-{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}})/2}.
\nonumber\end{aligned}$$ With only radial information available, the prior for ${{\bm{\beta}}}_{{\scriptscriptstyle K}}$ must be uniform on the $K$-hypersphere, $$\begin{aligned}
\label{ncg}
p({{\bm{\beta}}}_{{\scriptscriptstyle K}}{\,|\,}r)
&= \frac{\Gamma(K/2)\,\delta(r - {\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle K}}{\parallel})}
{2\pi^{K/2}\,r^{K-1}},\end{aligned}$$ with the Dirac delta function constraining the vector ${{\bm{\beta}}}_{{\scriptscriptstyle K}}$ to the surface of the $K$-sphere of radius $r$, while spherical symmetry on the $L$-hypersphere with radius $q$ requires $$\begin{aligned}
\label{nch}
p({{\bm{\beta}}}_{{\scriptscriptstyle L}}{\,|\,}q)
&= \frac{\Gamma(L/2)\,\delta(q - {\parallel}{{\bm{\beta}}}_{{\scriptscriptstyle L}}{\parallel})}
{2\pi^{L/2}\,q^{L-1}}.\end{aligned}$$ As shown in [@dekock2017bayesian], the conditioned evidence (\[ncd\]) can be expressed in closed form, $$\begin{aligned}
\label{nci}
p({{\bm{y}}}{\,|\,}r,q,{{\mathcal{H}}}_{{\scriptscriptstyle K}})
&= C
\exp\left[-\frac{1}{2}\left(F_{{\scriptscriptstyle K}}^2 {+} r^2\right)\right]
{}_0F_1\left[\frac{K}{2}\bigg| \frac{r^2 F_{{\scriptscriptstyle K}}^2}{4} \right]
\nonumber\\
&\quad \times
\exp\left[-\frac{1}{2}\left(\chi^2 {+} q^2\right)\right]
{}_0F_1\left[\frac{L\,}{2}\bigg| \frac{q^2\chi^2}{4} \right].\end{aligned}$$ We now turn to the radial priors. To capture the generic information $|\beta_{{\scriptscriptstyle L}}| \simeq 0$, we presuppose that ${{\bm{\beta}}}_{{\scriptscriptstyle L}}\sim {{\mathcal N}}(0,{\mathbb{I}}_{{\scriptscriptstyle L}}\delta^2)$, a normal distribution with zero mode and variance $\delta^2$, so that the prior for the radius $q$ is a chi-squared distribution or Gamma Distribution in $q^2/2\delta^2$ with hyperparameter $\delta$ and $L = N-K$ as usual, $$\begin{aligned}
\label{ncj}
p(q{\,|\,}\delta,L)
&= \frac{q}{\delta^2}\left(\frac{q^2}{2\delta^2}\right)^{(L/2)-1}
\frac{e^{-q^2/2\delta^2}}{\Gamma(L/2)}.\end{aligned}$$ Likewise projecting a $K$-dimensional normal distribution $p({{\bm{\beta}}}_{{\scriptscriptstyle K}}|\Delta,{{\bm{\mu}}}) =
{{\mathcal N}}({{\bm{\mu}}},{\mathbb{I}}_{{\scriptscriptstyle K}}\Delta^2)$ with nonzero mode ${{\bm{\mu}}}=
(\mu_1,\ldots,\mu_{{\scriptscriptstyle K}})$ and variance $\mathbb{I}_{{\scriptscriptstyle K}}\Delta^2$ onto radius $r$ results in a noncentral Gamma Distribution with radial hyperparameter $\gamma = [\sum_k \mu_k^2]^{1/2}$, $$\begin{aligned}
\label{nck}
p(r{\,|\,}\Delta,\gamma,K)
&= \frac{r}{\Delta^2} \left(\frac{r^2}{2\Delta^2}\right)^{(K/2)-1}
\frac{e^{-(r^2+\gamma^2)/2\Delta^2}}
{\Gamma(K/2)}
\;\; {}_0F_1\left[\frac{K}{2}\bigg| \frac{\gamma^2 r^2}{4\Delta^4}\right].\end{aligned}$$ Later, we shall interpret $\gamma$ as a signal-to-noise ratio, and with $\gamma\to 0$, $p(r{\,|\,}\Delta,\gamma,K)$ consistently reverts to the ordinary Gamma Distribution characteristic of noise. The noncentral Gamma Distribution results from the projection of $\mathcal{N}({{\bm{\mu}}},{\mathbb{I}}_{{\scriptscriptstyle K}}\Delta^2)$ onto the squared radius $p(r^2| \Delta,{{\bm{\mu}}}) = \int {\mathrm{d}}\,{{\bm{\beta}}}_{{\scriptscriptstyle K}}p(r^2|{{\bm{\beta}}}_{{\scriptscriptstyle K}})\, p({{\bm{\beta}}}_{{\scriptscriptstyle K}}| {{\bm{\mu}}},\Delta)$ and using the integral representation of the Dirac delta function $$\begin{aligned}
\label{ncl}
p(r^2{\,|\,}{{\bm{\beta}}}_{{\scriptscriptstyle K}}) =
\delta(r^2 - \textstyle\sum_k\beta_k^2)
&= \frac{1}{2\pi i}
\int_{-i\infty}^{i\infty} {\mathrm{d}}s\,\exp[s(r^2 - \textstyle\sum_k\beta_k^2)],\end{aligned}$$ whereby $$\begin{aligned}
\label{ncm}
p(r^2{\,|\,}\Delta,\gamma)
&= \frac{1}{(2\pi\Delta^2)^{K/2}}
\int {\mathrm{d}}{{\bm{\beta}}}_{{\scriptscriptstyle K}}\,\delta(r^2 - {\textstyle\sum_k}\beta_k^2)\,
\exp\left[-\sum_k\frac{(\beta_k-\mu_k)^2}{2\Delta^2} \right]
\\
\label{ncn}
&= \frac{1}{2\pi i}
\int_{-i\infty}^{i\infty} {\mathrm{d}}s\,[1 + 2s\Delta^2]^{-K/2}
\,\exp\left[-\frac{s\gamma^2}{(1{+}2s\Delta^2)}\right],\end{aligned}$$ becomes the Noncentral Gamma Distribution with the help of [@bateman_vol2] $$\begin{aligned}
\label{nco}
{}_0F_1[b;z]
&= \frac{\Gamma(b)}{2\pi i}
\int_{-i\infty}^{+i\infty} {\mathrm{d}}u\, u^{-b} \exp\left[u + \frac{z}{u}\right].\end{aligned}$$ Inserting Eqs. (\[nci\]), (\[ncj\]) and (\[nck\]) into (\[ncc\]) and using [@bateman_vol2] $$\begin{aligned}
\label{ncp}
\int_0^\infty \frac{e^{-r}r^{a-1}}{\Gamma[a]}
\; {}_0F_1[a ; x r] \; {}_0F_1[a ; y r] \; {\mathrm{d}}r
&= e^{x+y} \, {}_0F_1[a; xy],\end{aligned}$$ the evidence is found to be $$\begin{aligned}
\label{ncq}
p({{\bm{y}}}{\,|\,}\delta,\Delta,\gamma,K) &= \frac{C}{
(1+\delta^2)^{L/2}\,
(1+\Delta^2)^{K/2}
}\,
\exp\left[
-\frac{\chi^2}{2(1+\delta^2)}
-\frac{(F_{{\scriptscriptstyle K}}^2\Delta^2 + \gamma^2)}{2\Delta^2}
\right]
\nonumber\\
&\quad\times
\Psi_{(2)}\left[
\frac{K}{2}, \frac{K}{2}, \frac{K}{2}
\; ;\;
\frac{\Delta^2 F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)}
\; , \;
\frac{\gamma^2}{2\Delta^2(1{+}\Delta^2)}\right],\end{aligned}$$ where the Humbert function is defined in terms of Pochhammer symbols $(x)_y = \Gamma(x{+}y)/\Gamma(x)$ as [@bateman_vol1] $$\begin{aligned}
\label{ncr}
\Psi_{(2)}(a,b,c; x,y)
&= \sum_{m,n=0}^\infty\frac{(a)_{m+n}}{(b)_m (c)_{n}} x^m y^n,\end{aligned}$$ which for equal arguments reduces to $\Psi_{(2)}(a,a,a; x,y) = e^{x+y}\,{}_0F_1(a;xy)$, and so $$\begin{aligned}
\label{ncs}
p({{\bm{y}}}{\,|\,}\delta,\Delta,\gamma,K) &= \frac{C}{
(1{+}\delta^2)^{L/2}\,
(1{+}\Delta^2)^{K/2}
}\,
\exp\left[
-\frac{\chi^2}{2(1{+}\delta^2)}
-\frac{(\gamma^2 {+} F_{{\scriptscriptstyle K}}^2)}{2(1{+}\Delta^2)}
\right]
{}_0F_1\left[\frac{K}{2}; \frac{\gamma^2 F_{{\scriptscriptstyle K}}^2}{4(1{+}\Delta^2)^2}\right].\end{aligned}$$ Unlike the [<span style="font-variant:small-caps;">aic</span>]{} derivation, we have no reason to maintain the distinction between noise and model prior variances and can set $\delta = \Delta$, so the evidence reduces via Eq. (\[pth\]) to $$\begin{aligned}
\label{nct}
p({{\bm{y}}}{\,|\,}\Delta,\gamma,K) &= C (1{{+}}\Delta^2)^{-N/2}
\exp\left[-\frac{\gamma^2{+}z^2}{2(1{+}\Delta^2)}\right]
{}_0F_1\left[\frac{K}{2}; \frac{\gamma^2 F_{{\scriptscriptstyle K}}^2}{4(1{+}\Delta^2)^2}\right].\end{aligned}$$ If signal-to-noise ratios are known beforehand, $\gamma$ can be set to a fixed number; otherwise, it must remain indeterminate and integrated out. Aiming to have a maximally uniform but proper prior for $\gamma$, we use a half-Gaussian with arbitrarily large variance $\sigma_\gamma^2$, $$\begin{aligned}
\label{ncu}
p(\gamma{\,|\,}\sigma_\gamma)
&= \sqrt{\frac{2}{\sigma_\gamma^2\pi}}\exp\left[-\frac{\gamma^2}{2\sigma_\gamma^2}\right]
\qquad 0 < \gamma < \infty,\end{aligned}$$ yielding the evidence $$\begin{aligned}
\label{ncv}
p({{\bm{y}}}{\,|\,}\Delta,\sigma_\gamma,{{\mathcal{H}}}_{{\scriptscriptstyle K}})
&= \int {\mathrm{d}}\gamma\,p({{\bm{y}}}{\,|\,}\Delta,\gamma,{{\mathcal{H}}}_{{\scriptscriptstyle K}}) \,p(\gamma{\,|\,}\sigma_\gamma)
\\
&= C \frac{(1{{+}}\Delta^2)^{-(N-1)/2}}{\sqrt{1{+}\sigma_\gamma^2{+}\Delta^2}}
\exp\left[-\frac{z^2}{2(1{+}\Delta^2)}\right]
{}_1F_1\left[\frac{1}{2}; \frac{K}{2};
\frac{\sigma_\gamma^2 F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)(1{+}\Delta^2{+}\sigma_\gamma^2)}\right]
\nonumber\end{aligned}$$ and Bayes Factor $$\begin{aligned}
BF[K; N]
&= {}_1F_1\left[\frac{1}{2}; \frac{K}{2};
\frac{\sigma_\gamma^2 F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)(1{+}\Delta^2{+}\sigma_\gamma^2)}\right]
\biggl/ {}_1F_1\left[\frac{1}{2}; \frac{N}{2};
\frac{\sigma_\gamma^2 z^2}{2(1{+}\Delta^2)(1{+}\Delta^2{+}\sigma_\gamma^2)}\right].
\nonumber\end{aligned}$$ We can now take the limit $\sigma_\gamma\rightarrow\infty$ to obtain $$\begin{aligned}
\label{ncx}
BF[K; N]
&= {}_1F_1\left[\frac{1}{2}; \frac{K}{2}; \frac{ F_{{\scriptscriptstyle K}}^2}{2(1{+}\Delta^2)}\right]
\biggl/
{}_1F_1\left[\frac{1}{2}; \frac{N}{2}; \frac{z^2}{2(1{+}\Delta^2)}\right].\end{aligned}$$ The role of $\gamma$ has been to differentiate model and noise parameter behaviour. For finite $\Delta$, integration over both $\gamma$ in Eq. (\[ncv\]) and over $r$ in Eqs. (\[ncc\]) and (\[ncp\]) results, however, in redundancy which can safely be eliminated by letting $\Delta\rightarrow 0$: unlike the [<span style="font-variant:small-caps;">aic</span>]{}, our scales are set not by $\Delta$ but by $\gamma$ so we have no further need for it. The Bayes Factor hence simplifies to $$\begin{aligned}
\label{ncy}
BF[K; N]
&= {}_1F_1\left[\frac{1}{2}; \frac{K}{2}; \frac{ F_{{\scriptscriptstyle K}}^2}{2}\right]
\biggl/ {}_1F_1\left[\frac{1}{2}; \frac{N}{2}; \frac{z^2}{2}\right].\end{aligned}$$ Using the asymptotic properties of the confluent hypergeometric distribution, [@bateman_vol1], $$\begin{aligned}
\label{ncz}
{}_1F_1\left[a; c; x\right] \sim \frac{\Gamma[c]}{\Gamma[a]} e^{x}x^{a-c},\end{aligned}$$ we obtain for large $N$ a robust version of the [<span style="font-variant:small-caps;">bic</span>]{}, $$\begin{aligned}
\label{ndc}
\mathrm{{\textsc{bic}}_{(robust)}} = -2\log BF[K;N]
&\approx \chi^2 + (K-1)\log[F_{{\scriptscriptstyle K}}^2/2] - 2 \log \Gamma[K/2],\end{aligned}$$ which for large $K$ reduces further to $$\begin{aligned}
\label{ndf}
\mathrm{{\textsc{bic}}_{(robust)}} \simeq \chi^2 + K\log [F_{{\scriptscriptstyle K}}^2/K] + K.\end{aligned}$$ The three new forms of the [<span style="font-variant:small-caps;">bic</span>]{} in Eqs. (\[ncy\]), (\[ncz\]) and (\[ndf\]) are our central result.
We now show that, in the appropriate limits, the robust version approaches the [<span style="font-variant:small-caps;">aic</span>]{} and [<span style="font-variant:small-caps;">bic</span>]{}. The prior expectation value of $q^2$ for the Gamma Distribution (\[ncj\]) is $E[q^2] = L\delta^2/2$, while $E[r^2] = \gamma^2 + K
\Delta^2/2 $ for the Noncentral Gamma Distribution (\[nck\]), so that via Eq. (\[ncl\]) each parameter scales on average as $$\begin{aligned}
\label{nde}
E[\hat{\beta}_j^2]
&\simeq \begin{cases} \Delta^2/2 + \gamma^2/K, & j \leq K, \\
\delta^2/2, & j > K.\end{cases}\end{aligned}$$ As a result, the expectation value of the partial sum $F_j^2 =
\sum_{i=1}^j \hat{\beta}_i^2$ for $j=1,\ldots,N$ scales as $$\begin{aligned}
E[F_j^2]
&\simeq \sum_{i=1}^j E[\hat{\beta}_i^2]
= \begin{cases}
j \Delta^2/2 + j\gamma^2/K, & j < K, \\
K\Delta^2/2 + \gamma^2 + (K-j)\delta^2/2, & j \geq K.
\end{cases}\end{aligned}$$ In the [<span style="font-variant:small-caps;">bic</span>]{} limit, $\Delta^2 = N$ and $\delta=0$, so that $E[F_{{\scriptscriptstyle K}}^2/K] =
N/2 + \mbox{constant}$ and so for $N\gg K$, the robust [<span style="font-variant:small-caps;">bic</span>]{} becomes the [<span style="font-variant:small-caps;">bic</span>]{} up to a constant. In the [<span style="font-variant:small-caps;">aic</span>]{} limit, $\Delta^2=\delta^2=1$ so that $E[F_{{\scriptscriptstyle K}}^2/K] = 1 + \mathcal{O}(\gamma^2/K)$ and NIC reduces to approximately $\chi^2 + K + \mathcal{O}(\gamma^2/K)$, close to the [<span style="font-variant:small-caps;">aic</span>]{}’s $\chi^2+2K$.
Results {#sec:smm}
=======
To test the performance of the our robust [<span style="font-variant:small-caps;">bic</span>]{}, we present in this section a numerical simulation, followed by semi-analytical estimates of the salient quantities.
In the first part, we tested the success rate of information criteria in correctly identifying the number of parameters $S$ for competing models with varying parameter number $K$. Data sampling points $x_n$ were spread evenly over the interval $[0,\pi]$, $$\begin{aligned}
\label{smc}
x_n &= \frac{(2n{-}1)\pi}{2N} \quad n = 1,\ldots,N,\end{aligned}$$ and we generated $N=32$ data points throughout, setting the experimental uncertainties to $\sigma_n=1$. For each model ${{\mathcal{H}}}_{{\scriptscriptstyle S}}$ constructed from ${{S}}$ simulation parameters, data ${{\bm{z}}}({{S}}) = (z_1({{S}}),\ldots,z_{{\scriptscriptstyle N}}({{S}}))$ was generated as the sum of an “ideal data” term, a cosine series $$\begin{aligned}
\label{smd}
f_k(x_n)
&= \begin{cases}
\sqrt{\frac{1}{N}}& k=1, \\[8pt]
\sqrt{\frac{2}{N}} \cos[(k{-}1)x_n] & k = 2,\ldots,{{S}},
\end{cases} \end{aligned}$$ whose amplitude $a$ was varied randomly by an additive term $b\phi_k$ with $\phi_k$ drawn from the standardised Gaussian distribution $\phi_k\sim \mathcal{N}(0,1)$ and $b \geq 0$ an adjustable parameter. Conceptually, $a$ represents the signal strength while $b$ controls the variance of the signal. To simulate randomness associated with the experimental uncertainty normally captured in $\sigma_n$, a second random term $\varepsilon_n \sim \mathcal{N}(0,1)$ was added, so that the 32 data points generated from the true parameter model ${{\mathcal{H}}}_{{\scriptscriptstyle S}}$ are $$\begin{aligned}
\label{sme}
z_n({{S}})
&= \left[\sum_{k=1}^{{S}}f_k(x_n)\,(a + b \phi_k)\right] + \varepsilon_n({{S}}).\end{aligned}$$ For quenched values of $\phi_k$ and $\varepsilon_n$, one dataset was generated for each ${{S}}= 1,2,\ldots N = 32$. All datasets were efficiently computed in terms of $(N{\times}N)$-dimensioned matrices $$\begin{aligned}
\label{smf}
{\mathbb{D}}&= {\mathbb{X}}_{{\scriptscriptstyle N}}(a\,{\mathbb{I}}+ b\,{\mathbb{F}}){\mathbb{A}}+{\mathbb{E}},\end{aligned}$$ where ${\mathbb{D}}$ contains the $N$ column vectors ${{\bm{z}}}({{S}})$, one for each ${{S}}$, matrix ${\mathbb{X}}_{{\scriptscriptstyle N}}$ has elements $({\mathbb{X}}_{{\scriptscriptstyle N}})_{nk}=f_k(x_n)$, ${\mathbb{I}}$ is the diagonal matrix, noise matrix ${\mathbb{F}}$ is diagonal with elements $\phi_k$ and ${\mathbb{E}}$ contains the $N^2$ gaussian random numbers $\varepsilon_n({{S}})$ with unit variance. To limit the $k$-sum in Eq. (\[sme\]) to ${{S}}$, one must include an upper-triangular matrix with components ${\mathbb{A}}_{{{\scriptscriptstyle K}},{{\scriptscriptstyle S}}} = \Theta(K,{{S}}) = 1$ for integers $K \geq {{S}}$ and 0 otherwise. To calculate $F_{{\scriptscriptstyle K}}^2$ and $\chi^2$ for given ${{S}}$ for use in the Bayes Factor (\[ncy\]) and elsewhere, we must modify the notation to keep track of the “true” number of parameters ${{S}}$ to be compared to the number of model parameters $K$. We therefore write ${{\bm{\hat{\beta}}}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}$ for the parameter mode of the model with $K$ parameters for data simulated from ${{S}}$ parameters; correspondingly Eq. (\[rsq\]) becomes $F_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^2 = {{\bm{\hat{\beta}}}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^{{\scriptscriptstyle\sf T}}{{\bm{\hat{\beta}}}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^{\,} =
\sum_{k=1}^K \hat{\beta}_{k|{{\scriptscriptstyle S}}}^2$.
For model construction, we used the same cosine functions (\[smd\]) used for data generation, and since the cosines form an orthonormal system, the Hessian is diagonal, ${\mathbb{H}}_{{\scriptscriptstyle K}}= {\mathbb{X}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^{{\scriptscriptstyle\sf T}}{\mathbb{X}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}} = {\mathbb{I}}_{{\scriptscriptstyle K}}$, as are the rotation and eigenvalue matrices, so that with ${\bm{\hat{\alpha}}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}} = {\mathbb{H}}_{{\scriptscriptstyle K}}^{-1}{\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{z}}}({{S}})$ the mode simplifies to $$\begin{aligned}
\label{smi}
{{\bm{\hat{\beta}}}}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}
&= {\mathbb{L}}_{{\scriptscriptstyle K}}^{\!\!1/2}{\mathbb{S}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{\bm{\hat{\alpha}}}_{{\scriptscriptstyle K}}= {\mathbb{X}}_{{\scriptscriptstyle K}}^{{\scriptscriptstyle\sf T}}{{\bm{z}}}({{S}}).\end{aligned}$$ The $N{\times}K$ design matrix ${\mathbb{X}}_{{\scriptscriptstyle K}}$ is augmented by means of a projector ${\mathbb{P}}_{{\scriptscriptstyle K}}$ which contains 1’s along its first $K$ diagonal elements and 0 elsewhere; the truncated design matrix ${\mathbb{X}}_{{\scriptscriptstyle K}}= {\mathbb{X}}_{{\scriptscriptstyle N}}{\mathbb{P}}_{{\scriptscriptstyle K}}$ then contains zeros in the last $N{-}K$ columns of the $N{\times}N$ matrix and $f_k(x_n)$ elsewhere. The $N{\times}N$ matrix of modes ${\mathbb{B}}$ with elements ${\mathbb{B}}_{{{\scriptscriptstyle K}},{{\scriptscriptstyle S}}} = \hat{\beta}_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}$ is then compactly represented as $$\begin{aligned}
\label{smj}
{\mathbb{B}}&= {\mathbb{X}}_{{\scriptscriptstyle N}}^{{\scriptscriptstyle\sf T}}{\mathbb{P}}_{{\scriptscriptstyle K}}{\mathbb{D}}.\end{aligned}$$ Note that this matrix formulation is possible only for orthogonal basis functions; for nonorthogonal cases, the Hessian and its eigensystem must be recalculated for every $K$.
We now turn to the test results. Given a dataset with ${{S}}$ parameters, the $K$ value which minimises the AIC, BIC or the robust version in its Eq. (\[ndf\]) form is deemed a success if that $K$ correctly matches the data’s ${{S}}$. In Fig. \[fig:success\], we plot the number of successes as percentages of $2^{16}$ repetitions of 32 datasets as described above as a function of ${{S}}$ for four information criteria. The upper panel displays results for weak-signal data generated with $a=1,b=1$, while the strong-signal data shown in the lower panel used $a=5,b=1$. Red diamonds represent the robust ${\textsc{bic}}$ success rate, green circles the corresponding [<span style="font-variant:small-caps;">aic</span>]{} success rate and black triangles the [<span style="font-variant:small-caps;">bic</span>]{}. Also shown as blue squares is the success rate of the corrected [<span style="font-variant:small-caps;">aic</span>]{}, [@george2000calibration].
As expected, the [<span style="font-variant:small-caps;">bic</span>]{}performs poorly in the weak-signal environment where the models are badly specified but very well for strong signals, where the models are more successful. While the [<span style="font-variant:small-caps;">aic</span>]{} is more successful in the weak-signal case but underperforms the [<span style="font-variant:small-caps;">bic</span>]{} for strong signals. It is not a surprise that the corrected [<span style="font-variant:small-caps;">aic</span>]{}, which was designed for a particular subset of data scenarios, does very well in the mid-range of the strong-signal case but fails badly otherwise.[^3] By contrast, the robust [<span style="font-variant:small-caps;">bic</span>]{}matches or exceeds the performance of all other information criteria in both the strong and weak signal scenarios. The exact [<span style="font-variant:small-caps;">bic</span>]{} result (\[ncy\]) and the Eqs. (\[ndf\]) differ by less than one percent.
The general increase in success rates for ${{S}}$ near the simulation lower limit 1 and upper limit 32 are easily understood because there are fewer alternatives to ${{S}}$ at these edges. While this rise will persist for small ${{S}}$, it will for large ${{S}}$ shift with increasing $N$ and is therefore a nonpersistent “boundary effect”.
at (15\*[0.40]{},1.05\*[8]{}) [**Weak Signal**]{}; (0,-0.2) – (32\*[0.40]{},-0.2); (0,-0.2) – (0,1\*[8]{}); in [1,3,...,31]{} ([0.40]{},-0.2) – ([0.40]{},-0.35) node\[draw=none,anchor=north\] [****]{}; in [2,...,32]{} ([0.40]{},-0.2) – ([0.40]{},-0.35) node\[draw=none,anchor=north\] ; in [0,10,...,100]{} (0,/100\*[8]{}) – (-0.15,/100\*[8]{}) node\[draw=none,anchor=east\] [**%**]{}; at (16\*[0.40]{},0) [**Simulation parameters ${{S}}$**]{}; at (0,0.7\*[8]{}) [**Percentage of successes**]{}; (AIC-legend1) at (27\*[0.40]{},0.9\*[8]{}) ; (AIC-legend2) at (28\*[0.40]{},0.9\*[8]{}) ; (AIC-legend1) – (AIC-legend2); at (30\*[0.40]{},0.9\*[8]{}) [AIC]{}; /in [1/0.306,2/0.257,3/0.229,4/0.217,5/0.207,6/0.2,7/0.193,8/0.192,9/0.187, 10/0.182,11/0.185,12/0.182,13/0.178,14/0.177,15/0.176,16/0.177,17/0.174,18/0.172, 19/0.175,20/0.176,21/0.174,22/0.171,23/0.171,24/0.175,25/0.173,26/0.173,27/0.174, 28/0.175,29/0.177,30/0.189,31/0.201,32/0.236]{} [ (Aic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Aic) – (Aic); ]{} (BIC-legend1) at (27\*[0.40]{},0.85\*[8]{}) ; (BIC-legend2) at (28\*[0.40]{},0.85\*[8]{}) ; (BIC-legend1) – (BIC-legend2); at (30\*[0.40]{},0.85\*[8]{}) [BIC]{}; /in [1/0.27,2/0.19,3/0.149,4/0.125,5/0.108,6/0.096,7/0.086,8/0.078,9/0.072, 10/0.066,11/0.063,12/0.059,13/0.055,14/0.052,15/0.048,16/0.048,17/0.044,18/0.041, 19/0.039,20/0.039,21/0.037,22/0.035,23/0.033,24/0.032,25/0.031,26/0.029,27/0.029, 28/0.028,29/0.026,30/0.025,31/0.025,32/0.026]{}[ (Bic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Bic) – (Bic); ]{} (NIC-legend1) at (27\*[0.40]{},0.80\*[8]{}) ; (NIC-legend2) at (28\*[0.40]{},0.80\*[8]{}) ; (NIC-legend1) – (NIC-legend2); at (30\*[0.40]{},0.80\*[8]{}) [NIC]{}; /in [1/0.259,2/0.22,3/0.201,4/0.195,5/0.193,6/0.19,7/0.187,8/0.187,9/0.187, 10/0.184,11/0.187,12/0.187,13/0.186,14/0.185,15/0.185,16/0.189,17/0.188,18/0.187, 19/0.193,20/0.191,21/0.192,22/0.192,23/0.193,24/0.198,25/0.197,26/0.2,27/0.204, 28/0.209,29/0.216,30/0.232,31/0.258,32/0.318]{} [ (Mic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Mic) – (Mic); ]{} (AICc-legend1) at (27\*[0.40]{},0.75\*[8]{}) ; (AICc-legend2) at (28\*[0.40]{},0.75\*[8]{}) ; (AICc-legend1) – (AICc-legend2); at (30\*[0.40]{},0.75\*[8]{}) [AICc]{}; /in [1/0.328,2/0.287,3/0.285,4/0.21,5/0.176,6/0.155,7/0.145,8/0.106,9/0.086, 10/0.072,11/0.051,12/0.037,13/0.022,14/0.015,15/0.007,16/0.003,17/0.002,18/0, 19/0,20/0,21/0,22/0,23/0,24/0,25/0,26/0,27/0, 28/0,29/0,30/0,31/0,32/0]{} [ (Aicc) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Aicc) – (Aicc); ]{}
at (15\*[0.40]{},1.05\*[8]{}) [**Strong Signal**]{}; (0,-0.2) – (32\*[0.40]{},-0.2); (0,-0.2) – (0,1\*[8]{}); in [1,3,...,31]{} ([0.40]{},-0.2) – ([0.40]{},-0.35) node\[draw=none,anchor=north\] [****]{}; in [2,...,32]{} ([0.40]{},-0.2) – ([0.40]{},-0.35) node\[draw=none,anchor=north\] ; in [0,10,...,100]{} (0,/100\*[8]{}) – (-0.15,/100\*[8]{}) node\[draw=none,anchor=east\] [**%**]{}; at (16\*[0.40]{},0) [**Simulation parameters ${{S}}$**]{}; at (0,0.7\*[8]{}) [**Percentage of successes**]{}; (AIC-legend1) at (2\*[0.40]{},0.2\*[8]{}) ; (AIC-legend2) at (3\*[0.40]{},0.2\*[8]{}) ; (AIC-legend1) – (AIC-legend2); at (5\*[0.40]{},0.2\*[8]{}) [AIC]{}; /in [1/0.707,2/0.708,3/0.707,4/0.71,5/0.705,6/0.708,7/0.708,8/0.709,9/0.709, 10/0.709,11/0.708,12/0.71,13/0.71,14/0.708,15/0.709,16/0.707,17/0.713,18/0.707, 19/0.709,20/0.712,21/0.71,22/0.715,23/0.712,24/0.719,25/0.719,26/0.727,27/0.734, 28/0.736,29/0.753,30/0.783,31/0.841,32/0.994]{} [ (Aic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Aic) – (Aic); ]{} (BIC-legend1) at (2\*[0.40]{},0.15\*[8]{}) ; (BIC-legend2) at (3\*[0.40]{},0.15\*[8]{}) ; (BIC-legend1) – (BIC-legend2); at (5\*[0.40]{},0.15\*[8]{}) [BIC]{}; /in [1/0.906,2/0.904,3/0.904,4/0.906,5/0.906,6/0.902,7/0.903,8/0.905,9/0.904, 10/0.907,11/0.904,12/0.904,13/0.906,14/0.905,15/0.905,16/0.905,17/0.906,18/0.904, 19/0.905,20/0.905,21/0.904,22/0.904,23/0.903,24/0.906,25/0.906,26/0.906,27/0.905, 28/0.905,29/0.906,30/0.911,31/0.927,32/0.987]{} [ (Bic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Bic) – (Bic); ]{} (NIC-legend1) at (2\*[0.40]{},0.10\*[8]{}) ; (NIC-legend2) at (3\*[0.40]{},0.10\*[8]{}) ; (NIC-legend1) – (NIC-legend2); at (5\*[0.40]{},0.10\*[8]{}) [NIC]{}; /in [1/0.884,2/0.892,3/0.897,4/0.901,5/0.902,6/0.898,7/0.9,8/0.902,9/0.901, 10/0.903,11/0.901,12/0.901,13/0.904,14/0.902,15/0.902,16/0.901,17/0.903,18/0.902, 19/0.901,20/0.902,21/0.902,22/0.902,23/0.9,24/0.903,25/0.903,26/0.904,27/0.903, 28/0.903,29/0.903,30/0.909,31/0.925,32/0.987]{} [ (Mic) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Mic) – (Mic); ]{} (AICc-legend1) at (2\*[0.40]{},0.05\*[8]{}) ; (AICc-legend2) at (3\*[0.40]{},0.05\*[8]{}) ; (AICc-legend1) – (AICc-legend2); at (5\*[0.40]{},0.05\*[8]{}) [AICc]{}; /in [1/0.759,2/0.792,3/0.878,4/0.878,5/0.881,6/0.888,7/0.932,8/0.929,9/0.933, 10/0.955,11/0.948,12/0.958,13/0.949,14/0.958,15/0.932,16/0.932,17/0.917,18/0.877, 19/0.824,20/0.765,21/0.666,22/0.521,23/0.329,24/0.153,25/0.031,26/0.001,27/0, 28/0,29/0,30/0,31/0,32/0]{} [ (Aicc) at ([0.40]{},[8]{}) ; ]{} in [1,...,31]{} [ (Aicc) – (Aicc); ]{}
In the second test, we utilise the simple linear system of Eqs. (\[smf\])–(\[smj\]) to obtain analytical estimates of the squared signal and noise for a detailed but statistically approximate analysis of the shapes and sizes of the criteria’s $K$ vs ${{S}}$ curves. Because ${\mathbb{F}}^{{\scriptscriptstyle\sf T}}{\mathbb{F}}\simeq {\mathbb{I}}$ and ${\mathrm{var}}(\varepsilon)=1$, the squared data vectors scale approximately as $$\begin{aligned}
\label{smg}
z^2({{S}})
= \sum_n z_n^2({{S}})
&= \text{diag}({\mathbb{D}}^T{\mathbb{D}})
\ \simeq\ N + (a^2+b^2){{S}}.\end{aligned}$$ The squared signal is obtained from the diagonal elements of the squared mode matrix, $F_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^2 =
[{\mathbb{B}}^{{\scriptscriptstyle\sf T}}{\mathbb{B}}]_{{{\scriptscriptstyle S}},{{\scriptscriptstyle S}}}$. Inserting the explicit simulation model (\[smf\]) results in an approximate estimate of $$\begin{aligned}
\label{smk}
F_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^2
&\simeq (a^2 + b^2){\mathbb{A}}^{\!{{\scriptscriptstyle\sf T}}} {\mathbb{P}}_{{\scriptscriptstyle K}}{\mathbb{A}}+ {\mathbb{E}}^{{\scriptscriptstyle\sf T}}{\mathbb{P}}_{{\scriptscriptstyle K}}{\mathbb{E}}= (a^2 + b^2)\min(K,{{S}}) + K\end{aligned}$$ while the squared data vector and squared noise vector are obtained from $$\begin{aligned}
\label{sml}
z^2(S)
&= ({\mathbb{D}}^{{\scriptscriptstyle\sf T}}{\mathbb{D}})_{{{\scriptscriptstyle S}},{{\scriptscriptstyle S}}} = (a^2+b^2){{S}}+ N \\
\label{smm}
\chi^2_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}
&= z^2(S) - F_{{{\scriptscriptstyle K}}|{{\scriptscriptstyle S}}}^2 \nonumber\\
&\simeq N - K + (a^2 + b^2)[{{S}}- \min(K,{{S}})]\nonumber\\
&= \left\{
\begin{array}{ll}
N-K + (a^2+b^2)({{S}}-K) & \text{ for } K < {{S}},\\
N-K & \text{ for } K \geq {{S}}.\\
\end{array}
\right.\end{aligned}$$
(AIC-legend1) at (6.5\*[0.65]{},75\*[0.12]{}) ; (AIC-legend2) at (7.5\*[0.65]{},75\*[0.12]{}) ; (AIC-legend1) – (AIC-legend2); at (8.5\*[0.65]{},75\*[0.12]{}) [**AIC**]{}; (BIC-legend1) at (10.0\*[0.65]{},75\*[0.12]{}) ; (BIC-legend2) at (11.0\*[0.65]{},75\*[0.12]{}) ; (BIC-legend1) – (BIC-legend2); at (12.0\*[0.65]{},75\*[0.12]{}) [**BIC**]{}; (NIC-legend1) at (6.5\*[0.65]{},70\*[0.12]{}) ; (NIC-legend2) at (7.5\*[0.65]{},70\*[0.12]{}) ; (NIC-legend1) – (NIC-legend2); at (8.5\*[0.65]{},70\*[0.12]{}) [**NIC**]{}; (Chi-legend1) at (10.0\*[0.65]{},70\*[0.12]{}) ; (Chi-legend2) at (11.0\*[0.65]{},70\*[0.12]{}) ; (Chi-legend1) – (Chi-legend2); at (12.0\*[0.65]{},70\*[0.12]{}) [${{\bm{\chi}}}^2$]{}; (3.5\*[0.65]{},15\*[0.12]{}) – (12.5\*[0.65]{},15\*[0.12]{}); (3.5\*[0.65]{},15\*[0.12]{}) – (3.5\*[0.65]{},80\*[0.12]{}); at (8\*[0.65]{},15\*[0.12]{}) [**Model parameters $K$**]{}; in [4,...,12]{} ([0.65]{},15\*[0.12]{}) – ([0.65]{},15\*[0.12]{}-0.2) node\[draw=none,anchor=north\] [****]{}; in [15,20,...,75]{} (3.5\*[0.65]{},[0.12]{}) – (3.5\*[0.65]{}-0.2,[0.12]{}) node\[draw=none,anchor=east\] [****]{}; //in [4/32.0/0.693,5/30.0/0.693,6/28.0/0.693,7/26.0/0.693,8/24.0/0.693,9/23.0/0.636,10/22.0/0.588,11/21.0/0.547,12/20.0/0.511]{} [ (AIC) at ([0.65]{},[0.12]{}+2\*[0.12]{}) ; (BIC) at ([0.65]{},[0.12]{}+3.47\*[0.12]{}) ; (NIC) at ([0.65]{},[0.12]{}+ [0.12]{}+ [0.12]{}\*) ; (Chi) at ([0.65]{},[0.12]{}) ; ]{} in [4,...,11]{} [ (AIC) – (AIC); (BIC) – (BIC); (NIC) – (NIC); (Chi) – (Chi); ]{} //in [4/64.0/2.3,5/54.0/2.3,6/44.0/2.3,7/34.0/2.3,8/24.0/2.3,9/23.0/2.2,10/22.0/2.1,11/21.0/2.0,12/20.0/1.95]{} [ (AIC2) at ([0.65]{},[0.12]{}+2\*[0.12]{}) ; (BIC2) at ([0.65]{},[0.12]{}+3.47\*[0.12]{}) ; (NIC2) at ([0.65]{},[0.12]{}+ [0.12]{}+ [0.12]{}\*) ; (Chi2) at ([0.65]{},[0.12]{}) ; ]{} in [4,...,11]{} [ (AIC2) – (AIC2); (BIC2) – (BIC2); (NIC2) – (NIC2); (Chi2) – (Chi2); ]{}
These expressions provide instructive, if approximate, insights into the behaviour of the information criteria as a function of model parameter number $K$. Fig. \[fig:sim\] illustrates by example the shapes of the minima as functions of $K$ of the simplest robust [<span style="font-variant:small-caps;">bic</span>]{}form (\[ndf\]) as well as the [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}for fixed ${{S}}=8,b=0$ and strong-signal $a=3$ and weak-signal $a=1$ scenarios. The [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}do not depend on $F_{{\scriptscriptstyle K}}^2$ but only on $\chi^2$, which exhibits the well-known behaviour of steadily decreasing with $K$. Upper and lower branches of these curves denote the strong-signal and weak-signal cases respectively. Based on $\chi^2$ and the simple penalty terms, the [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}both exhibit a reasonably strong minimum at $K{=}{{S}}$ for $a=3$; for the weak-signal $a{=}1$, however, the [<span style="font-variant:small-caps;">aic</span>]{}remains flat while the [<span style="font-variant:small-caps;">bic</span>]{}has no minimum at all. This is reflected in the low [<span style="font-variant:small-caps;">bic</span>]{}success rate in Fig. \[fig:success\]. Since $\chi^2$ becomes independent of $a$ for $K \geq {{S}}$, the [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}do not distinguish between strong and weak scenarios in that region.
The robust [<span style="font-variant:small-caps;">bic</span>]{}, by contrast, is sensitive to the squared signal $F_{{\scriptscriptstyle K}}^2$, which lifts the degeneracy between strong and weak signal for $K\geq{{S}}$. Like the [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}, the robust [<span style="font-variant:small-caps;">bic</span>]{}has no trouble identifying $K=S$ for strong signal. For weak signal, it exhibits a minimum at the correct answer, albeit a shallow one. Shallow minima reflect, of course, the inherent uncertainty regarding the signal or noise character of the data. Details of Figure \[fig:sim\] and its discussion are, of course, specific to the model and numbers used and of illustrative value only.
Discussion and conclusions {#sec:dsc}
==========================
The robust version of the ${\textsc{bic}}$ introduced in this paper is based on three simple but novel ideas. Firstly, we have expanded Akaike’s original argument for a larger model space into a model space plus a fully-fledged noise space which together partition the entire data space. The resulting symmetries and scale behaviour of model and noise space provide a surprisingly unified and indeed beautiful framework for linear regression.
Secondly, building on the insight of earlier work [@dekock2017bayesian], we posit that both model parameter and noise parameter spaces should be projected onto a radial coordinate on the respective hypersphere. Unlike [@dekock2017bayesian], however, we now have not one but two hyperspheres reflecting the separate symmetries and scales of the model and noise spaces.
The third insight is that the crucial difference between model and noise parameters lies not in the scales $\delta$ and $\Delta$ — indeed we set these equal and eventually even set $\Delta = 0$ — but in explicitly taking into account that the maximum-likelihood parameter vector’s magnitude ${\parallel}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle K}}{\parallel}$ must, by the very definition of “signal”, be significantly nonzero, while ${\parallel}{{\bm{\hat{\beta}}}}_{{\scriptscriptstyle L}}{\parallel}\simeq 0$ for noise. This results in a Gamma Distribution for noise parameters arising from projection of a zero-mode Gaussian on the one hand, and a noncentral Gamma Distribution for model parameters arising from projection of a nonzero-mode Gaussian.
Together, these three insights have allowed us to calculate Bayes Factors for model comparison in closed form and construct a robust version of the [<span style="font-variant:small-caps;">bic</span>]{} which extends the robustness the [<span style="font-variant:small-caps;">aic</span>]{} has against model misspecification to the [<span style="font-variant:small-caps;">bic</span>]{}. Unlike the latter, the robust [<span style="font-variant:small-caps;">bic</span>]{}depends explicitly on the squared signal strength $F_{{\scriptscriptstyle K}}^2$, and as $F_{{\scriptscriptstyle K}}^2$ approaches the weak or the strong signal limit, the robust [<span style="font-variant:small-caps;">bic</span>]{} correspondingly approaches the [<span style="font-variant:small-caps;">aic</span>]{}and [<span style="font-variant:small-caps;">bic</span>]{}cases as limiting forms.
The noncentrality parameter $\gamma$ as a measure of signal strength appears to be the essence of the difference between signal and noise. Where the signal-to-noise ratio is known beforehand, $\gamma$ can be set to a fixed number or restricted to a limited interval. In the general case of unknown signal-to-noise ratio, however, it is better to integrate $\gamma$ over all possible values as implemented here.
We conclude with a few general remarks. Naturally, the scope of the numerical results presented here is limited, and this robust version of the [<span style="font-variant:small-caps;">bic</span>]{} should be tested and possibly improved when applied to a diversity of data scenarios. The present results should also be extended from the fixed experimental uncertainties ${{\bm{\sigma}}}$ to variable $\sigma$. The analysis was done in the context of linear regression and should strictly speaking be used only in that context. The degree of success for nonlinear situations cannot be estimated or guaranteed within the present framework. Our derivations presume that there is only one model per $K$. This limited approach is easily generalised to include more than one model for a given $K$ using, for example, indicator vectors as set out in [@Liang2008].\
**Acknowledgements**\
This work was supported in part by the South African National Research Foundation.
[^1]: The simplest way to ensure block-diagonality is to construct a complete orthogonal basis for all $K{=}N$ functions $f_k$ which would trivially fulfil these requirements. The block-diagonal form is, however, more widely applicable.
[^2]: The reasoning behind setting the model prior variances to $N$ is that the exponent of the likelihood $L[{{\bm{\alpha}}}_{{\scriptscriptstyle K}},{{\bm{\alpha}}}_{{\scriptscriptstyle L}}]$ scales roughly with $N$ as long as parameter-parameter correlations do not dominate, so that ${{\bm{\beta}}}\approx {{\bm{\alpha}}}/\sqrt{N}$ and ${{\bm{\beta}}}^{{\scriptscriptstyle\sf T}}{{\bm{\beta}}}\approx 1/N$.
[^3]: The corrected [<span style="font-variant:small-caps;">aic</span>]{} corrects the [<span style="font-variant:small-caps;">aic</span>]{}formula for small $N$ and therefore small ${{S}}$. In effect, this improves the [<span style="font-variant:small-caps;">aic</span>]{}for strong-signal cases, but destroys its performance for the weak-signal case and larger ${{S}}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We characterise the virtually abelian groups which are fundamental groups of compact Kähler manifolds and of smooth projective varieties. We show that a virtually abelian group is Kähler if and only if it is projective. In particular, this allows to describe the Kähler condition for such groups in terms of integral symplectic representations.'
author:
- |
Oliver Baues[^1]\
Institut für Algebra und Geometrie\
Universität Karlsruhe\
D-76128 Karlsruhe
- |
Johannes Riesterer[^2]\
Institut für Algebra und Geometrie\
Universität Karlsruhe\
D-76128 Karlsruhe
bibliography:
- 'kpgs.bib'
date: 'October 26, 2009'
title: Virtually abelian Kähler and projective groups
---
Introduction
============
A connection between Kähler geometry and topology is made by the topic of Kähler groups. A group is called Kähler, if it is isomorphic to the fundamental group of a compact Kähler manifold. Note that Kähler groups are finitely generated and even finitely presented, since the associated manifold is compact by assumption. Not all finitely presented groups are Kähler groups. Well known restrictions stem from Hodge-theory, see [@ABKT] for various results.
More restrictively, a compact complex manifold is called projective if it embeds holomorphically as a subvariety of complex projective space. A projective manifold is also a Kähler manifold, and its fundamental group is called a projective Kähler group. It is widely believed that the class of Kähler and projective groups coincide. An important remark is due to J.P. Serre [@serre prop. 15]:
\[serre\] Every finite group $G$ is the fundamental group of a smooth projective variety $X_G$.
We would like to understand how the class of Kähler groups behaves with respect to finite extension of groups. Since not much seems to be known about this particular question, the aim of this paper is to study the situation in a specific and accessible context: A group is called virtually abelian if it is an extension of an abelian group by a finite group. We shall examine the following question: *Which virtually abelian groups are Kähler and which are projective*?
Kähler and projective groups
----------------------------
Let $\Gamma$ be a virtually abelian group. Since we also assume $\Gamma$ to be finitely generated, there is an exact sequence $$\label{eq:fes} \xymatrix@1{
0\ar[r] & \mathbb{Z}^n \ar[r] & \Gamma \ar[r] & G\ar[r] & 1}\; ,$$ where $G$ is a finite group. Via conjugation in $\Gamma$, the extension defines the characteristic representation $\mu: G \to \GL(n,\mathbb{Z})$.
Recall that a complex structure on $\bbR^n$ is a linear map $J: \bbR^n {\rightarrow}\bbR^n$ with $J^2 = -\id$. It is called $\mu$-invariant, for a representation $\mu: G {\rightarrow}\GL(n,\bbR)$, if $\mu(g)J=J\mu(g)$, for all $g \in G$. We shall say that the virtualy abelian group $\Gamma$ admits a complex structure, if there exists a complex structure on $\bbR^n$, which is invariant by the characteristic representation $\mu$.
\[thm:Kaehler\] The group $\Gamma$ is Kähler if and only if it admits a complex structure.
Remark, as follows from elementary Hodge theory, if the torsionfree abelian group $\bbZ^n$ is Kähler then $n=2k$ is even. For a Kähler group $\Gamma$, $n$ is thus even in the above exact sequence. Theorem \[thm:Kaehler\] gives a more precise condition. It also shows that the Kähler condition for $\Gamma$ depends only on the characteristic representation $\mu$. In particular, it does not depend on the characteristic cohomology class of the extension .
We proceed now to provide a characterization of virtually abelian projective groups. For this, recall that a two-form $\omega \in \bigwedge^2 (\bbR^n)^*$ is called a polarisation for $J$ (with respect to the lattice $\bbZ^n$) if the following conditions are satisfied:
1. $\omega(J \cdot, J \cdot ) = \omega(\cdot, \cdot)$,
2. $g= \omega(J \cdot, \cdot)$ is positive definite,
3. $\omega(\bbZ^n, \bbZ^n) \subseteq \bbZ$.
Accordingly, if there exists a polarisation then the complex structure $J$ is called polarisable.
\[thm:projective\] The group $\Gamma$ is projective Kähler if and only if it admits a polarisable complex structure.
We call an integral representation symplectic if it preserves a non-degenerate alternating two-form $\omega \in \bigwedge^2 (\bbZ^n)^*$. We can add the following characterisation of the projectivity condition for $\Gamma$:
\[thm:symplectic\] The following conditions are equivalent:
1. $\Gamma$ admits a polarisable complex structure.
2. The characteristic representation $\mu$ for $\Gamma$ is $\SL_{n}(\bbZ)$-conjugate to an integral symplectic representation.
Note that condition 2. does not refer to the existence of a complex structure for $\Gamma$. The following consequence (see Theorem \[thm:algequivalence\]) is implied:
\[thm:kequivp\] The group $\Gamma$ is Kähler if and only if it is projective Kähler.
In particular, we obtain a characterisation of the Kähler condition in terms of integral symplectic representations.
\[cor:kequivp2\] The group $\Gamma$ is Kähler if and only if the characteristic representation $\mu$ is conjugate to an integral symplectic representation.
#### Fields of definition
We may further refine the projectivity condition for $\Gamma$, by taking into account the possible fields of definition for projective varieties with fundamental group $\Gamma$. In this direction, we remark:
\[thm:numberfields\] If the virtually abelian group $\Gamma$ is Kähler then there exists a smooth algebraic subvariety $X$ of ${P}^N \bbC$, which is defined over an algebraic number field and has fundamental group $\Gamma$.
### Flat and aspherical Kähler groups
Flat Kähler manifolds are constructed as quotients $M =\bbC^k/\Gamma$, where $\Gamma \leq \bbC^k \rtimes U(k)$ is a discrete and torsionfree subgroup of the isometry group of $\bbC^k$. By a famous theorem of Bieberbach (see, for example, [@wolf]), $\Lambda = \Gamma \cap \bbC^k$ is a lattice in $\bbC^k$ and of finite index in $\Gamma$. In particular, $\Gamma$ is a virtually abelian group. Conversely, if $\Gamma$ is a torsionfree virtually abelian Kähler group, it may as well be represented as the fundamental group of a flat Kähler manifold. In fact, we obtain:
\[cor:flatKaehler\] Let $\Gamma$ be torsionfree. The following conditions are equivalent:
1. $\Gamma$ is a torsionfree Kähler group.
2. $\Gamma$ is the fundamental group of a flat Kähler manifold.
3. $\Gamma$ is the fundamental group of a flat Kähler manifold which is projective.
The following concept is closely related. Recall that a manifold is called aspherical if its universal covering space is contractible. Henceforth, a Kähler group is called [*aspherical*]{} if it is the fundamental group of a compact aspherical Kähler manifold. Combining with results from [@bauescortes], we can state:
\[cor:aspKaehler\] The following conditions are equivalent:
1. $\Gamma$ is an aspherical Kähler group.
2. $\Gamma$ is a Kähler group which is torsionfree.
Deformation of complex representations to integral symplectic representations
-----------------------------------------------------------------------------
The above results translate into the following representation theoretic result:
\[thm:algequivalence\] Let $G$ be a finite group and $V$ a finite-dimensional $\bbQ[G]$-module. Then the following conditions are equivalent:
1. $V \tensor \bbR$ has a $G$-invariant complex structure $J$.
2. $V$ is isomorphic to a symplectic $\bbQ[G]$-module $W$.
3. There exists a $G$-invariant complex structure $J'$ on $V \tensor \bbR$, which is polarisable.
($W$ is called symplectic if it admits a non-degenerate $G$-invariant two-form.)
The method of our proof allows to show slightly more. Let $\mathfrak{D}(G)$ denote the set of $G$-invariant complex structures on $V \tensor \bbR$. Then the implication from 1. to 3. in Theorem \[thm:algequivalence\] will be implied by
\[thm:deformation\] Let $G$ be a finite group and $V$ a finite-dimensional $\bbQ[G]$-module such that $V \tensor \bbR$ has a $G$-invariant complex structure $J$. Then $J$ may be continuously deformed in $\mathfrak{D}(G)$ to a $G$-invariant complex structure $J'$, which is polarisable. Moreover, the subset $\mathfrak{D}(G)_{p}$ of polarisable elements in $\mathfrak{D}(G)$ is dense in $\mathfrak{D}(G)$.
### Deformation of complex structures on torus quotients
A complex torus $T = \bbC^k/\Lambda$ which is projective is an abelian variety. By the usual dictionary between complex tori (resp. abelian varieties) and complex structures (resp. polarisable complex structures), the previous results can be translated into statements about complex tori with a holomorphic $G$-action:\
Let $G$ be a finite group of automorphisms of a complex torus $T$, which acts freely on $T$. Let $X_{G} = T/G$ be the quotient complex manifold. Then we may view $\mathfrak{D}(G)$ as the deformation space of complex structures for $X_{G}$. Thus, Theorem \[thm:deformation\] implies
\[prop:deformation2\] The complex structure on $X_{G}$ may be continuously deformed to a projective structure.
We call a complex manifold *deformation rigid* if it does not admit any non-trivial continuous deformation of its complex structure. The following application concerning Kähler rigidity was suggested to us by F. Catanese:
\[cor:rigidity\] Let $X$ be a complex variety which is a finite quotient of a complex torus. If $X_{G}$ is deformation rigid then $X_{G}$ is a quotient of an abelian variety, in particular, $X_{G}$ is projective.
Notes
-----
Theorem \[thm:Kaehler\] appears as the main result of the diploma thesis [@riesterer]. The analogue of Theorem \[thm:Kaehler\] concerning flat Kähler manifolds, and the equivalence of 2) and 3) in Corollary \[cor:flatKaehler\] are the main theorems in Johnson’s paper [@Johnson]. Serre [@Serre2] constructed certain examples of subvarieties $X$ of complex projective space, which are defined over algebraic number fields, and whose fundamental groups are semi-direct extensions of $\bbZ^n$ by cyclic groups. Our Theorem \[thm:numberfields\] extends this construction without the restriction on the fundamental group. In [@Serre2], the construction is used to exhibit Galois-conjugate examples $X$ and $X'$ of this type with non-isomorphic fundamental group.
The equivalence of 1. and 3. in Theorem \[thm:algequivalence\] is also contained in [@Johnson]. The proof is based on a different approach, using the theory of complex multiplication. However, we will pick up this idea in the proof of Theorem \[thm:numberfields\]. In [@popovzarhin] irreducible symmetry groups of complex tori are investigated from the point of view of complex reflection groups. In particular, examples of non- polarisable tori with a non-trivial symmetry group are considered.
By Corollary \[cor:kequivp2\], virtually abelian Kähler groups of rank $n=2k$ can be classified in principle by the determination of the finite subgroups of $\Sp(2k, \bbZ)$. For $k=1$, this list comprises the finite cyclic groups of order 2,3,4,6 (see [@zieschang], for example). Fujiki [@fujiki] classifies the finite automorphism groups of abelian surfaces, the three dimensional case is covered in [@BGL].
Structure of the paper
----------------------
The proofs of Theorem \[thm:Kaehler\] and Theorem \[thm:projective\] can be found in sections 2 and 3, see Proposition \[vak\_prop\_2\] and Proposition \[prop:constructX\]. In section 4 we prove Theorem \[thm:algequivalence\] (see in particular §\[rationalpoints\]) and Theorem \[thm:deformation\]. The latter imply Theorem \[thm:symplectic\] and Theorem \[thm:kequivp\]. Section 5 is devoted to the proof of Theorem \[thm:numberfields\].\
The first named author learned several years ago from A. Szczepański about Johnson’s and Rees’ results on flat Kähler and projective manifolds. Fritz Grunewald informed us that he has characterised virtually abelian fundamental groups of projective algebraic varieties in terms of integral symplectic representations, this question playing a role in joint work with Ingrid Bauer and Fabrizio Catanese. We wish to thank all three of them for various discussions. Part of this paper was written up during a stay of the first named author at the research activity ’Groups in algebraic geometry’, Centro Ennio di Georgi, Pisa, 2008. He wishes to thank the organisers for the hospitality and support. Furthermore we wish to thank Eduard Loojenga for a helpful comment, as well as Vicente Cortés, Evelina Viada and Enrico Leuzinger for conversations regarding the context of this paper. We thank Stefan Kühnlein for commenting on a draft of section 5.\
The Picard variety and induced Hodge structure on the fundamental group {#hodge_structure}
=======================================================================
In this section, we prove that every virtually abelian Kähler group $\Gamma$ admits an invariant complex structure $J$. Furthermore we show, if $\Gamma$ is projective then it admits a polarised complex structure $J'$.
#### Conventions concerning cohomology groups
Let $X$ be a space, and $A$ an abelian group. Then $H^j_{sing}(X,A)$ denotes the singular cohomology groups with coefficients in $A$. Moreover, if $X$ is a manifold and $A$ is a subgroup of $\bbC$, we let $H^j(X,A)$ denote the *image* of $H^j_{sing}(X,A)$ in the de Rham cohomology group $H^j(X,\bbC) := H^j_{DR}(X,\bbC)$ under the de Rham isomorphism $H^j_{sing}(X,\bbC) {\rightarrow}H^j_{DR}(X,\bbC)$. Thus, in particular, $H^n(X, \mathbb{Z})$ denotes the full lattice in $H^n(X, \mathbb{R})$, which is the image of $H^n(X, \mathbb{Z})_{sing}$ under the coefficient homomorphism. For any continuous (respectively smooth) map $f: X {\rightarrow}X'$, there is a well defined induced map on cohomology groups $f^*: H^j(X', \mathbb{R}) \to H^j(X, \mathbb{R})$, such that $f^*(H^j(X', A)) \subset H^j(X, A)$.
Review of Picard varieties
--------------------------
We start by reviewing the definition of Picard varieties. Let $X$ be a Kähler manifold and $H^1(X,\mathbb{C}) = H^{1,0}(X) \oplus H^{0,1}(X)$ the Hodge decomposition. Then there is an associated isomorphism of real vector spaces $$\begin{aligned}
&&\Re: H^{0,1}(X) \to H^1(X,\mathbb{R}) , \\
&&\Re \, \omega = \frac {\omega + \bar \omega}{2}, \; \; \omega \in H^{0,1}(X) \; . \end{aligned}$$ Thus $J:= \Re I \Re^{-1}$ is a complex structure on $H^1(X,\mathbb{R})$, where $I$ denotes multiplication with $i$ on the complex vector space $H^{0,1}(X)$. The complex torus $$Pic^0(X):= (H^1(X,\mathbb{R}) / H^1(X, \mathbb{Z}),J)\; ,$$ is called the Picard variety of $X$. By construction, $Pic^0(X)$ is biholomorphic to $H^{0,1}(X)/ H^1(X, \mathbb{Z})$, which yields an equivalent definition. For details, see [@birken 11.11] or [@voisin 7.2.2].
\[vak\_prop\_1\] $Pic^0$ is functorial, that is, if $f:X \to X'$ is a holomorphic map between Kähler manifolds, then $f$ induces a holomorphic map $f^*: Pic^0 (X') \to Pic^0 (X)$.
Let $X, X'$ be Kähler manifolds and $f:X \to X'$ a holomorphic map. There is an induced map $f^*: H^1(X', \mathbb{R}) \to H^1(X, \mathbb{R})$ on the cohomology, such that $f^*(H^n(X, \mathbb{Z})) \subset H^n(X, \mathbb{Z})$. Hence, $f^*$ induces a homomorphism from $Pic^0(X')$ to $Pic^0(X)$.
Now we show, that $f^* : Pic^0(X') \to Pic^0(X)$ is holomorphic. For this consider the induced complex linear map $f^*: H^1(X', \mathbb{C}) \to H^1(X, \mathbb{C})$, given by $f^*(\omega \otimes z):= f^*\omega \otimes z$. Since $f$ is holomorphic, it preserves the Hodge decomposition (see [@Wells]), that is, $$\begin{aligned}
f^*(H^{1,0}(X')) \subset H^{1,0}(X) \; , \; \,
f^*(H^{0,1}(X')) \subset H^{0,1}(X) \; .\end{aligned}$$ Let $J$ be the complex structure of $Pic^0(X)$ and $J'$ of $Pic^0(X')$ respectively. Since $f^*: H^1(X',\mathbb{C}) \to H^1(X,\mathbb{C})$ preserves the real structures, $\Re f^* = f^*\Re$. We deduce, that $f^*J' = Jf^*$. Thus $f^*$ is holomorphic.
An abelian variety is a complex torus $\mathbb{C}^k / \Lambda$, which is also a projective manifold. By the criterion of Riemann [@mumford p.35], this is the case, if and only if there exists a positive definite hermitian form $h$ on $\bbC^k$ such that its imaginary part $\omega:= \Im h$ is integral on $\Lambda \times \Lambda$. The form $\omega$ is called a polarisation for $\Lambda$.
Now let $X$ be a projective manifold. Then there exists a polarisation $\omega$ for $X$. This means, $\omega$ is a Kähler form, such that its cohomology class $[\omega]$ is integral. Then the Hodge-Riemann pairing $$h(\alpha, \beta)= -2i \int_{X}\omega^{n-1} \wedge \alpha \wedge \bar{\beta} \; ,\; \alpha, \beta \in H^{0,1}(X)\; ,$$ defines a positive definite hermitian form on $H^{0,1}(X)$. Moreover, its imaginary part, $\omega= \Im h$ is integral on $H^1(X,\mathbb{Z})$. See [@voisin 7.2.2] for discussion. In particular, it follows
\[moduli\_prop\_picab\] Let $X$ be a projective manifold. Then $Pic^0(X)$ is an abelian variety.
Induced invariant Hodge structures on $\Gamma$
----------------------------------------------
Let $\Gamma$ be a virtually abelian Kähler group, and let $X$ be a compact Kähler manifold with $\pi_1(X)= \Gamma$. Let $\tilde{X}$ be the universal covering manifold of $X$. Associated to the exact sequence , with characteristic homomorphism $\mu:G \to\GL_{n}(\mathbb{Z})$, we consider the holomorphic covering $$p: \, X_{1} := \tilde{X}/ \mathbb{Z}^{n}\; {\longrightarrow}X \; .$$ Since this is a regular covering, $G= \Gamma / \mathbb{Z}^{n}$ identifies with the group of covering transformations of $p$. In particular, $G$ has an induced holomorphic action on the Picard variety $Pic^0(X_{1})$.
#### Group cohomology
If $\Gamma$ is a group, and $A$ is an abelian group, we let $H^j(\Gamma,A)$ denote the $j$-th cohomology group of $\Gamma$ with coefficients in $A$. In particular, $H^1(\Gamma, A) = {\rm Hom}(\Gamma, A)$. If $\phi: \Gamma {\rightarrow}\Gamma'$ is a homomorphism of groups there is a natural induced homomorphism of cohomology groups $\phi^*: H^j(\Gamma',A) {\rightarrow}H^j(\Gamma,A)$. (See [@Brown], for a general reference.) We can formulate now:
\[lemma:cohomologyisnat\] There exists a natural isomorphism $$Pic^0(X_{1}) \cong H^1(\mathbb{Z}^n, \mathbb{R})/ H^1(\mathbb{Z}^n, \mathbb{Z})$$ such that, for all $g \in G$, the diagram $$\label{vak_diagram_2}
\xymatrix@1{
Pic^0(X_{1}) \ar@1{->}[r]^(0.35){\cong} \ar[d]^{g^*} & H^1(\mathbb{Z}^n, \mathbb{R})/ H^1(\mathbb{Z}^n, \mathbb{Z}) \ar[d]^{\mu(g)^*} \\
Pic^0(X_{1}) \ar@1{->}[r]^(0.35){\cong} & H^1(\mathbb{Z}^n, \mathbb{R})/ H^1(\mathbb{Z}^n, \mathbb{Z}) }$$commutes.
Before proving the Lemma, we embark on the following
#### Topological digression
For the following, see [@maclane Prop 11.4, Thm. 11.5]. Let $X$ be a (reasonable) topological space, and $\Gamma$ a group which acts properly discontinuously on $X$. If $X$ is acyclic up to dimension $k$, that is, $$H_{j}(X) = \begin{cases} \bbZ \;, j = 0 \\ 0 \:\; , k \geq j \geq 1 \end{cases},$$ then, for all $k \geq j \geq 1$, there exist natural isomorphisms $$H^j_{sing}(X/\Gamma,A) \cong H^j(\Gamma,A)$$ such that the diagram $$\label{topo_diagram_1}
\xymatrix@1{
H^j_{sing}(X/\Gamma,A) \ar@1{->}[r]^(0.5){\cong}\ar[d]^{f^*} & \ H^j(\Gamma,A)\ar[d]^{\phi^*}\\
H^j_{sing}(X/\Gamma,A) \ar@1{->}[r]^(0.5){\cong} & \ H^j(\Gamma,A) }$$ commutes, for all mappings $f: X \rightarrow X $, and homomorphisms $\phi : \Gamma \rightarrow \Gamma$ with $$f( \gamma x) = (\phi \gamma)fx \; , \; \forall \gamma \in \Gamma, \; x \in X \; .$$ Now we give the proof of the above Lemma \[lemma:cohomologyisnat\].
Let $g \in G$ be a covering transformation of $p: X_{1} {\rightarrow}X$, which is represented by $\gamma \in \Gamma$, acting on $\tilde{X}$, such that the diagram $$\begin{aligned}
\label{vak_diagram_1}
\xymatrix@1{
\tilde{X} \ar[r]^{\gamma} \ar[d] & \ \tilde{X} \ar[d] \\
X_{1} \ar[r]^{g} & X_{1}
} $$ is commutative. Then, for all $a \in \bbZ^n$, $x \in \tilde{X}$, the relation $$\begin{aligned}
\gamma a x = \gamma a \gamma^{-1} \gamma x
= (\gamma a \gamma^{-1}) \gamma x
= \mu(g)(a) \gamma x \end{aligned}$$ holds. Since $\tilde{X}$ is simply connected, by the theorem of Hurewicz, $H_1(\tilde{M})= 0$, [@massey].) Applying (\[topo\_diagram\_1\]), it follows that there exists a natural isomorphism $$H^1_{sing}(X_{1},A) {\rightarrow}H^1(\bbZ^n,A) \; ,$$ such that the diagram $$\xymatrix@1{
H^1_{sing}(X_{1},A) \ar@1{->}[r]^{\cong}\ar[d]^{g^*} & \ H^1(\bbZ^n,A)\ar[d]^{\mu(g)^*}\\
H^1_{sing}(X_{1},A) \ar@1{->}[r]^{\cong} & \ H^1(\bbZ^n,A) }$$ is commutative. Since the change of coefficients is natural, the stated commutative diagram is induced by composing with the projection $H^{0,1}(X) {\rightarrow}H^1(X,\bbR)$, and the de Rham isomorphism $ H^1(X,\bbR) {\rightarrow}H^1_{sing}(X,\bbR) $.
\[vak\_prop\_2\] Let $\Gamma$ be a virtually abelian Kähler group. Then $\Gamma$ admits a complex structure $J$. If $\Gamma$ is projective then it also admits a polarisable complex structure $J'$.
Let $X$ be Kähler with $\pi_{1}(X) = \Gamma$. Note that $H^1(\bbZ^n, \bbR)$ is isomorphic to $\Hom(\bbZ^n, \bbR)$, which is just the dual of the module $\bbZ^n \otimes \bbR$. Therefore, according to Lemma \[lemma:cohomologyisnat\], the isomorphism $H^1(X,\bbR)^* \cong (\bbZ^n \otimes \bbR)$, is a $G$-module isomorphism, which is defined over $\bbQ$. Since $G$ acts by holomorphic transformations on $Pic^0(X_{1})$, the action of $G$ on $H^1(X,\bbR)$ admits an invariant complex structure. In particular, by duality, $\bbZ^n \otimes \bbR$ attains a $G$-invariant complex structure $J$. It thus defines a complex structure for $\Gamma$. Since all identifications are defined over $\bbQ$, the complex structure $J$ is also polarisable if $X$ is projective.
Construction of certain Kähler and projective manifolds {#construction}
=======================================================
Let $\Gamma$ be a virtually abelian group, which admits an invariant complex structure $J$. We construct compact Kähler manifolds with fundamental group $\Gamma$.
\[lemma:vakg\_emb\] There exists a homomorphism $\rho: \Gamma {\rightarrow}\mathbb{C}^k \rtimes U(k)$, which embeds the subgroup $\bbZ^n$ as a lattice $\Lambda = \rho(\bbZ^n)$ into $\bbC^k$. If, in addition, $J$ is polarisable then there exists a polarisation for the lattice $\Lambda$.
We form the pushout $\bar{\Gamma}$ of , with respect to the inclusion $\iota: \bbZ^n \hookrightarrow \bbR^n$. This gives a homomorphism of exact sequences: $$\label{eq:extequiv}
\begin{CD}
1 @>>> \bbZ^n @>>> \Gamma @>>> G @>>> 1\\
@. @VV{\iota}V @VV{}V @VV{\id_{G}}V \\
1 @>>> \bbR^n @>>> \bar \Gamma @>>> G @>>> 1
\end{CD} \; \, \; \; \; ,$$ \
(Explicitly, the construction of $\bar \Gamma$ is as follows: Since $\Gamma$ is an extension of $\bbZ^n$ by $G$, it is isomorphic to a group of the form $$\begin{aligned}
(\mathbb{Z}^{2k} \times G, \star); & (z_1,g_1) \star (z_2, g_2) = (z_1 + \mu (g_1)z_2 + \eta(g_1, g_2), g_1 g_2) \; ,\end{aligned}$$ with $\eta \in Z^2_{\mu}(G,\mathbb{Z}^{2k})$ a two-cocycle with respect to $\mu$. Then $\bar \Gamma$ is constructed as $$\begin{aligned}
\bar \Gamma := (\mathbb{R}^{2k} \times G, \star) ; & (u_1,g_1) \star (u_2, g_2) = (u_1 + \mu(g_1) u_2 + \iota \, \eta(g_1, g_2),g_1 g_2 )\; . \; \;)\end{aligned}$$ Now $H^{2}_{\mu}(G, \mathbb{C}^{k}) = 0$, since $G$ is finite (see [@Brown Chapter III,§10]). Therefore, the second exact sequence splits, and there exists a homomorphism $$\label{eq:splitting}
\begin{CD}
1 @>>> \bbR^n @>>> \bar \Gamma @>>> G @>>> 1 \\
@. @VV{\id_{\bbR^n}}V @VV{}V @VV{\id_{G}}V \\
1 @>>> \bbR^n @>>> \bbR^n \rtimes_{\mu} G @>>> G @>>> 1
\end{CD} \; \, \; \; \; .$$ \
The semi-direct product $\bbR^n \rtimes_{\mu} G$ maps naturally into $\bbR^n \rtimes \GL(n,\bbR)$, which, via the above, constructs a homomorphism $f: \Gamma {\rightarrow}\bbR^n \rtimes \GL(n,\bbR)$. Since $\mu(G)$ has an invariant complex structure, there exists $A \in \GL(n,\bbR)$ such that $\mu^A(G) \leq \GL(k,\bbC)$. Since $\mu^A(G)$ is finite, it admits an invariant hermitian metric, and we may as well assume that $\mu^A(G) \leq \U(k)$. Thus $\rho(g) = A f(g) A^{-1}$ defines the required homomorphism $$\rho: \Gamma {\rightarrow}\bbC^k \rtimes \U(k) \;$$ and $$\Lambda = \rho(\bbZ^n) = A(\bbZ^n)$$ is a lattice in $\bbC^k$. Assume furthermore that $\omega$ is a polarisation for $J$. Then $$A^* \omega \, (u,v) = \omega(A^{-1}u, A^{-1}v)$$ defines a polarisation for the lattice $\Lambda$ in $\bbC^k$.
Let $Y$ be a compact Kähler manifold with $\pi_1(Y)=G$. Such a manifold exists by [@serre Proposition 15], in fact, it may be chosen to be projective. Let $\tilde{Y}$ be the universal covering of $Y$, and put $$\begin{aligned}
\tilde{X}:= \; \mathbb{C}^k \times \tilde{Y} \; . \end{aligned}$$ Let $h: \Gamma {\rightarrow}G$ denote the quotient homomorphism. Since $G$ acts on $\tilde{Y}$ by covering transformations, we can form the diagonal action of $\Gamma$ on $\tilde X$: $$\begin{aligned}
\gamma \cdot (u,y) &=& ( \rho(\gamma) u, h(\gamma) y)\; . \end{aligned}$$ Note that $\rho(\Gamma)$ is a discrete subgroup of $\mathbb{C}^k \rtimes U(k)$ and acts with compact quotient on $\mathbb{C}^k$. In particular, the action of $\Gamma$ on $\tilde{X}$ is properly discontinuous. Moreover, $\ker \rho$ is a finite subgroup of $\Gamma$, which projects isomorphically onto $\ker \mu$. Therefore the action of $\Gamma$ on $\tilde{X}$ is also free, since $G$ acts freely on $\tilde Y$. Since $\Gamma$ acts by holomorphic transformations on $\tilde{X}$, the space of orbits $$X: = \; (\mathbb{C}^k \times \tilde{Y}) / \, \Gamma$$ is a compact complex manifold. Moreover, $X$ is also Kähler, since $\Gamma$ acts by isometries of the product metric on $\tilde{X}$. Put $T= \bbC^k/ \Lambda$. Then there is a sequence of holomorphic coverings $$\begin{aligned}
\label{vak_covering}
\xymatrix@1{
\tilde{X}= \mathbb{C}^k \times \tilde{Y} \ar[d]
\\
T \times \tilde{Y} \ar[d] \\
X = (T \times \tilde{Y}) / G
} \; , \end{aligned}$$ and also a locally trivial holomorphic fibering $$T {\longrightarrow}X {\longrightarrow}Y \; \; \, .$$
We have
\[prop:constructX\] The complex manifold $X$ is compact Kähler, and $\pi_{1}(X) = \Gamma$. If furthermore $J$ admits a polarisation then $X$ is a projective manifold.
We already established that $X$ is compact Kähler. If $J$ is polarised, then by Lemma \[lemma:vakg\_emb\], $T$ is an abelian variety. By [@serre Proposition 15], we may choose $\tilde{Y}$ as a projective manifold. Therefore, $T \times \tilde Y$ is projective. By the quotient lemma (cf. Lemma \[lemma:fquotients\] below), $X$ is projective.
Remark, if $\Gamma$ is torsionfree then the homomorphism $$\Gamma {\longrightarrow}\bbC^k \rtimes U(k)$$ constructed above is injective, and moreover, $\Gamma$ acts freely on $\bbC^k$. Therefore, $$\begin{aligned}
\label{flat_covering}
\xymatrix@1{
T \ar[d] \\
M = \bbC^k/ \Gamma
} \end{aligned}$$ is a holomorphic covering of flat Kähler manifolds, where $\pi_{1}(M) = \Gamma$. If $J$ has a polarisation, $T$ is an abelian variety, and thus by Lemma \[lemma:fquotients\] the complex manifold $M = T/G$ is projective. This proves Corollary \[cor:flatKaehler\].
#### The quotient lemma for Kähler manifolds and projective varieties
In the following form the quotient Lemma is an immediate application of Kodaira’s projective embedding theorem. (For the algebraic geometry variant, see section \[sect:fieldsofdef1\].)
\[lemma:fquotients\] Let $M,\bar{M}$ be compact complex manifolds and $p:\bar{M} \to M$ be a finite, holomorphic covering. Then $\bar{M}$ is projective, if and only if $M$ is projective.
Assume that $\bar{M}$ is projective. Then there is a Kähler metric such that $[\omega] \in H^2(\bar{M}, \mathbb{Q})$, where $\omega$ is the Kähler form. Let $G:= Deck(\bar{M}, p)$ denote the decktransformation group of the covering, which is a finite group. Now put $\bar{\theta}:= \sum_{g \in G} g^* \omega$, where $g^* \omega$ denotes the pull back of $\omega$ by the action of $g \in G$ on $\bar{M}$. Then $\bar \theta$ is a $G$-invariant two-form on $\bar{M}$, and it is also a Kähler form, since the covering transformations are holomorphic maps. Its cohomology class $[\bar{\theta}] \in H^2(\bar{M}, \mathbb{Q})$ is invariant by $G$. Let $\theta$ be the unique two-form on $M$ which satisfies $p^*\theta = \bar{\theta}$, and put $ H^2(\bar{M}, \mathbb{\bbR})^G$ for the $G$-invariant cohomology classes. Recall (cf. [@maclane Prop. 11.14]) that the map $p^*: H^2(\bar{M}, \mathbb{\bbR})^G \cong H^2(M,\mathbb{\bbR})$ is an isomorphism. Since this map is also compatible with the rational structures, it follows that $[\theta] \in H^2(M, \mathbb{Q})$. By Kodaira’s theorem (cf. [@Wells Chapter VI]), $M$ is a projective manifold.
For the converse, assume that $M$ is a projective manifold and let $[\theta] \in H^2(M,\mathbb{Q})$ be the Kähler class of a Kähler metric on $M$. Then $\omega= p^*\theta$ is a Kähler form on $\bar M$ with rational Kähler class. It follows that $\bar M$ is projective.
Deformation spaces of torus quotients and existence of polarisations
====================================================================
The local deformation space for a complex torus $$X= T^{k}$$ is represented by the space of all complex structures $${\mathfrak D} = \{ J: \bbR^{2k} \to \bbR^{2k} \mid J^2 = -\id \} =
\GL(2k,\bbR) \big/ \GL(k,\bbC) \; .$$ (Compare [@KodairaSpencerI-II p.408ff] or [@ShimizuUeno §2.1] for details.) Now let $$X_{G} = X/G$$ be a finite quotient, and $\mu: G \to \GL(k,\bbC)$ be the holonomy homomorphism associated to the action of $G$ on $X$ (see section \[construction\]). Then $${\mathfrak D}(G) = \{ J \in {\mathfrak D} \mid A J A^{-1} =J \text{, for all $A$ in $\mu(G)$} \}$$ describes the local deformation space for $X_{G}$.\
The main result of this section is:
\[prop:contdeformation\] Let $J \in {\mathfrak D}(G)$. Then there exists a continuous deformation $\bar J = (J_{t}): I= [0,1] \to {\mathfrak D}(G)$, with $J_{0}= J$, such that $J_{1}$ is polarisable.
Moreover, we will show in Proposition \[prop:density\] below that the set $ {\mathfrak D}(G)_{p}$ of polarisable complex structures is dense in $ {\mathfrak D}(G)$. This proves Theorem \[thm:deformation\] in the introduction.\
We deduce
Let $X_{G} = T/G$ be a finite quotient manifold of a complex torus $T$. Then $X_{G}$ may be continuously deformed to a projective manifold.
For the proof of Proposition \[prop:contdeformation\], we will study the cone of Kähler forms $\Omega(G)$ which is associated to $G$ and relate it to the deformation space ${\mathfrak D}(G)$.
Kähler forms and Siegel upper half space
----------------------------------------
Let $J: \bbR^{2k} {\rightarrow}\bbR^{2k}$ be a complex structure. Recall that a non-degenerate two-form $\omega \in \bigwedge^2 (\bbR^{2k})^*$ is called compatible with $J$ (or $J$-hermitian), if $$\omega( J \cdot, J \cdot ) = \omega( \cdot,\cdot) \; .$$ Thus, equivalently, $\omega$ is $J$-hermitian if and only if $J \in \Sp(\omega)$. If $\omega$ is $J$-hermitian and the associated symmetric bilinear form $g_{\omega,J} = \omega(J \cdot, \cdot)$ is positive definite then $\omega$ is called a *Kähler form* for $J$.
### Complex structures with common Kähler form
We define $${\mathfrak D}^+_{\omega}= \{ J \in {\mathfrak D} \mid \text{$\omega$ is a K\"ahler form for $J$} \},$$ to be the set of all complex structures which admit $\omega$ as Kähler form. Note that ${\mathfrak D}^+_{\omega}$ is non-empty and $\Sp(\omega)$ acts on ${\mathfrak D}^+_{\omega}$ by conjugation. In fact, this action is transitive:
\[Jswithomega\] Let $J_{0} \in {\mathfrak D}^+_{\omega}$ be a complex structure which has Kähler form $\omega$, and let $J \in {\mathfrak D}$. Then $J \in {\mathfrak D}^+_{\omega}$ if and only if there exists $A \in \Sp(\omega)$ such that $J= A J_{0} A^{-1}$.
This is a consequence of Lemma \[Jconjugacy\] below. We obtain
\[moduli\_siegel\] Let $J \in {\mathfrak D}^+_{\omega}$. There is a bijective correspondence $$\begin{aligned}
\xymatrix@1{
\mathfrak{S} := \; \Sp(\omega) \big/ \left(\Sp(\omega) \cap\GL(J)\right) \ar@1{<->}[d]^{~}
\\ \mathfrak{D}^+_{\omega} } \end{aligned}$$
The claimed correspondence is established by the map $$\begin{aligned}
[A] \in \mathfrak{S} \to A J A^{-1} \; .\end{aligned}$$ The correspondence is bijective by Proposition \[Jswithomega\].
Note that $\Sp(\omega) \cap \GL(J)$ is the unitary group of the associated hermitian form $g_{\omega,J}$. It is a maximal compact subgroup of $\Sp(\omega)$, since $J \in {\mathfrak D}^+_{\omega}$. The space $ \mathfrak{S}$ is customarily called the Siegel upper half plane, and it is a Riemannian symmetric space in the sense of É. Cartan (cf. [@helgason]).
### Conjugacy classes in $\Sp(\omega)$
Let $J \in \Sp(\omega)$. We put ${\rm sign}\, (\omega,J) $ for the signature of $g_{\omega,J}$. The signature classifies the conjugacy classes of complex structures $J \in \Sp(\omega)$.
\[Jconjugacy\] Let $J $, $J'$ be complex structures such that $J$ and $J'$ are in $\Sp(\omega)$. Then ${\rm sign}\, (\omega,J) = {\rm sign} \, (\omega,J')$ if and only if there exists $A \in \Sp(\omega)$ such that $J'= A J A^{-1}$.
Clearly, ${\rm sign}\, (\omega,J) = {\rm sign} \, (\omega,J')$ if $J'= A J A^{-1}$, for some $A \in \Sp(\omega)$. We show now that $J$ and $J'$ are conjugate in $\Sp(\omega)$ if ${\rm sign}\, (\omega,J) = {\rm sign} \, (\omega,J')$. We choose $B \in \GL(2k,\mathbb{R})$ such that $J'= BJB ^{-1}$. Then $$({B^{-1}})^* \omega \, ( \cdot, \cdot): = \omega(B^{-1} \cdot ,B^{-1} \cdot)$$ defines a compatible sympletic form for $J'$. Since $$g_{ ({B^{-1}})^* \omega ,J'} =
({B^{-1}})^* g_{\omega,J} \; ,$$ we have ${\rm sign}\, (J', ({B^{-1}})^* \omega ) = {\rm sign} \, (J',\omega)$. By the existence of a unitary basis, all hermitian symmetric forms for $J'$ of the same signature are equivalent under a $J'$-linear transformation. Hence, there exists $ \tilde{B} \in \GL(2k,J')$, such that $ ({B^{-1}})^* \omega= {\tilde{B}}^*\omega$. Now we have $$({\tilde{B} B})^*\omega = B^* (\tilde{B}^* \omega) = B^*( ({B^{-1}})^* \omega) = \omega \; .$$ Thus, $\tilde{B}B \in \Sp(\omega)$. Since $J'= \tilde{B} J' \tilde{B}^{-1} = \tilde{B} B J (\tilde{B} B)^{-1}$, our claim follows.
The cone of Kähler forms with respect to $G$
--------------------------------------------
Let $\Omega \subset \bigwedge^2 (\bbR^{2k})^*$ be the set of all non-degenerate two forms. Then $\Omega$ is an open cone in the vector space $\bigwedge^2 (\bbR^{2k})^*$. For a subgroup $G \leq \GL(2k,\bbR)$, we let $ \left(\bigwedge^2 (\bbR^{2k})^*\right)^G$ denote the vector space of $G$-invariant two forms. We then define $$\Omega(G) = \; \Omega \, \cap \,\left({\bigwedge}^2 (\bbR^{2k})^*\right)^G \; .$$
Let $J \in {\mathfrak D}$ be a complex structure. If $G$ is a finite (or compact) subgroup of $\GL(J)$, there always exists a non-degenerate $G$-invariant two-form:
\[lemma:Kaehlercones1\] Let $G \leq \GL(J)$ be a finite subgroup. Then there exists a $G$-invariant Kähler form $\omega$ for $J$.
Let $\omega'$ be a Kähler form for $J$. Since $G$ is finite, we may find (by averaging $\omega'$ over $G$) a $G$-invariant Kähler form $\omega$.
In particular, if $G$ is finite the cone $\Omega(G)$ is a non-empty and open subset in the vector space $\left(\bigwedge^2 (\bbR^{2k})^*\right)^G$.\
\[lemma:Kaehlercones2\] Let $\omega \in \Omega(G)$. If $G$ is finite then there exists a $G$-invariant complex structure $J \in {\mathfrak D}(G)$ which has Kähler form $\omega$.
By our assumption $G \leq \Sp(\omega)$. Then $G$ acts by conjugation on the space ${\mathfrak S}={\mathfrak D}(G)_{\omega}$ of complex structures with Kähler form $\omega$, as is described in Corollary \[moduli\_siegel\]. By Cartan’s theorem (see §\[Cartan\]), $G$ has a fixed point $J$ in ${\mathfrak D}(G)_{\omega}$.
We call $\Omega(G)$ the cone of Kähler forms for $G$.
### Rational points in $\Omega(G)$ {#rationalpoints}
Next let $$\Omega(G)(\bbQ) = \; \Omega(G) \, \cap \, {\bigwedge}^2 (\bbQ^{2k})^*$$ denote the set of rational points in $\Omega(G)$. We remark
\[lemma:rationaldense\] Let $G \leq \GL(2k,\bbQ) \cap \GL(J)$ be a finite subgroup. Then the set of rational points $\Omega(G)(\bbQ)$ is dense in $\Omega(G)$. In particular, $\Omega(G)(\bbQ)$ is non-empty.
Since $G \leq \GL(2k,\bbZ)$, the vector subspace $ \left({\bigwedge}^2 (\bbR^{2k})^*\right)^G $ of $\bigwedge^2(\bbR^{2k})^*$ is defined by rational equations, and it has a $\bbQ$-structure given by the vector space of rational solutions $\left(\bigwedge^2 (\bbQ^{2k})^*\right)^{G}$.
This proves, in particular:
\[prop:symplectic\] Let $G \leq \GL(2k,\bbZ)$ be a finite subgroup, which has an invariant complex structure $J$. Then there exists $A \in \GL(2k,\bbQ)$ such that $A \, G \, A^{-1} \leq \Sp(2k,\bbQ)$.
We are ready now for the
The equivalence of 1. and 3. is a consequence of Theorem \[thm:deformation\]. The implication from 1. to 2. is implied by the previous Proposition \[prop:symplectic\], and its converse by Lemma \[lemma:Kaehlercones2\].
#### Kähler cones and polarisations
Let $H^{1,1}(J,\bbR)$ denote the vector subspace of $J$-hermitian forms in $\bigwedge^2 (\bbR^{2k})^*$. The set of Kähler forms $\kappa(J)$ for $J$ is an open convex cone in $H^{1,1}(J,\bbR)$, which is called the Kähler cone for $J$. Consider $$H^{1,1}(J,\bbR)^G = H^{1,1}(J,\bbR) \cap
\left({\bigwedge}^2 (\bbR^{2k})^*\right)^G \;$$ we then call $$\kappa(J,G) = \kappa(J) \cap
H^{1,1}(J,\bbR)^G$$ the Kähler cone for $J$ and $G$. (It is, in fact, the Kähler cone $\kappa(X_{G})$ of $X_{G}$.) Note that the complex structure $J$ admits a polarisation if and only if $\kappa(J,G)$ contains a rational point, that is, if $$\kappa(J,G) \cap {\bigwedge}^2 (\bbQ^{2k})^* \, \neq \emptyset \; .$$ By Lemma \[lemma:Kaehlercones2\], we have that $$\Omega(G) = \bigcup_{J \in {\mathfrak D}(G)} \kappa(J,G) \; .$$ Therefore, we may deduce from Lemma \[lemma:rationaldense\]:
Let $G \leq \GL(J)$ be a finite subgroup. Then there exists $J' \in {\mathfrak D}(G)$ such that $\kappa(J',G)$ has a rational point.
In the remainder of this section, we shall show that such $J'$ may be obtained by a continuous deformation in ${\mathfrak D}(G)$ starting from $J$.
Lifting of curves from $\Omega(G)$ to ${\mathfrak D}(G)$
--------------------------------------------------------
Let $J \in {\mathfrak D}$ be a complex structure, and let $\omega \in \Omega$ be a Kähler form for $J$. We may lift curves in $\Omega$ starting in $\omega$ to curves in ${\mathfrak D}(G)$, starting in $J$:
\[lemma:Kaehlerlifting1\] Let $\bar \omega = (\omega_{t}): I {\rightarrow}\Omega$ be a continuous curve of non-degenerate two-forms, and let $J \in {\mathfrak D}^+_{\omega_{0}}$. Then there exists a continuous curve $\bar J= (J_{t}): I {\rightarrow}{\mathfrak D}$, with $J_{0}=J$, such that $\omega_{t}$ is a Kähler form for $J_{t}$.
Note that the transitive action of $\GL(2k,\bbR)$ on $\Omega$ gives an identification $$\Omega = \GL(2k,\bbR) / \Sp(\omega_{0}) \; .$$ We may thus consider a lift (horizontal lift of the corresponding $ \Sp(\omega_{0})$-principal bundle) of $\bar \omega$, to obtain a continuous curve $\bar A =(A_{t}): I {\rightarrow}\GL(2k,\bbR)$ such that $$\omega_{t} =(A_{t})^* \omega_{0}.$$ Now let $J_{0}$ be any complex structure, which has $\omega_{0}$ as a Kähler form. Then $$J_{t} = A_{t}^{-1}J_{0} A_{t}$$ defines the desired lift of $\omega_{t}$ to complex structures.
We now provide a lifting construction for curves in $\Omega(G)$.
\[lemma:Kaehlerlifting2\] Let $\bar \omega = (\omega_{t}): I {\rightarrow}\Omega(G)$ be a continuous curve of non-degenerate $G$-invariant two-forms, and let $J \in {\mathfrak D}(G)$, such that $\omega_{0}$ is a Kähler form for $J$. Then there exists a continuous curve $\bar J= (J_{t}): I {\rightarrow}{\mathfrak D}(G)$ of $G$-invariant complex structures, with $J_{0} =J$, such that $\omega_{t}$ is a Kähler form for $J_{t}$.
Write $\omega_{t} = {A_{t}}^* \omega_{0}$, as in the proof of Lemma \[lemma:Kaehlerlifting1\]. Then consider $$\mu_{t}: I {\rightarrow}\Hom(G, \Sp(\omega_{0})) \; ,$$ which is given by $\mu_{t}(g) = A_{t} g A_{t}^{-1}$, $g \in G$. Note here, since $$G \subset \Sp(\omega_{t}) =
A_{t}^{-1}\, \Sp(\omega_{0}) A_{t} ,$$ we have that $$\mu_{t}(G) \subset \Sp(\omega_{0}) \; .$$ Now let $$\bar C= (C_{t}): I {\rightarrow}{\mathfrak D}^+_{\omega_{0}}$$ be the continuous curve of barycenters for $\mu_{t}(G) J$, with $C_{0}= J$, as is constructed in Proposition \[prop:Cartancontinuous\]. The corresponding complex structures $C_{t}$ are $\mu_{t}(G)$-invariant, and have Kähler form $\omega_{0}$. Then it is clear that $$J_{t} = A_{t}^{-1} C_{t} A_{t}$$ is a family of complex structures, such that $J_{t} \in {\mathfrak D}(G)$ and $\omega_{t}$ is a Kähler form for $J_{t}$. (Remark that $J_{t}$ is actually the $G$-barycenter of $A_{t}^{-1} J_{0} A_{t}$ in ${ \mathfrak D}_{\omega_{t}}$.)
### Deformation to polarisable complex structures
We are now ready for the
Choose $\omega_{0} \in \Omega(G)$, such that $\omega_{0}$ is a Kähler form for $J$. Let $\omega \in \Omega(G)(\bbQ)$ be a rational two-form near $\omega_{0} $, and $\bar \omega = (\omega_{t}): I {\rightarrow}\Omega(G)$ a continuous path in $\Omega(G)$, such that $\omega_{1} = \omega$. Then define $\bar J =(J_{t})$ to be the lift of $\bar \omega$, with $J_{0} = J$, as in Lemma \[lemma:Kaehlerlifting2\].
### Density of polarisable complex structures
We may slightly refine the proof of Proposition \[prop:contdeformation\] to obtain:
\[prop:density\] The set of polarisable complex structures ${\mathfrak D}(G)_{p}$ is dense in ${\mathfrak D}(G)$.
Let $J_{0} \in {\mathfrak D}(G)$ with Kähler form $\omega_{0}$. Let $\ {\mathfrak D}(G) \cap U$ be a neighbourhood of $J_{0}$, where $U$ is open in $ {\mathfrak D}$. We show that ${\mathfrak D}(G)_{p} \cap U \neq \emptyset$ as follows:\
i) Consider the action of $\GL(n,\bbR)$ on $ {\mathfrak D}$ by conjugation. Choose neighbourhoods $W$ of $1 \in \GL(n,\bbR)$ and $U_{1} \subset {\mathfrak D}$ of $J_{0} $ such that $W U_{1} \subset U$.\
ii) By continuity of the symmetric space distance function $d$ on $ {\mathfrak D}^+_{\omega_{0}}$ (cf. §\[Cartan\]), choose a metric ball $U_{\epsilon}= B_{\epsilon}(J_{0}) \subset {\mathfrak D}^+_{\omega_{0}}$ with diameter $r$ such that the ball $B_{r+\epsilon}(J_{0})$ is contained in $U_{1}$.\
iii) For $g \in G$, consider the map $\pi_{g}: \GL(n,\bbR) {\rightarrow}{\mathfrak D}$, $A \mapsto A g A^{-1} J_{0} A g^{-1} A^{-1}$. Since $J_{0} \in {\mathfrak D}(G)$, $\pi_{g}(1) = J_{0}$. Choose an open neighbourhood $W_{\epsilon} \subset W$ of $1 \in \GL(n,\bbR)$ such that, for all $g \in G$, $\pi_{{g}}(W_{\epsilon}) \cap {\mathfrak D}^+_{\omega_{0}}
\subset U_{\epsilon}$.\
iv) Consider the quotient map $\pi_{{\Omega}}: \GL(n,\bbR) {\rightarrow}{\Omega}$, where $\pi_{{\Omega}}(1) = \omega_{0}$. Since $\pi_{{\Omega}}$ is locally a projection, we may assume that $W_{\epsilon} = Z \times V_{\epsilon}$, where $V_{\epsilon} \subset \Omega$ is a neighbourhood of $\omega_{0}$ and $Z$ is a neighbourhood of the identity in $\Sp(\omega_{0})$.\
Now let $\omega_{1} \in V_{\epsilon} \cap \Omega(G)(\bbQ)$ be a rational form, and join $\omega_{0}$ and $\omega_{1}$ by a path $(\omega_{t}): I {\rightarrow}\Omega(G) \cap V_{\epsilon}$. Using iv), we have a lift $(A_{t}): I {\rightarrow}V_{\epsilon} \subset W_{\epsilon}$. Let $\mu_{t}(G) = A_{t} G A_{t}^{-1}$. Then the orbit $\mu_{t}(G) J_{0}$ is contained in ${\mathfrak D}^+_{\omega_{0}}$. It follows by iii) that $\mu_{t}(G) J_{0} \subset U_{\epsilon}$. Using the remark following Proposition \[prop:Cartancontinuous\] it follows from ii) that the associated curve of barycenters $\bar C$ is contained in $U_{1}$. Moreover, since $A_{t} \in W$, it follows by i) that the curve of $G$-invariant complex structures $J_{t}: I {\rightarrow}{\mathfrak D}(G)$ is contained in $U$. Therefore, $J_{1} \in U \cap {\mathfrak D}(G)$ is polarisable with Kähler form $\omega_{1}$.
The barycenters of a continuous deformation {#Cartan}
-------------------------------------------
Let $G/K$ be a symmetric space, where $G$ is a real semisimple Lie group with finite center, and $K$ is a maximal compact subgroup. The following fact is due to É. Cartan (see [@helgason]\[Chapter I, Theorem 13.5\]): Let $\mu \leq G$ be a finite (or compact) subgroup. Then there exists a point $s \in G/K$ such that $\mu s = s$.\
We shall need the following refinement of Cartan’s result.
\[prop:Cartancontinuous\] Let $\varphi_{t}: I {\rightarrow}\Hom(\mu,G)$ be a continuous deformation, where $\varphi_{0} (g) =g$, for all $g \in \mu$. Let $q \in G/K$ be a fixed point for $\mu$. Then there exists a canonical continuous curve $\bar q = (q_{t}): I {\rightarrow}G/K$ with $\varphi_{t}(\mu) q_{t} = q_{t}$, and $q_{0}=q$.
For the proof of Proposition \[prop:Cartancontinuous\], we need two elementary lemmata:
\[Lemma1\] Let $X$ be a compact metric space, and $f: I \times X \to \bbR$ a continuous function, where $I= [0,1]$. For $t \in I $, define $$\bar f (t) = \min \{ f(t,q) \mid q \in X \} \, \; .$$ Then $\bar f: I \to \bbR$ is continuous.
As an infimum of continuous functions, $\bar f$ is upper semi-continuous. It remains to show that $\bar f$ is lower semi-continuous. That is, we show that given any sequence $t_{m} {\rightarrow}t$, $\liminf \bar f(t_m) \geq \bar f(t)$. Assume to the contrary that $\lim \bar f(t_m) < \bar f(t)$, for some sequence $t_{m} {\rightarrow}t$. We use the fact, that there exists a subsequence converging to $\liminf$. Since the minimum of $f(t_{m}, \cdot)$ is assumed on $X$, there exist $q_{m} \in X$ such that $f(t_{m},q_{m}) = \bar f(t_{m})$. By compactness of $X$, we may therefore also assume that $q_{m} \to p \in X$. By continuity of $f$, $\bar f(t_{m})=f(t_{m}, q_{m}) \to f(t,p) \geq \bar f(t)$. A contradiction.
\[Lemma2\] With the assumptions of Lemma \[Lemma1\], assume that the minimum $\bar f_{}(t)$ is attained at a unique point $q_{t} \in X$. Then the curve $t \to q_{t}$ is continuous.
It is clearly enough to show that every sequence $t_{m} {\rightarrow}t$, has a subsequence $s_{m}$ such that $q_{s_{m}} \to q_{t}$. By Lemma \[Lemma1\], $\bar f$ is continuous. Therefore, $ f(t_{m},q_{t_{m}})= \bar f(t_{m}) \to \bar f(t)= f(t,q_{t})$. Since $X$ is compact, there exists a subsequence $s_{m}$, such that $q_{s_{m}} \to p \in X$. By continuity of $f$, $f(t,p) = \bar f(t)$, and, therefore, $f(t, \cdot)$ attains its minimum at $p$. Since, by our assumption, the minimum point is unique, we conclude that $p= q_{t}$. Therefore, we have $q_{s_{m}} \to q_{t}$.
Let $d$ denote the distance function on the symmetric space $G/K$. By [@helgason]\[Chapter I,Theorem 13.5\]), given $p_{0} \in X$, the continuous function $$f: X {\rightarrow}\bbR \, , \,
f(q) = \sum_{g \in \mu} d^2(q, g p_{0})$$ attains its minimum at a *unique* point $q_{0} \in G/K$ (the barycenter of the orbit $ \mu \, p_{0})$. Moreover, $q_{0} \in C$, where $C$ is a compact ball around $p_{0}$ containing $\mu \, p_{0}$. Therefore, the function $$f(t, q) = \sum_{g \in \mu} d^2(q, \varphi_{t}(g) p_{0})$$ satisfies the assumptions of the above two lemmata. Then the minimum point $q_{t}$ for $\bar f_{t}$ is the barycenter for the orbit $ \varphi_{t}(\mu) p_{0}$, and $q_{t}$ is a fixed point for $ \varphi_{t}(\mu)$. Thus, by Lemma \[Lemma2\], the curve of barycenters $t \to q_{t}$ is continuous. Choosing, $p_{0}=q$, we obtain $q_{0} = q$, as required.
The ball $C = C(p_{0})$, which appears in the proof may be chosen as follows. Let $B_{r}(p_{0})$ be a metric ball of radius $r$, which contains the orbit $ \mu \, p_{0}$, and let $\ell$ denote its diameter. Then $f(q) > f(p_{0})$, for all $q \notin B_{r+\ell}(p_{0})$. Thus, we may take $C$ to be the closure of $B_{r+\ell}(p_{0})$.
Fields of definition
====================
As we have seen, the set of polarisable complex structures $ \mathcal{D}_{p}(G)$ is dense in the space of all $G$-invariant complex structures $\mathcal{D}(G)$. In the following, we explicitly construct points in $\mathcal{D}_{p}(G)$ which have additional symmetry. For this, we employ the approach used by Johnson in [@Johnson] (also implicit in [@Serre2]). As an advantage, we can deduce information about the fields of definition of the abelian varieties, which appear in the constructions in section \[construction\]. This leads us to the proof of Theorem \[thm:numberfields\]
Preliminaries
-------------
We collect some basic material from the theory of abelian varieties. Our principal references are [@Lang] and [@Shimura].
#### Field of definition for an abelian variety {#sect:fieldsofdef1}
Let $T= \mathbb{C}^g /\Lambda$ be a complex torus, and $$End(T):= \{f \in End(\mathbb{C}^g)\; | \; f \Lambda \subset \Lambda \}$$ the endomorphism ring of $T$. Recall that $T$ is called an abelian variety if it embeds holomorphically as a subvariety of a projective space. An abelian variety defined over a subfield $k$ of $\bbC$ is a complex projective variety $X$ defined over $k$, which has an algebraic group structure defined over $k$. If $T$ is biholomorphic to an abelian variety $X$ defined over $k$, then we will say that $T$ is defined over $k$. Correspondingly, the endomorphism ring $End(T)$ of $T$ is said to be defined over $k$ if its elements correspond to $k$-defined algebraic automorphisms of $X$.
#### The quotient of a projective variety by a finite group
Let $X$ be a complex projective variety defined over a subfield $k$ of $\bbC$. We let $\Aut_k(X)$ denote the group of $k$-defined automorphisms of $X$.
\[lemma:algebraicquotient\] Let $Z$ be a complex projective variety defined over $k$ and let $G \subset
\Aut_k(Z)$ be a finite subgroup. Then the space of orbits $Z/G$ admits the structure of a $k$-defined projective variety, and $Z \to Z/G$ is a finite morphism defined over $k$.
For the standard proof see [@serre §13] or [@Shimura Ch. 4.3, Prop 16]. We briefly sketch a more specialised argument which works in the context of projective varieties. Let $R$ be the homogeneous coordinate ring of $Z$. Then $R= R_k \otimes \mathbb{C}$, where $R_k$ is a finitely generated graded $k$-algebra. According to Noether’s theorem, the algebra of invariants $R_k^G $ is a finitely generated graded $k$-algebra. It follows that $R^G = R_k^G \otimes \mathbb{C}$ is a finitely generated $\mathbb{C}$-algebra with $k$-structure. To any such algebra $A$ one can assign a scheme, called $proj(A)$, such that $proj(A)$ is a projective variety defined over $k$. See [@Mumford2 p. 282]. Then $Z/G= proj(R^G)$ is the desired quotient variety.
Note that $\pi: Z {\rightarrow}Z/G$ is a quotient in the categorical sense. This means, given any $k$-defined morphism $f: Z {\rightarrow}Y$, where $Y$ is a variety defined over $k$, and $f$ is constant on $G$-orbits, there exists a unique $k$-defined morphism $\bar f: Z/G {\rightarrow}Y$ such that $f = \bar f \pi$.
#### Abelian varieties with complex multiplication
A number field $F$ is called a CM-field, if $F$ is a totally imaginary quadratic extension of a totally real number field $E$. In particular, we have $[F:\mathbb{Q}] = 2g$ for $g=[E: \mathbb{Q}]$ and $F= E(a)$ for a totally imaginary $a \in F$ with $a^2 \in E$. A choice of $g$ non conjugate embeddings $\varphi_i:
F \hookrightarrow \mathbb{C}$, $i = 1, \ldots ,g$, is called a CM-type. It gives rise to an embedding $\varphi= (\varphi_i)_{i=1, \ldots ,g} : F \to \mathbb{C}^g$ which extends to an isomorphism $\Phi: F \otimes_{\mathbb{Q}} \mathbb{R} \to \mathbb{C}^g $.
Let $T$ be a complex torus, and $End_{\mathbb{Q}}(T):= End(T) \otimes \mathbb{Q}$. Then $T$ is said to have complex multiplication by a CM-field $F$, if there exists an embedding $\iota: F
\to
End_{\mathbb{Q}}(T)$ such that $2g= 2 dim (T)$. For any CM-field $F$ one can construct an abelian variety with complex multiplication by $F$. Let $\Lambda$ be a lattice in $F$. Then $\Phi(\Lambda) \subset
\mathbb{C}^g$ is a lattice in $\mathbb{C}^g$ and $$X_F:= \mathbb{C}^g
/\Phi(\Lambda)$$ is a complex torus with $F \subset
End_{\mathbb{Q}}(X_F)$. Since $X_{F}$ admits a polarisation (see [@Lang Thm. 4.1] or [@Shimura Ch. 6]), it is an abelian variety with complex multiplication by $F$. Since $X_{F}$ has complex multiplication by a CM-field $F$, $X_{F}$ and $End(X_{F})$ may be defined over an algebraic number field $k$ (see [@Lang Ch.5; Prop. 1.1] or [@Shimura]).
$\mathbb{Q}[G]$-modules with complex multiplication
---------------------------------------------------
Let $G$ be a finite group and let $V$ be a finite dimensional $\mathbb{Q}[G]$-module. If there exists a homomorphism of rings $F {\rightarrow}End_{\mathbb{Q}[G]}(V)$, where $F$ is a CM-field, we say that $V$ has complex multiplication by $F$.
\[prop:specialJ\] Let $V$ be a $\mathbb{Q}[G]$-module, which has complex multiplication by $F$. Then there exists a complex structure $J_{F}$ on $V \otimes_{\mathbb{Q}} \mathbb{R}$, such that $F \subset End(V \otimes_\mathbb{Q} \mathbb{R},J_{F})$ and $J_{F} \in End_{\mathbb{R}[G]}(V\otimes_{\mathbb{Q}} \mathbb{R})$.
Choose a CM-type for $F$. By pulling back via $\Phi: F \otimes_{\mathbb{Q}} \mathbb{R} \to \mathbb{C}^g$ we obtain a complex structure on $F \otimes_{\mathbb{Q}} \mathbb{R}$. Since $V$ is an $F$-module, $V \cong F^m$, for some $m \in \mathbb{N}$. We let $J_{F}$ denote the product complex structure on $V \otimes_{\mathbb{Q}} \mathbb{R}$. Then $F \subset End(V \otimes_\mathbb{Q} \mathbb{R},J_{F})$.
Write $F = E(a)$, as above, and put $b = a^2 \in E$. Let $E_i:= \varphi_i (E)$. Note that $b$ acts on $F \otimes_{E_i} \mathbb{R}$ by scalar multiplication with $\varphi_i(b)$. We have a $G$-invariant decomposition $V \otimes_{\mathbb{Q}} \mathbb{R} = \bigoplus_{i=1}^g V \otimes_{E_i}
\mathbb{R}$. With respect to this decomposition $J_{F}$ acts as multiplication with $a \otimes \frac{1}{\sqrt{\varphi_i(b)}}$ on each factor. Since $a
\in End_{\mathbb{Q}[G]}(V)$, it follows that $J_{F}$ is $G$-invariant.
If $V$ is an irreducible ${\mathbb{Q}[G]}$-module then $D= End_{\mathbb{Q}[G]}(V) $ is a finite dimensional division algebra. Moreover, this division algebra admits a positive involution [@Johnson Proof of Prop. 3.2]. By a classification of Albert either $D$ contains a CM-field $F$ or $D$ is a totally real division algebra. (See [@Johnson §1; §3; Prop. 3.1], [@Shimura2] and [@Albert]). In the latter case, $End_{\mathbb{Q}[G]}(V \oplus V)$ has complex multiplication by a CM-field $F$. This has the following consequence:
\[prop:CMdecomposition\] Let $V$ be a $\mathbb{Q}[G]$-module such that $V \tensor_{\bbQ} \bbR$ admits a $G$-invariant complex structure. Then there exists a decomposition of $V$ into $G$-submodules $W_{j}$, such that $W_{j}$ admits complex multiplication by a CM-field $F_{j}$.
In fact, $V = \bigoplus V_i$ where $V_i$ are irreducible $\mathbb{Q}[G]$-modules. Since $V \otimes_{\mathbb{Q}} \mathbb{R}$ admits a $G$-invariant complex structure $J$, the multiplicity of the $V_i$ with totally real centraliser $End_{\mathbb{Q}[G]}(V_i)$ is even in $V$.
Kummer varieties
----------------
Let $\Lambda \subset V$ be a lattice, and $J$ a complex structure on $V \tensor_{\bbQ} \bbR$. We denote by $$T_{\Lambda,J}:= (V\otimes_{\mathbb{Q}} \mathbb{R},J)/ \Lambda \;$$ the associated complex torus.
\[prop:geometricconsequence1\] Let $V$ be a $\mathbb{Q}[G]$-module, which has complex multiplication by $F$. Let $\Lambda \subset V$ be a $G$-invariant lattice. Then the associated complex torus $T_{\Lambda,J_{F}}$ and its endomorphism ring $End(T_{\Lambda,J_{F}})$ may be defined over an algebraic number field $k$. In particular, the induced linear action of $G$ on $T_{\Lambda,J_{F}}$ is defined over $k$.
As in the proof of Proposition \[prop:specialJ\], write $V$ as a sum of 1-dimensional $F$-vector subspaces $F_{j}$. By construction, $F_{j} \tensor_{ \bbQ }\bbR$ is a $J_{F}$-subspace of $V \tensor_{ \bbQ} \bbR$. Let $\Lambda_j = \Lambda \cap F_{j}$. Then $\tilde{\Lambda}= \bigoplus_{j=1}^m \Lambda_j$ is a lattice in $V$, which is contained in $\Lambda$. Define $X_{j} = T_{\Lambda_{j},J_{F}}$. Then $X_{j}$ is an abelian variety with complex multiplication by $F$, and $$X_{\tilde{\Lambda}} = \bigoplus_{j =1}^m X_{j}$$ decomposes as a direct product of abelian varieties, which are defined over an algebraic number field. Therefore, $X_{\tilde{\Lambda}}$ and its endomorphism ring $End(X)_{\tilde{\Lambda}}$ may be defined over an algebraic number field $k_{0}$.
Consider the covering $X_{\tilde{\Lambda}} \to X_{\Lambda}$. Since the elements of the finite kernel are algebraic over $\bbQ^{a}$, $X_{\Lambda}$ and the covering map may be defined over an algebraic number field $k$ (using the quotient lemma), $k$ containing $k_{0}$. Since $G \subset End_{\mathbb{Q}}(X_{\tilde{\Lambda}})$, there exists an $\ell \in \mathbb{N}$, such that $\ell g \in End(X_{\tilde{\Lambda}})$, and, therefore, the linear action of $\ell g$ on $X_{\tilde{\Lambda}}$ is defined over $k_{0}$. Hence we get a diagram $$\begin{aligned}
\xymatrix@1{
X_{\tilde{\Lambda}} \ar[r]^{\ell g} \ar[d] & X_{\tilde{\Lambda}} \ar[d] \\
X_{\Lambda} \ar[r]^{\ell g} & X_{\Lambda}
} \; ,\end{aligned}$$ where the upper and the downward maps are $k$-defined. By the universal property of the quotient, $\ell g \in End_k(X_{\Lambda})$. By the Galois-criterion for rationality, it follows that also $g \in End_k(X_{\Lambda})$.
Thus, the following is a consequence of Proposition \[prop:CMdecomposition\].
\[prop:geometricconsequence2\] Let $V$ be a $\mathbb{Q}[G]$-module, such that $V \tensor_{\bbQ} \bbR$ admits a $G$-invariant complex structure. Let $\Lambda \leq V$ be a $G$-invariant lattice. Then there exists a complex structure $J \in {\mathcal D}(G)_{p}$ such that the complex torus $T_{\Lambda,J}$ and its endomorphism ring $End(T_{\Lambda,J})$ may be defined over an algebraic number field $k$.
Note, as follows from the quotient lemma, that the geometric quotient $$K = T_{\Lambda,J} /G$$ attains the structure of a $k$-defined projective variety. It is called a Kummer-variety for $G$.\
Combining the last result with the proof of Proposition \[prop:constructX\] we arrive at:
Let $\Gamma$ be a virtually abelian Kähler group. Then there exists a non-singular complex projective variety $X$ defined over a number field $k$ with $\pi_1(X)=\Gamma$.
Let $\Gamma$ be such a Kähler group, and $\Lambda \cong \bbZ^k$ a normal subgroup with $G= \Gamma / \Lambda$. The characteristic representation $\mu$ of $G$ turns $V = \Lambda \tensor_{\bbZ} \bbQ$ into a $\bbQ[G]$-module. Since $\Gamma$ is Kähler, $V \tensor_{\bbQ } \bbR$ has a $G$-invariant complex structure. We remark that $\Gamma$ can be mapped homomorphically to $V \ltimes_{\mu} G$ (analogously as in and ).
As shown in Proposition \[prop:constructX\], there exists a complex manifold $X$ with $\pi_1(X)= \Gamma$, such that $$\begin{aligned}
\xymatrix@1{
\tilde{X}= \mathbb{C}^k \times \tilde{Y} \ar[d]
\\
T \times \tilde{Y} \ar[d] \\
X = (T \times \tilde{Y}) / G
} \; \end{aligned}$$ are holomorphic coverings, where $T = \mathbb{C}^k / \Lambda$ is a complex torus. Note that $G$ acts on $T$ by affine transformations. The linear part of the action is given by the representation $\mu: G {\rightarrow}End(T)$. Using Proposition \[prop:geometricconsequence2\], we may assume in the construction that $T = T_{\Lambda,J}$ is an abelian variety defined over an algebraic number field $k_{0}$, such that the linear parts of $G$ act by elements in $\Aut_{k_{0}}(T)$. Moreover, as remarked above, since $\Gamma$ splits over $\bbQ$, the translation parts of the $G$-action may be assumed to be contained in the torsion group of $T_{\Lambda,J}$. Therefore, there exists an algebraic number field $k$, such that $G$ act by elements in $\Aut_{k}(T)$.
Note now that $\tilde{Y}$ is also a complex projective variety defined over $\bbQ$, with $G \subset \Aut_{\bbQ}(\tilde{Y})$, see the proofs given in [@serre] or [@Shafa]. By Lemma \[lemma:algebraicquotient\], $X$ is a projective variety, which may be defined over the number field $k$.
[^1]: [email protected]
[^2]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we analyse Belief Propagation over a Gaussian model in a dynamic environment. Recently, this has been proposed as a method to average local measurement values by a distributed protocol (“Consensus Propagation”, Moallemi & Van Roy, 2006), where the average is available for read-out at every single node. In the case that the underlying network is constant but the values to be averaged fluctuate (“dynamic data”), convergence and accuracy are determined by the spectral properties of an associated Ruelle-Perron-Frobenius operator. For Gaussian models on Erdős-Rényi graphs, numerical computation points to a spectral gap remaining in the large-size limit, implying exceptionally good scalability. In a model where the underlying network also fluctuates (“dynamic network”), averaging is more effective than in the dynamic data case. Altogether, this implies very good performance of these methods in very large systems, and opens a new field of statistical physics of large (and dynamic) information systems.'
author:
- Erik Aurell
- René Pfitzner
title: Gaussian Belief with dynamic data and in dynamic network
---
Message-passing algorithms have over the last two decades turned out to be an important paradigm in fields as distant as iterative decoding, image processing and AI, see [@Pearl1988] for the intuition behind Belief Propagation (BP) in AI, and [@Cow1998; @Kschischang2001; @YedFreWei2003] for more recent reviews. It has been realized that systems where the message-passing algorithms are effective can often be assimilated to disordered systems in statistical physics, and that the message-passing algorithms themselves are closely related to the Bethe approximation [@MezardMontanari2009]. Most applications pursued concern [*inference*]{} in [*static models*]{}; how to do this effectively (if approximately), and when these methods work. In another direction, Consensus Propagation (CP) has been proposed as a message-passing scheme to average measurement values in a network of connected nodes [@MoaVan2006]. This is a naturally [*dynamic*]{} setting, where, in large networks, and in many scenarios of interest, one must allow the measurement values, and perchance the network itself, to change on the same time scale as the averaging process. The two strands of inquiry are connected by the fact that CP is equivalent to BP on a class of Gauss-Markov random fields [@MoaVan2006; @MJW2006].\
\
In this contribution we study CP both in a static network with changing measurement values (dynamic data), and in a network where the strengths of the interconnections also change (dynamic network). We will show that the method has very good scalability, *i.e.* that its performance degrades very slowly as the systems grow. In a sense, to be made precise below, performance does not degrade with size at all. This should make CP a very interesting method for aggregation tasks in large and dynamic networks, possibly competitive to alternative schemes such as gossiping [@JeMoBa2005]. From a physics perspective the salient points are the following: (i) CP with dynamic data is (after a transient) a linear averaging process; (ii) the kernel of this averaging process, being the linearization of Gaussian BP, is related to the second variation of the Bethe free energy of the Gauss-Markov random field; (iii) the leading eigenvalue of the kernel is a self-averaging quantity in Erdős-Rényi networks, which in addition does not depend on the network size; (iv) CP with dynamic network and dynamic data functions as well (or better) as CB with dynamic data only. Points (ii) and (iii) imply that we identify a new random matrix construction with unexpected properties, and possibly important practical consequences. Point (iv) means concretely that dynamic data is the slow stable (and also flat) manifold of the kernel, while dynamic network spans the fast stable manifold. Perturbations in the dynamic network directions hence relax faster than dynamic data, which explains the good properties.\
\
[*Belief Propagation (BP) and Consensus Propagation (CP):*]{} BP is an algorithm to infer marginal probability distributions of a joint probability functions [@YedFreWei2003]. It works via distributed message passing from each node of the underlying graph of the model to every neighboring node (FIG. \[fig:network2\]). It is correct on tree-like graph topologies and has been shown to often yield good results in topologies including loops [@MezardMontanari2009; @MurWeiJo]. The messages in BP can be seen as 1-node marginal conditional probability distributions, which implies that BP works best computationally when the size of local state space is limited, e.g. for Ising spins. BP on Gauss-Markov random field is a special case, since Gaussianity is preserved under convolution, and the BP messages can be parametrized by two real values corresponding to (conditional) mean and (conditional) variance. A further very special property of BP on Gaussian models is that it is [*exact*]{} for the modes of the marginal distributions, in a very wide class of models [@WeFree2001; @MJW2006].\
\
A special instance of Gaussian Belief Propagation is *Consensus Propagation* [@MoaVan2006]. This algorithm aims to solve the problem of calculating the average $\overline{y}$ of some values $y_i$ (gathered by nodes $i$ in a network $G$) in a distributed way. The Gaussian model associated to CP is [@MoaVan2006]: $$\label{eqn:CP}
p(x,\beta)=\frac{1}{Z}\exp(-\|x-y\|^2-\beta \sum_{{i,j} \in E}Q_{ij}(x_i-x_j)^2)$$ In , $Z$ is a normalization, $\beta$ is a global and $Q_{ij}$ are local coupling parameters. BP on is guaranteed to converge for any finite $\beta$, and the modes of any one-node marginals computed by BP converge to the average $\overline{y}$ as $\beta$ tends to infinity (as follows from [@MJW2006]). In this way estimates of $\overline{y}$ can be obtained, where a trade-off must be made between convergence time and accuracy.
![Illustrated Belief Propagation message passing scheme in a 6-node network. In the Consensus Propagation case the messages $m_{ij}$ are decomposed into messages $K_{ij}$ and $\mu_{ij}$.[]{data-label="fig:network2"}](network2.eps){width="30.00000%"}
#### The algorithm – {#the-algorithm .unnumbered}
The following message update rules define Consensus Propagation: $$\begin{aligned}
K_{ij}^{(t+1)} &=\frac{1+\sum_{k \in N(i)\backslash j}K_{ki}^{(t)}}{1+\frac{1}{\beta Q_{ij}}(1+\sum_{k \in N(i)\backslash j}K_{ki}^{(t)})} \label{eqn:K-message} \\
\mu_{ij}^{(t+1)} &=\frac{y_i+\sum_{k \in N(i)\backslash j}K_{ki}^{(t)}\mu_{ki}^{(t)}}{1+\sum_{k \in N(i)\backslash j}K_{ki}^{(t)}} \label{eqn:mu-message}\end{aligned}$$ This parametrization of BP yields two-dimensional real-valued messages consisting of a *topology message* $K$ and a *local state update* $\mu$. The notation $X_{ij}^{(t)}$ means that message $X$ is sent from originating node $i$ to target node $j$ at iteration step $t$: $N(i)$ is the set of all neighbors of node $i$, and $N(i)\backslash j$ is the set of all neighbours of $i$ except $j$. The algorithm is said to have attained *consensus*, if the messages are fixed points of and . A belief for the average $\overline{y}$ at time $t$ and node $i$ is obtained via the CP output rule: $$\label{eqn:y_average}
\overline y_i = \frac{y_i+\sum_{k \in N(i)}K_{ki}^{(t)}\mu_{ki}^{(t)}}{1+\sum_{k \in N(i)}K_{ki}^{(t)}}$$ The consensus beliefs , with $K$ and $\mu$ at a fixed point of and , are the Belief Propagation predictions of the modes of the one-node marginals of the probability distribution .
#### Convergence for different initializations – {#convergence-for-different-initializations .unnumbered}
![Convergence of the $\overline{y}$ belief at one node in a random network with 500 nodes. The solid line is the CP performance, the dashed line indicates the correct y-average. Insert shows behaviour in 1-10 rounds; main figure shows behaviour in $10^3-10^4$ rounds. Node values were generated randomly, and then scaled by $90\%$ in round $5*10^3$. At iteration step $n=0$ all CP-messages were initialized uniformly to zero. Fast convergence followed by an overshoot (damped oscillation) is observed. After the reset at $n=5*10^3$ the CP-messages were left unchanged, and a markedly slower convergence, but without an overshoot, is observed.[]{data-label="fig:CP_500_1"}](CP_500_1.eps){width="50.00000%"}
Figure \[fig:CP\_500\_1\] shows performance of Consensus Propagation after initializing all messages to zero: the algorithm shows first an oscillating behaviour with *fast* convergence to a good approximation of the correct mean $\overline{y}$. After changing every node value and NOT re-initializing the messages, the algorithm exhibits a steady, yet much *slower*, convergence. This second behaviour corresponds to the case of *dynamic data*, where the topology message ($K$-message) and local state update ($\mu$-message) start at their converged values before the perturbation. Once a fixed point $K^*$ is reached, the topology messages will not change if only the local measurement values $y_i$ change, since is an equation only involving topology messages. Except for an initial transient, the dynamic data case can hence be completely understood by the linear operator expressed by the right-hand side of (see below). Generally, it seems that the topology messages converge much faster than the local state messages, and that therefore the linear theory (explained below) also bounds the behaviour of dynamic data, where both local values $y_i$ and local couplings $Q_{ij}$ change. Before we turn to the linear analysis, let us however point out the observation that different initializations of the messages yield different performance, and, perhaps surprisingly, that initializing with $K=\mu=0$ seems to be the superior choice.
![Convergence behaviour of Consensus Propagation on a random graph for different initial messages. As model we used a random graph with 80 nodes, all edges present with probability 0.1, $\beta=100$ and $Q_{ij}$ chosen i.i.d. random variables uniform between $0.5$ and $2$. The plot shows the time evolution of the deviation of two messages from their converged values ($K^\ast-K$ and $\mu^\ast -\mu$), sent from node 15 to node 10 during 500 iterations. The messages were initialized proportional to their fixed point values, and start at the top right corner of each trajectories in the figure. The trajectories exhibit an initial fast decay of the error in topology messages (abscissa) followed by a slower decay of the local state message (ordinate). The only exception is when the messages are initialized as $K=\mu=0$ in which case the trajectory seems to fall into the (more) stable manifold of the fixed point (a “direct hit”), with the second slow process along the ordinate absent. For a graphical illustration of the conjectured behaviour, see \[fig:convergence\_scheme\]. The fixed point in this example have $K_{15,10}^\ast=62.61$ and $\mu_{15,10}^\ast=3.95$.[]{data-label="fig:initialization"}](initializations.eps){width="50.00000%"}
The observations of Fig. \[fig:initialization\] contradict a conjecture put forward in [@MoaVan2006] that convergence times for $K^{(0)}=0$ and $K^{(0)}=K^\ast$ are equivalent. In fact, initializing with $K=0$ improves convergence dramatically. Let us note that if $K$ be re-initialized to zero, then the re-initialization of $\mu$ is arbitrary, since by (\[eqn:mu-message\]), $\mu^{(1)}$ will then be equal to $y_i$, *i.e* independent of $\mu^{(0)}$. In a scenario where many measurement values (and/or also the underlying network) change simultaneously, re-starting Consensus Propagation using $K=0$ may therefore by a valid option. We stress that this is not obvious, but follows if the dynamic behaviour is as illustrated in Fig. \[fig:convergence\_scheme\]. This may not be true in all underlying topologies. However, in the case that the underlying topology is locally tree-like, as is the case for the random graphs in Fig. \[fig:initialization\], a heuristic explanation for the faster convergence of Consensus Propagation, when initializing with $K=0$, is the following: as was shown by [@MoaVan2006] Consensus Propagation yields the exact node average on tree-like graphs with the global coupling constant $\beta=\infty$ and $K^{(0)}=0$. Initializing CP on a random graph with $K=0$ and a large value of $\beta$ will yield nearly the same messages, after a finite number of iterations, as initializing CP with $K=0$ on a computational tree associated with the graph (using the construction of [@MJW2006]) at $\beta=\infty$. This explains the improved behaviour of starting with $K=0$ *qualitatively*, but does not explain it *quantitatively*, *i.e.* the apparent complete absence of the slow process in Fig. \[fig:initialization\].
#### Theory of Consensus Propagation – {#theory-of-consensus-propagation .unnumbered}
Consensus Propagation can be considered as non-linear dynamical system in a multidimensional space spanned by all $K$- and $\mu$- messages: $$\begin{pmatrix}
\mu_{ij}^{(n+1)}\\
...\\
K_{ij}^{(n+1)}\\
...
\end{pmatrix}
=\mathbb{R}
\begin{pmatrix}
\mu_{ij}^{(n)}\\
...\\
K_{ij}^{(n)}\\
...
\end{pmatrix}$$
![Convergence scheme for Consensus Propagation. The $K$-subspace is a fast, the $\mu$-subspace a slow stable manifold.[]{data-label="fig:convergence_scheme"}](convergence_scheme.eps){width="30.00000%"}
The numerical experiments above indicate that the $\mu$-message subspace spans a slow stable and the $K$-message subspace a fast stable manifold (see Fig. \[fig:convergence\_scheme\]). We will use the eigenvalues of a linearized version of $\mathbb{R}$ to verify this. Following [@MJW2006], we refer to the non-linear iterated map transfer operator $\mathbb{R}$ as a *Ruelle-Peron-Frobenius Operator*. The linear part of $\mathbb{R}$ has the matrix representation: $$\begin{aligned}
\begin{pmatrix}
\mu_{ij}^{(n+1)}\\
...\\
K_{ij}^{(n+1)}\\
...
\end{pmatrix}&=
\begin{pmatrix}
\mu_{ij}^\ast + \Delta \mu_{ij}^{(n+1)}\\
...\\
K_{ij}^\ast + \Delta K_{ij}^{(n+1)}\\
...
\end{pmatrix}\\
&=
\begin{pmatrix}
\mu_{ij}^\ast\\
...\\
K_{ij}^\ast\\
...
\end{pmatrix}+
\mathbb{R}'
\begin{pmatrix}
\Delta \mu_{ij}^{(n)}\\
...\\
\Delta K_{ij}^{(n)}\\
...
\end{pmatrix}\end{aligned}$$ $\mathbb{R}'$ is the linearized transfer operator. The matrix representation of this operator can be decomposed into four quadratic submatrices: $$R'=
\begin{pmatrix}
A&C\\
0&B
\end{pmatrix}
\label{eqn:Rprime-def}$$
![Length of projections of (normlength) eigenvectors on the $\mu$-subspace for each eigenvalue $\lambda$ of the linearized Ruelle-Peron-Frobenius-Operator $\mathbb{R}'$ for a $G(N=20,c=8)$ Erdős-Rényi model. $\beta=100$, and $Q_{ij}$ randomly generated as in Fig \[fig:initialization\].[]{data-label="fig:proj2"}](proj2.eps){width="50.00000%"}
Submatrix $A$ is the transfer matrix in the *dynamic data* case, when the topology messages have converged, submatrix $B$ is the linear part of the transfer matrix acting in the *dynamic network* on the topology messages alone, and submatrix $C$ is the linear action of the topology messages on the local state variables. Around the fixed point, we can verify that topology messages converge faster than local state updates, by comparing the size of the eigenvalues of $\mathbb{R}'$ to the projection of the corresponding eigenvectors on the subspace spanned by the $\mu$-messages. As shown in Fig. \[fig:proj2\], the (isolated) largest eigenvalue lies in the subspace of local state updates. In addition, most of the other eigenvectors in the subspace of local updates also have eigenvalues larger than all the eigenvalues projecting on the topology messages. Table \[tab:lambda\_A\_B\] compares the leading eigenvalues of submatrices $A$ and $B$ for four different Erdős-Rényi graphs, reinforcing the observation from Fig. \[fig:proj2\].
N c $\lambda_{max}(A)$ $\lambda_{max}(B)$
---- ---- -------------------- --------------------
20 18 0.99949152 0.00054415
30 14 0.99924356 0.00083034
40 10 0.99895415 0.00119833
50 8 0.99851962 0.00186674
: Comparison of leading eigenvalues of linearized matrices $A$ and $B$ in four Erdős-Rényi graphs $G(N,p=c/N)$. The much smaller eigenvalues of matrix $B$ imply much faster convergence of the topology message s.[]{data-label="tab:lambda_A_B"}
In linear theory, the limiting factor on convergence is therefore the dynamics of the local state updates.
#### Dynamic data case – {#dynamic-data-case .unnumbered}
The case when topology messages have converged is also of interest when data to be measured keep on changing: in this scenario CP is a *linear averaging process*. Indeed, the local state update equation (\[eqn:mu-message\]) is then a linear equation of one free vector $\mathbb{\mu}=(...\mu_{ij}..)$ and can be expressed in linear operator form: $$\boldsymbol{\mu}^{(n+1)}=\boldsymbol{b}+A\boldsymbol{\mu}^{(n)}
\label{eqn:linearized-averaging}$$ The operator $A$ in (acting on $\mu$-messages) is the same as the submatrix $A$ of the operator $\mathbb{R}'$ in , and its spectral properties are as described in Fig. \[fig:proj2\] (rightmost set of eigenvalues, all completely in the subspace of local state updates). The vector $\boldsymbol{b}$ is the $\mu$-independent part of , and in particular depends on the set of local measurement values $\boldsymbol{y}$. If these change in time, is obviously a linear averaging process with kernel $A$. If the $\boldsymbol{y}$ do not change, and the $\mu$-messages are initialized in some manner, we expect from Fig. \[fig:proj2\] that convergence will eventually be dominated by the largest (isolated) eigenvalue of $A$. Fig. \[fig:q\_lambda\_plot\] shows that this is indeed the case, for several different Erdős-Rényi graphs. In these models, we always find an isolated largest eigenvalue (data not shown).
![Convergence ratios of the linear averaging process compared to leading eigenvalue of operator $A$. and numerical calculated leading eigenvalues $\lambda_{max}$ in four examples of Erdős-Rényi models $G(N,p=c/N)$. $q$: convergence ration, $\lambda_{max}$: largest eigenvalue. The solid line represents $q=\lambda_{max}$. Cases: 1: $c=8$, $N=50$, 2: $c=10$, $N=40$, 3: $c=14$, $N=30$, 4: $c=18$, $N=20$. $\beta=100$, and $Q_{ij}$ randomly generated as in Fig \[fig:initialization\] for all cases.[]{data-label="fig:q_lambda_plot"}](q_lambda_plot.eps){width="50.00000%"}
#### Scalability of CP in Erdős-Rényi graphs – {#scalability-of-cp-in-erdős-rényi-graphs .unnumbered}
The above discussion leads up to the conclusion that the largest eigenvalue of operator $A$ of is a quantity of major importance to understand the performance of CP in dynamic environments – both dynamic data only, and also dynamic network. The scaling properties of this largest eigenvalue therefore determines how effective the CP averaging procedure can be in a large network. Following the general principles of random graph theory, we should compare random graphs of increasing size $N$, but with the same average node degree $c$. This means that every link is present in the graph with probability $p=\frac{c}{N}$ (up to corrections decaying with $N$). Table \[tab:E-R-test1\] shows that in a family of Erdős-Rényi graphs with asymptotic average node degree $c=8$ the largest of eigenvalue $A$ seems to converge to a finite value less than one. In the experiments, the local couplings $Q_{ij}$ are generated randomly between $0.5$ and $2$. The fifth column gives (for the smaller instances) the standard deviation of the largest eigenvalue computed from $100$ experiments (independent realizations of the random graphs, and independent realizations of the local coupling constants $Q_{ij}$). The decay of the standard deviation with $N$ indicates that the leading eigenvalue is a self-averaging quantity in this ensemble.
N p $c_{exp}$ $[c_{exp}]_{100}$ $\sigma [c_{exp}]_{100}$ $\lambda_{max}$ $[\lambda_{max}]_{100}$
------- -------- ----------- ------------------- -------------------------- ----------------- -------------------------
20 0.4 7.3 7.6 0.6 0.9984 0.9985
40 0.2 7.7 7.8 0.5 0.9985 0.9985
80 0.1 7.8 7.9 0.4 0.9986 0.9985
160 0.05 8.2 8.0 0.3 0.9987 0.9985
5000 0.0016 0.99850
10000 0.0008 0.99851
20000 0.0004 0.99850
: Convergence ratios for Erdős-Rényi graphs. Coupling constant $\beta=100$ and $Q_{ij}$ were chosen randomly uniform between $0.5$ and $2$. All instances have a theoretical average node degree $c=8$. The table shows the outcome of a single experiment ($c_{exp}$, $\lambda_{max}$) and for small graphs of 100 experiments ($[c_{exp}]_{100}$, $\sigma [c_{exp}]_{100}$, $[\lambda_{max}]_{100}$).[]{data-label="tab:E-R-test1"}
Let us note that our results concur with (and extend) a result of [@MoaVan2006] for regular graphs, where the authors showed that convergence time is not dependent on the graph size. If this be true, the leading eigenvalue in that ensemble must also be a self-averaging quantity, independent of graph size.\
Fig. \[fig:E-R-test2\] shows the dependence of the leading eigenvalue on the node degree $c$, for a number of graphs with $20$ nodes. The eigenvalue shows an increasing trend, in this range fairly well approximated by a logarithmic behaviour.
![Dependence of convergence ratio $q$ on average node degree $c$ in an Erdős-Rényi graph with 20 nodes. The solid line is a data fit: q= $0.001046 \cdot \log (c)+0.9965$[]{data-label="fig:E-R-test2"}](q_c_plot.eps){width="50.00000%"}
#### Summary – {#summary .unnumbered}
Statistical physics has contributed very significantly in recent years to the understanding of Belief Propagation approaches to *inference*, which have very important applications to *e.g.* iterative decoding [@MezardMontanari2009]. In this contribution, we have looked at a Belief Propagation-based algorithm for *averaging*, with potentially numerous applications to network management. We showed that this Consensus Propagation algorithm, in a dynamic environment, is a dynamical system which can be fruitfully analysed by the tools of statistical physics and nonlinear dynamics. We showed that CP responds quickly to changes in the network topology, and more slowly to changing data. This can be understood intuitively as a dynamic network improves mixability, which should not be a disadvantage when computing an average (or an estimate of an average). In a real world application, CP is therefore not limited by a changing network structure but by dynamic data. Secondly, and of interest to statistical physicists, we exhibited an interesting self-averaging property of the leading eigenvalue of the transfer matrix describing the the dynamic data case. Perhaps surprisingly, this leading eigenvalue seems to be asymptotically independent of network size.
#### Acknowledgement – {#acknowledgement .unnumbered}
R.P. was partially funded by the European “Life Long Learning Program” under project number DE-2008-ERA/MOB-KonsZuV01-CP07. E.A. acknowledges support from the Swedish Science Council through the KTH Linnaeus Centre ACCESS, and from the Academy of Finland.
[so]{}
Pearl, “Probabilistic Reasoning In Intelligent Systems”, Morgan Kaufmann, 1988
Cowell, “Advanced Inference in Bayesian Networks”, in “Learning in Graphical Models”, edited by Michael Jordan, 1998 Kluwer Academic Publishers
Kschischang, Frey and Loeliger, “Factor Graphs and the Sum-Product Algorithm” IEEE Transactions on Information Theory, vol. 47, pp. 498-519, February 2001
Yedidia, Freeman and Weiss, “Understanding Belief Propagation And Its Generalization” in “Exploring Artificial Intelligence in the New Millennium”, (Science & Technology Books, TR2001-022), January 2003
Mezard and Montanari, “Information, Physics, and Computation”, Oxford University Press, 2009
Moallemi and Van Roy, “Consensus Propagation”, IEEE Transactions on Information Theory, Vol. 52, No. 11, November 2006
Jelasity, Montresor, and Babaoglu “Gossip-based aggregation in large dynamic networks”, ACM Transactions on Computer Systems, 23(3):219-252, August 2005.
Weiss and Freeman “Correctness of Belief Propagation in Gaussian graphical models of arbitrary topology”, Neural Computation 13:2173-2200 (2001)
Malioutov, Johnson and Willsky, “Walk-Sums And Belief Propagation In Gaussian Graphical Models”, Journal of Machine Learning Research 7 (2006), pp. 2031-2064
Weiss and Freeman, “Correctness Of Belief Propagation In Gaussian Graphical Models Of Arbitrary Topology” Neural Computation, Vol. 13, No. 10, pp. 2173-2200, October 2001
Murphy, Weiss and Jordan, “Loopy Belief Propagation For Approximate Inference: An Empirical Study”, Proceedings of the 15th Annual Conference on Uncertainty in Artificial Intelligence (UAI-99), Morgan Freeman, pp. 467-47, 1999
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This paper proposes a solution to the problem of smooth path planning for mobile robots in dynamic and unknown environments. A novel concept of Time-Warped Grid is introduced to predict the pose of obstacles in the environment and avoid collisions. The algorithm is implemented using C/C++ and the CUDA programming environment, and combines stochastic estimation (Kalman filter), harmonic potential fields and a rubber band model, and it translates naturally into the parallel paradigm of GPU programming. In simple terms, time-warped grids are progressively wider orbits around the mobile robot. Those orbits represent the variable time intervals estimated by the robot to reach detected obstacles.
The proposed method was tested using several simulation scenarios for the Pioneer P3-DX robot, which demonstrated the robustness of the algorithm by finding the optimum path in terms of smoothness, distance, and collision-free, in both static or dynamic environments, and with large number of obstacles.
author:
- |
Siavash Farzan\
Institute for Robotics and Intelligent Machines\
Georgia Institute of Technology\
Atlanta, GA 30332\
`[email protected]` Guilherme N. DeSouza\
ViGIR Lab, EECS Department\
University of Missouri\
Columbia, MO 65211\
`[email protected]`\
bibliography:
- 'references.bib'
title: 'Path Planning in Dynamic Environments Using Time-Warped Grids and a Parallel Implementation'
---
Introduction
============
The problem of path planning of mobile robots consists of the trajectory planning of a mobile robot from a start state to a final state. The objective of such start-to-goal path planning is to find the best suitable path between two points while avoiding collisions with both static and moving obstacles. In this case, the suitability of the “best” path is determined by a function representing the occupancy of the space, while the selected robot trajectory corresponds to the optimal path on the surface of this occupancy function according to some criteria (e.g. smoothness, length, etc.).
In such problems, the trade off between performing quick, shallow searches for sub-optimal solutions and performing deep searches for the price of heavy computation has always been a practical and important issue. The application of parallel algorithms however, can minimize or even eliminate this issue by enabling deeper searches without losing accuracy or real-time applicability. In addition to saving time, parallel computation can lead to extra advantages [@Barney10] such as solving the same problems at larger scales, providing redundancy of control, overcoming memory constraints, and performing remote processing (e.g. tele-operation of robots). However, these advantages do not come with a low price tag. For effective use of parallel computing, the computational problem must be partitioned into discrete parts, or tasks, to be executed simultaneously. This partitioning into discrete parts is carried out by identifying independent and self-contained sub tasks where the overall result can be readily extracted from the sum of the individual results. As it will be explained in further details later, this can be a daunting operation.
In this paper, a solution to the problem of smooth path planning for mobile robots in dynamic and unknown environments is proposed. A novel concept of time-warped grid is introduced to predict the pose of obstacles in the environment while avoiding collisions with those obstacles. The concept combines stochastic estimation (Kalman filter), harmonic potential fields and a rubber band model, and it translates naturally into the parallel paradigm of GPU programming. The intuition behind the proposed method for dealing with dynamic environments is derived from the analogy of pedestrians “jaywalking” or crossing streets on which vehicles are driving. When jaywalking, pedestrians anticipate where the vehicles will be at future times and adjust their path accordingly. If a person were to consider only a snapshot of the street and assume a fixed position for each vehicle, then he/she would be in a significant risk of being run over if tried to cross the street. Similarly in this work, the robot predicts the future configuration of moving obstacles by using a combination of Kalman filter and time-warped grids. Consequently, a path is plotted for the future positions of the obstacles, rather than the current ones.
In simple terms, time-warped grids are progressively wider orbits around the mobile robot. Those orbits represent the variable time intervals estimated by the robot to reach detected obstacles. The main idea of time-warped grid is to acknowledge that the further the obstacle is, the more delayed is its impact on the path. The use of time-warped grid allows the system to address at the same time, the problems of convergence, speed, and moving obstacles in the calculation of a smooth path for the mobile robot without any prior knowledge regarding the environment. All assumptions made by the system derive from a laser sensor mounted on the robot and a localization system (e.g. vision-based landmark localization [@DeSKak02]) that provides distance to the goal, even when it lies outside the range of the robot sensors in large environments.
As the experiments performed demonstrate, the path obtained by the predictive aspect of our method is not only short, but it also contains no loops, no sharp turns, and no changes of speed of the robot, making it ideal for carrying of delicate materials or for wheelchair navigation. Our experiments also demonstrate the robustness of the method, which can always find an optimum path – i.e in terms of smoothness, distance, and collision-free – either in static or dynamic environments, even with a very large number of obstacles.
Related Work
============
Unlike *coverage path planning* where the emphasis is in sweeping out the space sensed by the robot – e.g. floor cleaning, lawn mowing, harvesting, etc. [@Cho01], the objective of *start-goal path planning* is to find suitable paths between two points while avoiding collision with static or moving obstacles. The ability to generate an optimal path from an initial point to a final destination in real-time is still one of the key challenges in Robotics. The area is becoming even more influential with the near deployment of self-driving autonomous vehicles [@Bast16].
Different techniques have been proposed in this domain, where the challenges become increasingly difficult with the size of the environment and the number of moving obstacles. From the early algorithms, the focus often turned into finding cost-minimal paths through the robot environment, and the use of maps became so attached to the problem that many confuse start-goal and map-based as synonymous. Indeed, most of the approaches in the literature rely on some sort of map or grid, and the large size of the environment and consequent number of grids rendered many of these methods to off-line use only. Moreover, the complexity and uncertainty of the path planning problem increase greatly in dynamic environments due to the change of the entire information in the environment along with the movement of obstacles.
An early and important algorithm to address the path planning problem in dynamic environments is known as diffusion process over grid-based maps, first presented by Steels [@Steels88], and further developed by Schmidt et al. [@SchAza92]. In diffusion method, the environment is discretized into a grid of cells, where the cell representing the goal applies an attracting force in succession to the neighbouring free cells all the way to the cell occupied by the robot. These attractive forces are modeled by fluid diffusion equations, hence the name of the method. The robot proceeds towards the cell with the highest level of force computed by a gradient-based iterative algorithm. The diffusion method can be implemented as a two layer cellular neural network [@Sie94], and use of an unsteady diffusion equation model [@SchAza92] makes it applicable to time-varying environments. More recently, Vázquez-Otero et al. [@VazMun12] used reaction-diffusion dynamic models based on biological processes with advantages of smoother trajectories and tolerance to noisy data.
Diffusion-based methods allow on-line path planning and result in short and collision-free paths, without suffering from problems with local minima. However, the algorithm is global and require the prior knowledge of the robot’s environment, thus cannot deal with more advanced navigation tasks such as exploration of unknown environment or multi-robot navigation. Moreover, the diffusion equation parameters are difficult to tune, and the implementation is highly computationally demanding, making the algorithm less favorable for real-time or large environment applications.
Borenstein and Koren [@BorKor91] presented an efficient algorithm for path planning called the vector field histogram (VFH) method, further improved by Ulrich and Borenstein [@UlrBor98] [@UlrBor00] by considering the robot size and choosing a safe and efficient steering direction. This approach uses a grid-based occupancy mapping, where occupancy information is described by a histogram representation of the robot’s environment. The free space and obstacles are localized by their angle as well as their distance relative to the robot. The final steering direction for the robot is selected (based on specified thresholds and proximity to obstacles) from the candidates closest to the goal direction. Due to the statistical nature of the algorithm, the VFH method is very robust to uncertainties in sensor readings and dynamic models, making it ideal for unknown and time-varying environments. While the method is computationally efficient, it does not result in globally optimal paths (since it is a local path planner) and is also prone to dead-end situations.
The dynamic window approach incorporating the motion dynamics of the robot was proposed in [@FoxThr97], and further generalized in [@BroKha99]. In this method, the control of the robot is carried out directly in the space of translational and rotational velocities, and the search space for admissible velocities is reduced in three steps over three windows given by: i) approximating trajectories by finite sequences of circular arcs, ii) considering only the next steering command, and iii) restricting the space to the velocities that can be reached within a short time interval. The final path for the robot is generated by maximizing an objective function formed by the intersection of the three windows. Using a dynamic window approach, the acceleration capabilities of the robot are considered, and obstacle avoidance can be performed at high velocities. However, the algorithm is still prone to local minima problems, since it only considers the goal heading without integrating the information about the free space.
Potential field methods, as proposed by [@Khatib86], determine the heading direction of the robot by representing the robot environment as a potential field, where the goal applies an attractive force and obstacles assert repulsive forces to the field. The robot trajectory is calculated by superposition of the two fields and following the low potential along the field according to a fastest descent optimization. Methods based on potential fields are computationally efficient and suitable for path planning with real-time conditions. However, they often fail to find trajectories in congested environments, and can also result in oscillatory motions in narrow passages [@KorBor91]. Another problem in start-goal planning is regarding the convergence to a solution. In potential fields [@HwaAhu92] and the A[\*]{} algorithm [@HarRap68], for example, a guarantee that the system will find a solution either can not be provided at all (potential fields) [@Latombe12], or it can be provided only if the heuristic is guaranteed to always be optimistic (A[\*]{}) – i.e. the true cost of a path is at least as large as the estimated cost. For the first case, Kim and Khosla [@KimKho92] proposed instead a harmonic potential function that eliminates the possibility of local minima in the potential fields, which prevented it from finding a path to the goal. In the latter case of A[\*]{} algorithm, the alternative is to make the heuristic more optimistic, which increases the method’s computational complexity, making it less likely to run in real-time.
As mentioned above, some of these methods require prior knowledge about the environment in a static setting. Violation of this requirement, i.e. existence of moving obstacles, leads to sudden changes and oscillations in the robot path, which can be aggravated by the sensitivity and inaccuracy of the robot sensors. While these consequences may be acceptable for a mobile robot, applications involving autonomous wheelchair navigation can become quite uncomfortable for human passengers. In [@HonDes10], it was proposed a robust method using a rubber band model to smoothen the path and reduce the number of sharp angles obtained from the use of harmonic potential fields alone. While that approach worked well for static environments, it did not address the case of moving obstacles.
In order to address the computational complexity of these methods, many researchers have recently developed parallel implementation of path planning algorithms on Graphics Processing Units (GPUs). In [@Ble08], the authors proposed a method for globally optimum path planning using a combination of the A[\*]{} and Dijkstra’s algorithms [@Dijkstra59]. The two algorithms were modified to take advantage of data parallelism of GPUs, which led to an implementation of edge lists using adjacency tables to reach a remarkable speed-up when compared to traditional C++ implementations.
Also exploiting the nature of these algorithms and the parallel paradigm of GPU computation, Kider et al. [@KidSaf10] proposed a randomized version of A[\*]{}, called R[\*]{}GPU search. Their main contributions are certainly the smaller memory requirements when compared to the original A[\*]{}, the avoidance of local minima by the use of randomly selected subgoals, and the scalability of the method to high-dimensional planning problems.
However, even when computational complexity is not an issue, the major drawback of any A[\*]{} based method remains in the difficulty in coping with dynamic environments. That is, A[\*]{} algorithm relies on the *optimism* of the heuristics. Since those heuristics derive from the values of the map cells, potential changes over time in these values or changes in topology of the graphs due to moving obstacles lead to inversions of those heuristics, and hence to loops and/or failure in converging to the goal. Another problem of the A[\*]{} algorithm can be seen in the proposed approach in [@Russell16]. The main drawback in this approach is that the A[\*]{} uses uniform grid representation which requires large amount of memory for regions that may never be traversed or may not contain any obstacles, affecting the efficiency of the method. This drawback can also happen in the dynamic version of the A[\*]{} algorithm called D[\*]{} [@Ste94], even though it indeed generates optimal trajectories in unknown environments.
In this paper, we address at the same time the problems of convergence, speed, and highly dynamic environments in the calculation of a smooth path for a mobile robot without any prior knowledge regarding the environment (static or dynamic). All assumptions made by the system derive from a laser sensor mounted on the robot and a localization system (here assumed to be a vision-based, landmark-based localization [@DeSKak02]) which needs to provide only location of the goal, even when it lies outside the range of the robot sensors in large environments.
Proposed Method
===============
The proposed method for the path planning problem tackles the limitations of other systems – namely the problems of convergence, fast processing, moving obstacles and sharp paths – by combining a few concepts. First, we rely on harmonic potentials to guarantee the calculation of a path if one exists[^1]. As it was pointed out earlier, the path produced by harmonic potentials can present sharp turns, which for many applications, such as autonomous wheelchair navigation, can produce an uncomfortable experience for the passenger. So, we re-introduce the idea of a rubber band model [@HonDes10] to smoothen the path created by the harmonic potentials. Since moving obstacles can also lead to unexpected changes in path, we propose the use of Kalman filter for stochastic estimation of the positions of the obstacles. While Kalman filter in itself has been widely used in the past, our main contribution here is in the combination of Kalman filter and a novel idea of time-warped grid, which will be explained later.
As the experiments performed will demonstrate, the path obtained by the predictive aspect of our method is not only the shortest possible path, but it also contains no loops[^2], no sharp turns, and no changes of speed of the robot, making it ideal for carrying of delicate materials or for wheelchair navigation. Since our method is also relatively computationally intensive, we resort to an efficient parallel implementation using GPUs over the time-warped grids.
Harmonic Potential Fields \[subsec:Potential-Rubber\]
-----------------------------------------------------
In grid-based maps, the idea is to represent the environment as a 2D grid. Such grid is basically the projection of all objects in the environment – in our case detected using a laser range sensor and a vision system – onto the ground plane. When potential fields are applied on top of such grids, obstacles are described by high potentials or *hills* that must be avoided, and start and goal points are the highest and smallest potentials, respectively. The path towards the destination is defined along the valleys of the potential field. Unfortunately, due to interaction between multiple objects, valleys are not unique in their potential values. Moreover, besides the possibility of multiple valleys, potential fields can also present local minima. These two conditions can cause the robot to fail to find a path to the goal. However, these same problems disappear when we use harmonic potential fields instead [@ConGru93; @Robert_08_harmonic].
Harmonic functions satisfy the min-max principle, and hence, spontaneous creation of local minima within the space is impossible. This principle is satisfied when the Laplaces equation constraint on the functions is true. In other words, a harmonic function $\phi$ on a domain $\Omega\subset R^{n}$ is a function that satisfies: $$\nabla^{2}\phi=\sum_{i=1}^{n}\frac{\partial^{2}\phi}{\partial x_{i}^{2}}=0\textrm{.}$$
This same function can be discretized and the numerical solution of Laplace’s equation becomes [@ConGru93]: $$\begin{array}{ccc}
\phi^{(k+1)}(x,y) & =&
\dfrac{1}{4}[(\phi^{(k)}(x+1,y)+\phi^{(k)}(x-1,y)\quad+ \\ & & \phi^{(k)}(x,y+1)+\phi^{(k)}(x,y-1)]\label{eq:discretelaplace}
\end{array}$$ where $\phi^{(k)}(x,\,y)$ represents the discrete sample of $\phi$ at coordinates $( x,\,y )$ of the $\mathbb{R}^{2}$ grid, and $k$ denotes the iteration number. At each iteration, a grid cell of $\phi$ is updated with the average value of its neighbors. On a sequential computer, this solution is usually implemented as follows: $$\begin{array}{ccc}
\phi^{(k+1)}(x,y) & = & \dfrac{1}{4}[\phi^{(k)}(x+1,y)+\phi^{(k+1)}(x-1,y)\quad+\\
& & \phi^{(k)}(x,y+1)+\phi^{(k+1)}(x,y-1)] \textrm{ .} \label{eq:harmonic-potential}
\end{array}$$
That is, the next values of the top and left neighbors of the current cell are updated and used in the calculation of that same cell. This speed up of the algorithm allows for the values of the next iteration to quickly propagate through the grid. However, it also slightly distorts the real value of the harmonic potentials [@HonDes10].
In order to explain the proposed method, a few basic elements need to be revisited. First, a *goal* is a grid cell with the lowest harmonic value ($\phi^{(k)}(x,\,y)=0$). This value is fixed and it will never be affected by its neighboring values. An *obstacle* is any cluster of cells blocking a potential path towards the goal. Its value is maximum ($\phi^{(k)}(x,\,y)=1$) and is also never affected by its neighbors. However, its position may change – e.g. in dynamic environments. *Free space* is any grid cell that does not contain an obstacle or the goal. The value of the harmonic potential in the free space is initialized with $0.5$ and is updated at each iteration.
A path to the goal is given by an index matrix, $M_{idx}(x,\,y)$, which for every coordinate $( x,\,y )$ contains the index of the neighbor with the lowest harmonic potential. That is, $$M_{idx}(x,y)=min[\phi^{(k)}(x+1,y),\phi^{(k+1)}(x-1,y),\phi^{(k)}(x,y+1),\phi^{(k+1)}(x,y-1)]\textrm{.} \label{eq:index-matrix}$$
Rubber Band Model
-----------------
As proposed in [@HonDes10], the rubber band model is employed to optimize the path obtained by the harmonic potential fields. This idea of rubber band was previously introduced in [@Hilgert_03], but mostly to define obstacle contours. Here, we combine the idea of rubber band model and harmonic potentials to define the path as a smoothed linked list of grid cells. The two immediately adjacent cells in the link, i.e. the previous and the next cells along the link from the current cell, exert internal forces on that same cell. Figure \[fig:Tensions-exerted\] illustrates this idea for the cell $i$ and its previous and next cells in the path, $i-1$ and $i+1$, respectively.
(-1.35,3.9) circle\[radius=0.14\] node at (-2,3.9) [${i-1}$]{}; (4.5,-1.65) circle\[radius=0.14\] node at (5.2, -1.65) [${i+1}$]{}; (0,0) circle\[radius=0.14\] node at (-0.25,-0.25) [${i}$]{}; (-3,0) – (5.1,0) node \[right,below\] [$x$]{}; (0,-2.2) – (0,4.2) node \[above,left\] [$y$]{}; (0,0) – (0,2.6) node \[right\] [$T_y^{i-1}$]{}; (0,0) – (0,-1.1) node \[left\] [$T_y^{i+1}$]{}; (0,0) – (-0.9,0) node \[below\] [$T_x^{i-1}$]{}; (0,0) – (3,0) node \[above\] [$T_x^{i+1}$]{}; (0,0) – (-0.9,2.6) node \[left\] [$T^{i-1}$]{}; (0,0) – (3,-1.1) node \[below\] [$T^{i+1}$]{}; (3,0) – (3,-1.1); (0,-1.1) – (3,-1.1); (-0.9,0) – (-0.9,2.6); (-0.9,2.6) – (0,2.6); (0,0) – (-1.35,3.9); (0,0) – (4.5,-1.65); (0,0) – (1.8,1.6) node \[right\] [$\vec{T}$]{}; (0,0) – (1.27,1.13); at (0.2,1.5) \[text width=1.2, align=center, red \][resultant force]{}; (1.8,4.2) – (0.9,3.4); (2.2,3.8) – (1.3,3.0); (2.6,3.4) – (1.7,2.6); (3.0,3.0) – (2.1,2.2); (3.4,2.6) – (2.5,1.8); (3.8,2.2) – (2.9,1.4); (4.2,1.8) – (3.3,1.0); (-0.6,-1.7) – (-1.5,-2.5); (-1.0,-1.3) – (-1.9,-2.1); (-1.4,-0.9) – (-2.3,-1.7); (-1.8,-0.5) – (-2.7,-1.3) node \[below, blue\] [$\vec{F}$]{}; (-2.2,-0.1) – (-3.1,-0.9); (-2.6,0.3) – (-3.5,-0.5);
Every cell in the path is affected by two kinds of forces: the internal tension (rubber band) forces $\vec{T}_{i\pm1}$, and the potential force $\vec{F}$. The position of a cell in the path is given by the pair $(x,\,y)$ that leads to the resultant forces to be minimum. That is: $$\begin{gathered}
(\hat{x},\hat{y})=\underset{(x,\,y)}{argmin}\,\left\Vert \vec{F}+\vec{T}_{i+1}+\vec{T}_{i-1}\right\Vert \label{eq:resultant} \\
(x,\,y)_{k+1}=(x,\,y)_{k}+\delta{(\hat{x},\hat{y})} \label{eq:drag}\end{gathered}$$ where $\delta{(\hat{x},\hat{y})}$ represents the direction toward the pair $(\hat{x},\hat{y})$ in Equation (\[eq:resultant\]), used to move the position of the current cell in the path at each iteration.
Let us assume that the current coordinates of the $i^{th}$ cell is $(x_{i},\,y_{i})$, and the coordinates of the two neighbors are $(x_{i-1},\,y_{i-1})$ and $(x_{i+1},\,$ $y_{i+1})$, respectively. The resultant of the forces on the cell $i$, as shown by Figure \[fig:Resultant-forces\], provides the direction and intensity with which the path should be moved in order for the forces to reach equilibrium.
The last component of these calculations is the force $\vec{F}$ derived from the harmonic potential. This force is calculated using: $$\vec{F}=(1-\phi^{(k)}(\hat{x},\,\hat{y}))^{-1}-(1-\phi^{(k)}(x_{i},\,y_{i}))^{-1} \label{eq:force}$$ where $\phi^{(k)}(x_{i},\,y_{i})$ represents the harmonic potential at the current position of the cell in the path given by Equation (\[eq:discretelaplace\]), and $\phi^{(k)}(\hat{x},\,\hat{y})$ represents the harmonic potential of the candidate position to which the cell $i$ is being dragged. At each step, the robot moves towards the next optimized position in the path. Figure \[fig:Resultant-forces\] summarizes the idea of the harmonic potentials and the internal tension (elastic) forces of the model. In the figure, red lines represent obstacles (walls) and the black area is the desired destination of the robot. The darker the color in free space, the lower the harmonic potential value. Figure \[fig:Resultant-forces\] also shows how the path obtained from the simple application of harmonic fields (full/blue path) compares to the one being optimized by the rubber band model (dotted/purple path).
(harmonic) at (0,0) ; (-0.4,1.15) – (-0.4,0.65) node \[right\] [$T^{i+1}$]{}; (-0.4,1.15) – (-0.85,1.45) node at (-0.55,1.7) [$T^{i-1}$]{}; (-0.4,1.15) – (-0.05,1.35) node at (0.15,1.45) [$F$]{}; (-0.9,0.85) – (-1.5,0.5) node at (-2.5,0.9) [moving direction]{};
Kalman Filter
-------------
A Kalman filter addresses the general problem of estimating the state $x\:\in\mathbb{R}^{n}$ of a discrete-time process, which in our case represents the position of a moving obstacle. This state evolves with the step $k$, governed by the following linear stochastic equation [@WelBis97]: $$x_{k}=A_{k}x_{k-1}+B_ku_{k-1}+w_{k-1}\label{eq:kalman_state}$$ where $A$, the $n\times n$ state matrix, relates the state at time step $k-1$ to the state at $k$; $B$, the $n\times l$ input matrix, relates the control input $u\in \mathbb{R}^{l}$ to the state $x$, and the random variable $w_{k}$ represents a process noise. Due to sensor limitations, the state of the system is observed by $z\:\in\mathbb{R}^{m}\;\;(m<n)$ according to a measurement model given by: $$z_{k}=H_{k}x_{k}+v_{k}\label{eq:kalman_measurement}$$ where the $m\times n$ matrix H is the observation matrix, and the random variable $v_{k}$ represents the uncertainty of the observation given the current state. Also, all random variables in the Kalman filter are assumed to be independent, normal distributions. That is, $p(x_{i}|z_{j})\sim N(\mu_{i|j},P_{i|j})$ for $i,j\in\{k-1,k\}$, $p(z_{i}|x_{j})\sim N(\lambda_{i|j},S_{i|j})$ for $i,j\in\{k-1,k\}$, $p(w)\sim N(0,Q)$, and $p(v)\sim N(0,R)$.
The Kalman filter is an iterative method that alternates between two phases: prediction and update. Since in our case, moving obstacles are assumed to move without requiring any input $u$, at each iteration, the Kalman filter predicts the system’s next state through Equations (\[eq:kalman\_time1\]) and (\[eq:kalman\_time2\]): $$\hat{x}_{k|k-1}=A_{k}\hat{x}_{k-1|k-1}\label{eq:kalman_time1}$$ $$P_{k|k-1}=A_{k}P_{k-1|k-1}A_{k}^{T}+Q_{k-1}\label{eq:kalman_time2}$$ where $\hat{x}_{k|k-1}$ is the estimation of the state $x$ based on the observations up to time $k-1$, and the estimate variance $P_{k|k-1}$ is the mean squared error in the estimate $\hat{x}_{k|k-1}$.
Then during the update phase, the state of the system is refined by: $$\hat{x}_{k|k}=\hat{x}_{k|k-1}+K_{k}(z_{k}-H_{k}\hat{x}_{k|k-1})\label{eq:kalman_measurement1}$$ $$P_{k|k}=(I-K_{k}H_{k})P_{k|k-1}\label{eq:kalman_measurement2}$$ where the Kalman gain, $K_{k}$ is defined as $$K_{k}=P_{k|k-1}H_{k}^{T}(H_{k}P_{k|k-1}H_{k}^{T}+R_{k})^{-1} \textrm{ .}\label{eq:kalman_measurement3}$$
The integration of Kalman filter and the time-warped grid will be discussed in Section \[subsec:Integrating-KF-HP-TWG\].
Time-Warped Grid\[subsec:Time-Warps\]
-------------------------------------
In a static environment, a path computed by combining the harmonic potentials and the rubber band model above is guaranteed to always exist and to be smooth. However, if obstacles are allowed to freely move in the environment, that same path must be constantly updated using the robot sensors (in our case the laser range sensor and the vision system). In that case, newly sensed positions of the obstacles will lead to new harmonic potentials and hence new smoothened paths to the goal from the rubber band model. This behavior is undesired, since it can force the robot to follow inefficient paths like loops, paths that sharply move the robot away from the goal, and/or paths that bring the robot dangerously close to obstacles. By using time-warped grid and Kalman filter to estimate obstacle velocities, our system is able to estimate future positions of the obstacles and calculate optimum paths despite of moving obstacles.
The idea of time-warped grid was inspired in general relativity, where the fabric of space-time is warped by large gravitational forces. Here, instead of gravitational forces, we use the velocities of the robot and of the obstacles to warp the grid in the environment map. The motivation for that, as we will explain in greater detail next, comes from the consequent assignment of a parallel processor for the computation of the potentials at each cell of the grid – and hence the path of the robot to the goal. Since the grid is warped by the velocities of all moving obstacles in the grid, the path itself becomes a function of those velocities, leading to a path plotted for the future position of those obstacles, rather than the current ones.
First, imagine a grid warped by the velocity of the robot alone. Such grid, depicted in Figure \[fig:warps\], has enlarged squares in front of the robot, i.e. in the direction of motion. If we think of these squares as the space traveled in one unit of time, it becomes obvious that the grids must be larger in front of the robot, smaller behind it, and the same on each side. In fact, in the direction of the motion of the robot, the warped grids stretch like ellipses centered around the position of the robot. Each of such ellipse can be numbered, representing the degree of warping. In order to estimate the future positions of moving obstacles, the robot should find the warps containing moving obstacles and label the obstacle with the corresponding warp number. As a result, if an obstacle moves towards the robot, its label decreases.
(grid) at (0,0) ; (robot1) at (-2.3,-2.3) ; (robot2) at (1.7,1.7) ; (robot3) at (-2.9,1.5) ; (-1.8,-2.3)–(-0.1,-2.3) node \[right\] [**main robot**]{}; (-2.6,1.8)–(-1.0,3.0) node \[above\] [**moving obstacles**]{}; (1.45,1.85)–(-0.5,3.0);
The time-warp of a moving obstacle with position of $( x_{obj},\,y_{obj} )$ is derived using the equation for a standard ellipse which has been rotated through an angle $\theta$: $$\begin{aligned}
& \frac{\big(\cos{(\theta)}(x_{obj}-x_c)+\sin{(\theta)}(y_{obj}-y_c)\big)^2}{r_x^2}+ \nonumber \\
& \frac{\big(\sin{(\theta)}(x_{obj}-x_c)-\cos{(\theta)}(y_{obj}-y_c)\big)^2}{r_y^2} = 1 \label{eq:ellipse}\end{aligned}$$ where $(r_x,\,r_y)$ are the major and minor axes of the ellipse respectively, and $\theta$ is the orientation of the main robot. The centers of ellipses $(x_c,\,y_c)$ are calculated based on the position of the main robot $(x_r,\,y_r)$, so the ellipses are larger and longer as they get farther from the robot, that is: $$x_c=x_r+0.9\,r_x\cos{\theta} \quad \textrm{and} \quad y_c=y_r+0.9\,r_x\sin{\theta}\textrm{.}$$
The elliptical grids are set for this application such that $r_x=4r_y$. Therefore, the time-warp number of a moving obstacle is calculated as the length of the major axis of the ellipse on which the moving obstacle is located: $$\begin{aligned}
r_x= & \Big(\big(\cos{(\theta)}(x_{obj}-x_c)+\sin{(\theta)}(y_{obj}-y_c)\big)^2 \nonumber \\ & +16\big(\sin{(\theta)}(x_{obj}-x_c)-\cos{(\theta)}(y_{obj}-y_c)\big)^2\Big)^{\frac{1}{2}} \label{eq:time-warp}\end{aligned}$$ After calculating the time-warps, Kalman filter estimates the future position of the moving obstacles for the next $j$’th step, where $j$ is the warp number of the grid occupied by the detected obstacle. Needless to say, Kalman filter algorithm should be initialized with respect to the velocity of the robot and the amount of noise in sensor readings.
The main idea of time-warped grid is to acknowledge that the further the obstacle is, the more delayed should be its impact in the path unless the obstacle also moves with great velocity towards the path. In summary: $$\textrm{Time-Warp}\,=:\,v \cdot t$$ where $v$ is the ratio of the velocity of the robot to the moving obstacle and $t$ is the warp number i.e. the corresponding ellipse number. $$( x,\,y )_{\textrm{future}}=( x,\,y )_{\textrm{estimated by Kalman filter based on time-warp}}$$
This estimated position of the obstacle is considered as the possible collision point with the robot and marked on the grid as if it were a fixed obstacle by increasing the corresponding harmonic value. Therefore, this obstacle would only affect the path of the robot if its future position lies in the vicinity of the robot path. Since the further the distance of the robot to the warped grid, the worst is Kalman filter predictions – i.e. the further in the future is the Kalman filter prediction, the less accurate it is – the assignment of “occupied” grids (high harmonic values) uses a Gaussian function. In other words, not only the estimated positions of the obstacles in the future are marked as occupied, but also their neighboring points. This Gaussian function is defined based on the calculated uncertainty given by Kalman filter at each step (Equation (\[eq:kalman\_measurement2\])) and a desired “safety” distance to obstacles.
At each iteration of the algorithm, the grid is cleared from harmonic potentials, and new values are updated onto the grid. This approach is justified since the estimations from the Kalman filter do not change drastically from one iteration to the next, and the values evolve slowly anyways. Also, at each iteration, the predicted harmonic potentials are used together with the rubber band model to determine a path for the robot to follow. It should be mentioned here that the robot does not have any prior knowledge about the environment, except for the position of the goal. It should be clear to the reader by now, that the time-warped grid method reduces and simplifies many calculations. For one, it eliminates the need to take the directions of the movement and the absolute value of the distance between the robot and the moving obstacles into consideration for the calculation of the path. In fact, the degree of warping assigned to the grids encodes both that direction and the time factors required for the calculations of the path.
In the next section, it will be further detailed the use of the time-warped grid combined with the Kalman filtering.
Integration of Kalman Filter with Time-Warps\[subsec:Integrating-KF-HP-TWG\]
----------------------------------------------------------------------------
As mentioned before, in this work the estimation of the position of moving obstacles in the future is done by Kalman filter. It is assumed that the current position of the obstacle is not known accurately due to noise and other errors in the vision-based localization system. In fact, the motion dynamics of the obstacles and their associated amount of noise are modeled in the implemented Kalman filter. The estimated positions of the obstacles are based on the warp number, the dynamic model and the previous positions of the obstacle using the iterative process described by Equations (\[eq:kalman\_state\] - \[eq:kalman\_measurement3\]).
For the experiments presented in Section \[sec:Experimental-Results\], the moving obstacles are simulated by the MobileSim software and the Aria API [@Whitbrook09], and they wander inside of a two-dimensional map with a mostly constant velocity[^3]. Therefore, the state $x$ contains $x_r$ and $y_r$ positions, and $\dot{x}_r$ and $\dot{y}_r$ velocities, and $A$ matrix in Equation (\[eq:kalman\_state\]) is formed as below: $$x=\left[\begin{array}{c}
x_r\\
y_r\\
\dot{x}_r\\
\dot{y}_r
\end{array}\right] , \; A=\left[\begin{array}{cccc}
1 & 0 & dt & 0\\
0 & 1 & 0 & dt\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & 1
\end{array}\right]$$
There is no input for this application so $B$ matrix is zero in Equation (\[eq:kalman\_state\]). The vision-based localization system provides measurements only from $x_r$ and $y_r$ position, not the velocities. Therefore, the $H$ matrix in Equation (\[eq:kalman\_measurement\]) becomes: $$H=\left[\begin{array}{cccc}
1 & 0 & 0 & 0\\
0 & 1 & 0 & 0
\end{array}\right] \textrm{ .}$$
Since the simulated moving obstacles usually have only linear motion (they turn only when they get close to walls and other obstacles), the process noise $w_{k}$ ($Q$) is assumed to be very small. On the other hand, observations can be noisy and since we have no knowledge about the exact position of the obstacles, the measurement noise $v_{k}$ ($R$) should be adequately set.
As we mentioned before, at each step, the Kalman filter uses the computed a-posteriori error covariance estimate ($P_{k}$) to adjust the uncertainty for the neighbors of the estimated position. In other words, the more uncertainty returned by the Kalman filter, the larger is the marked area on the grid around the estimated position of the obstacle.
Figure \[fig:good-explain\] illustrates how the estimated future positions based on the time-warped grid and Kalman filtering affect the planned path by harmonic potential fields and the rubber band model, in order to avoid collisions. The window in Figure \[fig:good-explain\](a) is the output of MobileSim with the actual robots, the laser range sensor information, and the walls of the environment. The window in Figure \[fig:good-explain\](b) is created by the main program using OpenGL to present the internal representation of the environment based on the sensed information. As the reader can notice, Figure \[fig:good-explain\](b) presents some small blue dots, besides the larger red dots. The first are the estimated and noisy observations made by the laser sensor and the vision-based localization system, while the latter are the predicted future positions of the moving obstacles by time-warps and Kalman filter. The detected static obstacles are marked with red lines. The solid black line represents the traversed path by the robot, while the dotted purple line shows the planned path by harmonic potentials and rubber band optimization. The goal position is represented by the darkest area on the map, indicating the lowest harmonic potential value.
[c]{}
(harmonic) at (0,0) ; (-1.25,-0.75)–(1.5,-0.75) node \[right\] [main robot]{}; (0.35,1.1)–(1.3,1.1) node \[right\] [moving obstacles]{};
[(a)]{}
(harmonic) at (0,0) ; (-1.5,2.4)–(1.6,2.4) node \[right\] [goal]{}; (-2.9,1.18)–(-5,1.18) node \[left\] [noisy observation]{}; (-2.5,0.35)–(-5,0.35) node \[left\] [planned path]{}; (-1,-0.025)–(0.2,-0.025) node \[right\] [estimated future position]{}; at (2.25,-0.5) [of the moving obstacle]{}; (-3.2,-0.5)–(-5,-0.5) node \[left\] [static obstacle]{}; (-2.5,-1.75)–(-5,-1.75) node \[left\] [traversed path]{};
GPU-Based Parallel Implementation
---------------------------------
All the above mentioned algorithms are highly computationally intensive especially when the map is large and/or there are many moving obstacles in the environment. The problem of using harmonic fields is that it requires repeated updates of the potential values at every cell of the grid. These updates are in turn a function of the potential of the neighboring cells, which leads to a recursive and quite-time consuming algorithm. Furthermore, the algorithm for optimizing the path using the rubber band model needs to calculate different forces at every cells of the grid in all iterations which is a heavy process. Estimating the future positions of moving obstacles requires another real-time process using time-warps and Kalman filter for each moving obstacle. An effective way to speed up these time consuming algorithms is through the use of parallel computation. Even more pertinent to our problem, since a fine and detailed grid of the environment may require millions of cells, we propose the parallelization through the use of General Purpose Graphics Processor Units (GPGPUs). Due to the use of grid-based maps, all these algorithms translate quite nicely into the parallel paradigm of GPU computing.
In that sense, several CUDA programs, namely kernel functions, are implemented to carry out the calculation of grid cells. The environment is divided into 10 cm by 10 cm cells, and each cell contains the input data for a GPU kernel called by a number of threads equal to the number of grid cells. In other words, each GPU thread is responsible for its own cell of the grid – e.g. for a 25.6 meter by 25.6 meter area, 256 $\times$ 256 threads (forming 16 by 16 thread blocks) are created to process the data in the grids. To this end, four GPU kernels are called in order to: 1) calculate the warp of each cell; 2) if there is a moving obstacle in the cell, run the Kalman filter algorithm to estimate the future position of the obstacle; 3) calculate the harmonic potential value of each cell; and 4) optimize the path using rubber band model. Each kernel function is multiply instantiated by the CUDA platform and the calculations of the various grid cells is performed in parallel by the GPU. Using different thread blocks provides the advantage of using GPU shared memory. The calculated harmonic values reside in the GPU global memory and are accessed by different threads. Since the neighbor cells have some shared harmonic values, by loading the harmonic values into the shared memory, the number of global memory accesses are reduced, which has a significant effect on the speed of the algorithm. An example of an environment divided into grid cells and their corresponding threads and thread blocks are shown in Figure \[fig:map\_squares\]. The figure illustrates this advantage of thread blocks and their shared memory.
Moreover, using GPUs as proposed enables the robot to handle many (if not all) moving obstacles at the same time. Some test scenarios with large number of obstacles can be found in Section \[sec:Experimental-Results\]. Final control of the robot and wandering the moving obstacles inside of the map are done by CPU.
(harmonic) at (0,0) ; at (-5.2,0.2) [Environment]{}; at (-5.2,-0.3) [map]{}; at (5.3,1.3) [Thread block]{}; at (4.9,-0.4) [GPU thread]{};
The proposed path planning procedure is outlined in Algorithm \[alg:algorithm\].
Read laser range sensor and vision-based localization data Detect static and dynamic obstacles **in parallel** Update the harmonic potential value: $\phi(x,y)=1$ Calculate the time-warp number of the moving obstacle via Eq. \[eq:time-warp\] Estimate the future position $(x,y)_{future}$ using Kalman filter via Eqs. \[eq:kalman\_state\]-\[eq:kalman\_measurement3\] Update the the harmonic potential value: [$\phi((x,y)_{future})=1$]{} **in parallel** Calculate harmonic potentials $\phi^{(k)}(x,y)$ via Eq. \[eq:harmonic-potential\] **in parallel** Update the index matrix with the position of the neighbor with lowest harmonic potential value via Eq. \[eq:index-matrix\] **in parallel** Calculate tension force $\vec{T}$ and potential force $\vec{F}$ via Eq. \[eq:force\] Update the index matrix with the position of the neighbor with the minimum resultant force (rubber band optimization via Eqs. \[eq:resultant\] and \[eq:drag\]) Generate the current path based on the optimized index matrix starting from the current position of the robot Move the robot to the next cell on the current path
Experimental Results \[sec:Experimental-Results\]
=================================================
Various test scenarios were performed using different maps and conditions of the environments, as well as the number of moving obstacles. All tests were performed for indoor navigation using a robot simulator for the Pioneer P3-DX robots (MobileSim). It is important to mention that all test scenarios were conducted using the NVIDIA GeForce GTX480 GPU and the CUDA programming environment.
For Figures \[fig:pp\_results\_1\]-\[fig:nopass\], the left window is the output from MobileSim, while the right window was created by the main program using OpenGL/CUDA to depict the internal representation of the sensed information form the environment, including the harmonic potentials, estimated position of obstacles, etc. Also in the left window of Figures \[fig:pp\_results\_1\] and \[fig:pp\_results\_2\], the dashed green lines illustrate the movement of obstacles, and the solid black lines in the right window show the path traversed by the mobile robot. The red squares show the estimated position of the moving obstacles in future, which are updated in real time, while the red lines demonstrate the static obstacles sensed by the laser sensor of the robot. The reader should keep in mind that these figures display simply a snapshot of the environment and the final traversed path at the end of the tests. However, the moving obstacles may have caused the path to shift, which cannot be easily seen by these snapshots. Simulation videos of the the performance of the proposed path planning algorithm can be found in the accompanying video for the paper, available at <http://vigir.missouri.edu/Time-Warp-Grid/>.
Figures \[fig:pp\_results\_1\] and \[fig:pp\_results\_2\] show some of these scenarios for different number of moving obstacles. The robot was able to find a smooth and short path while avoiding collisions in the majority of the scenarios (Table \[tab:success\]).
[cc]{} ![\[fig:pp\_results\_1\]Path planning results for 1, 2 and 4 moving obstacles (Left: actual environment, Right: program’s internal representation).](1_45-vis3 "fig:"){width="5.2cm"} & ![\[fig:pp\_results\_1\]Path planning results for 1, 2 and 4 moving obstacles (Left: actual environment, Right: program’s internal representation).](2-43-vis3 "fig:"){width="5.2cm"} & ![\[fig:pp\_results\_1\]Path planning results for 1, 2 and 4 moving obstacles (Left: actual environment, Right: program’s internal representation).](4_11-vis3 "fig:"){width="5.2cm"} &
[cc]{} ![\[fig:pp\_results\_2\]Path Planning results for 8, 12 and 16 moving obstacles (Left: actual environment, Right: program’s internal representation).](8-35-vis3 "fig:"){width="5.2cm"} & ![\[fig:pp\_results\_2\]Path Planning results for 8, 12 and 16 moving obstacles (Left: actual environment, Right: program’s internal representation).](12-26-vis3 "fig:"){width="5.2cm"} & ![\[fig:pp\_results\_2\]Path Planning results for 8, 12 and 16 moving obstacles (Left: actual environment, Right: program’s internal representation).](16-53-vis3 "fig:"){width="5.2cm"} &
Our first evaluations were exactly regarding these three metrics: success in avoiding collisions; length; and smoothness of the path. The proposed approach was tested over 400 times in different environments, all with size of $25.6 \times 25.6$ meters. The percentage of success for different numbers of moving obstacles is presented in Table \[tab:success\]. A very strict criterion was used to determine success or failure: if the robot collided with any obstacle before reaching the destination, or if any moving obstacle was forced to change direction because of the robot, the test was considered a failed attempt. It is important to mention that with the proposed algorithm, the robot can never stop and must always move with a constant velocity.
Besides success rate, the algorithm was also tested for shortness of the path found. Table \[tab:length\] reports the average length of the path for different maps and different number of obstacles. It would be reasonable to expect the length of the path to grow with the number of moving obstacles in the environment. However, this is not always the case. Sometimes, random movement of obstacles cause the path to change drastically and the final length increases even for a small number of obstacles.
The last property of the path to be evaluated was regarding its smoothness. Here, it should be mentioned that a path is usually regarded as smooth if it does not intersect itself and its tangent at each point varies continuously [@TayMan83]. Based on this definition, it can be stated that all paths found by the algorithm were smooth, as it is shown by Figures \[fig:pp\_results\_1\] and \[fig:pp\_results\_2\].
However, due to the target application of the proposed algorithm for power wheelchairs, we also evaluated the algorithm for large turning angles. So, in order to further quantify the smoothness of the optimized path in a dynamic environment, we computed the histogram of turning angles performed during navigation. Figure \[fig:angles\] shows the histograms for different numbers of moving obstacles. As these plots indicate, the turning angles are mostly concentrated on the small angular values. Moreover, even though increasing the number of moving obstacles leads to larger turning angles, these are very rare and likely due to some “deadlocked” situations.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------
 A scene when there is no planned path at the moment. (c, d) The algorithm can plan the path quickly.](nopass_1b "fig:"){width="0.35\columnwidth"}
[(a)]{} [(b)]{}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----
 A scene when there is no planned path at the moment. (c, d) The algorithm can plan the path quickly.](nopass_2b "fig:"){width="0.35\columnwidth"}
(c) (d)
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----
Indeed, in scenarios with a very large number of moving obstacles, it might happen that these obstacles completely block the path of the robot towards the goal. One such scenario is presented in Figure \[fig:nopass\] (a)-(b). However, even in those extreme cases, after a while, the robot is able to re-trace a path towards the goal as the obstacles move again away from the robot path (Figure \[fig:nopass\] (c)-(d)).
[Number of moving obstacles]{} [1]{} [2]{} [4]{} [8]{} [12]{} [16]{}
-------------------------------- ------------ ------------ ------------ ------------ ------------ ------------
[Percentage of success]{} [95.0 %]{} [95.0 %]{} [92.5 %]{} [87.5 %]{} [80.0 %]{} [75.0 %]{}
: [\[tab:success\]Percentage of success in path planning for different scenarios (out of 400 simulations).]{}
[Number of moving obstacles]{} [1]{} [2]{} [4]{} [8]{} [12]{} [16]{}
-------------------------------- ---------- ---------- ---------- ---------- ---------- ----------
[Map \#1]{} [4610]{} [5400]{} [5270]{} [4970]{} [5955]{} [7085]{}
[Map \#2]{} [6705]{} [6955]{} [7665]{} [6475]{} [6415]{} [6965]{}
[Map \#3]{} [4850]{} [4680]{} [4725]{} [6800]{} [7380]{} [8215]{}
[Map \#4]{} [4870]{} [4875]{} [5045]{} [5090]{} [5660]{} [6885]{}
: [\[tab:length\]Average length of the traversed path (cm).]{}
Here, we should mention that due to the limitation of the algorithm of not controlling the velocity of the robot, if the blocking situation persists for long enough, the robot has to turn around until it finds a new path, causing loops and sharp turns in the path (Figure \[fig:nopass3\]). These deadlock situations occur when the moving obstacles are aligned in a linear formation in a very dense area, or when multiple moving obstacles try to negotiate a narrow passage, e.g. a doorway. Deadlock situations can be avoided by adapting the algorithm to control the velocity of the mobile robot, which is beyond the scope of this work and will be the subject of future work.
[c]{}
Conclusion
==========
This paper presented a unique smooth path planning approach for dynamic and unknown environments for mobile robots. The algorithm relies on: harmonic potential fields to build a path; rubber band model to smoothen the path; and an integration of Kalman Filter and a new idea of Time-Warped Grid to estimate the position of moving obstacles and avoid collisions. The concept of time-warped grid reduces and simplifies many processes by eliminating the need to take the directions of movement and the absolute value of the distance between the robot and the moving obstacles into consideration during the estimation process. The proposed method was tested exhaustively using several simulation scenarios for the Pioneer P3-DX robot. The implementation of the algorithm was carried out using C/C++ and CUDA programming for real-time performance. As the simulations demonstrated, our approach is robust and is able to find the optimum path – i.e in terms of smoothness, distance, and collision-free – either in static or dynamic environments, with very high degree of success even with a very large number of moving obstacles.
Future Work
===========
In the course of this research project we have identified many directions for future work. Some are particular to our proposed approach, and some are broader in scope. The vision-based landmark localization system could be replaced by an algorithm to distinguish dynamic and static obstacles using only a laser sensor, which may result in a simpler system. Furthermore, using embedded GPUs like NVIDIA Jetson enables us to run the algorithm on the robot itself, rather than having an external server and consequent communications.
The dynamic obstacles in this work are assumed to have linear motion (they turn only when they get close to walls and other obstacles) and move with a constant velocity, resulting in a simple state matrix and dynamics used by the Kalman filter algorithm. In case of more complicated behavior for the moving obstacles, the Kalman filter should be built using the associated dynamic equations. Machine learning techniques such as neural networks can also be employed to learn the model and estimate the future positions of the moving obstacles and collision points.
Finally, an algorithm that could adopt the velocity of the mobile robot – or even stop it – should be investigated to solve the rare, but existing cases of momentary deadlocks.
[^1]: A path may not exist only if obstacles completely block the robot’s path to the goal.
[^2]: Unless a path momentarily does not exist due to a large number of obstacles (see Section \[sec:Experimental-Results\]).
[^3]: Interactions between moving obstacles can lead to change of velocity and/or direction of motion by one or all obstacles involved.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'As a warmup for studying dynamics and gravitons in loop quantum gravity, Varadajan showed that Wilson loops give operators on the Fock space for electromagnetism in Minkowski spacetime—but only after regularizing the loops by smearing them with a Gaussian. Unregularized Wilson loops are too singular to give densely defined operators. Here we present a rigorous treatment of unsmeared Wilson loops for vacuum electromagnetism on an arbitrary globally hyperbolic static spacetime. Our Wilson loops are not operators, but “quasioperators”: sesquilinear forms on the dense subspace of Fock space spanned by coherent states corresponding to smooth classical solutions. To obtain this result we begin by carefully treating electromagnetism on globally hyperbolic static spacetimes, addressing various issues that are usually ignored, such as the definition of Aharonov–Bohm modes when space is noncompact. We then use a new construction of Fock space based on coherent states to define Wilson loop quasioperators. Our results also cover “Wilson surfaces” in $p$-form electromagnetism.'
author:
- Miguel Carrión Álvarez
bibliography:
- 'dissertation.bib'
title: |
Loop Quantization\
versus\
Fock Quantization\
of $\mathrm{p}$-Form Electromagnetism\
on Static Spacetimes
---
I am indebted in one way or another to the following people.
Pilar Álvarez, porque madre no hay más que una; Pedro Carrión, who taught me to count (the rest just follows); Coral Duro, who introduced me to the Feynman Lectures on Physics at a tender age; Petra Solera, who sent me to the Math Olympiad; Luis Vázquez, who sent me on my Erasmus exchange; Guillermo García–Alcaine, who taught me quantum mechanics; Antonio Dobado, who taught me high-energy physics; Lee Smolin, who suggested that I study with John Baez; José Gaite, who supervised me on my first serious research; Fernando Bombal, who taught me functional analysis; Miguel Martín Díaz, who taught me probability theory; Ignacio Sols, who tried to teach me things about algebra that I had to rediscover on my own years later; John Baez, the best advisor this side of the Virgo cluster; Fotini Markopoulou, who was interested enough in my research to invite me to PI; and Barbara Helisová, who is just wonderful, and wonderfully patient too.
Without them I would never have come this far.
Introduction
============
This work is motivated by the open problem of representing gravitons in loop quantum gravity [@rovelli98], a proposed quantum theory of geometry and candidate for a theory of quantum gravity. The great virtue of loop quantum gravity is that it is manifestly background-free and diffeomorphism-invariant. Unfortunately, because the usual construction of the graviton Fock space depends explicitly on a background metric, it is difficult to say precisely how the notion of *graviton* arises in this formalism. At least at the kinematical level, in loop quantum gravity states of quantum geometry are described not in terms of gravitons but in terms of spin networks [@baez96], which had been invented independently by Penrose [@penrose71] and can be seen as a generalization of the Wilson loops introduced in the 1970’s for the study of non-abelian gauge theories [@wilson74]. However, describing the dynamics of quantum gravity in terms of spin networks remains a difficult open problem. So, we are not yet in a position to study how this dynamics reduces to that of gravitons in some limit, as presumably it should.
As a warmup, it is natural therefore to investigate the dynamics of Wilson loops in a gauge theory which is better understood: vacuum electromagnetism. However, until recently we were in the embarrassing situation of not even knowing the precise relation between the loop representation of electromagnetism and the usual Fock representation. Here, of course, the theory is linear and formulated on a fixed background metric, which drastically simplifies the situation. The technical problem is that the loop representation is based on a diffeomorphism-invariant vacuum, while the traditional Fock vacuum is tied to a particular background metric, which implies that photon (Fock) states are not part of the loop state space and Wilson loop states are not part of the Fock state space. In particular, with respect to the Fock vacuum, the photon 2-point correlation function blows up at short distances at such a rate that Wilson loops are not well-defined operators on Fock space.
Varadarajan [@varadarajan00; @varadarajan01] tackled this problem by “smearing” the loop $\gamma$ using Gaussian convolution in Minkowski space. Varadarajan’s procedure puts photons and Wilson loops in a common framework. Our goal in the present work is to understand electromagnetic Wilson loops without the need for smearing, and on general static, globally hyperbolic spacetimes. A related and important outstanding problem in loop quantum gravity is that spin network dynamics is poorly understood, and here we tackle the analogous problem of electromagnetic Wilson loop dynamics in the Fock representation.
The modern view of electromagnetism is that the electromagnetic potential $A$ is a connection on a $U(1)$ or ${\mathbb{R}}$ bundle over spacetime, and the electromagnetic field is the curvature of this connection. A *Wilson loop observable* is what mathematicians call the *holonomy* of the connection around a closed loop. In quantum theory, observables of a physical system are represented by operators on a Hilbert space of states of the system. In the case of electromagnetism in Minkowski spacetime, the state space of the electromagnetic field is the so-called Fock space. The main problem with the Wilson loop approach to quantum gauge field theories is that, even in the simple case of electromagnetism, Wilson loop operators are not defined on Fock space. Because in quantum field theory there is a correspondence between observables and states, this means that there are also no Wilson loop states in the Fock space of electromagnetism.
Quantum field theory on curved spacetimes is a famously problematic subject, as it combines the difficulties of quantum field theory, notably ultaviolet divergences, with a lack of a well-defined vacuum state due to the lack of global symmetries in a curved spacetime. For a free quantum field theory on a static spacetime, such as we are studying, these problems go away as there are no divegent interactions and there is a unique time-invariant vacuum state. Because of this, most physicist would say that vacuum electromagnetism on a static spacetime is well-understood. This is more or less true for scalar fields [@wald94], but then despite it being known [@wald94 § 4.7] that
> the requirement that the classical field equations have a well-posed initial value formulation in curved spacetime is a highly nontrivial restriction: the straightforward generalization to curved spacetime of the standard spin-$s$ field equations in flat spacetime do not admit a well posed initial value formulation for $s>1$
even researchers concerned only with electromagnetism and not with scalar fields work on the assumption that the mathematical theorems on scalar fields apply without modification to other fields [@dimock92].
> For globally hyperbolic manifolds, the usual classical linear field equations will have global solutions if they are well-behaved locally. We quote the result for scalar fields.
Part of the point of this thesis is to show that things are not so simple: there are subtleties involved due to gauge invariance and noncompact spacetimes which interact in unexpected ways. Our first goal is to clear this up and give a rigorous general treatment of vacuum electromagnetism on a static, globally hyperbolic spacetime. The subtleties arise mainly from the difference between the usual de Rham cohomology and a certain *twisted* $L^2$ cohomology arising from gravitational time-dilation. Indeed, in a careful treatment the electromagnetic vector potential is not a smooth $1$-form modulo exact smooth $1$-forms, but a normalizable $1$-form modulo exact normalizable $1$-forms. Similarly, the Aharonov–Bohm effect arises not from closed smooth modulo exact smooth vector potentials, but from closed normalizable modulo exact normalizable ones. This distinction would be inconsequential if space were compact, but this is not believed to be the case in physically realistic models of spacetime.
In Chapter \[sec:harmonic\] we present a rogues’ gallery of pathologies and counterexamples which illustrate how these subtleties can manifest themselves as physical effects, including the photon acquiring a mass due to the interaction of gravitational time dilation and the asymptotic geometry at spatial infinity.
When we quantize electromagnetism in Chapter \[chap:qed\], we will actually exclude the Aharonov–Bohm modes from our analysis. Chapter \[chap:linear\] describes our quantization procedure—essentially just Fock quantization, but done in a way that emphasizes the role of coherent states. The reason for this is that Wilson loop “operators” $${\textstyle\oint_\gamma\hat A}
\qquad\hbox{or}\quad
{\mathopen{:}{e^{i\oint_\gamma\hat A}}\mathclose{:}}$$ are *not* densely-defined operators on Fock space, but their matrix elements $${{\left\langle \phi\right\vert}\textstyle\oint_\gamma\hat A{\left\vert \psi\right\rangle}}
\qquad\hbox{or}\quad
{{\left\langle \phi\right\vert}{\mathopen{:}{e^{i\oint_\gamma\hat A}}\mathclose{:}}{\left\vert \psi\right\rangle}}$$ exist when $\phi,\psi$ are linear combinations of regular coherent states—that is, coherent states corresponding to sufficiently smooth classical solutions of Maxwell’s equations. Such regular coherent states span a dense subspace of Fock space, so they are sufficiently general to study Wilson loop dynamics. We are then able to prove formulas such as $${{\mathrm{d}}\over{\mathrm{d}}t}{\textstyle\oint_\gamma\hat
A}={\textstyle\oint_\gamma\hat E}$$ and $${{\mathrm{d}}\over{\mathrm{d}}t}{{{\left\langle X'\right\vert}{\mathopen{:}{e^{i\oint_\gamma\hat
A}}\mathclose{:}}{\left\vert X\right\rangle}}\over\langle X'\mid
X\rangle}=i{{{\left\langle X'\right\vert}{\textstyle\oint_\gamma\hat
E}{\left\vert X\right\rangle}}\over\langle X'\mid X\rangle}\exp i{{{\left\langle X'\right\vert}{\textstyle\oint_\gamma\hat
A}{\left\vert X\right\rangle}}\over\langle X'\mid X\rangle},$$ where ${\left\vert X\right\rangle},{\left\vert X'\right\rangle}$ are regular coherent states.
The plan of this dissertation is as follows: in Part $I$ we study classical vacuum electromagnetism, and in Part $II$ the quantization of vacuum electromagnetism. Part $I$ consists of three chapters. In Chapter $2$ we study ordinary vacuum electromagnetism in a $(3+1)$-dimensional static, globally hyperbolic spacetime. In Chapter $3$ we generalize our results to $(n+1)$-dimensional spacetimes and also consider theories where the electromagnetic potential is not a $1$-form but any $p$-form, including the massless scalar field ($p = 0$) and the Kalb-Ramond field ($p = 2$), which plays a role in string theory. Finally, in Chapter $4$ we survey the theory of $L^2$ cohomology and suggest physical interpretations of some of its main results. Part $II$ consists of two chapters. Chapter $5$ is where we describe our coherent-state quantization of linear dynamical systems and develop the concept of a quasioperator. Lastly, in Chapter $6$ this quantization method is applied to vacuum electromagnetism and used to make sense of unregularized Wilson loop quasioperators.
In this part we lay the classical groundwork for a a rigorous quantization of the vacuum Maxwell equations and the analogous equations for $p$-form electromagnetism with gauge group ${\mathbb{R}}$ on an arbitrary static, globally hyperbolic, $(n+1)$-dimensional spacetime. In other words, we assume that spacetime is invariant under time evolution and time reversal, and that the time evolution of fields in spacetime is completely determined by initial data. In fact, any such spacetime is topologically $M = {\mathbb{R}}\times
S$, and has a metric of the form $$g_M = e^{2\Phi} (-{\mathrm{d}}t^2 + g)
\index{$g_M$!spacetime metric}
\index{$\Phi$!Newtonian potential}
\index{${\mathrm{d}}$!exterior derivative}
\index{$t$!time coordinate}
\index{$g$!optical metric on space}$$ where $g$ is a complete metric on $S$, so that no lightlike geodesics run off to spatial infinity in a finite amount of their affine parameter.
Because the Lie algebras of ${\mathbb{R}}$ and $U(1)$ are canonically isomorphic, there is no difference between the versions of electromagnetism with either gauge group as far as the local formulation of the Maxwell equations is concerned. Globally there is a difference, though, because all ${\mathbb{R}}$-bundles are trivializable whereas $U(1)$-bundles may not be. In $3+1$ dimensions, the second Chern class of a nontrivial gauge bundle manifests itself as a topological magnetic charge whose field can be gauged away locally, but not globally. While topological charges are interesting, our primary goal is to study the effects of spatial non-compactness on quantization, and so we choose the gauge group ${\mathbb{R}}$ to eliminate the possibility of nontrivial bundles. When a principal ${\mathbb{R}}$-bundle is trivialized, connections on it are ordinary $1$-forms.
Technically, the subtlest aspects of our work arise from the function $\Phi$ appearing in the spacetime metric. This function measures the time dilation due to the gravitational field, and reduces to the Newtonian gravitational potential in the limit $\Phi \to 0$. When $\Phi = 0$, $p$-form electromagnetism uses rather familiar mathematics, mainly this portion of the $L^2$ de Rham cohomology complex: $$\begin{CD}
{L^2 \Omega^{p-1}_S}@>{{\mathrm{d}}_{p-1}}>>{L^2 \Omega^p_S}@>{{\mathrm{d}}_p}>>{L^2 \Omega^{p+1}_S}
\end{CD}
\index{$L^2\Omega^p_S$!square-integrable $p$-forms on~$S$}
\index{${\mathrm{d}}_p$!exterior derivative on~$p$-forms}$$ where $L^2 \Omega^p_S$ stands for the Hilbert space of square-integrable $p$-forms on $S$. The case $\Phi \ne 0$ requires some less familiar mathematics—except when $p+1$ is half the dimension of spacetime, in which case $p$-form electromagnetism is conformally invariant, allowing us to eliminate $\Phi$ by an appropriate rescaling of the fields. Even in the absence of conformal invariance, the most elegant approach is still to hide all the factors involving $\Phi$ by a field redefinition, and replacing the exterior derivative with the ‘twisted’ differential $$D_k = e^{{1\over 2}(n-2p-1)\Phi} {\mathrm{d}}_k e^{-{1\over 2}(n-2p-1)\Phi}
\index{$D_k$!twisted exterior derivative on~$k$-forms}
\index{$\Phi$!Newtonian potential}$$ obtained by conjugating the ordinary differential by the rescaling factor. This gives rise to a ‘twisted’ version of $L^2$ cohomology which, on a noncompact space, can differ from the usual $L^2$ cohomology which, in turn, can differ from the smooth de Rham cohomology.
With this machinery in place we model the phase space of classical $p$-form electromagnetism on $(n+1)$-dimensional spacetime as a real Hilbert space with continuous Hamiltonian and symplectic stucture. In the process, we address the Aharonov–Bohm effect in situations where the twisted $L^2$ cohomology differs from the usual de Rham cohomology, a subtle issue that is largely neglected in the literature.
Among the most rigorous published treatments of Maxwell’s equations on a fairly generic manifold stands that of Dimock [@dimock92], which however is restricted to $(3+1)$-dimensional spacetimes with compact Cauchy surfaces. At the time of his writing, he said “nothing that follows is particularly new, but it seems that the various pieces have not been put together”. A later paper reviewing the canonical and covariant formulations of the classical Maxwell theory on a generic globally hyperbolic spacetime is the one by Corichi [@corichi], again “intended to fill an existing gap in the literature”.
Dimock constructs the classical phase space from gauge equivalence classes of Cauchy data and the symplectic structure obtained from the Noether current. Gauge fixing appears as a technical step used to show that Maxwell’s equations are strictly hyperbolic, so that solutions are determined by their Cauchy data. Dimock uses “fundamental solutions” (essentially Green’s functions) to parameterize the phase space, a technique that only works for linear field equations. Time evolution enters the picture through symplectic transformations induced on phase space by changes in the choice of Cauchy surface. In fact, Dimock makes “no choice of Hamiltonian or special time coordinate”, following the covariant canonical formalism of [@crnkovic87]. Dimock points out how the field strength does not provide a complete set of observables when the first homology class of the Cauchy surfaces is nontrivial. In Chapter \[sec:3+1\] we relate this phenomenon to the Aharonov–Bohm effect and in Chapter \[sec:harmonic\] we present a thorough overview of the situation in the non-compact case. Dimock assumes a trivial $U(1)$-bundle saying “presumably our results can be extended to non-trivial bundles for which $A$ is only defined locally”, while we take the more drastic step of assuming an ${\mathbb{R}}$-bundle.
For the purposes of this Part, Dimock’s presentation of Maxwell’s equations does have a couple of important limitations. First, the restriction to compact Cauchy surfaces may be unphysical, and certainly excludes many cases of theoretical interest. We address the thorny analytic issues associated to allowing noncompact Cauchy surfaces in Chapter \[sec:3+1\], albeit with the additional assumption that spacetime is static, which Dimock does not need. The topological implications of noncompactness are discussed in Chapter \[sec:harmonic\]. Dimock’s use of compact Cauchy surfaces allows his to bring Hodge’s theorem to bear on the Cauchy data and, using the Kodaira decomposition, to show that the symplectic structure is non-degenerate. Although Hodge’s theorem does not hold on a noncompact space (see Chapter \[sec:harmonic\]), we are nevertheless able to prove a form of Kodaira’s decomposition in Chapter \[sec:3+1\].
Dimock also states without proof or reference that “for globally hyperbolic manifolds, the usual classical linear field equations will have global solutions if they are well-behaved locally. We quote the result for scalar fields”. We repaired this defect by reference to Chernoff’s work in Chapter \[sec:harmonic\]. In the proof of existence of solutions with given Cauchy data Dimock states “The equation \[above\] has principal part $g^{\mu\nu}\partial_\mu\partial_\nu$ and thus is strictly hyperbolic”; hyperbolicity easily follows from Chernoff’s work. Finally, the phase space constructed by Dimock does not have a topology other than that induced by imposing the continuity of the symplectic structure. Therefore, it is not a real inner-product space like ours is.
While not assuming compact Cauchy surfaces, Corichi’s paper is “not very precise about functional-analytic issues” in the author’s own words. The covariant formulation is, like Dimock’s, based on the formalism of [@crnkovic87], and differs mostly in the notation. The canonical formulation is written in a manifestly covariant way, in terms of the foliation generated by an arbitrary time coordinate function. Both formulations of classical electromagnetism are more general than ours, and the relationship between Corichi’s covariant and canonical descriptions of phase space is equivalent to Dimock’s treatment of Cauchy data in the covariant formalism.
The plan of this Part is as follows. We begin in Chapter \[sec:3+1\] by setting up classical electromagnetism with gauge group ${\mathbb{R}}$, leading up to Theorems \[thm:3+1\] and \[thm:3+1phys\], in which we make the phase space for this theory into a real Hilbert space on which the classical Hamiltonian is a continuous nonnegative quadratic form. In Chapter \[sec:N+1\] we generalize this work to $p$-form electromagnatism in $n+1$ dimensions using the twisted de Rham complex, leading up to the analogous Theorems \[thm:N+1\] and \[thm:n+1phys\]. In Chapter \[sec:harmonic\] we survey what is known about $L^2$ cohomology on noncompact spaces, and study a number of examples illustrating some of the associated subtleties.
Classical vacuum electromagnetism {#sec:3+1}
=================================
In this chapter we discuss the classical vacuum Maxwell equations on a $(3+1)$-dimensional static globally hyperbolic spacetime. In particular, we explain how the classical phase space of electromagnetism splits into two parts, one containing the oscillatory modes of the electromagnetic field and the other containing the ‘topological’ modes responsible for the ‘Aharonov–Bohm’ effect.
The plan of this chapter is as follows: we begin in Section \[sec:3+1-geometry\] by describing in detail our assumptions and notation concerning spacetime geometry, decompose spacetime in the form $M\cong{\mathbb{R}}\times S$, and confront a number of analytical issues arising from trying to define the exterior derivative on square-integrable differential forms. In Section \[sec:maxwell\] we give an overview of the stationary action formulation of classical mechanics, and use it to derive the Maxwell equations, Noether current, Hamiltonian and symplectic structure, as well as kinematical, dynamical and physical phase spaces. Finally, in Section \[physical.interpretation\] we describe the splitting on the physical phase space of classical vacuum electromagnetism into an sector consisting of oscillating modes, and a sector consisting of topological modes responsible for the Aharonov–Bohm effect.
After seeing that the spacetimes we are interested split in the form $M\cong{\mathbb{R}}\times S$, where $S$ is *space*, we define the exterior derivative ${\mathrm{d}}$ and coderivative ${\mathrm{d}}^*$ so that they act on square-integrable differential forms on space and satisfy $$\label{eqn:adjoint}
\int_Sg(\alpha,{\mathrm{d}}\beta){\mathrm{vol}}=\int_Sg({\mathrm{d}}^*\alpha,\beta){\mathrm{vol}}\index{$S$!space}
\index{$g$!optical metric on space}
\index{${\mathrm{d}}$!exterior derivative}
\index{${\mathrm{vol}}$!volume form of optical metric}
\index{${\mathrm{d}}^*$!exterior coderivative}$$ whenever $\alpha$ and $\beta$ are square-integrable differential forms of appropriate degrees. The key is to show that no ‘boundary terms at infinity’ appear in the integration by parts implicit in Equation (\[eqn:adjoint\]). This can be used to show that the Laplacian on square-integrable differential forms is essentially self-adjoint and nonnegative, properties necessary for rigorous quantization.
In the *temporal gauge* (vanishing electrostatic potential) the configuration space of classical electromagnetism on $M$ consists of ${\mathbb{R}}$-connections on $S$ modulo gauge transformations, and so is isomorphic to a space of $1$-forms modulo square-integrable exact $1$-forms on $S$. In physics, such a $1$-form is called a *vector potential*. We make the configuration space into a real Hilbert space by defining it as $${\mathbf{A}}={{\mathop{\mathrm{dom}}}\{{\mathrm{d}}{\colon}L^2\Omega^1_S\to L^2\Omega^2_S\}\over\overline{{\mathop{\mathrm{ran}}}}\{{\mathrm{d}}{\colon}L^2\Omega^0_S\to L^2\Omega^1_S\}}
\index{${\mathbf{A}}$!space of vector potentials}
\index{${\mathrm{d}}$!exterior derivative}
\index{$L^2\Omega^p_S$!square-integrable $p$-forms on~$S$}$$ with its natural real inner product. That is, ${\mathbf{A}}$ consists of equivalence classes of square-integrable $1$-forms with square-integrable exterior derivatives, modulo exact $1$-forms. This space is naturally a real Hilbert space.
The canonical conjugate of the vector potential $[A]$ is a divergenceless $1$-form $E$, called the *electric field*. The space of electric fields $${\mathbf{E}}=\ker\{{\mathrm{d}}^*{\colon}L^2\Omega^1_S\to L^2\Omega^0_S\}
\index{${\mathbf{E}}$!space of electric fields}
\index{${\mathrm{d}}^*$!divergence}
\index{$L^2\Omega^p_S$!square-integrable $p$-forms on~$S$}$$ is also naturally a real Hilbert space. The phase space of classical electromagnetism is, then, the real Hilbert space $${\mathbf{P}}={\mathbf{A}}\oplus{\mathbf{E}}.
\index{${\mathbf{P}}$!phase space}
\index{${\mathbf{A}}$!space of vector potentials}
\index{${\mathbf{E}}$!space of electric fields}$$ The spaces ${\mathbf{A}}$ and ${\mathbf{E}}$ are dual to each other by $$\bigl([A],E\bigr)=\int_S g\bigl(A,E){\mathrm{vol}},
\index{$(~,~)$!$p$-form inner product on~$S$}
\index{$S$!space}
\index{$g$!optical metric}
\index{$[A]$!gauge equivalence class of vector potentials}
\index{$E$!electric field}
\index{${\mathrm{vol}}$!volume form of~$g$}$$ which is independent of the representative $A$ chosen for $[A]$ because $E$ is divergenceless. The symplectic structure on ${\mathbf{P}}$ is constructed from this duality pairing by antisymmetrization: $$\omega\bigl([A]\oplus E,[A']\oplus E'\bigr)=\int_S\bigl[g(A,E')-g(A',E)\bigr]{\mathrm{vol}}.$$
Because of global hyperbolicity, any point $X=[A]\oplus E$ of the physical phase space determines a unique solution of Maxwell’s equations on all of $M$. Time evolution is given by a continuous one-parameter group of continuous symplectic transformations $T(t){\colon}{\mathbf{P}}\to{\mathbf{P}}$. Unlike the symplectic structure and the Hamiltonian, the natural Hilbert space norm on ${\mathbf{P}}$ is not preserved by this time evolution.
As a result of gauge-fixing, when restricted to the phase space the Laplacian on $1$-forms is $\Delta={\mathrm{d}}^*{\mathrm{d}}$. The assumption that spacetime is static then implies that time evolution commutes with $\Delta$, and so the phase space admits the decomposition $${\mathbf{P}}={\mathbf{P}}_o\oplus{\mathbf{P}}_f
\index{${\mathbf{P}}$!phase space}
\index{${\mathbf{P}}_o$!oscillating sector of phase space}
\index{${\mathbf{P}}_f$!free sector of phase space}$$ where ${\mathbf{P}}_f$ is the kernel of $\Delta$ in ${\mathbf{P}}$ and consists of generalized Aharonov–Bohm modes. From the point of view of dynamics, the direct summand ${\mathbf{P}}_o$ consists of ‘oscillating modes’ and ${\mathbf{P}}_f$ of ‘free modes’. Specifically, on ${\mathbf{P}}_o$ the Hamiltonian is a positive-definite quadratic form, and so that ‘sector’ of the electromagnetic field has the dynamics of an infinite-dimensional harmonic oscillator. The free sector ${\mathbf{P}}_f$ has dynamics analogous to those of a free particle. For the free sector one can successfully apply the algebraic approach to quantization of Chapter \[chap:linear\], but the existence of a Hilbert-space representation on which time evolution is unitarily implementable is not guaranteed unless ${\mathbf{P}}_f$ is finite-dimensional. As we shall see in Chapter \[sec:harmonic\], that may not be the case on a noncompact space even if it is topologically trivial.
$p$-form electromagnetism in $n+1$ dimensions {#sec:N+1}
=============================================
In this chapter we generalize the results of the last chapter to electromagnetism on spacetimes of arbitrary dimension $n+1$. As before, we take our spacetime to be of the form ${\mathbb{R}}\times S$, equipped with the Lorentzian metric $$g_M = e^{2\Phi}(-dt^2 + g)$$ where $g$ is a complete Riemannian metric on $S$. The only difference is that now $S$ is $n$-dimensional instead of 3-dimensional. However, this means that Maxwell’s equations are no longer conformally invariant, so the ‘gravitational potential’ $\Phi$ plays a more significant role. To see why, recall that the Maxwell action is still given by $${{\mathcal{S}}}[A_M] = -{1\over 2}\int_M g_M(F_M , F_M )\, {\mathrm{vol}}_M$$ where $$F_M = {\mathrm{d}}t\wedge(\partial_t A - {\mathrm{d}}A_0) + {\mathrm{d}}A.$$ By Equation (\[eq:stpart2\]), $$g_M(F_M,F_M)=
e^{-4\Phi} \, \bigl[ -g(\partial_t A - dA_0,\partial_t A - dA_0)
+ g(dA,dA) \bigr]$$ and, by Equation (\[eq:vol\]), $${\mathrm{vol}}_M = e^{(n+1)\Phi} {\mathrm{vol}}\wedge dt$$ where ${\mathrm{vol}}$ is the volume form on space. Hence, we have $${{\mathcal{S}}}[A,A_0] = {1\over 2}\int_M
\bigl [g(\partial_t A - {\mathrm{d}}A_0,\partial_t A - {\mathrm{d}}A_0) -
g({\mathrm{d}}A,{\mathrm{d}}A) \bigr]\, e^{(n-3)\Phi} {\mathrm{vol}}\wedge dt .$$ The factors involving $\Phi$ cancel only if $n = 3$, indicating conformal invariance. In other dimensions, the most elegant way to deal with these factors involving $\Phi$ is to *redefine* the fields $A_0$ and $A$ by setting $$A_M = e^{-{1\over 2}(n-3)\Phi} ({\mathrm{d}}t \wedge A_0+A),$$ and then to ‘twist’ the exterior derivative of differential forms on space, defining a new operator $$D = e^{-{1\over 2}(n-3)\Phi} \, {\mathrm{d}}\, e^{{1\over 2}(n-3)\Phi} .$$ The action is then $${{\mathcal{S}}}[A,A_0] = {1\over 2}\int_M
\bigl [g(\partial_t A - DA_0,\partial_t A - DA_0) -
g(DA,DA) \bigr] {\mathrm{vol}}\wedge {\mathrm{d}}t$$ which is formally just like equation (\[eq:action\]) was in the $(3+1)$-dimensional case, but with rescaled fields $A$ and $A_0$, and with the twisted exterior derivative $D$ replacing the usual ${\mathrm{d}}$.
With the help of this formal equivalence, the whole theory goes through almost exactly as before. In particular, if we let $L^2
\Omega^p$ be the Hilbert space consisting of all square-integrable $p$-forms on $S$, there are mutually adjoint operators $$\xymatrix{L^2\Omega^0\ar@<.5 ex>[r]^{D_0} &
L^2\Omega^1\ar@<.5 ex>[l]^{D_0^*}\ar@<.5 ex>[r]^{D_1} &
L^2\Omega^2\ar@<.5 ex>[l]^{D_1^*}}$$ Using the Kodaira decomposition for this sequence we obtain $$L^2\Omega^1=
\overline{{\mathop{\mathrm{ran}}}D_0}
\oplus
\ker L_1
\oplus
\overline{{\mathop{\mathrm{ran}}}D_1^*}$$ where now the Laplacian is replaced by the ‘twisted Laplacian’ $L_1$, a nonnegative self-adjoint operator on 1-forms given by $$L_1 = D_1^* D_1 + D_0 D_0^*.$$
In fact, having come this far, it would be a pity not to treat ‘$p$-form electromagnetism’, a generalization of Maxwell’s equations in which the electromagnetic vector potential is replaced by a $p$-form. The general case was treated by Henneaux and Teitelboim [@henneaux86]. For $p = 0$, this theory is just the massless neutral scalar field. For $p = 2$, it is the Kalb–Ramond field arising naturally in string theory [@green87 Section 3.4.5][@kalb74], while for $p = 3$ it plays a part in 11-dimensional supergravity [@duff99]. All our formulas generalize painlessly to these theories in the absence of charges. Starting with the $p$-form $A_M$ on spacetime, we define a field strength tensor $F_M = {\mathrm{d}}_M A_M$, and take the action of the theory to be $${{\mathcal{S}}}[A_M] = -{1\over 2}\int_M g_M(F_M , F_M )\, {\mathrm{vol}}_M .$$ This action gives equations of motion and gauge symmetries having the same form as in Maxwell theory. Furthermore, if we set $$A_M = e^{-{1\over 2}(n-2p-1)\Phi} ({\mathrm{d}}t \wedge A_0+A)$$ and define the twisted differential $D$ as follows: $$D = e^{{1\over 2}(n-2p-1)\Phi} {\mathrm{d}}e^{-{1\over 2}(n-2p-1)\Phi} ,$$ we obtain $${{\mathcal{S}}}[A,A_0] = {1\over 2}\int_{{\mathbb{R}}}
\bigl [(\partial_t A - DA_0,\partial_t A - DA_0) -
(DA,DA) \bigr]{\mathrm{d}}t$$ in complete analogy with ordinary Maxwell theory. This allows us to generalize all our results on Maxwell’s equations to the $p$-form case.
Hodge–de Rham theory on noncompact manifolds {#sec:harmonic}
============================================
As we have seen, the space of harmonic differential forms, consisting of closed and coclosed differential forms, plays a special role in the analysis of the phase space of Maxwell’s equations: it corresponds to the space of physical vector potentials with vanishing magnetic field (Aharonov–Bohm effect), and also to static electric fields with no finite sources (charge without charge).
When space is compact, it is well known that the Hodge-de Rham theorem identifies the square-integrable, smooth and real cohomologies of a space, and that the space of square-integrable harmonic forms coincides with the kernel of the Hodge Laplacian $\Delta={\mathrm{d}}\delta
+\delta{\mathrm{d}}$.
When space is noncompact everything becomes more complicated. To begin with, the definition of the codifferential $\delta$ involves integration by parts. As a result, unless space is complete in the optical metric it may be impossible to define the codifferential (and hence the Laplacian) without specifying boundary conditions at infinity. When the optical metric on space is complete, not only is there an unambiguous definition of the codifferential and Laplacian, but the space of $L^2$ harmonic forms is identified with the kernel of the Hodge Laplacian, and it has a square-integrable cohomology interpretation. However, the square-integrable cohomology is not a topological invariant, as it depends crucially on the geometry at infinity.
These are the main questions one can ask about the Laplacian $\Delta$ on a complete Riemannian manifold [@lott97; @carron01; @carron02]:
1. Is the dimension of $\ker\Delta_p$ finite or infinite? In physical terms, this is the dimension of the space of $p$-form Aharonov–Bohm modes.
2. What are sufficient conditions for $\ker\Delta_p$ to be trivial or finite-dimensional?
3. If $\ker\Delta_p$ is finite-dimensional, does it have a topological interpretation?
4. Is $0$ in the essential spectrum of $\Delta_p$? Physically, this signals the presence of *infrared divergences* for massless $p$-form fields. Conversely, if the essential spectrum is bounded away from $0$, we have a mass gap for a free massless field induced by the spatial geometry at infinity! Note that it is possible for $0$ to be in the essential spectrum of the Laplacian even if $\ker\Delta_p$ is trivial, and that the most familiar example of this is Euclidean ${\mathbb{R}}^n$.
The answer to all of these questions depends on the behaviour of the curvature of the optical metric at infinity, so even a massless field may acquire an ‘effective mass’. In this chapter we collect some known facts and open issues about the space $\ker\Delta_p$ of harmonic $p$-forms and the spectrum of the Laplacian on a complete Riemannian manifold $S$, and give physical interpretations of them. Although this chapter is a review, it points out how rich the subject is compared to the amount of attention it has received from physicists.
This chapter is based in part on the excellent review of harmonic forms on noncompact manifolds by Carron [@carron01] (in French), which includes his finite-dimensionality results [@carron99] obtained from Sobolev-type inequalities involving the curvature. Another paper of his [@carron02] (in English) contains a shorter overview, and a geometrical interpretation of the $L^2$ cohomology of manifolds with flat ends (which are known to have finite cohomologies). The $L^2$ cohomology of hyperbolic manifolds is described by Lott [@lott97]. The case of geometrically finite hyperbolic manifolds was obtained by Mazzeo and Phillips [@mazzeo90], including a calculation of the essential spectrum of the Laplacian. Mazzeo also calculated the cohomology and essential spectrum of the Laplacian for conformally compact metrics [@mazzeo88]. The $L^2$ cohomology for rotationally symmetric manifolds was obtained by Dodziuk [@dodziuk79].
An additional complication is the ‘twisting’ of the cohomology complex: $$\xymatrix{L^2\Omega^{k-1}_S\ar[r]^{{\mathrm{d}}}\ar[d]^{e^{{1\over
2}(n-2p-1)\Phi}}&L^2\Omega^k_S\ar[r]^{{\mathrm{d}}}\ar[d]^{e^{{1\over
2}(n-2p-1)\Phi}}&L^2\Omega^{k+1}_S\ar[d]^{e^{{1\over 2}(n-2p-1)\Phi}}\\
L^2\Omega^{k-1}_S\ar[r]^{D}&L^2\Omega^k_S\ar[r]^{D}&L^2\Omega^{k+1}_S}$$ For compact $S$ the (smooth) function $\Phi$ is bounded, multiplication by $e^{{1\over 2}(n-1-2p)\Phi}$ is bi-continuous on each $L^2\Omega^k$, the twisted $L^2$ cohomology coincides with the ordinary $L^2$ cohomology, and the latter with the de Rham cohomology by Hodge’s theorem. For non-compact $S$, however, $\Phi$ might be unbounded, in which case the twisted $L^2$ cohomology complex need not be isomorphic to the ordinary $L^2$ cohomology complex, which we know already can be very much unlike the de Rham cohomology complex for which we have some intuition. Note that if $n+1=2(p+1)$ (when $p$-form electromagnetism is conformally invariant) there is no twisting of the cohomology complex, so the only subtleties are the differences between the $L^2$ and de Rham cohomologies.
Since the twisted Laplacian $DD^*+D^*D$ has not been studied in nearly as much detail as the ordinary Hodge Laplacian, we know little about its behaviour. Therefore, when $\Phi$ is unbounded, most of what we will say in this chapter is directly applicable only to the cases where $p$-form electromagnetism is conformally invariant—[*i.e.*]{}, $p$-form electromagnetism in $2(p+1)$-dimensional spacetime, which includes the classical case of $1$-forms in $3+1$ dimensions.
The apparent truism that a quantum mechanical theory needs to be cast in classical language in order to correlate its predictions with our experience, a point that Niels Bohr made into a cornerstone of his philosophy of quantum mechanics, has practical consequences for the development of quantum descriptions of physical systems. This is because a physical system will be described operationally or geometrically in inevitably classical terms, and this information needs to be fashioned into a quantum theory whose predictions need to be, again, reexpressed in classical terms. In addition, the process of constructing a classical theory from operational or geometric data is so well-understood that it is convenient to construct the quantum theory by first constructing a classical theory from the data and then ‘quantizing’ it.
Quantization is a catch-all term for any process taking as input a classical mechanical system, and producing as output a quantum mechanical system reducing to the original classical system in an appropriate limit. Quantization would ideally be algorithmic or functorial, but it turns out to be neither, although formulating quantization in algebraic language seems to bring it closest to the goal of functoriality.
In algebraic terms, a classical mechanical system is defined by specifying a Poisson algebra of observables, while any associative algebra can play the role of algebra of observables for a quantum system. The Dirac quantization prescription [@dirac57 Chapter IV] ‘promotes’ the commuting classical observables to operators satisfying the Heisenberg commutation relations $$[\hat f,\hat g]=i\hbar\{\widehat{f,g}\},
\index{$\hat f$!quantized observable}
\index{[~,~]!commutator}
\index{$\hbar$!Planck's constant}
\index{$f$!observable}
\index{$\{~,~\}$!Poisson bracket}$$ where $\{f,g\}$ is the Poisson bracket of the classical observables $f$ and $g$, $[\hat f,\hat g]$ is the commutator of their quantum counterparts, and Planck’s constant $\hbar$ measures the departure from classical behaviour (where observables commute). It is not hard to convince oneself that, because the algebra of quantum observables is nonabelian, the operation $f\mapsto\hat f$ cannot be an algebra homomorphism. That is, $\widehat{fg}\neq\hat f\hat g$ in general. Physicists call this fact ‘operator ordering ambiguities’.
An operator algebra of quantum observables realizing the canonical commutation relations achieves quantization in a kinematical sense, but the physical and dynamical content of the theory comes about by means of a specific representation of the quantum observables as an algebra of (unbounded) linear operators on a Hilbert space of quantum states. Each representation is associated to a choice of ‘vacuum expectation’ on the algebra of observables and it is known that, for systems with infinitely many degrees of freedom, different states may lead to unitarily inequivalent representations. The choice of representation can be narrowed down by the need to recover an appropriate classical limit, and by requiring that physical symmetries be implemented unitarily.
The classical limit is encoded in the correspondence principle, by which we mean the following. The Poisson algebra of classical observables consists of smooth functions on a symplectic manifold (phase space) playing the role of state space for the classical theory. The correspondence principle requires that, for any phase space point $x\in{\mathbf{P}}$ and any observable $f$, there should be a quantum state ${\left\vert x\right\rangle}$ such that the expected value of $\hat f$ in the state ${\left\vert x\right\rangle}$ equals the classical value $f(x)$, if not exactly, at least in the limit $\hbar\to 0$. That is, $${{\left\langle x\right\vert}\hat f{\left\vert x\right\rangle}}=f(x)+O(\hbar).
\index{${{\left\langle x\right\vert}\hat f{\left\vert y\right\rangle}}$!matrix element of~$\hat f$}$$
There is one last requirement that a sensible quantization must satisfy, and that is that physical symmetries be represented by unitary operators on the Hilbert space of quantum states of the system.
In the case where the classical phase space is a vector space, the linear observables can be identified with the points of the phase space itself, and so the Heisenberg commutation relations can be implemented on the phase space. In Chapter \[chap:linear\] we develop the quantization of an abstract linear system and develop the concept of a quasioperator on Fock space, and in Chapter \[chap:qed\] we apply this to Maxwell’s equations for the electromagnetic field and express the dynamics of the quantized electromagnetic field in terms of Wilson loops quasioperators.
The work most closely akin to ours is that of Dimock [@dimock92]. Like us, Dimock constructs a $C^*$-algebra of observables for the electromagnetic field, but he does not exhibit any states or Hilbert-space representations. He notes in passing that “in any case such \[Hilbert-space\] representations exist, say by a Fock space construction”. We discuss below some ways in which a Fock space representations may fail to exist.
Because Dimock describes the classical theory in the covariant canonical formalism, he is forced to focus on “the algebraic structure of the theory, not in the specification of particular states”. In our terms, Dimock quantizes the electromagnetic field as a ‘general boson field’. He also constructs a classical Poisson bracket, and his quantization procedure is equivalent to our general linear quantization. Dimock does show that different Hilbert-space representations lead to $*$-isomorphic $C^*$-algebras of observables. This form of equivalence, however, obviates the possible physical consequences of unitary inequivalence of Hilbert-space representations, and for this reason Dimock’s paper suffers from what Earman and coauthors critically term “algebraic imperialism” in [@earman].
Dimock does not show that the classical canonical transformations associated to changes in the choice of Cauchy surface are implemented unitarily on the $C^*$-algebras of quantum observables, because that is simply not true. In fact, Torre and Varadarajan [@torre] show that, even in the case of free scalar fields on a flat spacetime of dimension higher than two, there is no unitary transformation between the Fock representations associated to arbitrary initial and final Cauchy surfaces. They point out that unitary implementability is easily obtained if the Cauchy surfaces are related by a spacetime isometry, though. They also mention related results of Helfer (no unitary implementation of the $S$-matrix if the ‘in’ and ‘out’ states are Hadamard states) [@helfer96], and of van Hove (only a small subgroup of the classical canonical transformations is unitarily implementable) [@hove51].
Another paper addressing specifically the quantization of the electromagnetic field is the one by Corichi [@corichi]. Corichi stresses that Fock quantization depends crucially on the linear structure of phase space, and characterizes the Fock quantization procedure as “completely elementary”.
Here we perform Fock quantization of Maxwell’s equations on a static, globally hyperbolic spacetime with a trivial ${\mathbb{R}}$ bundle on it. Presumably this can be extended to stationary spacetimes, but not beyond that because of the need for a nontrivial group of isometries. The treatment of nontrivial or ${\mathrm{U}}(1)$ bundles should require only straightforward modifications, but one of the lessons of our work is that sometimes there are surprises in store even for topics as well-understood as electromagnetism.
In chapter \[chap:qed\], because of the appearance of negative powers of the Laplacian $\Delta$ (or the twisted Laplacian $L_p$ in the general case) in the process, we will be forced to restrict Fock quantization to the space ${\mathbf{P}}_o$ of oscillating modes of the electromagnetic field. Also, for mathematical convenience one often assumes that $\Delta\ge\epsilon>0$ for some $\epsilon$, which is true when space is compact but not necessarily when it is noncompact. However, we do not do this as one cannot exclude the possibility that the spectrum of $L_p$ or $\Delta$ reach all the way to $0$ because that is the case in physically interesting situations such as Minkowski space.
Coherent-state quantization of linear systems {#chap:linear}
=============================================
In this chapter we present a rigorous framework for quantization of linear dynamics based on the ideas of Irving Segal.
Segal pioneered the idea of of formalizing quantum mechanics in terms of algebras of observables, making Hilbert spaces play the subordinate role of supporting linear representations of them. These Hilbert spaces can, in fact, be constructed from the abstract algebra of observables by means of the Gel’fand–Naĭmark–Segal construction using a single *state* or, in physics parlance, *vacuum expectation*.
Implicit in the work of Segal is a concept of *general boson field* associated to any linear phase space, which formalizes the Heisenberg commutation relations among field operators in terms of exponentiated field operators, using the so-called *Weyl relations*. This has the advantage of avoiding the technicalities of unbounded operators. In addition, physical symmetries are readily implemented as automorphisms of the Weyl algebra.
Segal introduced the related concept of *free boson field*, which can be constructed from a phase space equipped with a compatible complex structure. Segal’s free boson field axiomatizes the properties of the usual of Fock space, and the axiomatic approach makes it transparent that the Fock, Schrödinger and Bargmann–Segal representations of linear quantum fields are all unitarily equivalent. Within this framework, Segal also studied the problem of representing time evolution unitarily on Fock space, and the stability of the generator of unitary time evolution, namely whether the quantum Hamiltonian is bounded below.
Here we put together both ideas, and the result is a new construction of the free boson field based on *coherent states*. In this construction we not only associate to each linear functional on phase space a field operator but, given a choice of vacuum state, we can associate to each point in phase space a coherent state. The collection of all coherent states indexed by points of phase space spans the Hilbert space of quantum states of the theory, and the result is what Segal called the *general boson field*. The free boson field, which as we have mentioned is unitarily equivalent to the Fock representation, is obtained by means of a GNS state with Gaussian statistics.
We find that the mathematical process of quantization can be understood with reference to three physical guiding principles: the canonical commutation relations, the correspondence principle, and the unitary implementation of physical symmetries.
We proceed as follows: we first construct the Weyl algebra of observables associated to a linear phase space, and then choose a compatible complex structure on the Phase space, which amounts to selecting a vacuum expectation on the Weyl algebra, with the help of the correspondence principle and the requirement that time evolution be unitarily and stably implemented.
Coherent states are most useful because many classical equations hold exactly between expectation values on coherent states. Thus, by using coherent states, our quantization procedure never loses sight of the correspondence principle. In addition, the vacuum expectation value acts as a generating function of the matrix elements of field operators between coherent states, not only for ordinary field operators but also for their Wick powers (called normal-ordered operators in physics). As an unexpected bonus, using matrix elements between coherent states one can define normal-ordered Wilson loops as quasioperators without the need for regularization.
Segal’s treatment of the free boson field is presented in [@BSZ]. A comprehensive physical treatment of the coherent states of the electromagnetic field can be found in [@mandel95 Chapter 11].
$p$-form Electromagnetism as a Free Boson Field {#chap:qed}
===============================================
In this chapter we show how the oscillating modes of $p$-form electromagnetism in $(p+1)$-dimensions have a free boson field representation, define certain physically interesting observables as quasioperators, and prove that suitable analogues of the classical equations of motion hold as quasioperator equations.
According to Section \[sec:freeBosonField\], in order to construct a free boson field representation we need a complex Hilbert space ${\mathbf{H}}$ consisting of classical observables of $p$-form electomagnetism. The same vector space with its real structure will be denoted ${\mathbf{P}}^*$ since the space of observables is the dual of the physical phase space ${\mathbf{P}}$. The complex inner product $\langle~,~\rangle$ on ${\mathbf{H}}$ must have as its imaginary part the classical symplectic structure $\omega$ on ${\mathbf{P}}^*$. The free boson field on ${\mathbf{H}}$ is the representation of ${\mathcal{W}}({\mathbf{P}}^*,\omega)$ produced by the GNS construction applied to a state $\langle~\rangle$ with characteristic functional $\mu(f)=\exp(-\|f\|^2/4)$ for all $f\in{\mathbf{H}}$. Now, in order for time evolution to be unitary, it is sufficient that $\mu$ be invariant under time evolution; in other words, $\mu$ and hence $\|~\|$ must be constants of the motion.
However, the analysis of the classical theory produces a real phase space, without a complex structure and not having necessarily even a real Hilbert space structure. That is, the starting point for quantization is a classical phase space ${\mathbf{P}}$ whose dual $({\mathbf{P}}^*,\omega)$ is a a real topological vector space with a continuous symplectic structure $\omega$. Time evolution acts on phase space as a strongly continuous one-parameter group of *bounded* operators $T(t)$ preserving the symplectic structure on ${\mathbf{P}}^*$. To quantize these symplectic dynamics involves constructing from $({\mathbf{P}}^*,\omega)$ and $T$ a complex Hilbert space ${\mathbf{H}}$ on which $T(t)$ is a strongly-continuous one-parameter group of *unitary* operators.
Ordinarily, for instance when quantizing a massive linear field such as the Klein–Gordon field, ${\mathbf{H}}$ would carry a weaker norm than ${\mathbf{P}}^*$, and so ${\mathbf{P}}^*$ would be contained in ${\mathbf{H}}$. However, as we shall see, when there are infrared divergences (as is the case for massless fields such as the Maxwell field) neither ${\mathbf{H}}$ nor ${\mathbf{P}}^*$ contain each other. However, there is a common subspace of both ${\mathbf{P}}^*$ and ${\mathbf{H}}$ on which all the mathematical objects we are discussing are well-defined. This space is constructed as a subspace of ${\mathbf{P}}^*$ in a well-prescribed way and then completed to obtain ${\mathbf{H}}$.
An additional complication is the existence of a nontrivial Aharonov–Bohm sector. We have seen that the dynamics in this sector are analogous to those of a free particle. In the case of the electromagnetic field, we will see that the definition of ${\mathbf{H}}$ involves negative powers of the Laplacian, and so the Aharonov–Bohm sector must be quantized in a different way, if at all. Accordingly, although we set out to quantize ${\mathbf{P}}\simeq{\mathbf{P}}_o\oplus {\mathbf{P}}_f$, we really only achieve a Fock quantization of the oscillating sector ${\mathbf{P}}_o$. We do not attempt to determine whether a free boson field representation of the free modes is possible; we expect this to be the case only when ${\mathbf{P}}_f$ is finite-dimensional. Moreover, only on ${\mathbf{P}}_o$ is it possible to find a (densely-defined) complex structure preserved by time evolution. On the space ${\mathbf{P}}_f$ of Aharonov–Bohm modes, the time evolution operator $T_f(t)$ is a shear, and there is no way to make it unitary.
The plan of this short chapter is as follows. In Section \[sec:freeBosonEM\] we construct the free boson field representation of the oscillating sector of $p$-form electromagnetism. In Section \[sec:punchline\] we use our quasioperator technology from Section \[sec:quasioperators\] to make sense of Wilson loop operators and their higher-dimensional generalizations, as well as electromagnetic field operators at a point, which are then shown to satisfy the Maxwell equations as quasioperator equations. Most importantly, we end with a description of the dynamics of the electromagnetic field in terms of Wilson loops, without any need for ‘regularizing’ or ‘smearing’ these loops as in the work of Varadarajan [@varadarajan00; @varadarajan01].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A cloud radio access network (Cloud-RAN) is a network architecture that holds the promise of meeting the explosive growth of mobile data traffic. In this architecture, all the baseband signal processing is shifted to a single baseband unit (BBU) pool, which enables efficient resource allocation and interference management. Meanwhile, conventional powerful base stations can be replaced by low-cost low-power remote radio heads (RRHs), producing a green and low-cost infrastructure. However, as all the RRHs need to be connected to the BBU pool through optical transport links, the transport network power consumption becomes significant. In this paper, we propose a new framework to design a green Cloud-RAN, which is formulated as a joint RRH selection and power minimization beamforming problem. To efficiently solve this problem, we first propose a greedy selection algorithm, which is shown to provide near-optimal performance. To further reduce the complexity, a novel group sparse beamforming method is proposed by inducing the group-sparsity of beamformers using the weighted $\ell_1/\ell_2$-norm minimization, where the group sparsity pattern indicates those RRHs that can be switched off. Simulation results will show that the proposed algorithms significantly reduce the network power consumption and demonstrate the importance of considering the transport link power consumption.'
author:
- 'Yuanming Shi, Jun Zhang, and Khaled B. Letaief, [^1]'
bibliography:
- '/Reference.bib'
title: 'Group Sparse Beamforming for Green Cloud-RAN'
---
Cloud-RAN, green communication, power consumption, greedy selection, group-sparsity.
Introduction
============
data traffic has been growing enormously in recent years, and it is expected that cellular networks will have to offer a 1000x increase in capacity in the following decade to meet this demand [@Soliman_CM13]. Massive MIMO [@Rusek_SPM2013] and heterogeneous and small cell networks (HetSNets) [@Soliman_CM13] are regarded as two most promising approaches to achieve this goal. By deploying a large number of antennas at each base station (BS), massive MIMO can exploit spatial multiplexing gain in a large-scale and also improve energy efficiency. However, the performance of massive MIMO is limited by correlated scattering with the antenna spacing constraints, which also brings high deployment cost to maintain the minimum spacing [@Soliman_CM13]. HetSNets exploit the spatial reuse by deploying more and more access points (APs). Meanwhile, as stated in [@Debbah_VTM11], placing APs based on the traffic demand is an effective way for compensating path-loss, resulting in energy efficient cellular networks. However, efficient interference management is challenging for dense small-cell networks. Moreover, deploying more and more small-cells will cause significant cost and operating challenges for operators.
Cloud radio access network (Cloud-RAN) has recently been proposed as a promising network architecture to unify the above two technologies in order to jointly manage the interference (via coordinated multiple-point process (CoMP)), increase network capacity and energy efficiency (via network densification), and reduce both the network capital expenditure (CAPEX) and operating expense (OPEX) (by moving baseband processing to the baseband unit (BBU) pool) [@mobile2011c; @Wu_WC2012]. A large-scale distributed cooperative MIMO system will thus be formed. Cloud-RAN can therefore be regarded as the ultimate solution to the “spectrum crunch" problem of cellular networks.
There are three key components in a Cloud-RAN: (i) a pool of BBUs in a datacenter *cloud*, supported by the real-time virtualization and high performance processors, where all the baseband processing is performed; (ii) a high-bandwidth low-latency optical transport network connecting the BBU pool and the remote radio heads (RRHs); and (iii) distributed transmission/reception points (i.e., RRHs). The key feature of Cloud-RAN is that RRHs and BBUs are separated, resulting a centralized BBU pool, which enables efficient cooperation of the transmission/reception among different RRHs. As a result, significant performance improvements through joint scheduling and joint signal processing such as coordinated beamforming or multi-cell processing[@Gesbert_JSAC10] can be achieved. With efficient interference suppression, a network of RRHs with a very high density can be deployed. This will also reduce the communication distance to the mobile terminals and can thus significantly reduce the transmission power. Moreover, as baseband signal processing is shifted to the BBU pool, RRHs only need to support basic transmission/reception functionality, which further reduces their energy consumption and deployment cost.
The new architecture of Cloud-RAN also indicates a paradigm shift in the network design, which causes some technical challenges for implementation. For instance, as the data transmitted between the RRHs and the BBU pool is typically oversampled real-time I/Q digital data streams in the order of Gbps, high-bandwidth optical transport links with low-latency will be needed. To support CoMP and computing resource sharing among BBUs, new virtualization technologies need to be developed to distribute or group the BBUs into a centralized entity [@mobile2011c]. Another important aspect is the energy efficiency consideration, due to the increased power consumption of a large number of RRHs and also of the transport links.
Conventionally, the transport network (i.e., backhaul links between the core network and base stations (BSs)) power consumption can be ignored as it is negligible compared to the power consumption of macro BSs. Therefore, all the previous works investigating the energy efficiency of cellular networks only consider the BS power consumption [@WeiYu_WC10; @Chang_ICC2013]. Recently, the impact of the backhaul power consumption in cellular network was investigated in [@Tombaz_GLOBECOM2011], where it was shown through simulation that the backhaul power consumption will affect the energy efficiency of different cellular network deployment scenarios. Subsequently, Rao *et al* in [@Rao_VT2013] investigated the spectral efficiency and energy efficiency tradeoff in homogeneous cellular networks when taking the backhaul power consumption into consideration.
In Cloud-RAN, the transport network power consumption will have a more significant impact on the network energy efficiency. Hence, allowing the transport links and the corresponding RRHs to support the sleep mode will be essential to reduce the network power consumption for the Cloud-RAN. Moreover, with the spatial and temporal variation of the mobile traffic, it would be feasible to switch off some RRHs while still maintaining the QoS requirements. It will be also practical to implement such idea in the Cloud-RAN with the help of centralized signal processing at the BBU pool. As energy efficiency is one of the major objectives for future cellular networks [@Wu_WC2012], in this paper we will focus on the design of green Cloud-RAN by jointly considering the power consumption of the transport network and RRHs.
Contributions
-------------
The main objective of this paper is to minimize the network power consumption of Cloud-RAN, including the transport network and radio access network power consumption, with a quality of service (QoS) constraint at each user. Specifically, we formulate the design problem as a joint RRH selection and power minimization beamforming problem, where the transport network power consumption is determined by the set of active RRHs, while the transmit power consumption of the active RRHs is minimized through coordinated beamforming. This is a mixed-integer non-linear programming (MINLP) problem, which is NP-hard. We will focus on designing low-complexity algorithms for practical implementation. The major contributions of the paper are summarized as follows:
1. We formulate the network power consumption minimization problem for the Cloud-RAN by enabling both the transport links and RRHs to support the sleep mode. In particular, we provide a group sparse beamforming (GSBF) formulation of the design problem, which assists the problem analysis and algorithm design.
2. We first propose a greedy selection (GS) algorithm, which selects one RRH to switch off at each step. It turns out that the RRH selection rule is critical, and we propose to switch off the RRH that *maximizes the reduction in the network power consumption* at each step. From the simulations, the proposed GS algorithm often yields optimal or near-optimal solutions, but its complexity may still be prohibitive for a large-size network.
3. To further reduce the complexity, we propose a three-stage group sparse beamforming (GSBF) framework, by adopting the weighted mixed $\ell_{1}/\ell_{p}$-norm to induce the group sparsity for the beamformers. In contrast to all the previous works applying the mixed $\ell_{1}/\ell_{p}$-norm to induce group sparsity, we exploit the additional prior information (i.e., transport power consumption, power amplifier efficiency, and instantaneous effective channel gain) to design the weights for different beamformer coefficient groups, resulting in a significant performance gain. Two GSBF algorithms with different complexities are proposed: namely, a bi-section GSBF algorithm and an iterative GSBF algorithm.
4. We shall show that the GS algorithm always provides near-optimal performance. Hence, it would be a good option if the number of RRHs is relatively small, such as in clustered deployment. With a very low computational complexity, the bi-section GSBF algorithm is an attractive option for a large-scale Cloud-RAN. The iterative GSBF algorithm provides a good tradeoff between complexity and performance, which makes it a good candidate for a medium-size network.
Related Works
-------------
A main design tool applied in this paper is optimization with the group sparsity induced norm. With the recent theoretical breakthrough in compressed sensing [@Donoho_TIT2006; @Tao_IT06], the sparsity patterns in different applications in signal processing and communications have been exploited for more efficient system design, e.g., for pilot aided sparse channel estimation [@Berger_CM10]. The sparsity inducing norms have been widely applied in high-dimensional statistics, signal processing, and machine learning in the last decade [@Bach_ML2011]. The $\ell_{1}$-norm regularization has been successfully applied in compressed sensing [@Donoho_TIT2006; @Tao_IT06]. More recently, mixed $\ell_{1}/\ell_{p}$-norms are widely investigated in the case where some variables forming a group will be selected or removed simultaneously, where the mixed $\ell_{1}/\ell_{2}$-norm [@Ming_SM06] and mixed $\ell_{1}/\ell_{\infty}$-norm [@Wainwright_IT2011] are two commonly used ones to induce group sparsity for their computational and analytical convenience.
In Cloud-RAN, one RRH will be switched off only when all the coefficients in its beamformer are set to zeros. In other words, all the coefficients in the beamformer at one RRH should be selected or ignored simultaneously, which requires group sparsity rather than individual sparsity for the coefficients as commonly used in compressed sensing. In this paper, we will adopt the mixed $\ell_{1}/\ell_{p}$-norm to promote group sparsity for the beamformers instead of $\ell_1$-norm, which only promotes individual sparsity. Recently, there are some works [@Z.Q.Luo_JSAC2013; @TonyQ.S._WC2013; @Mehanna_SP2013] adopting the mixed $\ell_{1}/\ell_{p}$-norm to induce group-sparsity in a large-scale cooperative wireless cellular network. Specifically, Hong [*et al.*]{} [@Z.Q.Luo_JSAC2013] adopted the mixed $\ell_{1}/\ell_{2}$-norm and Zhao *et al.* [@TonyQ.S._WC2013] used the $\ell_{2}$-norm to induce the group sparsity of the beamformers, which reduce the amount of the shared user data among different BSs. The squared mixed $\ell_{1}/\ell_{\infty}$-norm was investigated in [@Mehanna_SP2013] for antenna selection.
All of the above works simply adopted the un-weighted mixed $\ell_{1}/\ell_{p}$-norms to induce group-sparsity, in which, no prior information on the unknown signal is assumed other than the fact that it is sufficiently sparse. By exploiting the prior information in terms of system parameters, the weights for different beamformer coefficient groups can be more rigorously designed and performance can be enhanced. We demonstrate through simulations that the proposed three-stage GSBF framework, which is based on the weighted mixed $\ell_{1}/\ell_{p}$-norm minimization, outperforms the conventional unweighted mixed $\ell_{1}/\ell_{p}$-norm minimization based algorithms substantially.
Organization
------------
The remainder of the paper is organized as follows. Section II presents the system and power model. In section III, the network power consumption minimization problem is formulated, followed by some analysis. Section IV presents the GS algorithm, which yields near-optimal solutions. The three-stage GSBF framework is presented in Section V. Simulation results will be presented in Section VI. Finally, conclusions and discussions are presented in Section VII.
*Notations*: $\|\cdot\|_{\ell_{p}}$ is the $\ell_{p}$-norm. Boldface lower case and upper case letters represent vectors and matrices, respectively. $(\cdot)^{T}$, $(\cdot)^{\dag}$, $(\cdot)^{\sf{H}}$ and $\textrm{Tr}(\cdot)$ denote the transpose, conjugate, Hermitian and trace operators, respectively. $\mathfrak{R}(\cdot)$ denotes the real part.
System and Power Model
======================
System Model
------------
We consider a Cloud-RAN with $L$ remote radio heads (RRHs), where the $l$-th RRH is equipped with $N_{l}$ antennas, and $K$ single-antenna mobile users (MUs), as shown in Fig. [\[system\]]{}.
![The architecture of Cloud-RAN, in which, all the RRHs are connected to a BBU pool through transport links.[]{data-label="system"}](system){width="1\columnwidth"}
In this network architecture, all the base band units (BBUs) are moved into a single BBU pool, creating a set of shared processing resources, and enabling efficient interference management and mobility management. With the baseband signal processing functionality migrated to the BBU pool, the RRHs can be deployed in a large scale with low-cost. The BBU pool is connected to the RRHs using the common public radio interface (CPRI) transport technology via a high-bandwidth, low-latency optical transport network [@mobile2011c]. The digitized baseband complex inphase (I) and quadrature (Q) samples of the radio signals are transported over the transport links between the BBUs and RRHs. The key technical and economic issue of the Cloud-RAN is that this architecture requires significant transport network resources. As the focus of this paper is on network power consumption, we will assume all the transport links have sufficiently high capacity and negligible latency[^2].
Due to the high density of RRHs and the joint transmission among them, the energy used for signal transmission will be reduced significantly. However, the power consumption of the transport network becomes enormous and cannot be ignored. Therefore, it is highly desirable to switch off some transport links and the corresponding RRHs to reduce the network power consumption based on the data traffic requirement, which forms the main theme of this work.
Let $\mathcal{L}=\{1,...,L\}$ denote the set of RRH indices, $\mathcal{A}\subseteq\mathcal{L}$ denote the active RRH set, $\mathcal{Z}$ denote the inactive RRH set with $\mathcal{A}\cup\mathcal{Z}=\mathcal{L}$, and $\mathcal{S}=\{1,...,K\}$ denote the index set of scheduled users. In a beamforming design framework, the baseband transmit signals are of the form: $$\begin{aligned}
{\bf{x}}_{l}=\sum\limits_{k=1}^{K}{\bf{w}}_{lk}s_{k}, \forall l\in\mathcal{A},\end{aligned}$$ where $s_{k}$ is a complex scalar denoting the data symbol for user $k$ and ${\bf{w}}_{lk}\in\mathbb{C}^{N_{l}}$ is the beamforming vector at RRH $l$ for user $k$. Without loss of generality, we assume that $E[|s_{k}|^2]=1$ and $s_{k}$’s are independent with each other. The baseband signals ${\bf{x}}_l$’s will be transmitted to the corresponding RRHs, but not the data information $s_{k}$’s [@mobile2011c; @Shamai_TSP2013]. The baseband received signal at user $k$ is given by $$\begin{aligned}
y_{k}=\sum\limits_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{lk}s_{k}+\sum\limits_{i\ne
k}\sum\limits_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{li}s_{i}+z_{k}, k\in\mathcal{S},\end{aligned}$$ where ${\bf{h}}_{kl}\in\mathbb{C}^{N_{l}}$ is the channel vector from RRH $l$ to user $k$, and $z_{k}\sim\mathcal{CN}(0,\sigma_{k}^2)$ is the additive Gaussian noise.
We assume that all the users are employing single user detection (i.e., treating interference as noise), so that they can use the receivers with low-complexity and energy-efficient structure. Moreover, in the low interference region, treating interference as noise can be optimal [@Jafar_IT2008]. The corresponding signal-to-interference-plus-noise ratio (SINR) for user $k$ is hence given by $$\begin{aligned}
{\rm{SINR}}_{k}={{|\sum_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{lk}|^2}\over{{\sum\nolimits_{i\ne
k}|\sum_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{li}|^2}}+\sigma_{k}^2},
\forall k\in\mathcal{S}.\end{aligned}$$ Each RRH has its own transmit power constraint $$\begin{aligned}
\label{powerconstraint1}
\sum\limits_{k=1}^{K}\|{\bf{w}}_{lk}\|_{\ell_{2}}^2\le P_{l}, \forall l\in\mathcal{A}.\end{aligned}$$
Power Model
-----------
The network power model is critical for the investigation of the energy efficiency of Cloud-RAN, which is described as follows.
### RRH Power Consumption Model
We will adopt the following empirical linear model [@Auer_WC2011] for the power consumption of an RRH: $$\begin{aligned}
P_{l}^{\textrm{rrh}} = \left\{ \begin{array}{ll}
P_{a,l}^{\textrm{rrh}}+{1\over{\eta_{l}}} P_{l}^{\textrm{out}}, & \textrm{if $P_{l}^{\textrm{out}}>0$},\\
P_{s,l}^{\textrm{rrh}}, & \textrm{if $P_{l}^{\textrm{out}}=0$}.
\end{array} \right.\end{aligned}$$ where $P_{a,l}^{\rm{rrh}}$ is the active power consumption, which depends on the number of antennas $N_{l}$, $P_{s,l}^{\rm{rrh}}$ is the power consumption in the sleep mode, $P^{\textrm{out}}$ is the transmit power, and $\eta_{l}$ is the drain efficiency of the radio frequency (RF) power amplifier. For the Pico-BS, the typical values are $P_{a,l}^{\textrm{rrh}}=6.8 W$, $P_{s,l}^{\textrm{rrh}}=4.3W$, and $\eta_{l}=4$ [@Auer_WC2011]. Based on this power consumption model, we conclude that it is essential to put the RRHs into sleep if possible.
### Transport Network Power Consumption Model
Although there is no superior solution to meet the low-cost, high-bandwidth, low-latency requirement of transport networks for the Cloud-RAN, the future passive optical network (PON) can provide cost-effective connections between the RRHs and the BBU pool [@Kani_ICP2012]. PON comprises an optical line terminal (OLT) that connects a set of associated optical network units (ONUs) through a single fiber. Implementing a sleep mode in the optical network unit (ONU) has been considered as the most cost-effective and promising power-saving method [@Dhaini_ITN2013] for the PON, but the OLT cannot go into the sleep mode and its power consumption is fixed[@Dhaini_ITN2013]. Hence, the total power consumption of the transport network is given by [@Dhaini_ITN2013] $$\begin{aligned}
P^{\textrm{tn}}=P_{\textrm{olt}}+\sum_{l=1}^{L}P_{l}^{\textrm{tl}},\end{aligned}$$ where $P_{\textrm{olt}}$ is the OLT power consumption, $P_{l}^{\textrm{tl}}=P_{a,l}^{\textrm{tl}}$ and $P_{l}^{\textrm{tl}}=P_{s,l}^{\textrm{tl}}$ denote the power consumed by the ONU $l$ (or the transport link $l$) in the active mode and sleep mode, respectively. The typical values are $P_{\textrm{olt}}=20 W$, $P_{a,l}^{\textrm{tl}}=3.85W$ and $P_{s,l}^{\textrm{tl}}=0.75W$ [@Dhaini_ITN2013]. Thus, we conclude that putting some transport links into the sleep mode is a promising way to reduce the power consumption of Cloud-RAN.
### Network Power Consumption
Based on the above discussion, we can define a network power consumption model for the Cloud-RAN. Define $P_{l}^{a}=P_{a,l}^{\textrm{tl}}+P_{a,l}^{\textrm{tl}}$ $(P_{l}^{s}=P_{s,l}^{\textrm{rrh}}+P_{s,l}^{\textrm{tl}})$ as the active (sleep) power consumption when both the RRH and the corresponding transport link are switched on (off). For convenience, denote $P_{l}^c=P_{l}^a-P_l^s$. In the following, we will omit the constants $P_{\textrm{olt}}$ and $\sum_{l=1}^{L}P_{l}^s$ , which will not affect the system design. Thus, the network power consumption is given by$$\begin{aligned}
\label{power_con1}
p(\mathcal{A},{\bf{w}})=\sum\limits_{l\in\mathcal{A}}\sum\limits_{k=1}^{K}{1\over{\eta_{l}}}\|{\bf{w}}_{lk}\|_{\ell_2}^2+\sum\limits_{l\in\mathcal{A}}P_{l}^{c},\end{aligned}$$ where ${\bf{w}}=[{\bf{w}}_{11}^{T},\dots,{\bf{w}}_{1K}^{T},\dots,{\bf{w}}_{L1}^{T},\dots,{\bf{w}}_{LK}^{T}]^{T}$. In the following discussion, we shall refer to $P_{l}^c$ as the *transport link power consumption* for simplification. Therefore, the first part of (\[power\_con1\]) is the total transmit power consumption and the second part is the total transport network power consumption.
Problem Formulation and Analysis
================================
Based on the power model, we will formulate the network power consumption minimization problem in this section.
Power Saving Strategies and Problem Formulation
-----------------------------------------------
The network power consumption model $(\ref{power_con1})$ indicates the following two strategies to reduce the network power consumption:
- Reduce the transmission power consumption;
- Reduce the number of active RRHs and the corresponding transport links.
However, the two strategies conflict with each other. Specifically, in order to reduce the transmission power consumption, more RRHs are required to be active to exploit a higher beamforming gain. On the other hand, allowing more RRHs to be active will increase the power consumption of transport links. As a result, the network power consumption minimization problem requires a joint design of RRH (and the corresponding transport link) selection and coordinated transmit beamforming.
In this work, we assume perfect channel state information (CSI) available at the BBU pool. With target SINRs ${\boldsymbol{\gamma}}=(\gamma_{1},\dots,\gamma_{K})$, the network power consumption minimization problem can be formulated as $$\begin{aligned}
\label{global1}
\mathscr{P}:
\mathop {\rm{minimize}}_{\{{\bf{w}}_{lk}\}, \mathcal{A}}&& p(\mathcal{A}, {\bf{w}})\nonumber\\
\rm{subject~to}&& {{|\sum_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{lk}|^2}\over{{\sum\nolimits_{i\ne
k}|\sum_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{li}|^2}}+\sigma_{k}^2}\ge\gamma_{k},
\nonumber\\
&&\sum\nolimits_{k=1}^{K}\|{\bf{w}}_{lk}\|_{\ell_2}^2\le P_{l}, l\in\mathcal{A}.\end{aligned}$$ Problem $\mathscr{P}$ is a joint RRH set selection and transmit beamforming problem, which is difficult to solve in general. In the following, we will analyze and reformulate it.
Problem Analysis
----------------
We first consider the case with a given active RRH set $\mathcal{A}$ for problem $\mathscr{P}$, resulting a network power minimization problem $\mathscr{P}(\mathcal{A})$. Let ${\bf{w}}_{k}=[{\bf{w}}_{lk}^{T}]^{T}\in\mathbb{C}^{\sum_{l\in\mathcal{A}}N_{l}}$ indexed by ${l\in\mathcal{A}}$, and ${\bf{h}}_{k}=[{\bf{h}}_{lk}^{T}]^{T}\in\mathbb{C}^{\sum_{l\in\mathcal{A}}N_{l}}$ indexed by ${l\in\mathcal{A}}$, such that ${\bf{h}}_{k}^{\sf{H}}{\bf{w}}_{k}=\sum_{l\in\mathcal{A}}{\bf{h}}_{kl}^{\sf{H}}{\bf{w}}_{lk}$. Since the phases of ${\bf{w}}_{k}$ will not change the objective function and constraints of $\mathscr{P}(\mathcal{A})$ [@Shamai_SP2006], the SINR constraints are equivalent to the following second order cone (SOC) constraints: $$\begin{aligned}
\label{SOC1}
\mathcal{C}_{1}(\mathcal{A}):
\sqrt{\sum\nolimits_{i\ne k}|{\bf{h}}_{k}^{\sf{H}}{\bf{w}}_{i}|^2+\sigma_{k}^2}\le {1\over{\sqrt{\gamma_{k}}}}\mathfrak{R}({\bf{h}}_{k}^{\sf{H}}{\bf{w}}_{k}), k\in\mathcal{S}. \end{aligned}$$ The per-RRH power constraints (\[powerconstraint1\]) can be rewritten as $$\begin{aligned}
\label{SOC2}
\mathcal{C}_{2}(\mathcal{A}):
\sqrt{\sum\nolimits_{k=1}^{K}\|{\bf{A}}_{lk}{\bf{w}}_{k}\|_{\ell_2}^2}\le \sqrt{P_{l}}, l\in\mathcal{A},\end{aligned}$$ where ${\bf{A}}_{lk}\in\mathbb{C}^{\sum_{l\in\mathcal{A}}N_{l}\times \sum_{l\in\mathcal{A}}N_{l}}$ is a block diagonal matrix with the identity matrix ${\bf{I}}_{N_{l}}$ as the $l$-th main diagonal block square matrix and zeros elsewhere. Therefore, given the active RRH set $\mathcal{A}$, the network power minimization problem is given by $$\begin{aligned}
\label{op1}
\mathscr{P}(\mathcal{A}): \mathop {\rm{minimize}}\limits_{{\bf{w}}_{1},\dots, {\bf{w}}_{K}} &&\sum_{l\in\mathcal{A}}\left(\sum_{k=1}^{K}{1\over{\eta_{l}}}\|{\bf{A}}_{lk}{\bf{w}}_{k}\|_{\ell_2}^2+P_{l}^c\right)\nonumber\\
{\rm{subject~to}} && \mathcal{C}_{1}(\mathcal{A}), \mathcal{C}_{2}(\mathcal{A}), \end{aligned}$$ with the optimal value denoted as $p^{\star}(\mathcal{A})$. This is a second-order cone programming (SOCP) problem, and can be solved efficiently, e.g., via interior point methods [@boyd2004convex].
Based on the solution of $\mathscr{P}(\mathcal{A})$, the network power minimization problem $\mathscr{P}$ can be solved by searching over all the possible RRH sets, i.e., $$\begin{aligned}
\label{exhaustive1}
p^{\star}=\mathop{\rm{minimize}}_{Q\in\{J, \dots, L\}}~p^{\star}(Q),\end{aligned}$$ where $J\ge1$ is the minimum number of RRHs that makes the network support the QoS requirements, and $p^{\star}(Q)$ is determined by $$\begin{aligned}
\label{greedy_1}
p^{\star}(Q)=\mathop {\rm{minimize}}_{\mathcal{A}\subseteq\mathcal{L}, |\mathcal{A}|=Q}~p^{\star}({\mathcal{A}}),\end{aligned}$$ where $p^{\star}(\mathcal{A})$ is the optimal value of the problem $\mathscr{P}(\mathcal{A})$ (\[op1\]) and $|\mathcal{A}|$ is the cardinality of set $\mathcal{A}$. The number of subsets $\mathcal{A}$ of size $m$ is $L \choose m$, which can be very large. Thus, in general, the overall procedure will be exponential in the number of RRHs $L$ and thus cannot be applied in practice. Therefore, we will reformulate this problem to develop more efficient algorithms to solve it.
Group Sparse Beamforming Formulation
------------------------------------
One way to solve problem $\mathscr{P}$ is to reformulate it as a MINLP problem [@Yuanming_Globecom2013], and the generic algorithms for solving MINLP can be applied. Unfortunately, due to the high complexity, such an approach can only provide a performance benchmark for a simple network setting. In the following, we will pursue a different approach, and try to exploit the problem structure.
We will exploit the group sparsity of the optimal aggregative beamforming vector ${\bf{w}}$, which can be written as a partition: $$\begin{aligned}
\label{vector1}
{\bf{w}}=[\underbrace{{\bf{w}}_{11}^{T},\dots,{\bf{w}}_{1K}^{T}}_{\tilde{\bf{w}}_{1}^{T}},\dots,\underbrace{{\bf{w}}_{L1}^{T},\dots,{\bf{w}}_{LK}^{T}}_{\tilde{\bf{w}}_{L}^{T}}]^{T},\end{aligned}$$ where all the coefficients in a given vector $\tilde{\bf{w}}_{l}=[{\bf{w}}_{l1}^{T},\dots,{\bf{w}}_{lK}^{T}]^{T}\in\mathbb{C}^{KN_{l}}$ form a group. When the RRH $l$ is switched off, the corresponding coefficients in the vector $\tilde{\bf{w}}_{l}$ will be set to zeros simultaneously. Overall there may be multiple RRHs switched off and the corresponding beamforming vectors will be set to zeros. That is, ${\bf{w}}$ has a group sparsity structure, with the priori knowledge that the blocks of variables in $\tilde{\bf{w}}_{l}$’s should be selected (the corresponding RRH will be switched on) or ignored (the corresponding RRH will be switched off) simultaneously.
Define $N=K\sum_{l=1}^{L}N_{l}$ and an index set $\mathcal{V}=\{1,2,\dots,N\}$ with its power-set as $2^{\mathcal{V}}=\{\mathcal{I}, \mathcal{I}\subseteq\mathcal{V}\}$. Furthermore, define the sets $\mathcal{G}_{l}=\{K\sum_{i=1}^{l-1}N_{i}+1,\dots,K\sum_{i=1}^{l}N_{i}\}, l=1,\dots, L$, as a partition of $\mathcal{V}$, such that $\tilde{\bf{w}}_{l}=[w_{i}]$ is indexed by $i\in\mathcal{G}_{l}$. Define the support of beamformer $\bf{w}$ as $$\begin{aligned}
\mathcal{T}({\bf{w}})=\{i, w_{i}\ne0\},\end{aligned}$$ where ${\bf{w}}=[w_{i}]$ is indexed by $i\in\mathcal{V}$. Hence, the transport link power consumption can be written as $$\begin{aligned}
\label{transportnetwork1}
F(\mathcal{T}({\bf{w}}))=\sum\limits_{l=1}^{L}P_{l}^c I(\mathcal{T}({\bf{w}})\cap\mathcal{G}_{l}\ne\emptyset),\end{aligned}$$ where $I(\mathcal{T}\cap\mathcal{G}_{l}\ne\emptyset)$ is an indicator function that takes value 1 if $\mathcal{T}\cap\mathcal{G}_{l}\ne\emptyset$ and 0 otherwise. Therefore, the network power minimization problem $\mathscr{P}$ is equivalent to the following group sparse beamforming (GSBF) formulation $$\begin{aligned}
\label{networkpower1}
\mathscr{P}_{\textrm{sparse}}:
\mathop {\rm{minimize}}_{\bf{w}}\!\!\!\!\!\!\!&&T({\bf{w}})+F(\mathcal{T}(
{\bf{w}}))\nonumber\\
{\rm{subject~ to}}\!\!\!\!\!\!\! && \mathcal{C}_{1}(\mathcal{L}), \mathcal{C}_{2}(\mathcal{L}),\end{aligned}$$ where $T({\bf{w}})=\sum_{l=1}^{L}\sum_{k=1}^{K}{1\over{\eta_{l}}}\|{\bf{w}}_{lk}\|_{\ell_2}^2$ represents the total transmit power consumption. The equivalence means that if ${\bf{w}}^{\star}$ is a solution to $\mathscr{P}_{\textrm{sparse}}$, then $(\{{\bf{w}}_{lk}^{\star}\}, \mathcal{A}^{\star})$ with $\mathcal{A}^{\star}=\{l:\mathcal{T}({\bf{w}}^{\star})\cap\mathcal{G}_{l}\ne\emptyset\}$ is a solution to $\mathscr{P}$, and vice versa.
Note that the group sparsity of ${\bf{w}}$ is fundamentally different from the conventional sparsity measured by the $\ell_{0}$-norm of ${\bf{w}}$, which is often used in compressed sensing [@Tao_IT06; @Donoho_TIT2006]. The reason is that although the $\ell_{0}$-norm of ${\bf{w}}$ will result in a sparse solution for ${\bf{w}}$, the zero entries of ${\bf{w}}$ will not necessarily align to a same group $\tilde{\bf{w}}_{l}$ to lead to switch off one RRH. As a result, the conventional $\ell_{1}$-norm relaxation [@Tao_IT06; @Donoho_TIT2006] to the $\ell_{0}$-norm will not work for our problem due to the group sparsity of ${\bf{w}}$. Therefore, we will adopt the mixed $\ell_{1}/\ell_{p}$-norm [@Bach_ML2011] to induce group sparsity for ${\bf{w}}$. The details will be presented in Section V.
Since obtaining the global optimization solutions to problem $\mathscr{P}$ is computationally difficult, in the following sections, we will propose two low-complexity algorithms to solve it. We will first propose a greedy algorithm in Section IV, which can be viewed as an approximation to the iteration procedure of (\[exhaustive1\]). In order to further reduce the complexity, based on the GSBF formulation $\mathscr{P}_{\textrm{sparse}}$, a three-stage GSBF framework will then be developed based on the group-sparsity inducing norm minimization in Section V.
Greedy Selection Algorithm
==========================
In this section, we develop a heuristic algorithm to solve $\mathscr{P}$ based on the backward greedy selection, which was successfully applied in spare filter design [@Oppenheim_SP2010] and has been shown to often yield optimal or near-optimal solutions. The backward greedy selection algorithm iteratively selects one RRH to switch off at each step, while re-optimizing the coordinated transmit beamforming for the remaining active RRH set. The key design element for this algorithm is the selection rule of the RRHs to determine which one should be switched off at each step.
Greedy Selection Procedure
--------------------------
Denote the iteration number as $i=0,1,2,\dots$. At the $i$th iteration, $\mathcal{A}^{[i]}\subseteq\mathcal{L}$ shall denote the set of active RRHs, and $\mathcal{Z}^{[i]}$ denotes the inactive RRH set with $\mathcal{Z}^{[i]}\cup\mathcal{A}^{[i]}=\mathcal{L}$. In iteration $i$, an additional RRH $r^{[i]}\in\mathcal{A}^{[i]}$ will be added to $\mathcal{Z}^{[i]}$, resulting in a new set $\mathcal{Z}^{[i+1]}=\mathcal{Z}^{[i]}\cup\{r^{[i]}\}$ after this iteration. We initialize by setting $\mathcal{Z}^{[0]}=\emptyset$. In our algorithm, once an RRH is added to the set $\mathcal{Z}$, it cannot be removed. This procedure is a simplification of the exact search method described in Section III-B. At iteration $i$, we need to solve the network power minimization problem $\mathscr{P}(\mathcal{A}^{[i]})$ (\[op1\]) with the given active RRH set $\mathcal{A}^{[i]}$.
### RRH Selection Rule
How to select $r^{[i]}$ at the $i$th iteration is critical for the performance of the greedy selection algorithm. Based on our objective, we propose to select $r^{[i]}$ to maximize the decrease in the network power consumption. Specifically, at iteration $i$, we obtain the network power consumption $p^{\star}(\mathcal{A}_{m}^{[i]})$ with $\mathcal{A}_{m}^{[i]}\cup\{m\}=\mathcal{A}^{[i]}$ by removing any $m\in\mathcal{A}^{[i]}$ from the active RRH set $\mathcal{A}^{[i]}$. Thereafter, $r^{[i]}$ is chosen to yield the smallest network power consumption after switching off the corresponding RRH, i.e., $$\begin{aligned}
\label{selection1}
r^{[i]}=\arg\min_{m\in\mathcal{A}^{[i]}} p^{\star}(\mathcal{A}_{m}^{[i]}).\end{aligned}$$ We assume that $p^{\star}(\mathcal{A}_{m}^{[i]})=+\infty$ if problem $\mathscr{P}(\mathcal{A}_{m}^{[i]})$ is infeasible. The impact of switching off one RRH is reducing the transport network power consumption while increasing the total transmit power consumption. Thus, the proposed selection rule actually aims at minimizing the impact of turning off one RRH at each iteration.
Denote $\mathcal{J}$ as the set of candidate RRHs that can be turned off, the greedy selection algorithm is described as follows:
**Step 0:** Initialize $\mathcal{Z}^{[0]}=\emptyset$, $\mathcal{A}^{[0]}=\{1,\dots,L\}$ and $i=0$;\
**Step 1:** Solve the optimization problem $\mathscr{P}(\mathcal{A}^{[i]})$ (\[op1\]);\
1. [**[If]{}**]{} (\[op1\]) is feasible, obtain $p^{\star}(\mathcal{A}^{[i]})$;\
- [**[If]{}**]{} $\forall m\in\mathcal{A}^{[i]}$, problem $\mathscr{P}(\mathcal{A}_{m}^{[i]})$ is infeasible, obtain $\mathcal{J}=\{0,\dots, i\}$, [**go to Step 2**]{};\
- [**[If]{}**]{} $\exists m\in\mathcal{A}^{[i]}$, make problem $\mathscr{P}(\mathcal{A}_{m}^{[i]})$ feasible, find the $r^{[i]}$ according to (\[selection1\]) and update the set $\mathcal{Z}^{[i+1]}=\mathcal{Z}^{[i]}\cup\{r^{[i]}\}$ and the iteration number $i\leftarrow i+1$, **go to Step 1**;
2. **If** (\[op1\]) is infeasible, when $i=0$, $p^{\star}=\infty$, **go to End**; when $i>0$, obtain $\mathcal{J}=\{0,1, \dots, i-1\}$,\
**go to Step 2**;
**Step 2:** Obtain the optimal active RRH set $\mathcal{A}^{[j^{\star}]}$ with $j^{\star}=\arg\min_{j\in\mathcal{J}} p^{\star}(\mathcal{A}^{[j]})$ and the transmit beamformers minimizing $\mathscr{P}(\mathcal{A}^{[j^{\star}]})$;\
**End**
Complexity Analysis
-------------------
At the $i$-th iteration, we need to solve $|\mathcal{A}^{[i]}|$ SCOP problems $\mathscr{P}({\mathcal{A}}_{m}^{[i]})$ by removing the RRH $m$ from the set $\mathcal{A}^{[i]}$ to determine which RRH should be selected. For each of the SOCP problem $\mathscr{P}(\mathcal{A})$, using the interior-point method, the computational complexity is $\mathcal{O}((K\sum_{l\in\mathcal{A}}N_{l})^{3.5})$ [@boyd2004convex]. The total number of iterations is bounded by $L$. As a result, the total number of SOCP problems required to be solved grows *quadratically* with $L$. Although this reduces the computational complexity significantly compared with the mixed-integer conic programming based algorithms in [@leyffer_2012mixed] and [@Cheng_SP2013], the complexity is still prohibitive for large-scale networks. Therefore, in the next section we will propose a group sparse beamforming framework to further reduce the complexity.
Group Sparse Beamforming Framework
==================================
In this section, we will develop two low-complexity algorithms based on the GSBF formulation $\mathscr{P}_{\textrm{sparse}}$, namely a bi-section GSBF algorithm and an iterative GSBF algorithm, for which, the overall number of SOCP problems to solve grows *logarithmically* and *linearly* with $L$, respectively. The main motivation is to induce group sparsity in the beamformer, which corresponds to switching off RRHs.
In the bi-section GSBF algorithm, we will minimize the [*weighted*]{} mixed $\ell_{1}/\ell_{2}$-norm to induce group-sparsity for the beamformer. By exploiting the additional prior information (i.e., power amplifier efficiency, transport link power consumption, channel power gain) that available in our setting, the proposed bi-section GSBF algorithm will be demonstrated through rigorous analysis and simulations to outperform the conventional *unweighted* mixed $\ell_{1}/\ell_{p}$-norm minimization substantially[@Z.Q.Luo_JSAC2013; @TonyQ.S._WC2013; @Mehanna_SP2013]. By minimizing the *re-weighted* mixed $\ell_{1}/\ell_{2}$-norm iteratively to enhance the group sparsity for the beamformer, the proposed iterative GSBF algorithm will further improve the performance.
The proposed GSBF framework is a three-stage approach, as shown in Fig. [\[GSBF\]]{}. Specifically, in the first stage, we minimize a weighted (or re-weighted) group-sparsity inducing norm to induce the group-sparsity in the beamformer. In the second stage, we propose an ordering rule to determine the priority for the RRHs that should be switched off, based on not only the (approximately) sparse beamformer obtained in the first stage, but also some key system parameters. Following the ordering rule, a selection procedure is performed to determine the optimal active RRH set, followed by the coordinated beamforming. The details will be presented in the following subsections.
![A Three-Stage GSBF Framework.[]{data-label="GSBF"}](GSBF){width="1\columnwidth"}
Preliminaries on Group-Sparsity Inducing Norms
----------------------------------------------
The mixed $\ell_{1}/\ell_{p}$-norm has recently received lots of attention and is shown to be effective to induce group sparsity [@Bach_ML2011], which is defined as follows:
Consider the vector ${\bf{w}}=[{\bf{w}}_{lk}]$ indexed by $l\in\mathcal{L}$ and $k\in\mathcal{S}$ as define in (\[vector1\]). Its mixed $\ell_{1}/\ell_{p}$-norm is defined as follows: $$\begin{aligned}
\label{mixed1}
\mathcal{R}({\bf{w}})=\sum\limits_{l=1}^{L}\beta_{l}\|\tilde{\bf{w}}_{l}\|_{\ell_{p}},
~p>1,\end{aligned}$$
where $\beta_{1}, \beta_{2},\dots, \beta_{L}$ are positive weights.
Define the vector ${\bf{r}}= [\|\tilde{\bf{w}}_{1}\|_{\ell_p}, \dots, \|\tilde{\bf{w}}_{L}\|_{\ell_p}]^{T}$, then the mixed $\ell_{1}/\ell_{p}$-norm behaves as the $\ell_{1}$-norm on the vector ${\bf{r}}$, and therefore, inducing group sparsity (i.e., each vector $\tilde{\bf{w}}_{l}$ is encouraged to be set to zero) for ${\bf{w}}$. Note that, within the group $\tilde{\bf{w}}_{l}$, the $\ell_{p}$-norm does not promote sparsity as $p>1$. By setting $p=1$, the mixed $\ell_{1}/\ell_{p}$-norm becomes a weighted $\ell_{1}$-norm, which will not promote group sparsity. The mixed $\ell_{1}/\ell_{2}$-norm and $\ell_{1}/\ell_{\infty}$-norm are two commonly used norms for inducing group sparsity. For instance, the mixed $\ell_{1}/\ell_{2}$-norm is used with the name *group least-absolute selection* and *shrinkage operator* (or *Group-Lasso*) in machine learning [@Ming_SM06]. In high dimensional statistics, the mixed $\ell_{1}/\ell_{\infty}$-norm is adopted as a regularizer in the linear regression problems with sparsity constraints for its computational convenience [@Wainwright_IT2011].
Bi-Section GSBF Algorithm
-------------------------
In this section, we propose a binary search based GSBF algorithm, in which, the overall number of SOCP problems required to be solved grows logarithmically with $L$, instead of quadratically for the GS algorithm.
### Group-Sparsity Inducing Norm Minimization
With the combinatorial function $F(\cdot)$ in the objective function $p({\bf{w}})=T({\bf{w}})+F(\mathcal{T}(\bf{w}))$, the problem $\mathscr{P}_{\textrm{sparse}}$ becomes computationally intractable. Therefore, we first construct an appropriate convex relaxation for the objective function $p(\bf{w})$ as a surrogate objective function, resulting a weighted mixed $\ell_{1}/\ell_{2}$-norm minimization problem to induce group sparsity for the beamformer. Specifically, we first derive its tightest positively homogeneous lower bound $p_{h}({\bf{w}})$, which has the property $p_{h}(\lambda {\bf{w}})=\lambda p_{h}({\bf{w}}), 0<\lambda< \infty$. Since $p_{h}(\bf{w})$ is still not convex, we further calculate its Fenchel-Legendre biconjugate $p_{h}^{**}({\bf{w}})$ to provide a tightest convex lower bound for $p_{h}(\bf{w})$. We call $p_{h}^{**}({\bf{w}})$ as the *convex positively homogeneous lower bound* (the details can be found in [@Obozinski_arXiv2012]) of function $p(\bf{w})$, which is provided in the following theorem:
The tightest convex positively homogeneous lower bound of the objective function in $\mathscr{P}_{\textrm{sparse}}$, denoted as $p({\bf{w}})$, is given by $$\begin{aligned}
\label{plower}
\Omega({\bf{w}})=2\sum_{l=1}^{L}\sqrt{P_{l}^c\over{\eta_{l}}}\|\tilde{\bf{w}}_{l}\|_{\ell_2}.\end{aligned}$$
Please refer to Appendix A.
This theorem indicates that the group-sparsity inducing norm (i.e., the weighted mixed $\ell_{1}/\ell_{2}$-norm) can provide a convex relaxation for the objective function $p(\bf{w})$. Furthermore, it encapsulates the additionally prior information in terms of system parameters into the weights for the groups. Intuitively, the weights indicate that the RRHs with a higher transport link power consumption and lower power amplifier efficiency will have a higher chance being forced to be switched off. Using the weighted mixed $\ell_{1}/\ell_{2}$-norm as a surrogate for the objective function, we minimize the weighted mixed $\ell_{1}/\ell_{2}$-norm $\Omega({\bf{w}})$ to induce the group-sparsity for the beamformer ${\bf{w}}$: $$\begin{aligned}
\label{sp1}
\mathscr{P}_{\textrm{GSBF}}: \mathop {\rm{minimize}}_{\bf{w}} \!\!\!\!\!\!\!&&\Omega({\bf{w}})\nonumber\\
{\rm{subject~ to}}\!\!\!\!\!\!\! && \mathcal{C}_{1}(\mathcal{L}), \mathcal{C}_{2}(\mathcal{L}),\end{aligned}$$ which is an SOCP problem and can be solved efficiently.
### RRH Ordering
After obtaining the (approximately) sparse beamformer $\hat{{\bf{w}}}$ via solving the weighted group-sparsity inducing norm minimization problem $\mathscr{P}_{\textrm{GSBF}}$, the next question is how to determine the active RRH set. We will first give priorities to different RRHs, so that an RRH with a higher priority should be switched off before the one with a lower priority. Most previous works [@Z.Q.Luo_JSAC2013; @TonyQ.S._WC2013; @Mehanna_SP2013] applying the idea of group-sparsity inducing norm minimization directly map the sparsity to their application, e.g., in [@Mehanna_SP2013], the transmit antennas corresponding to the smaller coefficients in the group (measured by the $\ell_{\infty}$-norm) will have a higher priority to be switched off. In our setting, one might be tempted to give a higher priority for an RRH $l$ with a smaller coefficient $r_{l}=(\sum_{k=1}^{K}\|\hat{\bf{w}}_{lk}\|_{\ell_{2}}^2)^{1/2}$, as the RRH $l$ with a smaller coefficient $r_{l}$ may provide a lower beamforming gain and should be encouraged to be turned off. It turns out that such ordering rule is not a good option and will bring performance degradation.
To get a better performance, the priority of the RRHs should be determined by not only the beamforming gain but also other key system parameters that indicate the impact of the RRHs on the network performance. In particular, the channel power gain $\kappa_{l}=\sum\nolimits_{k=1}^{K}\|{\bf{h}}_{kl}\|_{\ell_2}^2$ should be taken into consideration. Specifically, by the broadcast channel (BC)-multiple-access channel (MAC) duality [@Goldsmith_TIT2003], we have the sum capacity of the Cloud-RAN as: $$\begin{aligned}
C_{\textrm{sum}}=\log\det({\bf{I}}_{N}+{\sf{snr}}\sum_{k=1}^{K}{\bf{h}}_{k}{\bf{h}}_{k}^{\sf{H}}),\end{aligned}$$ where we assume equal power allocation to simplify the analysis, i.e., ${\sf{snr}}=P/\sigma^2, \forall k=1,\dots,K$. One way to upper-bound $C_{\textrm{sum}}$ is through upper-bounding the capacity by the total receive SNR, i.e., using the following relation $$\begin{aligned}
\log\det({\bf{I}}_{N}+{\sf{snr}}\sum_{k=1}^{K}{\bf{h}}_{k}{\bf{h}}_{k}^{\sf{H}})&\le&{\rm{Tr}}({\sf{snr}}\sum_{k=1}^{K}{\bf{h}}_{k}{\bf{h}}_{k}^{\sf{H}})\nonumber\\
&=&{\sf{snr}}\sum_{l=1}^{L}\kappa_{l},\end{aligned}$$ which relies on the inequality $\log(1+x)\le x$. Therefore, from the capacity perspective, the RRH with a higher channel power gain $\kappa_{l}$ contributes more to the sum capacity, i.e., it provides a higher power gain and should not be encouraged to be switched off.
Therefore, different from the previous democratic assumptions (e.g., [@Z.Q.Luo_JSAC2013; @TonyQ.S._WC2013; @Mehanna_SP2013]) on the mapping between the sparsity and their applications directly, we exploit the prior information in terms of system parameters to refine the mapping on the group-sparsity. Specifically, considering the key system parameters, we propose the following ordering criterion to determine which RRHs should be switched off, i.e., $$\begin{aligned}
\label{sparse_selection1}
\theta_{l}:=\sqrt{{\kappa_{l}\eta_{l}}\over{P_{l}^c}}\left(\sum_{k=1}^{K}\|\hat{\bf{w}}_{lk}\|_{\ell_{2}}\right)^{1/2}, \forall l=1,\dots,L,\end{aligned}$$ where the RRH with a smaller $\theta_{l}$ will have a higher priority to be switched off. This ordering rule indicates that the RRH with a lower beamforming gain, lower channel power gain, lower power amplifier efficiency, and higher transport link power consumption should have a higher priority to be switched off. The proposed ordering rule will be demonstrated to improve the performance of the GSBF algorithm significantly through simulations.
### Binary Search Procedure
Based on the ordering rule (\[sparse\_selection1\]), we sort the coefficients in the ascending order: $\theta_{\pi_{1}}\le\theta_{\pi_{2}}\le\cdots\le\theta_{\pi_{L}}$ to fix the final active RRH set. We set the first $J$ smallest coefficients to zero, as a result, the corresponding RRHs will be turned. Denote $J_{0}$ as the maximum number of RRHs that can be turned off, i.e., the problem $\mathscr{P}(\mathcal{A}^{[i]})$ is infeasible if $i> J_{0}$, where $\mathcal{A}^{[i]}\cup\mathcal{Z}^{[i]}=\mathcal{L}$ with $\mathcal{Z}^{[i]}=\{{\pi_{0}},{\pi_{1}},\dots,{\pi_{i}}\}$ and $\pi_{0}=\emptyset$. A binary search procedure can be adopted to determine $J_{0}$, which only needs to solve no more than $1+\lceil\log(1+L)\rceil$ SOCP problems. In this algorithm, we regard $\mathcal{A}^{[J_{0}]}$ as the final active RRH set and the solution of $\mathscr{P}(\mathcal{A}^{[J_{0}]})$ is the final transmit beamformer.
Therefore, the bi-section GSBF algorithm is presented as follows:
**Step 0:** Solve the weighted group-sparsity inducing norm minimization problem $\mathscr{P}_{\textrm{GSBF}}$;\
1. [**If**]{} it is infeasible, set $p^{\star}=\infty$, [**go to End**]{};
2. [**If**]{} it is feasible, obtain the solution $\hat{\bf{w}}$, calculate ordering criterion (\[sparse\_selection1\]), and sort them in the\
ascending order: $\theta_{\pi_{1}}\le\dots\le\theta_{\pi_{L}}$, [**go to Step 1**]{};
**Step 1:** Initialize $J_{\textrm{low}}=0$, $J_{\textrm{up}}=L$, $i=0$;\
**Step 2:** Repeat
1. Set $i\leftarrow\lfloor{{J_{\textrm{low}}+J_{\textrm{up}}}\over{2}}\rfloor$;\
2. Solve the optimization problem $\mathscr{P}(\mathcal{A}^{[i]})$ (\[op1\]): if it is infeasible, set $ J_{\textrm{low}}=i$; otherwise, set $J_{\textrm{up}}=i$;
**Step 3:** Until $J_{\textrm{up}}-J_{\textrm{low}}=1$, obtain $J_{0}=J_{\textrm{low}}$ and obtain the optimal active RRH set $\mathcal{A}^{\star}$ with $\mathcal{A}^{\star}\cup\mathcal{J}=\mathcal{L}$ and $\mathcal{J}=\{{\pi_{1}},\dots, {\pi_{J_{0}}}\}$;\
**Step 4:** Solve the problem $\mathscr{P}(\mathcal{A}^{\star})$, obtain the minimum network power consumption and the corresponding transmit beamformers;\
**End**
Iterative GSBF Algorithm
------------------------
Under the GSBF framework, the main task of the first two stages is to order the RRHs according to the criterion (\[sparse\_selection1\]), which depends on the sparse solution to $\mathscr{P}_{\textrm{GSBF}}$, i.e., $\{\hat{\bf{w}}_{lk}\}$. However, when the minimum of $r_{l}=(\sum_{k=1}^{K}\|\hat{\bf{w}}_{lk}\|_{\ell_{2}}^2)^{1/2}>0$ is not close to zero, it will introduce large bias in estimating which RRHs can be switched off. To resolve this issue, we will apply the idea from the majorization-minimization (MM) algorithm [@hunter2004tutorial] (please refer to appendix B for details on this algorithm), to enhance group-sparsity for the beamformer to better estimate which RRHs can be switched off.
The MM algorithms have been successfully applied in the re-weighted $\ell_{1}$-norm (or mixed $\ell_{1}/\ell_{2}$-norm) minimization problem to enhance sparsity [@Boyd_2008enhancing; @TonyQ.S._WC2013; @Mehanna_SP2013]. However, these algorithms failed to exploit the additional system prior information to improve the performance. Specifically, they used the un-weighted $\ell_{1}$-norm (or mixed $\ell_{1}/\ell_{p}$-norm) minimization as the start point of the iterative algorithms and re-weighted the $\ell_{1}$-norm (or mixed $\ell_{1}/\ell_{p}$-norm) only using the estimate of the coefficients obtained in the last minimization step. Different from the above conventional re-weighted algorithms, we exploit the additionally system prior information at each step (including the start step) to improve the estimation on the group-sparsity of the beamformer.
### Re-weighted Group-Sparsity Inducing Norm Minimization
One way to enhance the group-sparsity compared with using the weighted mixed $\ell_{1}/\ell_{2}$ norm $\Omega({\bf{w}})$ in (\[plower\]) is to minimize the following combinatorial function directly: $$\begin{aligned}
\label{enhancesparse}
\mathcal{R}({\bf{w}})=2\sum_{l=1}^{L}\sqrt{P_{l}^c\over{\eta_{l}}}I(\|\tilde{{\bf{w}}}_{l}\|_{\ell_2}>0),\end{aligned}$$ for which the convex function $\Omega({\bf{w}})$ in (\[plower\]) can be regarded as an $\ell_{1}$-norm relaxation. Unfortunately, minimizing $\mathcal{R}(\bf{w})$ will lead to a non-convex optimization problem. In this subsection, we will provide a sub-optimal algorithm to solve (25) by adopting the idea from the MM algorithm to enhance sparsity.
Based on the following fact in [@sriperumbudur2011majorization] $$\begin{aligned}
\lim_{\epsilon\rightarrow0}{{{\rm{log}}(1+x\epsilon^{-1})}\over{\rm{log}}(1+\epsilon^{-1})}
= \left\{ \begin{array}{ll}
0 & \textrm{if $x=0$},\\
1 & \textrm{if $x>0$},
\end{array} \right.\end{aligned}$$ we rewrite the indicator function in (\[enhancesparse\]) as $$\begin{aligned}
\label{mm_power}
I(\|\tilde{{\bf{w}}_{l}}\|_{\ell_2}>0)=\lim_{\epsilon\rightarrow
0}{{{\rm{log}}(1+\|\tilde{\bf{w}}_{l}\|_{\ell_{2}}\epsilon^{-1})}\over{{\rm{log}}(1+\epsilon^{-1})}}, \forall l\in\mathcal{L}.\end{aligned}$$ The surrogate objective function $\mathcal{R}({\bf{w}})$ can then be approximated as $$\begin{aligned}
\label{mmo}
f({\bf{w}})=\lambda_{\epsilon}\sum_{l=1}^{L}\sqrt{P_{l}^{c}\over{\eta_{l}}}{\rm{log}}(1+\|\tilde{\bf{w}}_{l}\|_{\ell_{2}}\epsilon^{-1}),\end{aligned}$$ by neglecting the limit in (\[mm\_power\]) and choosing an appropriate $\epsilon>0$, where $\lambda_{\epsilon}={2\over{\rm{log}}(1+\epsilon^{-1})}$. Compared with $\Omega({\bf{w}})$ in (\[plower\]), the log-sum penalty function $f({\bf{w}})$ has the potential to be much more sparsity-encouraging. The detail explanations can be found in [@Boyd_2008enhancing].
Since ${\textrm{log}}(1+x),
x\ge0$, is a concave function, we can construct a majorization function for $f$ by the first-order approximation of $\textrm{log}(1+\|\tilde{\bf{w}}_{l}\|_{\ell_{2}}\epsilon^{-1})$, i.e., $$\begin{aligned}
f({\bf{w}})\le \lambda_{\epsilon} \sum_{l=1}^{L}\sqrt{P_{l}^{c}\over{\eta_{l}}}\left(\underbrace{{{\|\tilde{\bf{w}}_{l}\|_{\ell_{2}}}\over{\|\tilde{\bf{w}}_{l}^{[m]}\|_{\ell_{2}}+\epsilon}}+c({\bf{w}}^{[m]})}_{g({\bf{w}}|{\bf{w}}^{[m]})}\right),\end{aligned}$$ where ${\bf{w}}^{[m]}$ is the minimizer at the $(m-1)$-th iteration, and $c({\bf{w}}^{[m]})={\rm{log}}(1\!+\!\|\tilde{\bf{w}}_{l}^{[m]}\|_{\ell_{2}})\!-\!{{\|\tilde{\bf{w}}_{l}^{[m]}\|_{\ell_2}}/({\|\tilde{\bf{w}}_{l}^{[m]}\|_{\ell_{2}}\!+\!\epsilon}})$ is a constant provided that ${\bf{w}}^{[m]}$ is already known at the current $m$-th iteration.
By omitting the constant part of $g({\bf{w}}|{\bf{w}}^{[m]})$ at the $m$-th iteration, which will not affect the solution, we propose a re-weighted GSBF framework to enhance the group-sparsity: $$\begin{aligned}
\label{wGSBF}
\!\!\!\!\!\!\!\!\!\! \mathscr{P}_{\textrm{iGSBF}}^{[m]}\!:\!\{\tilde{{\bf{w}}}_{l}^{[m+1]}\}_{l=1}^{L}\!=\!\arg\min\!\!\!\!\!\!\!\!\!\!\!&& \sum_{l=1}^{L}\beta_{l}^{[m]}\|\tilde{\bf{w}}_{l}\|_{\ell_{2}}\nonumber\\
{\rm{subject~to}}\!\!\!\!\!\!\!\!\!\! &&\mathcal{C}_{1}(\mathcal{L}), \mathcal{C}_{2}(\mathcal{L}),\end{aligned}$$ where $$\begin{aligned}
\label{upweights}
\beta_{l}^{[m]}=\sqrt{P_{l}^c\over{\eta_{l}}}{{1}\over{(\|\tilde{\bf{w}}_{l}^{[m]}\|_{\ell_{2}}+\epsilon)}},\forall l=1,\dots, L,\end{aligned}$$ are the weights for the groups at the $m$-th iteration. At each step, the mixed $\ell_{1}/\ell_{2}$-norm optimization is re-weighted using the estimate of the beamformer obtained in the last minimization step.
As this iterative algorithm cannot guarantee the global minimum, it is important to choose a suitable starting point to obtain a good local optimum. As suggested in [@TonyQ.S._WC2013; @Boyd_2008enhancing; @Mehanna_SP2013], this algorithm can be initiated with the solution of the unweighted $\ell_{1}$-norm minimization, i.e., $\beta_{l}^{[0]}=1, \forall l=1,\dots,L$. In our setting, however, the prior information on the system parameters can help us generate a high quality stating point for the iterative GSBF framework. Specifically, with the available channel state information, we choose the $\ell_{2}$-norm of the initial beamformer at the $l$-th RRH $\|\tilde{{\bf{w}}}_{l}^{[0]}\|_{\ell_2}$ to be proportional to its corresponding channel power gain $\kappa_{l}$, arguing that the RRH with a low channel power gain should be encouraged to be switched off as justified in section V-B. Therefore, from (\[upweights\]), we set the following weights as the initiation weights for $\mathscr{P}_{\textrm{iGSBF}}^{[0]}$: $$\begin{aligned}
\label{initiation1}
\beta_{l}^{[0]}=\sqrt{P_{l}^{c}\over{\eta_{l}\kappa_{l}}}, \forall l=1,\dots,L.\end{aligned}$$ The weights indicate that the RRHs with a higher transport power consumption, lower power amplifier efficiency and lower channel power gain should be penalized more heavily.
As observed in the simulations, this algorithm converges very fast (typically within 20 iterations). We set the maximum number of iteration as $m_{\textrm{max}}=L$ in our simulations.
### Iterative Search Procedure
After obtaining the (approximately) sparse beamformers using the above re-weighted GSBF framework, we still adopt the same ordering criterion (\[sparse\_selection1\]) to fix the final active RRH set.
Different from the aggressive strategy in the bi-section GSBF algorithm, which assumes that the RRH should be switched off as many as possible and thus results a minimum transport network power consumption, we adopt a conservative strategy to determine the final active RRH set by realizing that the minimum network power consumption may not be attained when the transport network power consumption is minimized.
Specifically, denote $J_{0}$ as the maximum number of RRHs that can be turned off, the corresponding inactive RRH set is $\mathcal{J}=\{{\pi_{0}},{\pi_{1}},\dots,{\pi_{J_{0}}}\}$. The minimum network power consumption should be searched over all the values $\mathscr{P}^{*}(\mathcal{A}^{[i]})$, where $\mathcal{A}^{[i]}=\mathcal{L}\setminus\{\pi_{0},\pi_{1},\dots,\pi_{i}\}$ and $0\le i\le J_{0}$. This can be accomplished using an iterative search procedure that requires to solve no more than $L$ SOCP problems.
Therefore, the overall iterative GSBF algorithm is presented as Algorithm 3:
**Step 0:** Initialize the weights $\beta_{l}^{[0]}, l=1,\dots, L$ as in (\[initiation1\]) and the iteration counter $m=0$;\
**Step 1:** Solve the weighted GSBF problem $\mathscr{P}_{\textrm{iGSBF}}^{[m]}$ (\[wGSBF\]): [**if**]{} it is infeasible, set $p^{\star}=\infty$ and [**go to End**]{}; otherwise, set $m=m+1$, [**go to Step 2**]{};\
**Step 2:** Update the weights using (\[upweights\]);\
**Step 3:** [**If**]{} converge or $m=m_{\textrm{max}}$, obtain the solution $\hat{\bf{w}}$ and calculate the selection criterion (\[sparse\_selection1\]), and sort them ; otherwise, **go to Step 1**;\
**Step 4:** Initialize $\mathcal{Z}^{[0]}=\emptyset$, $\mathcal{A}^{[0]}=\{1,\dots,L\}$, and $i=0$;\
**Step 5:** Solve the optimization problem $\mathscr{P}(\mathcal{A}^{[i]})$ (\[op1\]);
1. [**If**]{} (\[op1\]) is feasible, obtain $p^{*}(\mathcal{A}^{[i]})$, update the set $\mathcal{Z}^{[i+1]}=\mathcal{Z}^{[i]}\cup\{\pi_{i+1}\}$ and $i=i+1$, [**go to Step 5**]{};
2. [**If**]{} (\[op1\]) is infeasible, obtain $\mathcal{J}=\{0,1,\dots, i-1\}$, [**go to Step 6**]{};\
[**Step 6:**]{} Obtain optimal RRH set $\mathcal{A}^{[j^{\star}]}$ and beamformers minimizing $\mathscr{P}(\mathcal{A}^{[j^{\star}]})$ with $j^{\star}=\arg\min_{j\in\mathcal{J}}p^{*}(\mathcal{A}^{[j]})$;\
**End**
Simulation Results
===================
In this section, we simulate the performance of the proposed algorithms. We consider the following channel model $$\begin{aligned}
{\bf{h}}_{kl}=10^{-L(d_{kl})/20}\sqrt{\varphi_{kl}s_{kl}}{\bf{g}}_{kl},\end{aligned}$$ where $L(d_{kl})$ is the path-loss at distance $d_{kl}$, , as given in Table \[parameter1\], $s_{kl}$ is the shadowing coefficient, $\varphi_{kl}$ is the antenna gain and ${\bf{g}}_{kl}$ is the small scale fading coefficient. We use the standard cellular network parameters as showed in Table \[parameter1\].
--------------------------------------------------------------------------------------------------------------
Parameter Value
------------------------------------------------------------------ -------------------------------------------
Path-loss at distance $d_{kl}$ (km) 148.1+37.6 ${\textrm{log}}_{2}({d}_{kl})$
Standard deviation of log-norm shadowing $\sigma_{s}$ 8 dB
Small-scale fading distribution ${\bf{g}}_{kl}$ $\mathcal{CN}({\bf{0}},
{\bf{I}})$
Noise power $\sigma_{k}^{2} $ [@Soliman_CM13] (10 MHz bandwidth) -102 dBm
Maximum transmit power of RRH $P_{l}$ [@Soliman_CM13] 1 W
Power amplifier efficiency $\alpha_{l}$ [@Auer_WC2011] 4
Transmit antenna power gain 9 dBi
--------------------------------------------------------------------------------------------------------------
: Simulation Parameters[]{data-label="table_example"}
\[parameter1\]
Each point of the simulation results is averaged over 50 randomly generated network realizations.
The proposed algorithms are compared to the following algorithms:
- **Coordinated beamforming (CB) algorithm**: In this algorithm, all the RRHs are active and only the total transmit power consumption is minimized [@WeiYu_WC10].
- **Mixed-integer nonlinear programming (MINLP) algorithm**: This algorithm [@leyffer_2012mixed; @Cheng_SP2013] can obtain the global optimum. Since the complexity of the algorithm grows exponentially with the number of RRHs $L$, we only run it in a small-size network.
- **Conventional Sparsity pattern (SP) based algorithm**: In this algorithm, the unweighted mixed $\ell_{1}/\ell_{p}$-norm is adopted to induce group sparsity as in [@Z.Q.Luo_JSAC2013] and [@Mehanna_SP2013]. The ordering of RRHs is determined only by the group-sparsity of the beamformer, i.e., $\theta_{l}:=(\sum_{k=1}^{K}\|\hat{\bf{w}}_{lk}\|_{\ell_{2}})^{1/2},
\forall l=1,\dots,L$, instead of (\[sparse\_selection1\]). The complexity of the algorithm grows logarithmically with $L$.
- **Relaxed mixed-integer nonlinear programming (RMINLP) based algorithm**: In this algorithm, a deflation procedure is performed to switch off RRHs one-by-one based on the solutions obtained via solving the relaxed MINLP by relaxing the integers to the unit intervals [@Cheng_SP2013]. The complexity of the algorithm grows linearly with $L$.
Network Power Consumption versus Target SINR
---------------------------------------------
Consider a network with $L=10$ 2-antenna RRHs and $K=15$ single-antenna MUs uniformly and independently distributed in the square region $[-1000~1000]\times[-1000~1000]$. We set all the transport link power consumption to be $P_{l}^{c}=(5+l) W, l=1,\dots,L$, which is to indicate the inhomogeneous power consumption on different transport links. Fig. [\[smallnetworktransport\]]{} demonstrates the average network power consumption with different target SINRs.
This figure shows that the proposed GS algorithm can always achieve global optimum (i.e., the optimal value from the MINLP algorithm), which confirms the effectiveness of the proposed selection rule for the greedy search procedure. With only logarithmic complexity, the proposed bi-section GSBF algorithm achieves almost the same performance as the RMINLP algorithm, which has a linear complexity. Moreover, with the same complexity, the gap between the conventional SP based algorithm and the proposed bi-section GSBF algorithm is large. Furthermore, the proposed iterative GSBF algorithm always outperforms the RMINLP algorithm, while both of them have the same computational complexity. These confirm the effectiveness of the proposed GSBF framework to minimized the network power consumption. Overall, this figure shows that our proposed schemes have the potential to reduce the power consumption by $40\%$ in the low QoS regime, and by $20\%$ in the high QoS regime.
Network Power Consumption versus Transport Links Power Consumption
------------------------------------------------------------------
![Average network power consumption versus target SINR.[]{data-label="smallnetworktransport"}](simulation1){width="1\columnwidth"}
Consider a network involving $L=20$ 2-antenna RRHs and $K=15$ single-antenna MUs uniformly and independently distributed in the square region $[-2000~2000]\times[-2000~2000]$ meters. We assume all the transport links have the same power consumption, i.e., $P_{c}=P_{l}^{c}, \forall l=1,\dots, L$ and set the target SINR as 4 dB. Fig.[\[transportlink\]]{} presents average network power consumption with different transport links power consumption.
This figure shows that both the GS algorithm and the iterative GSBF algorithm significantly outperform other algorithms, especially in the high transport link power consumption regime. Moreover, the proposed bi-section GSBF algorithm provides better performance than the conventional SP based algorithm and is close to the RMINLP based algorithm, while with a lower complexity. This result clearly indicates the importance of considering the key system parameters when applying the group sparsity beamforming framework.
![Average network power consumption versus transport links power consumption.[]{data-label="transportlink"}](simulation2){width="1\columnwidth"}
Network Power Consumption versus the Number of Users
----------------------------------------------------
Consider a network with $L=20$ 2-antenna RRHs uniformly and independently distributed in the square region $[-2000~2000]\times[-2000~2000]$ meters. We assume all the transport links have the same power consumption, i.e., $P_{l}^{c}=20 W, \forall l=1,\dots, L$ and set the target SINR as 4 dB. Fig. [\[mobileuser\]]{} presents the average network power consumption with different numbers of MUs, which are uniformly and independently distributed in the same region.
![Average network power consumption versus the number of mobile users.[]{data-label="mobileuser"}](simulation3){width="1\columnwidth"}
Overall, this figure further confirms the following conclusions:
1. With the $\mathcal{O}(L^{2})$ computational complexity, the proposed GS algorithm has the best performance among all the low-complexity algorithms.
2. With the $\mathcal{O}(L)$ computational complexity, the proposed iterative GSBF algorithm outperforms the RMINLP algorithm, which has the same complexity.
3. With $\mathcal{O}(\log(L))$ computational complexity, the proposed bi-section GSBF algorithm has almost the same performance with the RMINLP algorithm and outperforms the conventional SP based algorithm, which has the same complexity. Therefore, the bi-section GSBF algorithm is very attractive for practical implementation in large-scale Cloud-RAN.
Conclusions and Discussions
===========================
In this paper, we proposed a new framework to improve the energy efficiency of cellular networks with the new architecture of Cloud-RAN. It was shown that the transport network power consumption can not be ignored when designing green Cloud-RAN. By jointly selecting the active RRHs and minimizing the transmit power consumption through coordinated beamforming, the overall network power consumption can be significantly reduced, especially in the low QoS regime. The proposed group sparse formulation $\mathscr{P}_{\textrm{sparse}}$ serves as a powerful design tool for developing low complexity GSBF algorithms. Through rigorous analysis and careful simulations, the proposed GSBF framework was demonstrated to be very effective to provide near-optimal solutions. Especially, for the large-scale Cloud-RAN, the proposed bi-section GSBF algorithm will be a prior option due to its low complexity, while the iterative GSBF algorithm can be applied to provide better performance in a medium-size network. Simulation also showed that the proposed GS algorithm can always achieve nearly optimal performance, which makes it very attractive in the small-size clustered deployment of Cloud-RAN.
This initial investigation demonstrated the advantage of Cloud-RAN in terms of the network energy efficiency. More works will be needed to exploit the full benefits and overcome the main challenges of Cloud-RAN. Future research directions include more efficient beamforming algorithms for very large scale Cloud-RAN deployment, joint beamforming and compression by considering the limited-capacity transport links, joint user scheduling, and effective CSI acquisition methods.
Proof of Proposition 1
======================
We begin by deriving the tightest positively homogeneous lower bound of $p({\bf{w}})$, which is given by [@Rockafellar1997convex; @Obozinski_arXiv2012] $$\begin{aligned}
p_{h}({\bf{w}})=\inf_{\lambda>0}{{p(\lambda{\bf{w}})}\over{\lambda}}=\inf_{\lambda>0}{{\lambda}}T({\bf{w}})+{1\over{\lambda}}F(\mathcal{T}({\bf{w}})).\end{aligned}$$ Setting the gradient of the objective function to zero, the minimum is obtained at $\lambda=\sqrt{F(\mathcal{T}({\bf{w}}))/T({\bf{w}})}$. Thus, the positively homogeneous lower bound of the objective function becomes $$\begin{aligned}
\label{lowerbound}
p_{h}({\bf{w}})=2\sqrt{T({\bf{w}})F(\mathcal{T}({\bf{w}}))},\end{aligned}$$ which combines two terms multiplicatively.
Define diagonal matrices ${\bf{U}}\in\mathbb{R}^{N\times
N}$, ${\bf{V}}\in\mathbb{R}^{N\times N}$ with $N=K\sum_{l=1}^{L}N_{l}$, and the $l$-th block elements are $\eta_{l}{\bf{I}}_{KN_{l}}$ and ${1\over{\eta_{l}}}{\bf{I}}_{KN_{l}}$, respectively. Next, we calculate the convex envelope of $p_{h}({\bf{w}})$ via computing its conjugate: $$\begin{aligned}
p_{h}^{*}({\bf{y}})&=&\sup_{{\bf{w}}\in\mathbb{C}^{N}} \left({\bf{y}}^{T}{\bf{U}}^{T}{\bf{V}}{\bf{w}}-2\sqrt{T({\bf{w}})F(\mathcal{T}({\bf{w}}))}\right),\nonumber\\
&=&\sup_{\mathcal{I}\subseteq\mathcal{V}}\sup_{{\bf{w}}_{\mathcal{I}}\in\mathbb{C}^{|\mathcal{I}|}}
\!\!\!\left({\bf{y}}_{\mathcal{I}}^{T}{\bf{U}}_{\mathcal{I}\mathcal{I}}^{T}{\bf{V}}_{\mathcal{I}\mathcal{I}}{\bf{w}}_{\mathcal{I}}\!-\!2\sqrt{T({\bf{w}}_{\mathcal{I}})F(\mathcal{I})}\right)\nonumber\\
& =& \left\{ \begin{array}{ll}
0 & {\textrm{if}}~ \Omega^{*}({\bf{y}})\le 1 \\
\infty, & {\textrm{otherwise}}.
\end{array} \right.\end{aligned}$$ where ${\bf{y}}_{\mathcal{I}}$ is the $|\mathcal{I}|$-dimensional vector formed with the entries of ${\bf{y}}$ indexed by ${\mathcal{I}}$ (similarly for ${\bf{w}}$), and ${\bf{U}}_{\mathcal{I}\mathcal{I}}$ is the $|\mathcal{I}|\times|\mathcal{I}|$ matrix formed with the rows and columns of ${\bf{U}}$ indexed by $\mathcal{I}$ (similarly for ${\bf{V}}$), and $\Omega^{*}({\bf{y}})$ defines a dual norm of $\Omega({\bf{w}})$: $$\begin{aligned}
\label{dual1}
\Omega^{*}({\bf{y}})=\sup\limits_{\mathcal{I\subseteq\mathcal{V}}, \mathcal{I}\ne\emptyset}~
{\|{\bf{y}}_{\mathcal{I}}{\bf{U}}_{\mathcal{I}}\|_{\ell_2}\over{2\sqrt{F(\mathcal{I})}}}={1\over{2}}\max_{l=1,\dots,L}
\sqrt{{\eta_{l}}\over{P_{l}^{c}}}{{\|{\bf{y}}_{\mathcal{G}_{l}}\|_{\ell_2}}}.\end{aligned}$$ The first equality in (\[dual1\]) can be obtained by the Cauchy-Schwarz inequality: $$\begin{aligned}
{\bf{y}}_{\mathcal{I}}^{T}{\bf{U}}_{\mathcal{I}\mathcal{I}}^{T}{\bf{V}}_{\mathcal{I}\mathcal{I}}{\bf{w}}_{\mathcal{I}}&\le&\|{\bf{y}}_{\mathcal{I}}{\bf{U}}_{\mathcal{I}}\|_{\ell_2}\cdot\|{\bf{V}}_{\mathcal{I}\mathcal{I}}{\bf{w}}_{\mathcal{I}}\|_{\ell_2}\nonumber\\
&=&\|{\bf{y}}_{\mathcal{I}}{\bf{U}}_{\mathcal{I}}\|_{\ell_2}\cdot\sqrt{T({\bf{w}}_{\mathcal{I}})}.\end{aligned}$$ The second equality in (\[dual1\]) can be justified by $$\begin{aligned}
\Omega^{*}({\bf{y}})&\ge& \sup_{\mathcal{I}\subseteq\mathcal{V},\mathcal{I}\ne\emptyset}\left(
{1\over{2\sqrt{F(\mathcal{I})}}}\max_{l=1,\dots,L}\|{\bf{y}}_{\mathcal{I}\cap\mathcal{G}_{l}}{\bf{U}}_{\mathcal{I}\cap\mathcal{G}_{l}}\|_{\ell_2}\right)\nonumber\\
&=& {1\over{2}}\max_{l=1,\dots,L}\sqrt{{\eta_{l}}\over{P_{l}^{c}}} {{\|{\bf{y}}_{\mathcal{G}_{l}}\|_{\ell_2}}},\end{aligned}$$ and $$\begin{aligned}
\Omega^{*}({\bf{y}})&\le& \sup_{\mathcal{I}\subseteq\mathcal{V},\mathcal{I}\ne\emptyset}\left(
{{\|{\bf{y}}_{\mathcal{I}}{\bf{U}}_{\mathcal{I}}\|_{\ell_2}}\over{2\min_{l=1,\dots,L}\sqrt{F(\mathcal{I}\cap\mathcal{G}_{l})}}}\right)\nonumber\\
&=& {1\over{2}}\max_{l=1,\dots,L}\sqrt{{\eta_{l}}\over{P_{l}^{c}}} {{\|{\bf{y}}_{\mathcal{G}_{l}}\|_{\ell_2}}}.\end{aligned}$$ Therefore, the tightest convex positively homogeneous lower bound of the function $p({\bf{w}})$ is $$\begin{aligned}
\Omega({\bf{w}})\!\!\!&=&\!\!\!\!\!\!\!\sup_{ \Omega^{*}({\bf{y}})\le1} {\bf{w}}^{T}{\bf{y}}\nonumber\\
\!\!\!\!\!&\le&\!\!\!\!\!\!\!\sup_{\Omega^{*}({\bf{y}})\le1}\sum_{l=1}^{L}\|{\bf{w}}_{\mathcal{G}_{l}}\|_{\ell_2}\|{\bf{y}}_{\mathcal{G}_{l}}\|_{\ell_2}
\nonumber\\
&\le&\!\!\!\!\!\!\!\! \sup_{\Omega^{*}({\bf{y}})\le1} \!\!\left(\!\sum_{l=1}^{L}\sqrt{P_{l}^c\over{\eta_{l}}}\|{\bf{w}}_{\mathcal{G}_{l}}\|_{\ell_2}\right)\!\!\!\left(\!\!\max_{l=1,\dots,L}
\sqrt{{\eta_{l}}\over{P_{l}^{c}}}{{\|{\bf{y}}_{\mathcal{G}_{l}}\|_{\ell_2}}}\right)
\nonumber\\
&=&\!\!\!\!\! 2\sum_{l=1}^{L}{\sqrt{P_{l}^{c}\over{\eta_{l}}}}\|{\bf{w}}_{\mathcal{G}_{l}}\|_{\ell_2}.\end{aligned}$$ This upper bound actually holds with equality. Specifically, we let $\bar{\bf{y}}_{\mathcal{G}_{l}}=2\sqrt{P_{l}^c\over{\eta_{l}}}{{\bf{w}}_{\mathcal{G}_{l}}^{\dagger}\over{\|{\bf{w}}_{\mathcal{G}_{l}}^{\dagger}}\|_{\ell_2}}$, such that $\Omega^{*}(\bar{\bf{y}})=1$. Therefore, $$\begin{aligned}
\Omega({\bf{w}})&=&\sup_{ \Omega^{*}({\bf{y}})\le1} {\bf{w}}^{T}{\bf{y}}\nonumber\\
&\ge& \sum_{l=1}^{L}{\bf{w}}_{\mathcal{G}_{l}}^{T}\bar{\bf{y}}_{\mathcal{G}_{l}}=2\sum_{l=1}^{L}{\sqrt{P_{l}^{c}\over{\eta_{l}}}}\|{\bf{w}}_{\mathcal{G}_{l}}\|_{\ell_2}.\end{aligned}$$
Preliminaries on Majorization-Minimization Algorithms
=====================================================
The majorization-minimization (MM) algorithm, being a powerful tool to find a local optimum by minimizing a surrogate function that majorizes the objective function iteratively, has been widely used in statistics, machine learning etc., [@hunter2004tutorial]. We introduce the basic idea of MM algorithms, which allows us to derive our main results.
Consider the problem of minimizing $f({\bf{x}})$ over $\mathcal{F}$. The idea of MM algorithms is as follows. First, we construct a majorization function $g({\bf{x}}|{\bf{x}}^{[m]})$ for $f({\bf{x}})$ such that $$\begin{aligned}
\label{mm1}
g({\bf{x}}|{\bf{x}}^{[m]})\ge f({\bf{x}}), \forall~{\bf{x}}\in\mathcal{F},\end{aligned}$$ and the equality is attained when ${\bf{x}}={\bf{x}}^{[m]}$. In a MM algorithm, we will minimize the majorization function $g({\bf{x}}|{\bf{x}}^{[m]})$ instead of the original function $f({\bf{x}})$. Let ${\bf{x}}^{[m+1]}$ denote the minimizer of the function $g({\bf{x}}|{\bf{x}}^{[m]})$ over $\mathcal{F}$ at $m$-th iteration, i.e., $$\begin{aligned}
\label{mm2}
{\bf{x}}^{[m+1]}=\arg\min_{{\bf{x}}\in\mathcal{F}}~g({\bf{x}}|{\bf{x}}^{[m]}),\end{aligned}$$ then we can see that this iterative procedure will decrease the value of $f({\bf{x}})$ monotonically after each iteration, i.e., $$\begin{aligned}
f({\bf{x}}^{[m+1]})\le g({\bf{x}}^{[m+1]}|{\bf{x}}^{[m]})\le g({\bf{x}}^{[m]}|{\bf{x}}^{[m]})=f({\bf{x}}^{[m]}),\end{aligned}$$ which is a direct result from the definitions (\[mm1\]) and (\[mm2\]). The decreasing property makes an MM algorithm remarkable numerical stability. More details can be found in a tutorial on MM algorithms [@hunter2004tutorial] and references therein.
[^1]: The authors are with the Department of Electronic and Computer Engineering, the Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong. E-mail: $\{\textrm{yshiac, eejzhang, eekhaled}\}[email protected].
[^2]: The impact of limited-capacity transport links on compression in Cloud-RAN was recently investigated in [@Shamai_TVT2013; @Shamai_TSP2013], and its impact in our setting is left to future work.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Mesoscopic low frequency noise in electrical characteristics of disordered conductors is a result of dynamic quantum interference pattern due to motion of defects. This has been firmly established by demonstrating the characteristic partial suppression of the noise amplitude by the dephasing effect of a weak external magnetic field. The spatial correlation of the quantum interference pattern in disordered normal state conductors is invariably limited by the exponential phase relaxation due to inelastic processes. In this paper we develop a quantitative theory of the mesoscopic noise in the s-wave superconducting phase of a strongly disordered superconductor (such that the superconducting coherence length is much longer than the mean free path). We find that the superconducting coherence length limits the quantum interference effects in superconductors. However, in contrast to the normal phase, the decay of the phase relaxation on the scale of the superconducting coherence length is non-exponential. This unusual slow relaxation manifests in the enhanced amplitude of the mesoscopic noise in superconductors and a peculiar non-linear scaling of the amplitude with the strength/number of mobile defects in very thin superconducting films and wires (effectively 2D and 1D with respect to the superconducting coherence length). Mesoscopic noise sets a natural limit on the quality of kinetic inductance elements.'
author:
- 'K. Kechedzhi'
title: 'Slow non-exponential phase relaxation and enhanced mesoscopic kinetic inductance noise in disordered superconductors'
---
Introduction
============
Development of superconducting circuits with very low levels of noise is largely motivated by their potential applications for quantum computing devices[@Nori2005; @Clarke2008], ultra-sensitive detectors[@Zmuidzinas2004; @100years] and magnetometers[@ClarkeBook]. Performance of all these devices is limited by the level of intrinsic noise of various types. A dramatic progress was achieved recently with the elimination of the amorphous insulators and/or suppression of the charge noise associated with them[@Martinis2005; @Schreier2008; @Manucharyan2009]. This requires minimizing the use of arrays of Josephson junctions as inductance elements in these circuits. The use of magnetic self-inductance of wires and coils is not feasible due to size and geometry restrictions in these devices. High kinetic inductance appears naturally in disordered superconductors which makes them an obvious candidate for these elements. Here we show that disordered superconducting wires show significant fluctuations of kinetic inductance due to electron interference induced by defect motion; we develop a quantitative analytical theory of this quantum interference effect and calculate the amplitude of the kinetic inductance fluctuations and noise with accurate numerical coefficients. The noise due to these fluctuations provides a natural limit for the quality of kinetic inductance elements[@WallraffSchoelkopf2004; @Klapwijk2013; @AnnunziataCatelani2010; @Zmuidzinas2012].
Kinetic inductance noise is caused by local fluctuations of impurities invariably present in metallic wires. Quantum interference of electrons moving diffusively in the potential of impurities significantly enhances the effect of these local fluctuations. This is because the macroscopic interference pattern in a conductor is very sensitive to the position of individual impurities. Motion of a single impurity, for example an impurity jumping between two stable spatial configurations, results in a substantial fluctuation in macroscopic (or rather mesoscopic) properties of the conductor. This effect was analyzed in great detail for the case of metals in the normal state[@AltshulerSpivak; @Feng86] in which the quantum interference leads to mesoscopic noise, i.e. a substantial enhancement of the noise in electronic characteristics due to local fluctuations in the impurity potential. The quantum interference pattern in the normal state is spatially correlated up to the length scale, $L_{\varphi}$, that limits the coherent propagation of electrons. $L_{\varphi}$ is typically set by low energy electron-electron or electron-phonon scattering[@LeeStoneFukuyama; @AAKH]. In superconductors, where electrons at low energies are bound into Cooper pairs similarly to the case of normal metals macroscopic characteristics demonstrate mesoscopic fluctuations[@SpivakZyuzin; @Koshelev2012]. However, the spatial correlations of mesoscopic fluctuations in a superconductor are distinct from those in normal metals. Paired state of electrons in presence of disorder is characterized by a superconducting coherence length $\zeta\equiv\sqrt{D(1/\Delta)}$ where $D$ is the diffusion coefficient and $\Delta$ is the superconducting gap, which describes a typical diffusion path of an electron during a time $1/\Delta$, a lifetime of a virtual excitation with the energy of order $\Delta$. It is reasonable to expect that the role of the dephasing length in superconductors is played by the superconducting correlation length, i.e. $\zeta$ defines the scale of exponential phase relaxation and limits the quantum interference responsible for the variation of macroscopic characteristics. However, we will show that this intuition is not fully consistent with microscopic calculations. Instead, we find that the superconducting gap $\Delta$ limits the coherence of electronic states however the phase relaxation has a slow non-exponential character. The slow phase relaxation results in an enhanced amplitude of noise in superconductors and non-linear scaling of the noise amplitude with the strength/number of locally fluctuating impurities in thin films and wires. This behavior is in contrast to the exponential phase relaxation in the case of weak localization corrections to the superfluid density of s-wave superconductors[@SmithAmbegaokar].
Experimentally, the noise in the normal state conductivity associated with quantum interference is a well established phenomena[@Birge1989; @Birge1990; @Birge1993]. In contrast, the noise in kinetic inductance of Josephson circuits was reported to be absent in early experiments[@Wellstood1987] and observed only very recently in Ref. . This work[@McDermott] also reports a surprising degree of correlations between the inductance noise and flux noise in the superconducting devices that remain poorly understood[@Kechedzhi_2011].
The main result of this paper is the quantitative prediction of the average amplitude $\delta_{K}^{2}\equiv\left\langle (K_{u'}-K_{u})^{2}\right\rangle /(2\left\langle K\right\rangle ^{2})$ of the mesoscopic noise in the kinetic inductance $K$ in superconducting wires. Here $K_{u}$ and $K_{u'}$ correspond to the value of kinetic inductance given impurity potential $u$ and $u'$ respectively. In realistic measurement the two distinct impurity potential realization $u$ and $u'$ represent different moments in time, and the amplitude $\delta_{K}^{2}$ is averaged over long periods of time, longer than any characteristic time of the electronic system, and longer than the characteristic fluctuation time of impurities. Practically, this amplitude can be connected to the amplitude of the noise spectral density.[^1]
We consider a finite size rectangular piece of a superconductor of a mesoscopic size $L_{x}\times L_{y}\times L_{z}$. In other words the dimensions of the superconductor are not too much larger than the superconducting coherence length. We show that in three dimensional samples, $\zeta\ll L_{x},\, L_{y},\, L_{z}$ the result is given by,
$$\begin{gathered}
\delta_{K}^{2}\approx\widetilde{C}_{3}\gamma_{T}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{3}}{V},\label{3Dnoise}\end{gathered}$$
where $p_{F}$, is Fermi wave vector, $\ell$ is the mean free path, $\zeta=\sqrt{D/\Delta}$ coherence length of the superconducting electrons, $V=L_{x}L_{y}L_{z}$ is the wire volume and the numerical coefficient is $\widetilde{C}_{3}\approx60$. The relative density of thermally activated defects is defined by $\gamma_{T}=\Gamma_{T}/\Gamma$, where $\Gamma=1/\tau$ is the elastic relaxation rate of electrons and $\Gamma_{T}$ is the part due to fluctuating defects. The ratio $\Gamma_{T}/\Gamma\approx T\kappa$ is roughly linear in temperature and is determined by $\kappa$, the relative density of states of thermally fluctuating defects that is only weakly material dependent[@Yu1988]. The remaining parameters that determine the strength of the fluctuations $p_{F}\ell$ and $\zeta$ can be determined from independent measurements. The noise amplitude grows rapidly as the device becomes smaller and more disordered. As a result, this mechanism is likely to be the dominant source of inductance noise in small and highly disordered devices. Eq. (\[3Dnoise\]) holds for three dimensional wires with thickness $L_{\perp}>\zeta$, we will see that the fluctuations get rather larger for very thin, two and one dimensional, wires $\ell<L_{\perp}<\zeta$.
Note that throughout this paper we distinguish the sample to sample fluctuations of kinetic inductance defined as,
$$\frac{\left\langle \delta K^{2}\right\rangle }{\left\langle K\right\rangle ^{2}}\equiv\frac{\left\langle (K-\left\langle K\right\rangle )^{2}\right\rangle }{\left\langle K\right\rangle ^{2}}\approx C_{3}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta\ell^{2}}{V},\label{eq:StSInductance}$$
where the angular brackets mean averaging over disorder realizations. This quantity is the saturation value of the noise amplitude $\delta_{K}^{2}$ defined above when the two disorder configurations $u$ and $u'$ in Eq. (\[3Dnoise\]) are completely uncorrelated. In other words all of the impurities have changed their positions in Eq. (\[eq:StSInductance\]) as opposed to only a fraction in Eq. (\[3Dnoise\]). In this way Eq. (\[eq:StSInductance\]) is the upper limit of the noise amplitude Eq. (\[3Dnoise\]). The result in the right hand part is valid in three dimensional wires $\zeta\ll L_{x},\, L_{y},\, L_{z}$, and we will see that $C_{3}\approx25$.
Up to a numerical factor, the results Eqs. (\[3Dnoise\]) and (\[eq:StSInductance\]) can be derived from the following qualitative argument. We first estimate the amplitude of the sample-to-sample fluctuations. Consider a small cubic piece of a superconductor of the size $L^{3}$ with $L\lesssim\zeta.$ At these scales the coherence of electrons is weakly affected by the superconductivity. Optical sum rule $\int\sigma(\omega)d\omega=const$ and Anderson theorem of gap disorder independence imply that superfluid response is directly related to normal state conductivity [@DeGennes]. Therefore the fluctuations of the normal state conductivity $\sigma$ roughly translate into fluctuations of the superfluid response, $\delta_{K}^{2}\sim\delta_{\sigma}^{2}$. Conductance fluctuations of a small piece of metal have a universal value $\left\langle \delta\sigma^{2}\right\rangle \approx\left(2e^{2}/h\right)^{2}$. Thus, one expects that the maximal change in the interference pattern correspond to the relative change of the superfluid response by $\left\langle \delta K^{2}\right\rangle /\left\langle K\right\rangle ^{2}=1/(p_{F}^{2}\ell L)^{2}$ in three dimensional wires. This is the amplitude of sample-to-sample fluctuations of a mesoscopic size piece of superconductor. We now estimate the fluctuation of the superfluid response during the time $t\gg1/\Delta,\, L^{2}/D$. During the time $t$ a small number of defects change their position in space, which means that in contrast to the sample-to-sample fluctuations only a small number of electronic paths are affected, resulting in a smaller value of the fluctuation of the superfluid response. We expect the fluctuation to be proportional to the number of paths affected by the motion of the defects. A typical path of a diffusive electron that enters and exits a cube of size $L$ has a length $(L/\ell)^{2}\ell$, the probability that the concentration $n_{T}$ of randomly positioned thermally activated impurities affect this path is $\sigma_{T}n_{T}(L/\ell)^{2}\ell$ where $\sigma_{T}$ is the scattering cross-section of the fluctuating impurities. Assuming that the fluctuating and static impurities are roughly equivalent we can relate $n_{T}\sigma_{T}=\gamma_{T}n_{i}\sigma=\gamma_{T}/\ell$, where $\sigma$ is the average scattering cross-section for all impurities and $n_{i}$ is the total impurity concentration. Combining all these factors together we get $\delta_{K}^{2}=\gamma_{T}/(p_{F}l)^{4}$.
We assume for the sake of this estimate that in a larger sample $L>\zeta$ the information of the single electron phases is lost on the scale of the coherence length $\zeta$. This implies that the regions of the size $\zeta$ fluctuate independently. Electromagnetic response of the whole sample is obtained by adding these regions as independent resistor network, adding $\zeta/L$ independent factors we get Eq. (\[3Dnoise\]) for the fluctuating part of the kinetic inductance. The fraction of the paths affected has to be modified as well $\gamma_{T}(\zeta/\ell)^{2}$. Combining these factors we obtain Eq. (\[3Dnoise\]). We will see that the assumption of uncorrelated fluctuations on the scale of $\zeta$ is violated in the case of thin films and wires.
In the next Section \[sec:Sample-to-sample\] we calculate sample-to-sample fluctuations of the superfluid response. After that we use this result to calculate the noise amplitude in Section \[sec:Noise-amplitude\] and conclude in Section \[sec:Conclusion\]. Details of construction of diagramatic perturbation theory are given in Appendix \[sec:Details-of-diagrams\]. Calculation of Fourier integrals of the superconducting coherence factors is shown in Appendix \[sec:FourierT0\].
Sample-to-sample fluctuation of the superfluid response\[sec:Sample-to-sample\]
===============================================================================
We now turn to the analytical computation. We focus on the properties of a superconducting wire with the simplest geometry: a rectangle of total volume $V=L_{x}\times L_{y}\times L_{z}$ connected to two leads that carry spatially uniform supercurrent. A small supercurrent is injected into the wire, along $Oz$ axis, by an external source and the resulting phase difference is measured, for example with Josephson junctions. This geometry is very similar to the one used in Ref. to measure SQUID inductance. Local electro-magnetic response of a superconductor is given by the kernel defined as $j_{\alpha}(r)=\int dr'S_{\alpha\beta}(r,r')A_{\beta}(r')$. This supercurrent response can be thought of as a superfluid density, $S$. Since the definition of the kernel is a local form of London equation. Kinetic inductance, $K$, is a result of the work done by electromagnetic field to accelerate the Cooper pairs and is therefore inversely proportional to the superfluid density response $S$, $K\sim1/S$. Therefore $\left\langle \delta K^{2}\right\rangle /\left\langle K\right\rangle {}^{2}\approx\left\langle \delta S^{2}\right\rangle /\left\langle S\right\rangle {}^{2}$ and in the following we will discuss fluctuations of $S$. The fluctuation in the total response of the wire of the volume $V$ is given by the spatial average, $$\langle\delta S^{2}\rangle=\int\frac{\prod_{i=1}^{4}d\mathbf{r_{i}}}{L_{z}^{4}}\left[\left\langle S(\mathbf{r_{1},r_{2}})S(\mathbf{r_{3},r_{4}})\right\rangle -\left\langle S\right\rangle ^{2}\right].$$
We introduce the exact single particle eigenstates of the disordered system, $$\begin{aligned}
\left(-\frac{1}{2m}\nabla^{2}-\mu+u(\mathbf{r})\right)\phi_{\xi}(r) & = & \xi\phi_{\xi}(r),\end{aligned}$$ where we assumed Gaussian delta-correlated disorder $\langle u(\mathbf{r})u(\mathbf{r'})\rangle=n_{i}u_{0}^{2}\delta(\mathbf{r-r'})$ in the wire characterized by a momentum relaxation rate $\tau^{-1}=2\pi\nu n_{i}u_{0}^{2}$, where $\nu$ is the density of single electron states and $n_{i}$ density of impurities. We assume that $\langle u(\mathbf{r})\rangle=0$ without loss of generality as non-zero value would result in a shift of the chemical potential $\mu$ that can be absorbed into its definition. The response kernel of a superconductor to electromagnetic field can be expressed in terms of these eigenstates, $\phi_{\xi}(r)$[@DeGennes], $$\begin{gathered}
S_{\alpha\beta}(r,r')=-\frac{e^{2}}{4m^{2}}\int d\xi_{1}d\xi_{2}p_{\xi_{1}\xi_{2}}^{\alpha}(r)p_{\xi_{1}\xi_{2}}^{\beta}(r')\mathcal{L}_{\xi_{1}\xi_{2}},\label{SinEigenStates}\\
\mathbf{p}_{\xi_{1}\xi_{2}}(r)\equiv\phi_{\xi}(r)\overleftrightarrow{\nabla}\phi_{\xi'}(r)=\phi_{\xi}(r)\nabla\phi_{\xi'}(r)-\phi_{\xi'}(r)\nabla\phi_{\xi}(r).\label{eq:Momentum}\end{gathered}$$ where we introduced a coherence factor[@BCS], $$\begin{aligned}
& \mathcal{L}_{\xi\xi'}=-\left[\frac{1}{2}\frac{EE'-\Delta^{2}-\xi\xi'}{EE'(E+E')}\left(1-f_{E}-f_{E'}\right)\right.\nonumber \\
& \left.-\frac{1}{2}\frac{f_{E}-f_{E'}}{E-E'}\frac{EE'+\Delta^{2}+\xi\xi'}{EE'}-\frac{f_{\xi'}-f_{\xi}}{\xi-\xi'}\right].\label{mathcalL}\end{aligned}$$ Here $E\equiv\sqrt{\xi^{2}+\Delta^{2}}$ and $f_{\xi}$ is the Fermi-Dirac distribution. The 3rd term in Eq. (\[mathcalL\]) is obtained taking $\Delta\rightarrow0$ in the first two terms. This term represents the normal state diamagnetic part of the response function such that the right hand side of Eq. (\[SinEigenStates\]) vanishes in the normal state. Note that taking $\Delta\rightarrow0$, leads to $E\rightarrow|\xi|$ whereas the eigenstate energy $\xi\rightarrow\pm|\xi|$ can be both positive and negative, and one has to keep in mind that $f_{-|\xi|}=1-f_{|\xi|}$.
It is useful to rewrite Eq. (\[SinEigenStates\]) in terms of exact single particle Green functions, $$G_{\xi}^{R/A}(\mathbf{r,r'})=\sum_{n}\frac{\phi_{n}(\mathbf{r})\phi_{n}^{\ast}(\mathbf{r'})}{\xi-\xi_{n}\pm i\delta},$$ as follows, $$\overline{S}=\frac{e^{2}}{4m^{2}}\int\frac{d\mathbf{r}d\mathbf{r'}d\xi d\xi'}{(2\pi)^{2}L_{z}^{2}}\mathcal{L}_{\xi\xi'}\Delta G_{\xi}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}\Delta\overline{G}_{\xi'},\label{eq:ScDens}$$ where we introduced a notation, $\Delta G_{\xi}\equiv G_{\xi}^{R}(\mathbf{r,r'})-G_{\xi}^{A}(\mathbf{r,r'})$ and $\Delta\overline{G}_{\xi}\equiv G_{\xi}^{R}(\mathbf{r',r})-G_{\xi}^{A}(\mathbf{r',r})$, and the meaning of the double arrowed gradient symbol can be inferred from Eq. (\[eq:Momentum\]).
The variance of the superconducting density, $$\langle\delta S^{2}\rangle=\int d\xi_{1}d\xi_{2}d\xi_{3}d\xi_{4}\mathcal{L}_{\xi_{1}\xi_{2}}\mathcal{L}_{\xi_{3}\xi_{4}}\mathcal{F}\label{eq:deltaS}$$ using Eq. (\[SinEigenStates\]) can be written in terms of a correlator, $$\begin{aligned}
& \mathcal{F}\equiv\frac{W^{2}}{16}\left[\langle p_{\xi_{1}\xi_{2}}^{\alpha}p_{\xi_{1}\xi_{2}}^{\beta}p_{\xi_{3}\xi_{4}}^{\alpha'}p_{\xi_{3}\xi_{4}}^{\beta'}\rangle-\langle p_{\xi_{1}\xi_{2}}^{\alpha}p_{\xi_{1}\xi_{2}}^{\beta}\rangle\langle p_{\xi_{3}\xi_{4}}^{\alpha'}p_{\xi_{3}\xi_{4}}^{\beta'}\rangle\right],\label{eq:CorrelatorF}\\
& W\equiv\frac{e^{2}}{m^{2}}\frac{1}{(2\pi)^{2}L_{z}^{2}}.\nonumber \end{aligned}$$
![(a), (b) Diagrams contributing to mesoscopic sample-to-sample fluctuations of superconducting density. Solid lines represent disorder averaged Green function; short wavy lines stand for current vertexes. The shaded regions in (b) indicate appropriately renormalized Hikami boxes, see text and Fig. \[fig:Hikami\]. Long wavy lines stand for diffuson/Cooperon impurity ladders $\mathcal{D}(x_{1},x_{2})$, see discussion after Eq. (\[eq:DiamondDiagramTwoT\]).[]{data-label="fig:AARRDiagrams"}](DiagramsRA){width="0.9\columnwidth"}
A similar correlator to $\mathcal{F}$ appears in the variance of conductance fluctuations in the normal metal for which a perturbation theory in $k_{F}\ell\gg1$ was developed and used extensively[@LeeStoneFukuyama]. The main order contribution in this perturbation series may be written in terms of the diagrams shown in Fig. \[fig:AARRDiagrams\]. In the case of a superconductor the diagrams will have a similar form.
It is instructive therefore to compare expressions Eqs. (\[eq:ScDens\], \[eq:deltaS\], \[eq:CorrelatorF\]) with the normal case. The amplitude of normal state conductance fluctuations contains only the diagrams shown in Fig. \[fig:AARRDiagrams\]. It has been shown rigorously[@BarangerStone] that dissipative normal state conductance is fully determined by products of one retarded and one advanced Green functions averaged together $G^{R}G^{A}$, compare Eq. (\[eq:ScDens\]). Therefore expanding the normal state analog of the Eq. (\[eq:ScDens\]) in terms of $G^{R}$ and $G^{A}$ (in other words the formula for conductance) the products of two retarded or two advanced Green functions in Eq. (\[eq:ScDens\]) that describe non-dissipative diamagnetic currents vanish. A normal state analog of the correlator $\mathcal{F}$ in Eq. (\[eq:CorrelatorF\]) contains therefore only the averages of the type $\langle G^{R}G^{A}G^{R}G^{A}\rangle$ resulting in the diagrams in Fig. \[fig:AARRDiagrams\] being the only diagrams contributing to conductance fluctuations. In contrast to the normal phase, the superfluid density is a thermodynamic property of a superconductor and despite Anderson theorem relating it to the normal state conductivity the same arguments do not apply. One has to be cautious and take into account a number of additional diagrams shown in Fig. \[fig:34diags\] which vanish in a normal state yet give non-zero contribution in the superconducting phase. These diagrams originate form the terms of the form $G^{R}G^{R}$ in the superconducting density Eq. (\[eq:ScDens\]). A more detailed discussion of the standard procedure of constructing the diagramatic perturbation theory can be found in Appendix \[sec:Details-of-diagrams\].
![Diagrams obtained originating from the contributions of the form $G^{R}(x_{1},x_{2})G^{R}(x_{2},x_{1})$ to each of the superfluid densities Eq. (\[eq:ScDens\]) in the correlator Eq. (\[eq:deltaS\]). These diagrams cancel in the case of normal state conductance calculation yet contribute to the superfluid density.[]{data-label="fig:34diags"}](DiagramsRR){width="0.85\columnwidth"}
The diagrams in Figs. \[fig:AARRDiagrams\] and \[fig:34diags\] can be simplified in the standard way[@LeeStoneFukuyama; @AkkermansBook] by identifying two types of blocks characterized by distinct length scales: the mean free path $\ell$ and the coherent diffusion length scale $\zeta\gg\ell$. For the diagram in Fig. \[fig:AARRDiagrams\](a) such separation is done as follows, $$\begin{aligned}
& \mathcal{F}_{1a}\equiv W^{2}\int\prod_{i=1}^{4}dx_{i}j(x_{1},x_{2})j(x_{3},x_{4})\mathcal{D}(x_{1},x_{3})\mathcal{D}(x_{4},x_{2})\nonumber \\
& \times\exp(i\xi_{1}t_{1}-i\xi_{3}t_{3}+i\xi_{4}t_{4}-i\xi_{2}t_{2}),\label{eq:DiamondDiagram}\end{aligned}$$ where $x_{i}\equiv(\mathbf{r_{i}},t_{i})$. We introduced the ’Hikami boxes’ $j(x_{1},x_{2})$ and the diffuson/Cooperon propagators $\mathcal{D}(x_{1},x_{2})$, see Fig. \[fig:AARRDiagrams\]. In the main order in $p_{F}\ell\gg1$ and $\zeta/\ell\gg1$ these blocks can be averaged over disorder independently of each other.
The diffusion propagator is defined as a joint average of two Green functions $\mathcal{D}(x_{1},x_{2})=\frac{1}{2\pi\nu\tau^{2}}\langle G^{R}(x_{1},x_{2})G^{A}(x_{2},x_{1})\rangle$ and describes the long-range coherent diffusion of electrons on the scales $|\mathbf{r_{1}}-\mathbf{r_{2}}|\gg\ell$[@AkkermansBook]. The diffuson propagators are represented diagramatically by impurity ladders of diffuson and Cooperon type[@AkkermansBook] and satisfy Dyson-like equations Figs. \[fig:Dyson+TLS\](a). In the diffusive regime Figs. \[fig:Dyson+TLS\](a) reduces to the standard diffusion equation. To model a realistic conductor this diffusion equation has to be supplemented with boundary conditions. We set $\mathcal{D}=0$ at the contacts and $\nabla\mathcal{D}=0$ at the surface of the wire. The diffusion propagators can be written in terms of the eigenmodes of the diffusion equation $\Phi_{\mathbf{q}}(\mathbf{r})=\sqrt{\frac{2^{3}}{L_{x}L_{y}L_{z}}}\sin q_{x}x\cos q_{y}y\cos q_{z}z$, $\mathbf{q}=\left[\frac{\pi n_{x}}{L_{x}},\frac{\pi n_{y}}{L_{y}},\frac{\pi n_{z}}{L_{z}}\right]$, $n_{x}=1,2,..$ $n_{y},n_{z}=0,1,2...$, $$\begin{gathered}
\mathcal{D}(t,\mathbf{r,r'})=\frac{1}{2\pi\nu\tau^{2}}\sum_{q}e^{-Dq^{2}t}\Phi_{q}(\mathbf{r})\Phi_{q}(\mathbf{r'}),\end{gathered}$$ where $t\geq0$ stands for the diffusion time, $D$ is the diffusion coefficient for electrons and $\nu$ is the density of electronic states. Averaging over the wire volume in Eq. (\[eq:DiamondDiagram\]) of the product of orthogonal eigenfunctions $\Phi_{\mathbf{q}}(\mathbf{r})$ gives rise to the momentum conservation condition (and this is true also for all other diagrams). As a result all diffusion propagators in each diagram in Figs. \[fig:AARRDiagrams\], \[fig:34diags\] have the same momentum $q$. Sums over $q$ may be approximated by integral in the case $\zeta/L_{i}\ll1,\, i=x,\, y,\, z$, $$\sum_{q}\rightarrow\frac{\Omega_{d}}{\pi^{d}\zeta^{d}}\int_{0}^{\infty}dk_{i},$$ where $\Omega_{d}=\left[V,\, L_{z}L_{y},\, L_{z}\right]$ for $d=3,2,1$. We define dimensionality $d$ of the sample with respect to the coherence length, i.e. $d=3$ corresponds to $L_{x},\, L_{y},\, L_{z}\gg\zeta$, $d=2$ corresponds to $L_{x},\, L_{z}\gg\zeta>L_{y}$ and $d=1$ corresponds to $L_{z}\gg\zeta>L_{x},\, L_{y}$.
The following relation will be useful, $$\sum_{q}q^{2m}e^{-Dq^{2}t}\approx\frac{\kappa_{m}}{2^{d}\pi^{d/2}}\frac{1}{(t\Delta){}^{d/2+m}}\frac{\Omega_{d}}{\zeta^{d}}\frac{1}{\zeta^{2m}},\label{eq:IntOverQDiffuson}$$ where $m=0,1,2$ and $\kappa_{m}=1,\frac{1}{2},\frac{3}{4}$ respectively. In the absence of magnetic fields the Cooperon and diffuson propagators are identical and we will not distinguish them in the following, instead including a factor of $2$ in front of all the diagrams.
![Dyson equations for diffusons/Cooperons (shaded regions) in time reversal symmetric case. Vertical dashed lines stand for the disorder correlator $\langle u(\mathbf{r})u(\mathbf{r'})\rangle=\frac{1}{2\pi\nu\tau}\delta(\mathbf{r-r'})$, solid lines correspond to disorder averaged Green functions. (a) Corresponds to static impurity configuration. (b) Corresponds to the case when a fraction of impurities have moved between different measurements. The additional “diamond” vertex corresponds to the correlator $\langle\left(u_{\mathbf{r_{i}}+\delta\mathbf{r}}-u_{\mathbf{r_{i}}}\right)u_{\mathbf{r_{i}}}\rangle$.[]{data-label="fig:Dyson+TLS"}](Diagrams3){width="0.5\columnwidth"}
The second type of blocks that we introduced in Eq. (\[eq:DiamondDiagram\]) are Hikami boxes $j(x_{1},x_{2})$. There are four types shown in Fig. \[fig:Hikami\], where straight lines represent disorder averaged Green functions. $j(x_{1},x_{2})$ decays exponentially on the short length scales of the order of the mean free path $|r|\sim\ell\ll L_{x},\, L_{y}$ and therefore can be approximated by a delta function $j(x_{1},x_{2})\approx j_{0}\delta(x_{1}-x_{2})$. The constant factor $j_{0}$ can be calculated in momentum space using disorder averaged Green functions. The result of this calculation for each of the vertexes in Fig. \[fig:Hikami\](a)-(d) reads, $$\begin{gathered}
j_{a}\approx\frac{4\pi\nu\tau^{3}p_{F}^{2}}{d},\\
j_{b}\approx\frac{2\pi\nu\tau^{3}p_{F}^{2}}{d},\\
j_{c}\approx-\frac{2\pi\nu\tau^{3}p_{F}^{2}}{d},\\
j_{d}^{\alpha}\approx\pm\frac{4\pi\nu\tau^{3}p_{F}^{2}}{d}\frac{q_{\alpha}}{m},\end{gathered}$$ where the expression for the last vertex depends explicitly on the momentum of the diffusion eigenmodes $\mathbf{q}$.
![Four types of Hikami boxes appearing in the diagrams in Figs. \[fig:AARRDiagrams\] and \[fig:34diags\].[]{data-label="fig:Hikami"}](HikamiBoxes){width="0.8\columnwidth"}
Using the short range character of the current vertexes we simplify Eq. (\[eq:DiamondDiagram\]) for the diagram in Fig. \[fig:AARRDiagrams\] (a),
$$\begin{aligned}
& \mathcal{F}_{1a}\equiv W^{2}j_{a}^{2}\int d\mathbf{r}d\mathbf{r'}dtdt'\mathcal{D}(t,\mathbf{r},\mathbf{r'})\mathcal{D}(t',\mathbf{r'},\mathbf{r})\nonumber \\
& \times\exp(-i(\xi_{1}-\xi_{4})t-i(\xi_{3}-\xi_{2})t').\label{eq:DiamondDiagramTwoT}\end{aligned}$$
Note that all energy dependence in the above expression is contained in Fourier factors. As a result after substituting Eq. (\[eq:DiamondDiagramTwoT\]) into Eq. (\[eq:deltaS\]) we can take integrals with respect to energy in Eq. (\[eq:deltaS\]). The expression for the diagram in Fig. \[fig:AARRDiagrams\](a) therefore can be rewritten using Eqs. (\[eq:deltaS\],\[eq:DiamondDiagramTwoT\],\[eq:IntOverQDiffuson\]),
$$\begin{gathered}
\langle\delta S^{2}\rangle_{1a}=\frac{2^{-d}\pi^{-\frac{d}{2}}W^{2}j_{a}^{2}\Delta^{2}}{(2\pi\nu\tau^{2})^{2}}\frac{\Omega_{d}}{\zeta^{d}}\int_{0}^{\infty}\frac{d\lambda d\lambda'\left(\Theta_{\lambda+\lambda'}^{(-)}\right)^{2}}{(\lambda+\lambda')^{d/2}},\label{eq:DiamFullExp}\end{gathered}$$
and introducing the Fourier integral,
$$\Theta_{\lambda+\lambda'}^{(-)}\equiv\frac{1}{\Delta}\int d\xi d\xi'\mathcal{L}_{\xi\xi'}e^{i\xi\frac{\lambda}{\Delta}-i\xi'\frac{\lambda'}{\Delta}}=-2\pi\int_{\lambda+\lambda'}^{\infty}dxK_{0}(x),\label{eq:Fourier0}$$
where $K_{n}(x),\, n=0,1,2,..$ is the modified Bessel function[@GR] and we introduced a dimensionless time $\lambda\equiv t\Delta$ measured in units of $1/\Delta$. See Appendix \[sec:FourierT0\] for details of Fourier integration. Note that the integral over $\lambda,\,\lambda'$ in the right hand side of Eq. (\[eq:DiamFullExp\]) is dimensionless and therefore simply represents a numerical coefficient.
It is instructive to show the calculation of the diagram in Fig. \[fig:AARRDiagrams\](b),
$$\begin{aligned}
& j_{b}^{2}\int d\mathbf{r}d\mathbf{r'}\mathcal{D}(t,\mathbf{r},\mathbf{r'})\mathcal{D}(t',\mathbf{r'},\mathbf{r}))\\
& \times\exp(-i(\xi_{1}-\xi_{4})t-i(\xi_{1}-\xi_{4})t').\end{aligned}$$
Including the coherence factors in Eq. (\[eq:deltaS\]) and using Eq. (\[eq:IntOverQDiffuson\]) we get,
$$\begin{gathered}
\langle\delta S^{2}\rangle_{1b}=\frac{2^{-d}\pi^{-\frac{d}{2}}W^{2}j_{b}^{2}\Delta^{2}}{(2\pi\nu\tau^{2})^{2}}\frac{\Omega_{d}}{\zeta^{d}}\int_{0}^{\infty}\frac{d\lambda d\lambda'}{(\lambda+\lambda')^{d/2}}\left(\Theta_{\lambda+\lambda'}^{(0)}\right)^{2},\label{eq:LDOSDiag}\end{gathered}$$
where crucially the Fourier integral has a distinct form from that in Eq. (\[eq:Fourier0\]),
$$\begin{aligned}
& \Theta_{\lambda+\lambda'}^{(0)}\equiv\frac{1}{\Delta}\int d\xi d\xi'\mathcal{L}_{\xi\xi'}e^{i\xi\frac{(\lambda+\lambda')}{\Delta}}\nonumber \\
& =-\pi\int_{\lambda+\lambda'}^{\infty}dx\left(K_{0}(x)+\frac{1-xK_{1}(x)}{x^{2}}\right).\label{eq:ThetaZero}\end{aligned}$$
Note that here we switched to dimensionless time $\lambda\equiv\Delta t$. The details of the integration are shown in Appendix \[sec:FourierT0\]. The rest of the diagrams can be computed in the analogous way noting that the diagrams in Fig. \[fig:34diags\] contain the third type of the Fourier transform of the coherence factor,
$$\begin{aligned}
& \Theta_{\lambda+\lambda'}^{(+)}\equiv\frac{1}{\Delta}\int d\xi d\xi'\mathcal{L}_{\xi\xi'}e^{i\xi\frac{\lambda}{\Delta}+i\xi'\frac{\lambda'}{\Delta}}\nonumber \\
& =-2\pi\int_{\lambda+\lambda'}^{\infty}dx\frac{1-xK_{1}(x)}{x^{2}}.\label{eq:ThetaPlus}\end{aligned}$$
Note that the long range asymptotic, $x\gg1$, of the Bessel function is proportional to an exponent $K_{\nu}(x)\approx\sqrt{\frac{\pi}{2x}}e^{-x}$. This means that the Fourier transform of the coherence factor Eq. (\[mathcalL\]) $\Theta_{\lambda}^{(-)}\sim e^{-\lambda}$ determines the roughly exponential phase relaxation in the diagram in Fig. \[fig:AARRDiagrams\](a), Eq. (\[eq:DiamFullExp\]). This diagram corresponds to the fluctuations in the transmission coefficient of electrons in the superconductor. In contrast, in the diagram in Fig. \[fig:AARRDiagrams\](b), Eq. (\[eq:LDOSDiag\]), $\Theta_{\lambda}^{(0)}\sim\frac{1}{\lambda}$ and therefore the coherent diffusion is only suppressed by a slow power law phase relaxation rather than an exponent. An analogous slow relaxation of diffusion modes is described by the diagrams in Fig. \[fig:34diags\]. Note that the non-exponential relaxation of the diffusion modes is precisely the reason why the diagrams in Fig. \[fig:34diags\] give non-zero contribution in contrast to the normal metal case. This can be seen by replacing each of the Fourier integrals $\Theta_{\lambda}^{(-)},\,\Theta_{\lambda}^{(0)},\,\Theta^{(+)}$ with an exponent $e^{-\lambda}$ in which case the sum of the diagrams in Fig. \[fig:34diags\] vanishes, see Appendix \[sec:Details-of-diagrams\] for details. The consequences of non-exponential relaxation is two-fold: (i) due to both the contribution of the numerous additional diagrams and slow-decaying non-exponential integrands the amplitude of the superfluid density fluctuations is enhanced; (ii) as we will see in the following the noise amplitude acquires non-linear dependence on the effective impurity strength in one and two dimensions.
Summing all the diagrams in Figs. \[fig:AARRDiagrams\],\[fig:34diags\] including the appropriate symmetry factors, see Appendix \[sec:Details-of-diagrams\], we arrive at, $$\begin{gathered}
\frac{\langle\delta S^{2}\rangle}{S^{2}}=C_{d}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{4-d}\ell^{2}\Omega_{d}}{V^{2}},\label{eq:FluctAmpResult}\\
C_{d}\equiv2^{1-d}9\pi^{-d/2}\left(\mathcal{I}^{RA}+2\mathcal{I}^{RR}\right).\label{eq:CoefficientUCF}\end{gathered}$$ where $\Omega_{d}=\left[V,L_{z}L_{y},L_{z}\right]$ for $d=3,2,1$. The explicit form of the integrals in Eq. (\[eq:CoefficientUCF\]) reads, $$\begin{gathered}
\mathcal{I}^{RA}=\int_{0}^{\infty}d\lambda_{1}d\lambda_{2}\frac{\left(\Theta_{\lambda_{1}+\lambda_{2}}^{-}\right)^{2}+\frac{1}{2}\left(\Theta_{\lambda_{1}+\lambda_{2}}^{0}\right)^{2}}{(\lambda_{1}+\lambda_{2})^{d/2}}A_{\lambda_{1}+\lambda_{2}},\label{eq:IRA}\end{gathered}$$ corresponding to the contribution of the diagrams in Fig. \[fig:AARRDiagrams\], and
$$\begin{gathered}
\mathcal{I}^{RR}=\int_{0}^{\infty}d\lambda_{1}d\lambda_{2}\frac{\left(\left(\Theta_{\lambda_{1}+\lambda_{2}}^{+}\right)^{2}+\frac{3}{2}\Theta_{\lambda_{1}+\lambda_{2}}^{+}\Theta_{\lambda_{1}+\lambda_{2}}^{0}\right)A_{\lambda_{1}+\lambda_{2}}}{(\lambda_{1}+\lambda_{2})^{d/2}}\nonumber \\
-2\int_{0}^{\infty}d\lambda_{1}d\lambda_{2}d\lambda_{3}\frac{\left(\Theta_{\lambda_{1}+\lambda_{2}+\lambda_{3}}^{+}\right)^{2}A_{\lambda_{1}+\lambda_{2}+\lambda_{3}}}{(\lambda_{1}+\lambda_{2}+\lambda_{3})^{d/2+1}}\nonumber \\
-2\int_{0}^{\infty}d\lambda_{1}d\lambda_{2}d\lambda_{3}\frac{\Theta_{\lambda_{1}+\lambda_{2}+\lambda_{3}}^{+}\Theta_{\lambda_{1}+\lambda_{2}+\lambda_{3}}^{0}A_{\lambda_{1}+\lambda_{2}+\lambda_{3}}}{(\lambda_{1}+\lambda_{2}+\lambda_{3})^{d/2+1}}\nonumber \\
+\int_{0}^{\infty}d\lambda_{1}d\lambda_{2}d\lambda_{3}d\lambda_{4}\frac{9\left(\Theta_{\lambda_{1}+\lambda_{2}+\lambda_{3}+\lambda_{4}}^{+}\right)^{2}A_{\lambda_{1}+\lambda_{2}+\lambda_{3}+\lambda_{4}}}{(\lambda_{1}+\lambda_{2}+\lambda_{3}+\lambda_{4})^{d/2+2}},\label{eq:IRR}\end{gathered}$$
corresponding to the contribution of the diagrams in Fig. \[fig:34diags\]. The factor $A_{\lambda}=1$ in Eq. (\[eq:FluctAmpResult\]). We estimate the coefficients in Eq. (\[eq:FluctAmpResult\]) by numerically taking the integrals in Eqs. (\[eq:IRA\], \[eq:IRR\]) $C_{3}\approx25,C_{2}\approx60,C_{1}\approx200$.
\[sec:Noise-amplitude\]Noise amplitude
======================================
A model of tunneling two level (bistable) defect, i.e. an impurity fluctuating between two spatial positions, is used to describe noise in a wide range of materials. For simplicity we use this model to give a precise meaning to the amplitude of the noise calculated here. All of the results can be easily generalized for a more generic dynamics of impurities [@AltshulerSpivak]. In the presence of a single bistable defect superfluid response demonstrates significant fluctuations. The noise amplitude for a single defect is given by the correlator, $$\begin{gathered}
\langle\delta S_{\delta\mathbf{r}}^{2}\rangle=\langle\delta S^{2}\rangle-\langle\delta S_{\mathbf{r_{i}+\delta}\mathbf{r}}\delta S_{\mathbf{r_{i}}}\rangle.\label{S_correlator}\end{gathered}$$ The first term on the right hand side of Eq. (\[S\_correlator\]) is simply the fluctuation amplitude Eq. (\[eq:FluctAmpResult\]), $\delta S\equiv S-\langle S\rangle$. The second term corresponds to the diagrams in Figs. \[fig:AARRDiagrams\] and \[fig:34diags\] with one of the response kernels $\delta S_{\mathbf{r_{i}+\delta}\mathbf{r}}$ containing the disorder configuration with one defect shifted from $\mathbf{r_{i}}$ by a distance $\mathbf{\delta r}$, and the other $\delta S_{\mathbf{r_{i}}}$ containing the bistable defect in its original position $\mathbf{r_{i}}$. This shift introduces an effective dephasing rate cutting off the diffusion modes. To show this we recalculate the diffuson/Cooperon in the presence of the shifted bistable defect. The Dyson equation for the diffuson/Cooperon has to be modified by including an additional vertex shown in Fig. \[fig:Dyson+TLS\](b) as an impurity line with a diamond[@Feng86; @Falko90]. This new vertex corresponds to the correlator $\langle\left(u_{\mathbf{r_{i}}+\delta\mathbf{r}}-u_{\mathbf{r_{i}}}\right)u_{\mathbf{r_{i}}}\rangle$. It is obtained in the perturbation expansion in the small parameter, $\langle\left(u_{\mathbf{r_{i}}+\delta\mathbf{r}}-u_{\mathbf{r_{i}}}\right)u_{\mathbf{r_{i}}}\rangle/\langle u_{\mathbf{r_{i}}}u_{\mathbf{r_{i}}}\rangle$. In momentum space the last term in Fig. \[fig:Dyson+TLS\](c) reads, $$u_{0}^{2}\int\frac{d\mathbf{p}d\mathbf{p'}}{(2\pi)^{6}}|G_{\mathbf{p}}^{R}|^{2}|G_{\mathbf{p'}}^{A}|^{2}\left(\cos\left[\mathbf{(p-p')\delta r}\right]-1\right)\mathcal{D}(\mathbf{q}),$$ where we keep only the main contribution in the gradient expansion of the diffusion mode. This gives, $$\begin{gathered}
\left(-i\omega+Dq^{2}+\tau_{eff}^{-1}\right)\mathcal{D}(q)=1,\\
\tau_{eff}^{-1}=\frac{1}{\tau}\alpha(k_{F}\delta r),\\
\alpha(k_{F}\delta r)\equiv\left(1-\frac{\sin^{2}\left(k_{F}\delta r/2\right)}{(k_{F}\delta r/2)^{2}}\right).\end{gathered}$$ In the presence of more than one thermally activated defect their contributions simply add up for small enough concentration of such defects $\gamma_{T}$,
$$\tau_{eff}^{-1}\approx\gamma_{T}\frac{1}{\tau}\langle\alpha(k_{F}\delta r)\rangle,\label{eq:Dephasing}$$
which includes an average over characteristics of bistable defects $\langle\alpha(k_{F}\delta r)\rangle$. For defects characterized by $k_{F}\delta r\gg1$, $\langle\alpha(k_{F}\delta r)\rangle\approx1$. The effect of bistable defects on the vertex parts of the diagrams $j(r_{1},r_{2})$ is small as $\ell/\zeta$, and therefore we neglect it in the following. The main order amplitude of the noise in the superfluid density Eq. (\[S\_correlator\]) is given by the same diagrams in Figs. \[fig:AARRDiagrams\], \[fig:34diags\]. However, in the case of the second term in Eq. (\[S\_correlator\]) we need to include the dephasing effect of thermally activated defects Eq. (\[eq:Dephasing\]). The result is,
$$\begin{gathered}
\delta_{K}^{2}=\widetilde{C}_{d}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{4-d}\ell^{2}\Omega_{d}}{V^{2}},\\
\widetilde{C}_{d}\equiv9\pi^{-d/2}2^{1-d}\left(\mathcal{I}^{RA}(\tau_{eff})+2\mathcal{I}^{RR}(\tau_{eff})\right),\label{eq:IntegralNoise}\end{gathered}$$
where the integrals $\mathcal{I}^{RA}(\tau_{eff}),\,\mathcal{I}^{RR}(\tau_{eff})$ are given by Eq. (\[eq:IRA\], \[eq:IRR\]) with the factor $A_{t}$ modified to include the effect of bistable defects, $$\begin{gathered}
A_{\lambda}\equiv1-e^{-\lambda/(\Delta\tau_{eff})}.\label{exponent}\end{gathered}$$ The first and the second term in Eq. (\[exponent\]) correspond to the first and the second terms in Eq. (\[S\_correlator\]). The factor $\Omega_{d}$ has the same meaning as in Eq. (\[eq:FluctAmpResult\]).
The amplitude of the noise in $d=3$ can be estimated by expanding the exponent in Eq. (\[exponent\]), $A_{\lambda}\approx\frac{\lambda}{\Delta\tau_{eff}}$, resulting in, $$\begin{gathered}
\delta_{K}^{2}\approx60\gamma_{T}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{3}}{V},\label{eq:Noise3DR}\end{gathered}$$ where we took the integrals in Eq. (\[eq:IntegralNoise\]) numerically to estimate the value of the coefficient.
In lower dimensions the simple expansion of the factor in Eq. (\[exponent\]) is not possible since the integral over $t$ diverges. Instead we keep only the main asymptotic contribution in the parameter $\Delta\tau_{eff}=3\zeta^{2}/(\gamma_{T}\ell^{2})\gg1$ which gives,
$$\delta_{K}^{2}=\frac{297}{4\pi}\gamma_{T}\log\left[\frac{3\zeta^{2}}{\gamma_{T}\ell^{2}}\right]\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{4}}{VL_{y}},\label{eq:Noise2DR}$$
in effective two dimensional superconductor, $L_{y}\lesssim\zeta\ll L_{x},L_{z}$, and
$$\delta_{K}^{2}=99\left(3\gamma_{T}\right)^{\frac{1}{2}}\frac{1}{(p_{F}\ell)^{4}}\frac{\zeta^{4}\ell}{VS_{\perp}},\label{eq:Noise1DR}$$
in effective one dimension, $L_{x},L_{y}\lesssim\zeta\ll L_{z}$. The results Eqs. (\[eq:Noise3DR\]-\[eq:Noise1DR\]) are valid only in the limit of very small relative concentration of bistable defects $\gamma_{T}\ll1$ such that the noise amplitude is much smaller than the sample-to-sample fluctuation Eq. (\[eq:FluctAmpResult\]) which sets the upper limit on the noise amplitude. At high concentration of fluctuating defects the noise amplitude saturates at the value of sample-to-sample fluctuations Eq. (\[eq:FluctAmpResult\]).
Conclusion\[sec:Conclusion\]
============================
We have performed the detailed analysis of kinetic inductance fluctuations caused by motion of impurities in the superconductor and enhanced by the mesoscopic quantum interference. We found that the effect is closely related to the universal conductance fluctuations in normal metals with an important distinction that the superconducting coherence length determines the scale of coherent diffusion, in contrast to the inelastic scattering length in the normal metal. We found that the phase relaxation in superconductors has slow non-exponential character which results in the enhanced amplitude of the noise and non-linear scaling of the amplitude with the density of fluctuating impurities. Our estimates of the magnitude of the noise suggest that the effect is likely to dominate inductance noise in small devices. Experimentally, the interference contribution to the inductance noise can be unambiguously identified by driving supercurrents close to critical in magnitude through the wire. In presence of strong supercurrent the superconducting order parameter phase changes by $2\pi$ on the scale of the coherence length which results in strong suppression of the quantum interference and therefore the noise amplitude.
The upper limit of the noise amplitude is given by the mesoscopic fluctuation amplitude. We estimate this amplitude for an Al wire with dimensions $100\textrm{nm}\times100\textrm{nm}\times10\mu\textrm{m}$ using Eq. (\[eq:FluctAmpResult\]), $\frac{\sqrt{\left\langle \delta K^{2}\right\rangle }}{K}\sim10^{-5}$.
In normal state metals the mesoscopic noise amplitude was found to be roughly temperature independent. This was attributed to a rough cancellation of the temperature dependence of the density of thermally activated defects $\gamma_{T}\propto T$ and the temperature smearing of the mesoscopic fluctuation which roughly suppresses the amplitude as $\propto1/T$ [@Birge1989]. In contrast, in the very low temperature regime of transport in disordered superconductors considered here the mesoscopic noise is limited by the superconducting coherence length which demonstrates very weak temperature dependence away from the transition temperature. As a result we expect the mesoscopic noise in the kinetic inductance of disordered superconducting wires in presence of thermally activated impurity dynamics to scale linearly with with temperature $\propto T$ in the effective $d=3$ case and $\propto T\log T$ in $d=2$ and $\propto\sqrt{T}$ in $d=1$. This prediction could be tested experimentally in superconducting thin films and wires.
The author is grateful to Lev Ioffe for motivating author’s interest in this problem and numerous comments and discussions. The author is also grateful to Vladimir Fal’ko, Robert Smith, and Igor Lerner for helpful discussions. This work is supported by LPS-CMTC.
[33]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1063/1.2155757) [****, ()](\doibase 10.1038/nature07128) [****, ()](http://dx.doi.org/10.1109/JPROC.2004.833670) and , eds., @noop [**]{} (, ) and , eds., @noop [**]{} (, ) [****, ()](http://dx.doi.org/10.1103/PhysRevLett.95.210503) [****, ()](http://dx.doi.org/10.1103/PhysRevB.77.180502) [****, ()](http://dx.doi.org/10.1126/science.1175552) [****, ()](\doibase 10.1038/nature02851) [ ()](http://arxiv.org/abs/1306.4992) [****, ()](\doibase
10.1088/0957-4484/21/44/445202) [****, ()](\doibase 10.1146/annurev-conmatphys-020911-125022) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.56.2772.2) [****, ()](\doibase 10.1103/PhysRevB.35.1039) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.85.214507) [****, ()](\doibase 10.1103/PhysRevB.45.2463) [****, ()](http://dx.doi.org/10.1103/PhysRevLett.62.195) [****, ()](http://dx.doi.org/10.1103/PhysRevB.42.2735) [****, ()](http://dx.doi.org/10.1103/PhysRevB.47.16667) (, ) pp. [****, ()](\doibase
10.1103/PhysRevLett.103.117001) [ ()](http://arxiv.org/abs/1102.3445) @noop [****, ()]{} @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRev.108.1175) [****, ()](\doibase 10.1103/PhysRevB.40.8169) and , eds., @noop [**]{} (, ) and , eds., @noop [**]{} (, ) @noop [****, ()]{}
Details of the diagrams evaluation\[sec:Details-of-diagrams\]
=============================================================
Fig. 1a 1b 2a 2b 2c 2d 2e 2f 2g
-------- ---- ---- ---- ---- ---- ---- ---- ---- ----
factor 2 4 4 8 8 8 16 8 12
: Combinatorial factors corresponding to diagrams of different topologies shown in Fig. \[fig:AARRDiagrams\], \[fig:34diags\] contributing to the kinetic inductance fluctuations and the noise amplitude.[]{data-label="table"}
Fluctuations of the superfluid response are given by the disorder, spatial and energy average of a correlator $\mathcal{F}$ (see Eq. (\[eq:deltaS\])) of four matrix elements of the current operator. This correlator can be rewritten in terms of four imaginary parts of the exact Green functions, see Eq. (\[eq:ScDens\]),
$$\begin{aligned}
& \mathcal{F}\propto\left\langle \Delta G\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}\Delta\overline{G}\Delta G\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}\Delta\overline{G}\right\rangle \label{eq:SvImG}\end{aligned}$$
The diagrams in Figs. \[fig:AARRDiagrams\], \[fig:34diags\] are constructed by pairing Green functions to form diffuson and Cooperon ladders, which are products of one retarded and one advanced Green functions averaged over disorder $\left\langle G^{R}G^{A}\right\rangle $. All possible such connections give rise to the diagrams in Figs. \[fig:AARRDiagrams\], \[fig:34diags\]. In particular, the average of the form,
$$\left\langle G^{R}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{A}G^{R}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{A}\right\rangle \label{eq:GRGA}$$
gives rise to the two diagrams shown in Fig. \[fig:AARRDiagrams\](a),(b). These are the diagrams that correspond to normal state conductance fluctuations[@LeeStoneFukuyama; @AkkermansBook; @BarangerStone]. The averages of the form,
$$\left\langle G^{A}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{A}G^{R}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{R}\right\rangle$$
give rise to the diagrams in Figs. \[fig:34diags\](a),(b),(d),(f) and (g). Finally, the averages of the form,
$$-\left\langle G^{R}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{A}G^{R}\overleftrightarrow{\nabla}\overleftrightarrow{\nabla'}G^{R}\right\rangle$$
give rise to the diagrams shown in Fig. \[fig:34diags\](c) and (e). Crucially, the latter expression comes with a different sign from expanding the expression in Eq. (\[eq:SvImG\]).
Each diagram in Figs. \[fig:AARRDiagrams\] and \[fig:34diags\] comes with a different “combinatorial factor” reflecting the number of times it occurs in the perturbation expansion of $\left\langle \delta S^{2}\right\rangle $. This factor counts the number of ways to choose pairs of $\langle G^{R}G^{A}\rangle$ such that the two Green functions are parts of different superfluid density loops. The corresponding factors are summarized in Table \[table\].
In the case of mesoscopic fluctuations of the normal state conductance it has been shown[@BarangerStone] that only the diagrams constructed from $G^{R}G^{A}$ loops (Eq. (\[eq:GRGA\]) and Fig. \[fig:AARRDiagrams\]) give non-zero contribution. Since the symmetry factors and the Hikami boxes in all of the diagrams constructed using $G^{R}G^{R}$ loops (Fig. \[fig:34diags\]) are the same as in the normal state we expect exact cancellation of the contribution of such diagrams to the normal state conductance fluctuations. To very that this contribution indeed vanishes we set all Fourier transformed coherence factors to be equal to an exponent $\Theta_{\lambda}^{(0)},\Theta_{\lambda}^{(+)},\Theta_{\lambda}^{(-)}\rightarrow e^{-\lambda/\tau_{\varphi}}$ describing the effect of decoherence due to inelastic electron-electron or electron-phonon collisions operational in the normal state. The resulting sum of the integrals vanishes exactly $\mathcal{I}{}^{RR}(\tau_{\varphi})=0$. Note that the non-exponential decay of $\Theta_{\lambda}^{(0)},\Theta_{\lambda}^{(+)},\Theta_{\lambda}^{(-)}$ is precisely the reason for non-vanishing contribution of the diagrams in Fig. \[fig:34diags\] to the superfluid response fluctuations.
Fourier integrals of the coherence factor \[sec:FourierT0\]
===========================================================
First type integral
-------------------
In the following we calculate the Fourier integral, $$\begin{gathered}
\Theta_{t+t'}^{(-)}=\frac{1}{\Delta}\int_{-\infty}^{\infty}d\xi d\xi'\delta\mathcal{L}_{\xi\xi'}e^{-i\xi t+i\xi't'},\end{gathered}$$ in Eq. (\[eq:DiamFullExp\]), see also Ref. . Note that in this appendix we perform the calculations using the real time as opposed to the dimensionless time $\lambda$. We introduce the latter at the end of the calculation. We introduce notations symmetric with respect to $\xi\rightarrow-\xi$,
$$\mathcal{L}_{\xi\xi'}=\frac{1}{2}\frac{F(\xi)-F(\xi')+\xi\xi'\left[G(\xi)-G(\xi')\right]}{\xi^{2}-\xi'^{2}}.\label{eq:LvFG}$$
where the symmetric functions $F(\xi)=F(-\xi)$ and $G(\xi)=G(-\xi)$ are defined as,
$$\begin{aligned}
& F(\xi)=\left(E+\frac{\Delta^{2}}{E}\right)\left[1-2f(E)\right]-\xi\left[1-2f(\xi)\right],\label{eq:F}\\
& G(\xi)=\frac{1}{E}\left[1-2f(E)\right]-\frac{1}{\xi}\left[1-2f(\xi)\right],\label{eq:G}\end{aligned}$$
note that $f(-\xi)=1-f(\xi)$.
Using Eqs. (\[eq:LvFG\],\[eq:F\],\[eq:G\]) we write, $$\begin{aligned}
& \Theta_{t+t'}^{(-)}=\frac{2}{\Delta}\int_{0}^{\infty}d\xi d\xi'\frac{\cos\xi t\cos\xi't'\left[F(\xi)-F(\xi')\right]}{\xi^{2}-\xi'^{2}}\nonumber \\
& +\frac{2}{\Delta}\int_{0}^{\infty}d\xi d\xi'\frac{\xi\xi'\sin\xi t\sin\xi't'\left[G(\xi)-G(\xi')\right]}{\xi^{2}-\xi'^{2}}\label{eq:ThetaMinusSym}\end{aligned}$$ We replace the integral in the above expression by two principal value (at point $\xi=\xi'$) integrals, $$\begin{gathered}
A_{1}=2\lim_{a\rightarrow0,b\rightarrow\infty}\mathcal{P}'\int_{a}^{b}\int_{a}^{b}d\xi d\xi'\frac{\cos\xi t\cos\xi't'F(\xi)}{\xi^{2}-\xi'^{2}},\label{eq:A1even}\\
A_{2}=2\lim_{a\rightarrow0,b\rightarrow\infty}\mathcal{P}'\int_{a}^{b}\int_{a}^{b}d\xi d\xi'\frac{\xi\xi'\sin\xi t\sin\xi't'G(\xi)}{\xi^{2}-\xi'^{2}}.\label{eq:A1odd}\end{gathered}$$ Since $F(\xi)\rightarrow0$ as $\xi\rightarrow\infty$ the value of the integral does not depend on the way the upper limit is approached. Therefore we can set $b\rightarrow\infty$ right away. Care must be taken with the lower limit, for the integral in Eq. (\[eq:A1even\]), $$\begin{gathered}
\int_{a}^{\infty}d\xi'\frac{\cos\xi't'}{\xi^{2}-\xi'^{2}}=\int_{0}^{\infty}d\xi'\frac{\cos\xi't'}{\xi^{2}-\xi'^{2}}-\int_{0}^{a}d\xi'\frac{1}{\xi^{2}-\xi'^{2}},\label{eq:PVint}\end{gathered}$$ where in the last expression we assumed $a$ small enough so that $\cos\xi't\approx1$. We evaluate the second term in Eq. (\[eq:PVint\]), $$\begin{gathered}
\int_{0}^{a}d\xi'\frac{1}{\xi^{2}-\xi'^{2}}=\frac{1}{2\xi}\ln\frac{\xi+a}{\xi-a},\end{gathered}$$ which is possible since $\xi>a$. In the first term in Eq. (\[eq:PVint\]) the principal value integral gives, $$\begin{gathered}
\mathcal{P}'\int_{0}^{\infty}d\xi'\frac{\cos\xi't'}{\xi^{2}-\xi'^{2}}=\frac{\pi}{2}\frac{\sin\xi t'}{\xi},\end{gathered}$$ A similar integral in Eq. (\[eq:A1odd\]) gives, $$\begin{gathered}
\mathcal{P}'\int_{0}^{\infty}d\xi'\frac{\xi'\sin\xi't'}{\xi^{2}-\xi'^{2}}=-\frac{\pi}{2}\cos\xi t'.\end{gathered}$$ So that $$\begin{aligned}
& A_{1}+A_{2}=-\lim_{a\rightarrow0}\int_{a}^{\infty}\frac{d\xi}{\xi}F(\xi)\ln\frac{\xi+a}{\xi-a}\\
& +2\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\cos\xi tF(\xi)\frac{\pi}{2}\frac{\sin\xi t'}{\xi}\\
& +2\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\xi\sin\xi tG(\xi)\left(-\frac{\pi}{2}\cos\xi t'\right).\end{aligned}$$ Similar expressions arise in Eq. (\[eq:ThetaMinusSym\]) where the integral over $\xi$ is taken first. Combining all the results we get, $$\begin{gathered}
\Theta_{t+t'}^{(-)}=-\frac{2}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}\frac{d\xi}{\xi}F(\xi)\ln\frac{\xi+a}{\xi-a}\\
+\frac{\pi}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi F(\xi)\cos\xi t\frac{\sin\xi t'}{\xi}\\
-\frac{\pi}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\xi G(\xi)\sin\xi t\cos\xi t'\\
+\frac{\pi}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi F(\xi)\cos\xi t'\frac{\sin\xi t}{\xi}\\
-\frac{\pi}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\xi G(\xi)\sin\xi t'\cos\xi t,\end{gathered}$$ where we have replaced $\xi'$ with $\xi$ in the last two terms. Simplifying this expression we get, $$\begin{aligned}
& \Theta_{t+t'}^{(-)}=-\frac{2}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}\frac{d\xi}{\xi}F(\xi)\ln\frac{\xi+a}{\xi-a}\\
& +\frac{\pi}{\Delta}\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\left[F(\xi)-\xi^{2}G(\xi)\right]\\
& \times\frac{\cos\xi t'\sin\xi t+\sin\xi t'\cos\xi t}{\xi}.\end{aligned}$$ In the first term in the latter, as we take the limit we can take $F(\xi)\rightarrow F(0)=2\Delta\left[1-2f(\Delta)\right]$ and calculate the integral, $$\begin{gathered}
\int_{1}^{\infty}dx\frac{1}{x}\ln\frac{x+1}{x-1}=\frac{\pi^{2}}{4}.\end{gathered}$$ In the second term we take into account that, $$\begin{gathered}
F(\xi)-\xi^{2}G(\xi)=\frac{2\Delta^{2}}{E}\left[1-2f(E)\right].\end{gathered}$$ We get, $$\begin{aligned}
& \Theta_{t+t'}^{(-)}=-\pi^{2}\Delta\left[1-2f(\Delta)\right]\\
& +2\Delta^{2}\pi\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi\frac{\left[1-2f(E)\right]}{E}\frac{\sin\xi(t+t')}{\xi}.\end{aligned}$$ where the limit operation can be dropped in the last expression. At $T=0$ this simplifies further to, $$\begin{aligned}
& \Theta_{t+t'}^{(-)}=-\pi^{2}\\
& +2\Delta\pi\int_{0}^{\infty}d\xi\frac{1}{E}\frac{\sin\xi(t+t')}{\xi}.\end{aligned}$$ The second term in the last expression can be expressed in terms of a modified Bessel function, $$\begin{gathered}
B_{1}(x)\equiv\int_{0}^{\infty}d\xi\frac{1}{\sqrt{\xi^{2}+\Delta^{2}}}\frac{\sin\xi x}{\xi},\end{gathered}$$ taking the derivative w.r.t. $x$, $$\begin{gathered}
\frac{\partial}{\partial x}B_{1}=\int_{0}^{\infty}d\xi\frac{1}{\sqrt{\xi^{2}+\Delta^{2}}}\cos\xi x=K_{0}(\Delta x).\end{gathered}$$ Therefore we get, $$\begin{gathered}
\Theta_{t+t'}^{(-)}=-\pi^{2}+2\Delta\pi\int_{0}^{t+t'}dxK_{0}(\Delta x)\\
=-2\pi\int_{\Delta(t+t')}^{\infty}dxK_{0}(x).\end{gathered}$$
Introducing the dimensionless time $\lambda\equiv\Delta t$ we arrive at the expression in the main text.
Second type integral
--------------------
We show details of the Fourier integral of the coherence factor shown in Eq (\[eq:LDOSDiag\]), $$\begin{gathered}
\Theta_{t}^{(0)}=\frac{1}{\Delta}\int d\xi d\xi'\mathcal{L}_{\xi\xi'}e^{-i\xi t},\end{gathered}$$ see also Ref. . As above we perform the calculations using the real time as opposed to the dimensionless time $\lambda$ and introduce the latter at the end of the calculation.
Using the symmetric notations for the coherence factor $\mathcal{L}_{\xi\xi'}$ we write, $$\begin{gathered}
\Theta_{t}^{(0)}=\frac{2}{\Delta}\int_{0}^{\infty}d\xi d\xi'\frac{F(\xi)-F(\xi')}{\xi^{2}-\xi'^{2}}\cos\xi t,\label{eq:Thet0Int}\end{gathered}$$ where the integral is taken in the principal value sense at $\xi=\xi'$,
$$\mathcal{P}\int_{0}^{\infty}d\xi d\xi'...=\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi d\xi'...$$ In the first term in Eq. (\[eq:Thet0Int\]) we can integrate right away, $$\begin{gathered}
\mathcal{P}'\int_{0}^{\infty}d\xi d\xi'\frac{F(\xi)}{\xi^{2}-\xi'^{2}}\cos\xi\tau=-\frac{\pi^{2}}{8}F(0).\label{eq:1stInt}\end{gathered}$$ The second term can be integrated over $\xi$, $$\begin{gathered}
\mathcal{P}'\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi d\xi'\frac{F(\xi')}{\xi^{2}-\xi'^{2}}\cos\xi\tau\nonumber \\
=-\lim_{a\rightarrow0}\int_{a}^{\infty}d\xi'F(\xi')\left[\frac{\pi}{2}\frac{\sin\xi'\tau}{\xi'}-\frac{1}{2\xi'}\ln\frac{\xi'+a}{\xi'-a}\right]\nonumber \\
=\frac{\pi^{2}}{8}F(0)-\frac{\pi}{2}\int_{0}^{\infty}d\xi'F(\xi')\frac{\sin\xi'\tau}{\xi'}.\label{eq:2ndInt}\end{gathered}$$ Substituting Eqs. (\[eq:1stInt\],\[eq:2ndInt\]) into Eq. (\[eq:Thet0Int\]) we get, $$\begin{gathered}
\Theta_{t}^{(0)}=-\frac{\pi^{2}}{2\Delta}F(0)+\frac{\pi}{\Delta}\int_{0}^{\infty}d\xi'F(\xi')\frac{\sin\xi't}{\xi'}.\end{gathered}$$
The last integral can be taken explicitly at $T=0$, simplifying,
$$\begin{aligned}
& \Theta_{t}^{(0)}=-\pi^{2}\\
& +\pi\Delta\int_{0}^{\infty}d\xi\left[\sqrt{\xi^{2}+1}-\xi+\frac{1}{\sqrt{\xi^{2}+1}}\right]\frac{\sin\left(\xi t\Delta\right)}{\xi}.\end{aligned}$$
Using modified Bessel functions to represent the integrals[@GR], $$\begin{gathered}
\int_{0}^{\infty}d\xi\frac{\cos\xi x}{\xi+\sqrt{\xi^{2}+1}}=\frac{1-xK_{1}(x)}{x^{2}},\\
\int_{0}^{\infty}dxK_{1}(x)=\frac{\pi}{2},\end{gathered}$$ we obtain, $$\begin{gathered}
\Theta_{t}^{(0)}=-\pi\int_{\Delta t}^{\infty}dx\left[K_{0}(x)+\frac{1-xK_{1}(x)}{x^{2}}\right].\end{gathered}$$ Introducing dimensionless time $\lambda\equiv\Delta t$ we arrive at Eq. (\[eq:ThetaZero\]).
The third type of the Fourier integral of the coherence factor arising in Eq. (\[eq:ThetaPlus\]) is obtained in the analogous way.
[^1]: Here we assume that the observation time $t\rightarrow \infty$ is long enough. The noise amplitude $\delta_K^2$ can be connected to the noise power spectra typically measured experimentally. We assume that scattering centers responsible for the noise can be modelled by a system of bistable tunneling defects with a Lorentzian noise spectrum characterized by a tunneling time $\tau$ with the distribution $g(\tau)$, $\mathcal{P}_K(\omega)=\delta_K^2\int d\tau g(\tau)\frac{2\tau}{1+(\omega\tau)^2}$. For a typically considered model of $g(\tau)=A/\tau$ for $\tau_{min}<\tau<\tau_{max}$ and the normalization coefficient given by $A=1/(\pi\ln(\tau_{max}/\tau_{min}))$. For this model the amplitude of the noise can be related to the power spectra, $
\mathcal{P}_K(\omega)=\delta_K^2\frac{1}{\ln(\tau_{max}/\tau_{min})} \frac{1}{\omega}.
$ This defines the meaning of $\gamma_T$ which is the density of defects integrated over a broad bandwidth.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We examine the relation between oxygen abundances in the narrow-line regions (NLRs) of active galactic nuclei (AGNs) estimated from the optical emission lines through the strong-line method (the theoretical calibration of @thaisa98), via the direct $T_e$-method, and the central intersect abundances in the host galaxies determined from the radial abundance gradients. We found that the $T_e$-method underestimates the oxygen abundances by up to $\sim$2 dex (with average value of $\sim 0.8$ dex) compared to the abundances derived through the strong-line method. This confirms the existence of the so-called “temperature problem” in AGNs. We also found that the abundances in the centres of galaxies obtained from their spectra trough the strong-line method are close to or slightly lower than the central intersect abundances estimated from the radial abundance gradient both in AGNs and Star-forming galaxies. The oxygen abundance of the NLR is usually lower than the maximum attainable abundance in galaxies ($\sim$2 times the solar value). This suggests that there is no extraordinary chemical enrichment of the NLRs of AGNs.'
date: 'Accepted 2015 Month 00. Received 2015 Month 00; in original form 2014 December 17'
title: 'On the central abundances of Active Galactic Nuclei and Star-forming Galaxies'
---
\[firstpage\]
galaxies: general – galaxies: evolution – galaxies: abundances – galaxies: formation– galaxies: ISM
Introduction
============
Active galactic nuclei (AGNs) and star-forming regions show prominent emission lines of heavy elements that can be easily measured, even for objects at large redshifts. The intensity of these emission lines depends on the metallicity, which makes them convenient tracers of the chemical evolution of the Universe.
The abundance of a given element in AGNs and in star-forming regions can be derived from measurements of the relative strengths of the emission lines of its ions and of the electron temperature and density of the gas [@osterbrock06]. Oxygen has been generally used as tracer of the metallicity ($Z$) of star-forming regions [e.g., @tremonti04; @Thuan2010; @nagao06a] and of AGNs [@izotov10; @groves06a among others]. Methods for the abundance determinations in star-forming regions have been discussed in many papers [e.g., @stasinska04]. The consensus is that [*bona-fide*]{} determinations of $Z$ are only obtained by using the $T_{\rm e}$-method [@hagele06; @hagele08]. This method is based on the determination of the electron temperature ($T_{\rm e}$) from the ratios between intensities of two emission lines originating in transitions from two levels with different excitation energies of a same ion, such as the ratios \[\]($\lambda$4959+$\lambda$5007)/$\lambda$4363, \[\]($\lambda$9069+$\lambda$9532)/$\lambda$6312, and \[\]($\lambda$6548+$\lambda$6584)/$\lambda$5755 (see @osterbrock06.) Unfortunately, these line ratios cannot easily be measured in distant objects and/or in objects with low excitation (e.g., @bresolin05), preventing the use of the $T_{\rm e}$-method.
Therefore, the oxygen abundances in H[ii]{} regions are usually estimated using the strong-line method pioneered by @pagel79 and @alloin79. The principal idea of the strong-line method is to establish the relation between the (oxygen) abundance in an H[ii]{} region and some combination of the intensities of strong emission lines in its spectrum, i.e., the position of the objects in the strong-line diagram is calibrated in terms of abundance. Therefore such a relation is often called a “calibration”. The calibrations of strong emission lines in terms of the oxygen abundances can be accomplished by three methods: (i) theoretical calibrations based on photoionization models [e.g., @kewley02; @dors11; @dors05], (ii) empirical calibrations based on H[ii]{} regions with abundances derived through the $T_{\rm e}$-method [@leonid00; @pilyugin01; @pilyugin2005; @perez09 among others], and (iii) hybrid calibrations where both photoionization models and H[ii]{} regions with $T_{e}$-based abundances are used as calibration data points [@pettini04]. The metallicity scales produced by these different methods may differ from each other by up 0.7 dex [@kewley08].
Here we use the metallicity scale in H[ii]{} regions defined by the abundances determined through the classical $T_{e}$-method. This is because in the $T_{e}$ method the physical conditions in the nebulae, which are essential ingredients in order to calculate the abundance, are derived directly from observations. Otherwise, if the physical conditions and consequently the abundances are derived by models they will be less certain. This is because of the parameters of the model such as the ionizing source, geometry, etc., tend not to be sufficiently realistic (see @dors05 [@kennicutt03; @viegas02]).
AGN metallicities are usually estimated through strong-line theoretical calibrations [e.g., @thaisa98]. Even solar or supersolar abundances have been found. @richardson14 established a sequence of photoionization models to reproduce the narrow-line regions (NLRs) of AGN spectra and found that models with a metallicity of $1.4 \: Z_{\odot}$ provide the best agreement with the observational data. @groves06a used the photoionization models to analyse the emission lines in the spectra of the NLRs from the Sloan Digital Sky Survey (see @tremonti04) and found supersolar metallicities for typical Seyfert galaxies (see also @batra14 [@du14; @wang11; @dhanda07; @baldwin03; @hamann02; @thaisa98; @ferland96; @hamann93; @hamann92]). High metallicities at the centres of spiral galaxies are also obtained by the extrapolation of radial abundance gradients to the central regions [@VilaCostas1992; @Zaritsky1994; @vanZee1998; @pily04; @pilyugin2007; @Gusev2012 among others].
In contrast, low metallicities in AGNs have been obtained when the $T_{\rm e}$ method was used [e.g., @zhang13]. The disagreement between the abundances obtained through the strong-line methods and through the $T_{\rm e}$-method is the so-called “temperature problem”. On the other hand, @izotov08 determined the abundances through the $T_{\rm e}$-method in the AGNs in four dwarf galaxies and found metallicities between 0.05 and 0.2 of the solar value, which is a typical abundance of those galaxies. @alloin92, also using the $T_{\rm e}$-method, estimated the abundance in the Seyfert 2 nucleus of the galaxy ESO138G1 classified as E/S0 [@lauberts82] and found $Z\approx0.4 \: Z_{\odot}$. Thus, the $T_{\rm e}$ method seems to produce realistic abundance estimations in the low metallicity range, i.e., the temperature problem seems to disappear at low metallicities. Finding the correct method for the determination of AGN metallicities and establishing a relation between the AGN metallicity and the metallicity of the host galaxy are very important challenges. It should be noted that the AGN metallicity is sometimes adopted as a surrogate metallicity of the host galaxy, mainly at high redshifts (e.g., @dors14 [@matsuoka09; @nagao06b]).
Nowadays, a large number of measured intensities of emission lines sensitive to the electron temperature are available for AGNs in the literature. This also holds for determinations of oxygen gradients in a large sample of spiral galaxies. This provides the possibility for an analysis of the metallicity determinations in AGNs. The main goals of the current study are: (i) To investigate the discrepancy between AGN abundances derived through the strong-line methods and through the $T_{\rm e}$ method. The magnitude of this discrepancy is well known for star-forming regions [@pilyugin03; @kewley08; @lopez10 among others], however, this problem has received less attention in the case of the NLRs of AGNs. (ii) To compare the abundances of the NLRs of AGNs and nuclear star-forming regions estimated from their spectra and central oxygen abundances of the host galaxies derived from the radial abundance gradients. Differences between these values could be evidence in favour of the existence of gas infall onto the center of a galaxy or extraordinary chemical enrichment of these objects.
The present study is organised as follows. In Section 2 we compare the NLR abundances determined through the $T_{e}$-method and the strong emission-line relations of @thaisa98 [hereafter SB98]. We examine the relation between the abundances in the NLRs estimated from their optical emission-line spectra and the central intersect abundances in the host galaxies determined from the radial abundance gradients in Section 3. The conclusions are given in Section 4.
Abundance determinations in NLRs: $T_{e}$-method vs. relations by SB98
======================================================================
AGN sample {#sdata}
----------
Intensities of narrow emission lines of AGNs were compiled from the literature. Our selection criterion was the presence of flux measurements of the narrow optical emission lines \[\]$\lambda$3727, \[\]$\lambda$4363, H$\beta$, \[\]$\lambda$5007, H$\alpha$, \[\]$\lambda$6584, and \[\]$\lambda\lambda$6717,6731. We only considered galaxies whose nuclei were classified as Seyfert 2 (Sy2) and Sy1.9. Sequences of photoionization models, not including the shock gas but only a power law radiation of the ionizing source, are able to reproduce measurements of strong narrow emission lines of Sy2 and Sy1.9 for a large spectral range (e.g., @riffel13a [@dors12; @groves06a]). Hence, the selection of these AGN types minimises the effects of shock gas excitations and ionizations, which are not considered in our abundance determinations.
We list in Table \[tab1\] the identification of the objects, the emission-line intensities (relative to H$\beta$=1.0), the spectral classification taken from the [ned]{}[^1], and bibliographic references from which the data were taken. The sample consists of 47 local AGNs (redshift $z<0.1$) observed with long-slit spectroscopy. In the cases where the reddening correction was not performed in the original works (indicated in Table. \[tab1\]), we dereddened the emission-lines comparing the observed H$\alpha$/H$\beta$ ratio with the theoretical value of 2.86 [@hummer87], obtained for an electron temperature of 10000 K and an electron density of $100\: \rm cm^{-3}$.
The emission-line flux errors are reported in the original works for only 4 out of the 47 objects of our sample. Therefore the errors are not given in Table \[tab1\]. Typical errors of the emission-line intensities are about 10 to 20 per cent (e.g., @kraemer94), which yield uncertainties in the oxygen abundance estimations of about 0.1 to 0.2 dex (e.g., @rosa14 [@hagele08; @dors11; @kennicutt03]). Hereafter we will assume that the oxygen estimations from the $T_{e}$-method and the relations by SB98 have an uncertainty of 0.15 dex, an average of the values above. For the objects which the flux of the line \[\]$\lambda$4959 is not available, this was calculated from the theoretical relation $I$\[\]$\lambda$4959=$I$\[\]$\lambda$5007/3.0.
Object \[\]$\lambda$3727 \[\]$\lambda$4363 \[\]$\lambda$4959 \[\]$\lambda$5007 H$\alpha$ \[\]$\lambda$6584 \[\]$\lambda$6717 \[\]$\lambda$6731 Type Ref.
---------------------- ------------------- ------------------- ------------------- ------------------- ----------- ------------------- ------------------- ------------------- ------- ------
NGC7674 1.29 0.12 3.94 12.55 3.70 3.68 0.54 0.64 Sy2 1
IZw92 2.63 0.32 3.52 10.12 2.42 0.97 0.37 0.40 Sy2 1
NGC3393$^{(a), (b)}$ 2.41 0.14 5.47 16.42 2.78 4.50 0.64 0.89 Sy2 2
Mrk176 3.54 0.32 4.69 14.36 2.81 2.99 0.56 0.54 Sy2 3
3c33 4.93 0.32 4.10 12.68 2.63 1.76 0.87 0.73 Sy2 3
Mrk3 3.52 0.24 3.99 12.67 3.10 3.18 0.73 0.82 Sy2 3
NGC1068 1.23 0.22 4.11 12.42 2.57 4.55 0.26 0.55 Sy2 3
Mrk573 2.92 0.18 3.89 12.12 2.95 2.47 0.75 0.80 Sy2 3
Mrk78 4.96 0.14 3.88 11.94 2.46 2.32 0.68 0.61 Sy2 3
Mrk34 3.43 0.15 3.68 11.46 2.99 2.18 0.82 0.80 Sy2 3
Mrk1 2.78 0.21 3.56 10.95 2.66 2.21 0.49 0.52 Sy2 3
3c433 6.17 0.41 3.27 9.44 3.38 5.13 1.58 1.13 Sy2 3
Mrk270 5.64 0.28 2.96 8.71 3.14 2.93 1.21 1.39 Sy2 3
3c452 4.81 0.18 2.40 6.85 2.98 3.58 1.10 0.77 Sy2 3
Mrk198 2.51 0.12 1.79 5.56 3.02 2.26 0.89 0.68 Sy2 3
Mrk268 3.75 0.25 1.55 4.82 3.38 4.94 1.28 1.08 Sy2 3
Mrk273 8.27 0.22 1.52 4.44 2.94 2.62 0.89 0.54 Sy2 3
NGC3227$^{(b)}$ 3.22 0.50 3.57 10.73 2.86 5.01 1.24 1.26 Sy2 4
Mrk6 2.45 0.28 3.37 10.13 2.79 1.79 0.62 0.63 Sy2 4
ESO138G1 2.35 0.34 2.93 8.71 3.01 0.68 0.47 0.48 Sy2 5
NGC5643$^{(a),(b) }$ 5.55 0.54 4.20 12.61 2.66 2.90 0.91 0.71 Sy2 6
NGC1667 11.50 0.59 2.78 9.20 2.80 6.96 1.37 1.17 Sy2 7
Mrk423 8.00 0.40 2.80 6.20 2.90 3.60 1.20 1.00 Sy1.9 8
Mrk609 1.80 0.50 1.80 5.00 2.80 2.60 0.60 0.50 Sy1.9 8
Mrk226SW 5.23 0.08 1.50 4.50 3.35 2.07 0.29 0.25 Sy2 9
NGC3081$^{(a)}$ 2.16 0.23 4.53 12.62 2.73 2.33 0.60 0.62 Sy2 10
NGC3281$^{(a)}$ 2.33 0.22 2.77 7.59 2.64 2.54 0.53 0.60 Sy2 10
NGC3982$^{(a)}$ 3.92 0.03 5.64 18.68 2.76 2.57 0.76 0.82 Sy2 10
NGC4388$^{(a)}$ 2.68 0.15 3.67 10.63 2.71 1.44 0.68 0.60 Sy2 10
NGC5135$^{(a)}$ 2.01 0.10 1.40 4.47 2.64 2.35 0.37 0.35 Sy2 10
NGC5643$^{(a)}$ 5.11 0.42 4.56 15.4 2.64 3.07 0.97 0.90 Sy2 10
NGC5728$^{(a)}$ 3.41 0.44 3.70 10.98 2.65 3.71 0.41 0.41 Sy2 10
NGC6300$^{(a)}$ 15.48 1.39 8.32 23.32 2.59 6.62 1.42 1.26 Sy2 10
NGC6890$^{(a)}$ 2.77 0.72 7.48 20.05 2.75 4.26 0.65 0.52 Sy2 10
IC5063$^{(a)}$ 5.06 0.28 3.37 10.31 2.77 2.67 0.69 0.60 Sy2 10
IC5135$^{(a)}$ 4.05 0.25 2.07 6.88 2.65 3.30 0.49 0.46 Sy2 10
Mrk744 2.38 0.33 3.18 8.84 2.47 3.62 2.83 2.83 Sy2 11
Mrk1066 3.34 0.08 1.22 3.84 2.76 2.42 0.51 0.55 Sy2 11
NGC5506 2.84 0.14 2.46 7.69 2.84 2.53 0.92 0.99 Sy1.9 12
NGC2110 4.38 0.26 1.61 4.76 2.66 3.76 1.52 1.38 Sy2 12
NGC3281$^{(a), (b)}$ 4.30 0.42 2.44 7.34 2.60 2.60 0.78 0.73 Sy2 13
Akn347$^{(a)}$ 2.98 0.42 4.95 15.01 2.65 3.23 0.75 0.75 Sy2 14
UM16$^{(a)}$ 2.90 0.22 4.62 14.00 2.72 1.70 0.45 0.45 Sy2 14
Mrk533$^{(a)}$ 1.59 0.13 4.03 12.23 2.72 2.72 0.39 0.45 Sy2 14
IZw92$^{(a)}$ 2.60 0.34 3.54 10.14 2.78 1.00 0.40 0.43 Sy2 14
Mrk612$^{(a)}$ 1.88 0.17 2.99 9.37 2.67 3.60 0.74 0.55 Sy2 14
Mrk622$^{(a)}$ 10.06 0.03 1.90 5.44 2.47 2.33 0.19 0.14 Sy2 14
References— (1) @kraemer94, (2) @contini12, (3) @koski78, (4) @cohen83, (5) @alloin92\
(6) @enrique94 (7) @radovich96, (8) @osterbrock81, (9) @osterbrock83,\
(10) @phillips83, (11) @goodrich83, (12) @shuder80, (13) @durret88,\
and (14) @shuder81. $^{(a)}$ Data corrected for reddening in this work. $^{(b)}$ Value of $I$\[\]$\lambda$4959 estimated from the theoretical relation $I$\[\]$\lambda$4959=$I$\[\]$\lambda$5007/3.0.\
Relations of SB98 {#thaisasec}
-----------------
SB98 carried out NLR model calculations using the photoionization code [Cloudy]{} [@ferland96] and suggested two relations for the abundance determinations in the NLRs of AGNs. The first one is $$\begin{aligned}
\begin{array}{lll}
{\rm (O/H)}_{{\rm SB98,1}}& = & 8.34 + (0.212 \, x) - (0.012 \, x^{2}) - (0.02 \, y) \\
& + & (0.007 \, xy) - (0.002 \, x^{2}y) +(6.52 \times 10^{-4} \, y^{2}) \\
& + & (2.27 \times 10^{-4} \, xy^{2}) + (8.87 \times 10^{-5} \, x^{2}y^{2}), \\
\end{array}
\label{equation:sb1}\end{aligned}$$ where $x$ = \[N[ii]{}\]$\lambda$$\lambda$6548,6584/H$\alpha$ and $y$ = \[O[iii]{}\]$\lambda$$\lambda$4959,5007/H$\beta$. The second relation of SB98 is $$\begin{aligned}
\begin{array}{lll}
{\rm (O/H)}_{{\rm SB98,2}} & = & 8.643 - 0.275 \, u + 0.164 \, u^{2} \\
& + & 0.655 \, v - 0.154 \, u v - 0.021 \, u^{2}v \\
& + & 0.288 v^{2} + 0.162 u v^{2} + 0.0353 u^{2}v^{2}, \\
\end{array}
\label{equation:sb2}\end{aligned}$$ where $u$ = log(\[O[ii]{}\]$\lambda$$\lambda$3727,3729/\[O[iii]{}\]$\lambda$$\lambda$4959,5007) and $v$ = log(\[N[ii]{}\]$\lambda$$\lambda$6548,6584/H$\alpha$). Both calibrations are valid for $\rm 8.4 \: \lid \: 12+log(O/H) \: \lid \: 9.4$.
The dependence of these relations on the density should be taken into account. This dependence is given by the expression considered by SB98 $$\label{necorr}
({\rm O/H})_{{\rm final}} = {\rm O/H} - 0.1\,\,\log(N_{e}/300),$$ where $N_{e}$ is the electron density in cm$^{-3}$ and the correction is valid for 100 cm$^{-3}$ $\la$ $N_{e}$ $\la$ 10$^{4}$ cm$^{-3}$. It should be noted that the value of this correction exceeds 0.1 dex for high-density objects only, i.e., for objects with electron densities $N_{e}$ $\ga$ 3$\times$10$^{3}$ cm$^{-3}$.
$T_{\rm e}$-method {#temed}
------------------
We use the emission-line intensities listed in Table \[tab1\] to determine the oxygen abundances (O/H)$_{T_{e}}$ of the narrow-line regions through the classic $T_{\rm e}$-method. We will follow the methodology described in @dors13 based on the equations given by @perez09, @hagele08, @perez07, and @perez03.
The electron temperature in the high ionization zone (referred to as $t_{3}$) for each object was obtained from the observed line-intensity ratio $R_{\rm O3}$=\[\]($\lambda4959\: + \: \lambda 5007)/\lambda4363$ using the expression $$\label{eqt3}
t_{3}=0.8254-0.0002415 R_{\rm O3}+\frac{47.77}{R_{\rm O3}},$$ with $t_{3}$ in units of $10^{4}$K. The electron temperature of the low ionization zone (referred to as $t_{2}$) was derived from the theoretical relation: $$\begin{aligned}
\label{eqo3}
t_{2}^{-1}\,=\,0.693\,t_{3}^{-1}+0.281.
$$ The $\rm O^{++}$ and $\rm O^{+}$ ionic abundances were computed through the relations: $$\begin{aligned}
12+\log(\frac{{\rm O^{++}}}{{\rm H^{+}}}) \!\!\!&=&\!\!\! \log \big( \frac{I(4959)+I(5007)}{I{\rm (H\beta)}}\big)+6.144 \nonumber\\
&&\!\!\!+\frac{1.251}{t_{3}}-0.55\log t_{3}.
\end{aligned}$$ and $$\begin{aligned}
12+\log(\frac{{\rm O^{+}}}{{\rm H^{+}}}) \!\!\!&=&\!\!\! \log \big( \frac{I(3727)}{I{\rm (H\beta)}}\big)+5.992 \nonumber\\
&&\!\!\!+\frac{1.583}{t_{2}}-0.681\log t_{2} +\log(1+2.3 n_{\rm e}).\end{aligned}$$
Finally, we assumed $$\begin{aligned}
\rm
\frac{O}{H}=\frac{O^{+}}{H^{+}}+\frac{O^{++}}{H^{+}} .\end{aligned}$$ for the determination of the total abundance.
Abundance results {#res}
-----------------
---------- ------------------------- ------------------------- -------------------- -------- --------
Object $N_{\rm e}$ (cm$^{-3}$) $T_{\rm e}$ ($10^{4}$K) $T_{\rm e}$-method SB98,1 SB98,2
NGC7674 1148.0 1.14 8.49 8.72 9.19
IZw92 822.0 1.95 7.90 8.44 8.67
NGC3393 2022.0 1.09 8.69 — 9.24
Mrk176 535.0 1.61 8.21 8.84 9.02
3c33 252.0 1.72 8.14 8.64 8.80
Mrk3 948.0 1.49 8.26 8.75 8.96
NGC1068 26993.0 1.44 8.40 — —
Mrk573 781.0 1.34 8.33 8.65 8.92
Mrk78 370.0 1.22 8.49 8.74 8.88
Mrk34 546.0 1.27 8.39 8.61 8.85
Mrk1 767.0 1.50 8.17 8.63 8.90
3c433 50.0 2.38 7.84 8.96 9.04
Mrk270 1027.0 1.97 7.99 8.62 8.75
3c452 50.0 1.76 7.95 8.80 8.95
Mrk198 111.0 1.59 7.88 8.61 8.84
Mrk268 260.0 2.68 7.52 8.76 8.93
Mrk273 50.0 2.60 7.73 8.67 8.72
NGC3227 647.0 2.49 7.75 8.98 9.15
Mrk6 647.0 1.81 7.95 8.55 8.84
ESO138G1 685.0 2.22 7.73 8.40 8.53
NGC5643 141.0 2.36 7.90 8.86 8.96
NGC1667 281.0 3.12 7.84 9.15 9.06
Mrk423 239.0 3.14 7.67 8.73 8.80
Mrk609 239.0 4.41 7.11 8.62 8.92
Mrk226SW 296.0 1.45 8.08 8.52 8.60
NGC3081 693.0 1.84 7.85 8.60 8.91
NGC3281 974.0 1.86 7.84 8.61 8.90
NGC3982 819.0 0.68 9.54 8.87 8.99
NGC4388 343 1.31 8.30 8.54 8.80
NGC5135 492.0 1.61 7.78 8.57 8.83
NGC5643 451.0 1.79 8.17 8.93 8.99
NGC5728 606.0 2.26 7.84 8.86 9.05
NGC6300 360.0 2.96 8.11 — 9.17
NGC6890 176.0 2.11 8.06 — 9.35
IC5063 311.0 1.79 8.06 8.51 8.65
IC5135 471.0 2.12 7.78 8.72 8.90
Mrk744 606.0 2.16 7.76 8.83 9.10
Mrk1066 839.0 1.56 7.90 8.54 8.69
NGC5506 809.0 1.46 8.09 8.60 8.85
NGC2110 395.0 2.78 7.54 8.73 8.87
NGC3281 471.0 2.87 7.61 8.66 8.83
Akn347 606.0 1.82 8.09 8.93 9.11
UM16 606.0 1.37 8.35 8.59 8.87
Mrk533 1046.0 1.18 8.44 8.72 9.12
IZw92 805.0 2.02 7.87 8.42 8.65
Mrk612 75.0 1.46 8.10 8.89 9.24
Mrk622 64.0 0.97 8.76 8.69 8.73
---------- ------------------------- ------------------------- -------------------- -------- --------
Electron densities ($N_{\rm e}$), electron temperatures ($T_{\rm e}$), and the oxygen abundances estimated using three different ways for the objects in our sample are reported in Table \[tab2\]. The oxygen abundances of some objects can not be determined through the relations given by SB98 because for those objects the estimated abundances or electron densities are outside of the validity ranges considered by these authors, i.e., beyond of the range of values for which the relations are defined. The electron densities, $N_{\rm e}$, for the objects in our sample were computed from the line intensity ratio \[\]$\lambda 6716/\lambda 6731$ using the [temden]{} routine of the nebular package of [iraf]{}[^2]. Most of our objects have values in the range $0.01<n_{\rm e}<0.12$, where $n_{\rm e}=N_{\rm e}/(10^{4} \rm cm^{-3})$. These values are somewhat higher than the ones derived for regions, which were found to be $n_{\rm e}<0.06$ (see @krabbe14 [@copetti00]). However, they are sufficiently low so that one does not need to consider the contribution from the collisional de-excitation [@rubin89] or the direct recombination of the forbidden lines used in the standard abundance determinations.
We found the electron temperature $t_{3}$ to be in the range of 6000 to 50000 K, with an average value of $\sim$20000 K. This value is somewhat higher than the one found by @zhang13 for Sy2s, who used data from SDSS DR7 and found an average value of 14000 K.
Fig. \[f4\] shows the comparison between oxygen abundances in NLRs determined through the two methods presented above. Inspection of Fig. \[f4\] reveals that the $T_{\rm e}$-method provides oxygen abundances lower than the ones obtained via the relations of SB98 by up to $\sim$2dex, with an average value of $\sim 0.8$ dex, which confirms the existence of the so-called “temperature problem” in AGNs. This discrepancy is much larger than the one found for regions. For example, @kennicutt03, using spectroscopy data of regions in spiral galaxies, found that O/H abundances computed from relations based on photoionization models of @kewley02 are higher by 0.2 to 0.5 dex than the ones from the $T_{\rm e}$-method.
The abundance discrepancy in AGNs could be caused by the presence of a secondary heating (ionizing) source in addition to the radiation from the gas accretion onto the AGN. This secondary source is probably related to the shock. Indeed, @zhang13 showed that the strong \[\]$\lambda$4363 flux in AGNs (and consequently high electron temperature) suggests the presence of some supplementary energy source(s), which could be due to the presence of shock waves (see also @contini12 [@prieto05]). Moreover, the presence of fluctuations of electron temperature [@peimbert67] in the gas phase of AGNs could also contribute for the discrepancies found in Fig. \[f4\].
In order to examine if Sy2 galaxies have a secondary heating source, we performed a simple test. In Fig. \[f2ac\], the intensity of \[\]$\lambda$4363/H$\beta$ versus the $R_{23}$ parameter defined as $R_{23}=[I([{\rm O\:II}]\lambda3727)+I([{\rm O\:III}]\lambda4959+\lambda5007)]/I({\rm H\beta})$ for the objects in our sample and the ones predicted by a grid of photoionization models are shown.
We consider two models to reproduce the observational data: (i) the AGN model to describe the NRL spectra, and (ii) the SF+AGN model to describe composite (star-forming regions + AGN) spectra. We use the version 08.00 of the Cloudy code [@ferland96] to construct NLR models similar to the ones used by @dors12, but considering the Table-AGN model [@mathews87] as the ionizing source. We considered a large range of values for the physical parameters: the number of ionizing photons, $51 \lid \: \log Q({\rm H}) \: \lid \:55$; electron densities, $1.0 \lid \: \log(N_{\rm e}) \: \lid 4.0$; and metallicities, $\rm 0.05 \: \lid \: (Z/Z_{\odot}) \: \lid \: 2$. The spectrum of the SF+AGN model is the sum of the predictions of the SF model and the AGN model. The SF model is constructed using the Cloudy code [@ferland96] assuming the ionizing source to be a stellar cluster with an age of 2.5 Myr and $\log Q({\rm H})=52.85$ whose spectrum is computed with the $STARBURST99$ code [@leitherer99]. The metallicity of the gas phase was considered to be solar with an electron density $N_{\rm e}=100\, \rm cm^{-3}$. These parameters are similar to the ones derived in circumnuclear star-formation regions observed in two galaxies (NGC1097 and NGC6951) containing a Seyfert 2 nucleus by @dors08.
We can see in Fig. \[f2ac\] that the ratio \[\]$\lambda$4363/H$\beta$ is underpredicted by the AGN and SF+AGN models when compared the observations. This line ratio depends strongly on the electron temperature, therefore the electron temperatures in AGNs predicted by our models are lower than the ones estimated using the $T_{\rm e}$-method. Thus, this discrepancy could be attributed to the presence of gas shock waves propagating at supersonic velocities through the NLRs, which does to enhance the intensity of \[\]$\lambda$4363 producing larger electron temperature values and, consequently, low (unrealistic) O/H values when using the $T_{\rm e}$-method (see @nagao01 and references therein).
Due to the result above, in our paper the $T_{\rm e}$-method is no longer used to compute the metallicity of AGNs.
SB98 pointed out that an averaged value of $Z$ obtained from both relations (Eq. \[equation:sb1\] and \[equation:sb2\]) should be used. Instead of doing that we use our models to choose which of the relations by SB98 yields more reliable abundances. For that, we use the standard O/H – $R_{23}$ empirical diagram. The $R_{23}$ parameter was suggested by @pagel79 as indicator of the oxygen abundance when the $T_{\rm e}$-method cannot be applied. Although $R_{23}$ has a strong dependence on metallicity, it also depends on the ionization degree of the nebulae, which should be taken into account [@pilyugin01; @leonid00]. We compare the predictions of our AGN models with the observed O/H – $R_{23}$ diagram for the objects in our sample, where the oxygen abundances are estimated using both relations of SB98 (see Fig. \[f3ac\]). Since the AGN and SF+AGN models predict similar values of $R_{23}$ in the zone of Fig. \[f2ac\] where our data are located, only the former models are shown in Fig. \[f3ac\]. In this figure we can see that a better agreement is obtained when the abundances are determined through the first relation of SB98. Hereafter this relation will be used.
Abundance at the centre of a galaxy vs. central intersect abundance
===================================================================
The data
--------
It is a widely accepted practice to specify the abundance at the center of a galaxy by the central intersect abundance obtained from the radial abundance gradient [e.g., @VilaCostas1992; @Zaritsky1994; @vanZee1998; @pily04; @pilyugin2007; @Gusev2012 among others]. The radial abundance gradients in the disks of nearby late-type galaxies were recently determined by @pilyugin2014 [@pilyugin2015]. The extrapolation of the radial abundance gradient to the zero galactocentric distance gives the central intersect abundance (O/H)$_{0}$ in galaxies. On the other hand, @ho97 obtained emission-line spectra of the central regions of many nearby galaxies. This provides a possibility to estimate the oxygen abundances at the centers of those galaxies. The comparison between central oxygen abundance estimated from the spectrum of the center of the galaxy with that obtained from the radial abundance gradient can tell us something about the chemical evolution of the central parts of galaxies.
We compare the central O/H abundances estimated from the spectra of @ho97 and the central intersect abundances obtained by @pilyugin2014 [@pilyugin2015]. We selected from the sample of @ho97 only the galaxies with central spectra classified by @ho97 as H[ii]{}-like regions or Seyferts and which are also in the @pilyugin2014 [@pilyugin2015] list. Our selected sample contains 45 objects (12 AGNs and 33 star-forming regions) whose emission-line intensities and derived oxygen abundances (see below) are listed in Table \[tab3\]. Fig. \[figure:bptho\] shows the positions of the 45 sample objects in a BPT diagnostic diagram [@baldwin81], together with a large sample of emission-line SDSS galaxies studied by @Thuan2010, which are plotted with cyan (grey) symbols. In this figure we also plotted the boundary lines obtained by @Kauffmann2003 and @Kewley2001 that separate the H[ii]{}-like objects and AGNs in the diagram.
[lccccc]{} Object & \[\]$\lambda$5007 & \[\]$\lambda$6584 & \[\]$\lambda$6717+$\lambda$6731 & 12+log(O/H) & 12+log(O/H)$_{0}$\
\
NGC 1058 & 3.64 & 3.50 & 1.95 & 8.58 & 8.62\
NGC 1068 & 12.10 & 2.17 & 0.68 & 8.54 & 8.64\
NGC 2336 & 2.84 & 5.12 & 2.99 & 8.74 & 8.80\
NGC 3031 & 4.07 & 6.38 & 3.89 & 8.78 & 8.58\
NGC 3227 & 5.92 & 3.82 & 1.95 & 8.62 & 8.64\
NGC 3486 & 4.50 & 3.01 & 2.64 & 8.56 & 8.60\
NGC 4258 & 10.20 & 2.29 & 2.61 & 8.51 & 8.54\
NGC 4395 & 6.23 & 0.93 & 2.04 & 8.34 & 8.19\
NGC 4501 & 5.15 & 6.02 & 2.66 & 8.80 & 8.92\
NGC 4725 & 6.68 & 3.27 & 1.82 & 8.59 & 8.83\
NGC 5033 & 4.59 & 6.76 & 2.99 & 8.85 & 8.64\
NGC 5194 & 8.09 & 8.26 & 2.28 & 9.01 & 8.88\
\
NGC 598 & 1.57 & 0.48 & 0.61 & 8.35 & 8.48\
NGC 783 & 0.19 & 1.03 & 0.55 & 8.65 & 8.68\
NGC 925 & 0.81 & 0.63 & 0.96 & 8.41 & 8.48\
NGC 1156 & 4.51 & 0.28 & 0.46 & 8.24 & 8.16\
NGC 2403 & 1.91 & 0.79 & 1.03 & 8.36 & 8.48\
NGC 2537 & 1.84 & 0.43 & 0.48 & 8.35 & 8.35\
NGC 2903 & 0.08 & 0.97 & 0.52 & 8.71 & 8.82\
NGC 2997 & 0.31 & 1.00 & 0.77 & 8.56 & 8.80\
NGC 3184 & 0.12 & 0.94 & 0.55 & 8.67 & 8.66\
NGC 3198 & 0.23 & 1.19 & 0.84 & 8.59 & 8.60\
NGC 3310 & 0.91 & 1.88 & 0.71 & 8.55 & 8.37\
NGC 3319 & 0.99 & 0.46 & 1.71 & 8.26 & 8.50\
NGC 3344 & 0.83 & 1.20 & 1.15 & 8.49 & 8.72\
NGC 3351 & 0.25 & 1.31 & 0.66 & 8.62 & 8.82\
NGC 3359 & 0.54 & 0.88 & 1.55 & 8.43 & 8.40\
NGC 3631 & 0.29 & 1.23 & 0.68 & 8.60 & 8.71\
NGC 3738 & 2.99 & 0.26 & 0.81 & 8.13 & 8.10\
NGC 3893 & 0.21 & 1.05 & 0.61 & 8.62 & 8.73\
NGC 4088 & 0.19 & 0.91 & 0.48 & 8.65 & 8.71\
NGC 4214 & 3.66 & 0.19 & 0.33 & 8.20 & 8.20\
NGC 4254 & 0.87 & 1.37 & 0.59 & 8.54 & 8.77\
NGC 4303 & 1.32 & 2.23 & 1.12 & 8.53 & 8.78\
NGC 4449 & 2.36 & 0.40 & 0.65 & 8.26 & 8.26\
NGC 4490 & 2.34 & 0.71 & 1.89 & 8.22 & 8.29\
NGC 4535 & 0.13 & 1.17 & 0.61 & 8.66 & 8.71\
NGC 4559 & 0.32 & 1.20 & 1.10 & 8.53 & 8.53\
NGC 4631 & 1.51 & 0.69 & 0.63 & 8.42 & 8.39\
NGC 4654 & 0.14 & 0.77 & 0.60 & 8.63 & 8.66\
NGC 4656 & 3.96 & 0.14 & 0.50 & 8.04 & 8.06\
NGC 5248 & 0.26 & 1.48 & 0.64 & 8.61 & 8.64\
NGC 5457 & 0.23 & 1.08 & 0.70 & 8.61 & 8.71\
NGC 5474 & 1.75 & 0.37 & 0.72 & 8.26 & 8.19\
NGC 6946 & 0.37 & 1.81 & 0.81 & 8.55 & 8.72\
@pilyugin2014 [@pilyugin2015] derived oxygen abundances from the published emission-line intensities of disk regions trough the $C$ method [see @pily12]. This is an empirical method based on the comparison between strong emission-line intensities in the spectrum of a target region and those in a set of reference regions with known abundances. The strong emission-lines considered in this method are \[\]$\lambda$3727, H$\beta$, \[\]$\lambda$5007, \[\]$\lambda$6584 and \[\]$\lambda$6716, $\lambda$6731. The optical data of @ho97 consist of long-slit spectroscopy of the nuclear region ($r \la 200$ pc) of a large sample of nearby galaxies covering the 4200-5200Å and 6200-6900Å spectral ranges with spectral resolutions of about 4 and 2.5 Å pixel$^{-1}$, respectively. We use again the first relation of SB98 to estimate the central abundances of the NLRs of the active galaxies \[(O/H)$_{\rm SB98,1}$\]. In the case of galaxies with central H[ii]{}-like regions we use the $C_{NS}$ method [@pily12; @pily13] to estimate the oxygen abundances \[(O/H)$_{C_{NS}}$\]. The $C_{NS}$ method is a variant of the $C$ method applicable when the \[\]$\lambda$3727 emission-lines are not available.
Abundance at the centre vs. central intersect abundance
-------------------------------------------------------
In the bottom panel of Figure \[figure:oho-ohho\] we show the comparison between the central intersect abundances in the selected galaxies (O/H)$_{0}$ obtained from the radial abundance gradients and central abundances (O/H)$_{\rm SB98,1}$ and (O/H)$_{C_{NS}}$ determined from the spectral measurements of the central regions of @ho97. This figure shows that the (O/H)$_{\rm SB98,1}$ abundances in the NLRs are close to or slightly lower than the central intersect abundances in the host galaxies. The central abundances in galaxies with central H[ii]{}-like regions show a similar behaviour. Thus the metallicity in the NLRs obtained through the relation of SB98 is close to the central metallicity of the host galaxy estimated from the radial abundance gradient. The mean difference (D) between the direct and intersect central oxygen abundances (see upper panel of Figure \[figure:oho-ohho\]) for H[ii]{}-like objects is $-0.11 \pm0.09 $ at high metallicities (12+log(O/H)$>$8.5) and $0.00 \pm0.08 $ at low metallicities (12+log(O/H)$<$8.5). The mean difference for AGNs is $-0.01 \pm0.13$ at high metallicities. The differences for the high metallicity AGNs seem to follow a linear regression with a slope of $-0.32(\pm0.33) \: \rm dex^{-1}$. The differences for H[ii]{}-like objects follow a regression with a slope of $-0.45(\pm0.23) \: \rm dex^{-1}$ at high metallicities, and with a slope of $-0.23(\pm0.15) \: \rm dex^{-1}$ at low metallicities. We also performed a linear regression considering all points in Fig. \[figure:oho-ohho\] due to the small sample we are using. It yields a slope of $-0.25(\pm0.06) \: \rm dex^{-1}$.
The differences are within the uncertainty in the abundance estimations (i.e., compatible with zero). However, the existence of a trend in the differences suggests that the direct central abundance in some high metallicity galaxies can be lower than the central intersect abundance. It should be noted that @Sanchez2014 have also found observational evidence of lower central oxygen abundances than that predicted by the central abundance extrapolation of the gradients in a number of spiral galaxies. A possible explanation of the trend of the differences with metallicity could be the accretion of metal-poor gas onto the centers of galaxies. In the case of the low-metallicity galaxies, where the metallicity of the accretion material is similar (or not so different) to that of the gas in the central region, the infalling gas should not significantly change the local metallicity. In contrast, in the case of high-metallicity galaxies this low-metallicity infalling gas would dilute the heavy element content of the gas in the central region. Therefore this effect would be more relevant in high-metallicity galaxies than in low-metallicity ones.
Observational evidence of the presence of gas flows from the outer parts (with low-metal content material) to the centre of the galactic disc (high metal content) has been found for isolated barred galaxies [@martin94; @Zaritsky1994] and interacting ones [@rosa14; @ellison11; @ellison10; @kewley10]. Likewise, there are several kinematical studies based on optical and infrared integral field spectroscopy that support the scenario where gas is infalling toward the central region of AGNs [e.g. @muller14; @muller11; @riffel13b; @riffel08; @thaisa07; @fathi06]. Moreover, @rupke10 performed numerical simulations of galaxy mergers and found that the central underabundances observed in this kind of systems could be accounted for by a radial infall of low-metallicity material coming from the outskirts of both galaxies involved.
Is there an extraordinary chemical enrichment of the NLRs?
----------------------------------------------------------
It was found by @pilyugin2006 [@pilyugin2007] that there is an upper limit to the oxygen abundances in galaxies, i.e., there is a maximum attainable oxygen abundance. These authors found that this maximum value of oxygen abundance in galaxies is about twice the solar abundance [adopting the solar oxygen abundance to be $\rm 12+\log(O/H)_{\odot}=8.69$; @allende01].
The abundances in the NLRs of AGNs taken from the list of @ho97 are in the range $\rm 8.5 \: < \: 12+log(O/H) \: <9.0$ or $ 0.6 \: < \: Z/Z_{\odot} \: < 2$, which does not exceed the maximum attainable abundance for galaxies. This suggests that there is no an extraordinary chemical enrichment of the NLRs. Similar AGN abundances were obtained by @matsuoka09 from the analysis of the C[iv]{}$\lambda$1549/He[ii]{}$\lambda$1640 – C[iii]{}\]$\lambda$1909/C[iv]{}$\lambda$1549 diagram calibrated in terms of gas abundance through photoionization models using the [Cloudy]{} code [@ferland96]. @dors14 also obtained similar abundances using a new index (C$_{43}$ = log\[(C[iv]{}$\lambda$1549 + C[iii]{}\]$\lambda$1909)/He[ii]{}$\lambda$1640\]) defined by them as a metallicity indicator for AGNs.
However, it should be noted that we found metallicities larger than twice solar up to $Z/Z_{\odot}\sim 6.5$ \[$12+\log(\rm O/H)_{SB98,1} \sim 9.5$ (see Fig. \[f4\]\]) for five AGNs from Table \[tab1\]. Similarly extra high metallicities for some NLRs were obtained by SB98. Further study of the galaxies with possible extra high metallicity AGNs should be carried out (in particular, radial abundance distributions and central intersect abundances should be obtained in those galaxies) in order to be able to draw solid conclusions about the upper limit to the oxygen abundances in NLRs.
Summary and conclusions
=======================
We compiled from the literature a sample of spectra of narrow-line regions of active galactic nuclei with available optical emission lines: \[\]$\lambda$3727, \[\]$\lambda$4363, H$\beta$, \[\]$\lambda$5007, H$\alpha$, \[\]$\lambda$6584, \[\]$\lambda$6717, and \[\]$\lambda$6731. We estimated the oxygen abundances in those NLRs through the classic $T_e$ method and through the strong-line method (the calibration of @thaisa98). We found that the abundances determined through the $T_e$ method are lower by up to $\sim$2 dex than the abundances estimated through the calibration of SB98, i.e., we confirmed the existence of the so-called “temperature problem” in AGNs.
We also considered a second sample of galaxies for which the emission-line spectra of the central regions were measured by @ho97 and the central intersect abundances were found by @pilyugin2014 [@pilyugin2015]. The direct central abundances of the AGNs and Star-forming regions in those galaxies were estimated through the calibration of @thaisa98 and through the $C_{NS}$ method [@pily12; @pily13], respectively. We found that the abundances of the NLRs and H[ii]{}-like objects estimated from the direct spectral measurements are close to or slightly lower than the central intersect abundances obtained from the radial abundance gradients. This may suggest that the infall of the low-metallicity gas onto the centers of the galaxies can take place in some galaxies where the central abundance estimated from the direct spectral measurements is lower than the central intersect abundance.
The abundances in the NLRs of the AGNs in our samples do not suggest that there is an extraordinary chemical enrichment of the narrow-line regions. There are only a few AGNs with oxygen abundances higher than the maximum attainable abundance for galaxies [$\sim$2 times the solar value; @pilyugin2006; @pilyugin2007]. Additional investigations of the galaxies with possible extra high metallicity AGNs are necessary to draw reliable conclusions on the upper limits of the oxygen abundances in NLRs.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to the anonymous referee for his/her useful comments and suggestions that helped us to substantially clarify and improve the manuscript. O.L.D. and A.C.K. are grateful to the FAPESP for support under grant 2009/14787-7 and 2010/01490-3, respectively.\
E.K.G. and L.S.P. acknowledge support within the framework of Sonderforschungsbereich (SFB 881) on “The Milky Way System” (especially subproject A5), which is funded by the German Research Foundation (DFG).\
O.L.D. thanks the hospitality of the Astronomisches Rechen-Institut at Heidelberg University and the Universidad Nacional de La Plata where part of this work was done.\
L.S.P. thanks the hospitality of the Astronomisches Rechen-Institut Heidelberg University where part of this investigation was carried out.\
This work was partly funded by the subsidy allocated to Kazan Federal University for the state assignment in the sphere of scientific activities (L.S.P.).\
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.\
[99]{} Allende Prieto C., Lambert D. L., Asplund M., 2001, ApJ, 556, L63 Alloin D., Bica E., Bonatto C., Prugniel P., 1992, A&A, 266, 117 Alloin D., Collin-Souffrin S., Joly M., Vigroux L., 1979, A&A, 78, 200 Baldwin J. A., Hamann F., Korista K. T., Ferland G. J., Dietrich M., Warner C., 2003, ApJ, 583, 649 Baldwin J. A., Phillips M. M., Terlevich R., 1981, PASP, 93, 5 Batra N. D., & Baldwin J. A., 2014, MNRAS, 439. 771 Bennert N., Jungwiert B., Komossa S., Haas M., Chini R., 2006a, A&A, 456, 953 Bennert N., Jungwiert B., Komossa S., Haas M., Chini R., 2006b, A&A, 459, 55 Bresolin F., Schaerer D., González Delgado R. M., Stasińska G., 2005, A&A, 441, 981 Cohen R. D., 1983, ApJ, 273, 489 Contini M., 2012, MNRAS, 425, 1205 Copetti M. V. F., Mallmann J. A. H., Schmidt A. A., Castañeda H. O., 2000, A&A, 357, 621 Dack S., McCall M. L., 2012, ApJ, 750, 167 Dhanda N., Baldwin J. A., Bentz M. C., Osmer P. S., 2007, ApJ, 658, 804 Dors O. L., Cardaci M. V., Hägele G. F., Krabbe A. C., 2014, MNRAS, 443, 1291 Dors O. L. et al., 2013, MNRAS, 432, 2512 Dors O. L. et al., 2012, MNRAS, 422, 252 Dors, O. L., Krabbe A. C., Hägele G. F., Pérez-Montero E., 2011, MNRAS, 415, 3616 Dors O. L., Storchi-Bergmann T., Riffel R. A., Schimdt A. A., 2008, A&A, 482, 59 Dors O. L., & Copetti M. V. F., 2005, A&A, 437, 837 Du P. et al., 2014, MNRAS, 438, 2828 Durret F., & Bergeron J., 1988, ApJSS, 75, 273 Ellison S. L., Patton D. R., Mendel J. T., Scudder J. M., 2011, MNRAS, 418, 2043 Ellison S. L. et al., 2010, MNRAS, 407, 1514 Esteban C., Peimbert M., Torres-Peimbert S., Escalante V., 1998, MNRAS, 295, 401 Fathi,K., Storchi-Bergmann T., Riffel R. A., 2006, ApJ, 641, 25L Ferland G. J. et al., 1996, ApJ, 461, 683 Ferland G. J., & Netzer H., 1983, ApJ, 264, 105 Goodrich R. W., & Osterbrock D., 1983, ApJ, 269, 416 Groves B. A., Heckman T. M., Kauffmann G., 2006, MNRAS, 371, 1559 Gusev A. S., Pilyugin L. S., Sakhibov F., Dodonov S. N., Ezhkova O. V., Khramtsova M. S., 2012, MNRAS, 424, 1930 Hamann F., Korista K. T., Ferland G. J., Warner C., Baldwin J., 2002, ApJ, 564, 592 Hamann F., & Ferland G. J., 1993, ApJ, 418, 11 Hamann F., & Ferland G. J., 1992, ApJ, 391, L53 Hägele G. F., Díaz A. I., Terlevich E., Terlevich R., Pérez-Montero E., Cardaci M. V., 2008, MNRAS, 383, 209 H[ä]{}gele G. F., P[é]{}rez-Montero E., D[í]{}az [A]{}. I., Terlevich E., Terlevich R., 2006, MNRAS, 372, 293 Ho L. C., Filippenko A. V., Sargent W. L. W., 1997, ApJS, 112, 315 Hummer D. G., Storey P. J., 1987, MNRAS, 224, 801 Izotov Y. I., Guseva N. G., Fricke K. J., et al., 2010, A&A, 517, 90 Izotov Y. I., & Thuan T. X., 2008, ApJ, 687, 133 Kauffmann et al., 2003, MNRAS, 346, 1055 Kennicutt R. C., Bresolin F., Garnett D. R., 2003, ApJ, 591, 801 Kewley L. J., Rupke D., Zahid H. J., Geller M. J., Barton E. J., 2010, ApJ, 721, L48 Kewley L. J., & Ellison S., 2008, ApJ, 681, 1183 Kewley L. J., Dopita M. A., 2002, ApJS, 142, 35 Kewley L.J., Dopita M. A., Sutherland R. S., Heisler C. A., Trevena J., 2001 ApJ, 556, 121 Koski A. T., 1978, ApJ, 223, 56 Krabbe A. C. et al., 2014, MNRAS, 435, 1155 Kraemer S. B., Wu C.-C., Crenshaw D. M., Harrington J. P., 1994, ApJ, 435, 171 Martin P., & Roy J. R., 1995, ApJ, 445, 161 Mathews W. G., & Ferland G. J., 1987, ApJ, 323, 456 Matsuoka K., Nagao T., Maiolino R., Marconi A., Taniguchi Y., 2009, A&A, 503, 721 Meyer D. M., Jura M., Cardelli J. A., 1998, ApJ, 493, 222 Nagao T., Maiolino R., Marconi A., 2006a, A&A, 447, 157 Nagao T., Maiolino R., Marconi A., 2006b, A&A, 447, 863 Osterbrock D. E., Ferland G., 2006, Astrophysics of Gaseous Nebulae and Active Galactic Nuclei. -2nd ed. Unviersity Science Books, Mill Valley, CA Osterbrock D. E., & Dahari, O., 1983, ApJ, 273, 478 Osterbrock D. E., 1981, ApJ, 249, 462 Lauberts A., 1982, ESO/Uppsala Survey of the ESO (B) Atlas, ed. ESO (Garching) Leitherer C. et al., 1999, ApJ, 123, 3 López-Sánchez Á. R., Dopita M. A., Kewley L. J., Zahid H. J., Nicholls D. C., Scharwächter J., 2012, MNRAS, 426, 2630 López-Sánchez Á. R., & Esteban C., 2010, A&A, 517, 85 Nagao T., Murayama T., Taniguchi Y, ApJ, 549, 155 Pagel B. E. J., Edmunds M. G., Blackwell D. E., Chun M. S., Smith G., 1979, MNRAS, 189, 95 Peimbert M., 1967, ApJ, 150, 825 Pérez-Montero E., & Contini T., 2009, MNRAS, 398, 949 Pérez-Montero E., Hägele G. F., Contini T., Díaz A. I., 2007, MNRAS, 381, 125 P[é]{}rez-Montero E., & D[í]{}az A. I., 2003, MNRAS, 346, 105 Pettini M., & Pagel B. E. J., 2004, MNRAS, 348, L59 Phillips M. M., Charles P. A., Baldwin J. A., 1983, ApJ, 266, 485 Pilyugin L. S., 2000, A&A, 362, 325 Pilyugin L. S., 2001, A&A, 369, 594 Pilyugin L. S., 2003, A&A, 399, 1003 Pilyugin L. S., & Thuan T.X., 2005, ApJ, 631, 231 Pilyugin L. S., Vílchez J. M., Contini T., 2004, A&A, 425, 849 Pilyugin L. S., Thuan T. X., Vílchez J. M., 2006, MNRAS, 367, 1139 Pilyugin L. S., Thuan T. X., Vílchez J. M., 2007, MNRAS, 376, 353 Pilyugin L. S., Grebel E. K., Mattsson L., 2012, MNRAS, 424, 2316 Pilyugin L. S., Lara-López M. A., Grebel E. K., Kehrig C., Zinchenko I. A., López-Sánchez Á.R., Vílchez J.M., Mattsson L., 2013, MNRAS, 432, 1217 Pilyugin L. S., Grebel E. K., Kniazev A. Y., 2014, AJ, 147, 131 Pilyugin L. S., Grebel E. K., Zinchenko I. A., 2015, MNRAS, 450, 3254 Prieto M. A., Marco, O., Gallimore, J., 2005, MNRAS, 364, L28 Radovich, M., & Rafanelli, P., 1996, A&A, 306, 97 Richardson C. T., Allen J. T.; Baldwin J. A., Hewett P. C., Ferland G. J., 2014, MNRAS, 437, 2376 Riffel R., Rodríguez-Ardila A., Aleman I. et al., 2013a, MNRAS, 430, 2002 Riffel, R. A., Storchi-Bergmann, T., Winge, C., 2013b, MNRAS, 430, 2249 Riffel, R. A., Storchi-Bergmann, T., Winge, C., et al., 2008, MNRAS, 385, 1129 Robertson P., Shields G.A., Blanc G. A., 2012, ApJ, 748, 48 Rosales-Ortega F. F., Díaz A. I., Kennicutt R. C., Sánchez, S. F., 2011, MNRAS, 415, 2439 Rosa D. A., Dors O. L., Krabbe A. C. et al., 2014, MNRAS, 444, 2005 Rupke D. S. N., Veilleux S., Baker A. J., 2008, ApJ, 674, 172 Rupke D. S. N., Kewley L. J., Barnes J. E., 2010, ApJ, 710, L156 Rubin R. H., 1989, ApJSS, 69, 897 Sánchez S. F. et al., 2014, A&A, 563, 49 Schmitt H. R., Storchi-Bergmann T., Baldwin J. A., 1994, ApJ, 423, 237 Shuder J. M., 1980, ApJ, 240, 32 Shuder J. M., & Osterbrock D. E., 1981, ApJ, 250, 55 Stasińska G., 2004, cmpe.conf, 115 Stasińska G., 2002, Rev. Mexicana Astron. Astrofis. Ser. Conf. 12, Ionized Gaseous Nebulae, ed. W. J. Henney et al. (Mexico, DF: UNAM), 62 Schnorr-Müller, A. et al., 2014, MNRAS, 437, 1708 Schnorr-Müller, A. et al., 2011, MNRAS, 413, 149 Storchi-Bergmann, T., Dors, O. L., Riffel, R. A., et al., 2007, ApJ, 670, 959 Storchi-Bergmann T., Schmitt H. R., Calzetti D., Kinney A. L., 1998, AJ, 115, 909 (SB98) Tremonti C. A., Heckman T. M., Kauffmann G., et al., 2004, ApJ, 613, 898 Thuan T. X., Pilyugin L. S., Zinchenko I. A., 2010, ApJ, 712, 1029 van Zee L., Salzer J. J., Haynes M. P., O‘Donoghue A. A., Balonek T. J., 1998, AJ, 116, 2805 Viegas S. M., 2002, RMxAc, 12, 219 Vila-Costas M. B., & Edmunds M. G., 1992, MNRAS, 259, 121 Wang J.-M. et al., 2011, ApJ, 739, 3 Zaritsky D., Kennicutt R. C., Huchra, J. P., 1994, ApJ, 420, 87 Zhang Z. T., Liang Y. C.; Hammer F., 2013, MNRAS, 430, 2605 Zurita A., & Bresolin F., 2012, MNRAS, 427, 1463
\[lastpage\]
[^1]: The NASA/IPAC Extragalactic Database ([ned]{}) is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
[^2]: Image Reduction and Analysis Facility, distributed by NOAO, operated by AURA, Inc., under agreement with the NSF.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[ We consider chance-constrained binary programs, where each row of the inequalities that involve uncertainty needs to be satisfied probabilistically.]{} Only the information of the mean and covariance matrix is available, and we solve distributionally robust chance-constrained binary programs (DCBP). Using two different ambiguity sets, we equivalently reformulate the DCBPs as 0-1 second-order cone (SOC) programs. We further exploit the submodularity of 0-1 SOC constraints under special and general covariance matrices, and utilize the submodularity as well as lifting to derive extended polymatroid inequalities to strengthen the 0-1 SOC formulations. We incorporate the valid inequalities in a branch-and-cut algorithm for efficiently solving DCBPs. [ We demonstrate the computational efficacy and solution performance using diverse instances of a chance-constrained bin packing problem.]{}'
author:
- |
**[Yiling Zhang]{}, **[Ruiwei Jiang]{}, and **[Siqian Shen]{}\
[Department of Industrial and Operations Engineering]{}\
[University of Michigan, Ann Arbor, MI 48109]{}\
[Email: {zyiling, ruiwei, siqian}@umich.edu]{}\
******
date:
title: '[**Ambiguous Chance-Constrained Binary Programs under Mean-Covariance Information**]{}[^1]'
---
[*Key words:*]{} Chance-constrained binary program; distributionally robust optimization; conic integer program; submodularity; extended polymatroid; bin packing
Introduction {#sec:intro}
============
We consider chance-constrained binary programs that involve a set of individual chance constraints. More specifically, we consider $I$ individual chance constraints and, for each $i \in [I] := \{1,\ldots,I\}$, let $y_i \in \{0, 1\}^J$ be a binary decision vector such that $y_i := [y_{i1}, \ldots, y_{iJ}]^{\top}$ and $\tilde{t}_i$ be the corresponding random coefficients such that $\tilde{t}_i := [\tilde{t}_{i1}, \ldots,\tilde{t}_{iJ}]^{\top}$. Then, we consider the following individual chance constraints: $$\mathbb{P} \Biggl\{ \tilde{t}_{i}^{\top} y_{i} \leq T_i\Biggr\} \ \geq \ 1 - \alpha_i \quad \forall i \in [I], \label{cc}$$ where $T_i \in \mathbb R$, $\mathbb{P}$ represents the joint probability distribution of $\{\tilde{t}_{ij}: i \in [I], \ j \in [J]\}$, and each $\alpha_i$ represents an allowed risk tolerance of constraint violation that often takes a small value (e.g., $\alpha_i = 0.05$).
The individual chance constraints have wide applications in service and operations management, providing an effective and convenient way of controlling capacity violation and ensuring high quality of service. For example, in surgery allocation, $y_i$ represents yes-no decisions of allocating $J$ surgeries in operating room (OR) $i$, for all $i \in [I]$. The operational time limit of each OR (i.e., $T_i$) is usually deterministic, but the processing time of each surgery (i.e., $\tilde{t}_{ij}$) is usually random due to the variety of patients, surgical teams, and surgery characteristics. Then, chance constraints make sure that each OR will not go overtime with a large probability, offering an appropriate “end-of-the-day” guarantee.
Chance-constrained programs are difficult to solve, mainly because the feasible region described by constraints is non-convex in general [@Prekopa2003]. Nonetheless, promising special cases have been identified to recapture the convexity of chance-constrained models. In particular, if $\{\tilde{t}_{ij}: j \in [J]\}$ are assumed to follow a Gaussian distribution with a known mean $\mu_i$ and covariance matrix $\Sigma_i$, then the chance constraints are equivalent to the second-order cone (SOC) constraints $$\mu_i^{\top} y_i + \Phi^{-1}(1 - \alpha_i) \sqrt{y_i^{\top} \Sigma_i y_i} \ \leq \ T_i \quad \forall i \in [I], \label{cc-normal}$$ where $\Phi(\cdot)$ represents the cumulative distribution function of the standard Gaussian distribution. [ In this case, feasible binary solutions $y_i$ to constraints can be quickly found by off-the-shelf optimization solvers.]{} In another promising research stream, the probability distribution $\mathbb{P}$ of $\tilde{t}_{ij}$ is replaced by a finite-sample approximation, leading to a sample average approximation (SAA) of the chance-constrained model [@CCP_luedtke-ahmed-2008; @CCPNEW_pagnoncelli2009computational]. The SAA model is then recast as a mixed-integer linear program (MILP), on which many strong valid inequalities can be derived to accelerate the branch-and-cut algorithm (see, e.g., [@Luedtke2010; @CCP_kucukyavuz2010mixing; @luedtke2014branch; @song2014chance; @liu2016decomposition]).
However, a basic challenge of the chance-constrained approach is that the perfect knowledge of probability distribution $\mathbb{P}$ may not be accessible. Under many circumstances, we only have a series of historical data that can be considered as samples taken from the true (while ambiguous) distribution. As a consequence, the solution obtained from a chance-constrained model can be sensitive to the choice of distribution $\mathbb{P}$ we employ in and hence perform poorly in out-of-sample tests. This phenomenon is often observed when solving stochastic programs and is called the *optimizer’s curse* [@smith2006optimizer]. A natural way of addressing this curse is that, instead of a single estimate of $\mathbb{P}$, we employ a set of plausible probability distributions, termed the ambiguity set and denoted $\mathcal{D}$. Then, from a robust perspective, we ensure that chance constraints hold valid with regard to all probability distributions belonging to $\mathcal{D}$, i.e., $$\inf_{\mathbb{P} \in \mathcal{D}} \ \mathbb{P} \Biggl\{ \tilde{t}_{i}^{\top} y_{i} \leq T_i\Biggr\} \ \geq \ 1 - \alpha_i \quad \forall i \in [I], \label{drcc}$$ and accordingly we call *distributionally robust chance constraints* (DRCCs).
[ In this paper, we consider distributionally robust chance-constrained binary programs (DCBP) that involve binary $y_i$-variables and DRCCs .]{} Without making the Gaussian assumption on $\mathbb{P}$, we show that a DCBP is equivalent to a 0-1 SOC program when $\mathcal{D}$ is characterized by the first two moments of $\tilde{t}_{i}$. Furthermore, [ building upon existing work on valid inequalities for submodular/supermodular functions]{}, we exploit the submodularity of the 0-1 SOC program to generate valid inequalities. As demonstrated in extensive computational experiments of bin packing instances, these valid inequalities significantly accelerate the branch-and-cut algorithm for solving the related DCBPs. Notably, the proposed submodular approximations and the resulting valid inequalities apply to general 0-1 SOC programs than DCBPs.
The remainder of the paper is organized as follows. Section \[sec:prior\] reviews the prior work related to optimization techniques used in this paper and stochastic bin packing problems. Section \[sec:model\] presents two 0-1 SOC representations, respectively, for DRCCs under two moment-based ambiguity sets. Section \[sec:algorithm\] utilizes submodularity and lifting to generate valid inequalities to strengthen the 0-1 SOC formulations. [ Section \[sec:result\] demonstrates the computational efficacy of our approaches for solving different 0-1 SOC reformulations of a DCBP for chance-constrained bin packing, with diverse problem sizes and parameter settings. Section \[conclusion\] summarizes the paper and discusses future research directions. ]{}
Prior Work {#sec:prior}
==========
Chance-constrained binary programs with uncertain technology matrix and/or right-hand side are computationally challenging, largely because (i) the non-convexity of chance constraints and (ii) the discrete variables. The majority of existing literature requires full distributional knowledge of the random coefficients and applies the SAA approach to approximate the models as MILPs. For example, [@song2014chance] considered a generic chance-constrained binary packing problem using finite samples of the random item weights, and derived lifted cover inequalities to accelerate the computation. For generic chance-constrained programs, the SAA approach and valid inequalities for the related MILPs have been well studied in the literature (see, e.g., [@CCP_luedtke-ahmed-2008; @Luedtke2010; @CCP_kucukyavuz2010mixing; @luedtke2014branch]). We also refer to [@ORA_Shylo_CC; @yan2013ORP; @shen-wang-CC-SS2013] for wide applications of chance-constrained binary programs, mainly in service systems and operations. As compared to the existing work, this paper waives the assumption of full distributional information and only relies on the first two moments of the uncertainty.
Distributionally robust optimization has received growing attention, mainly because it provides effective modeling and computational approaches for handling ambiguous distributions of random variables in stochastic programming by using available distributional information. Moment information has been widely used for building ambiguity sets in various distributionally robust optimization models (see, e.g., [@bertsimas2010models; @delage2010distributionally; @wiesemann2014distributionally]). Using moment-based ambiguity sets, [@el2003worst; @CCP_calafiore-el-ghaoui-2006; @wagner2008stochastic; @CCP_chen2010cvar; @zymler2013distributionally; @cheng2014distributionally; @jiang2013data] derived exact reformulations and/or approximations for DRCCs, often in the form of semidefinite programs (SDPs). In special cases, e.g., when the first two moments are *exactly* matched in the ambiguity set, the SDPs can further be simplified as SOC programs. While many existing ambiguity sets exactly match the first two moments of uncertainty (see, e.g., [@el2003worst; @wagner2008stochastic; @zymler2013distributionally]), [@delage2010distributionally] proposed a data-driven approach to construct an ambiguity set that can model moment estimation errors. In this paper, we consider both types of moment-based ambiguity sets. To the best of our knowledge, for the first time, we provide an SOC representation of DRCCs using the general ambiguity set proposed by [@delage2010distributionally].
Meanwhile, distributionally robust optimization has received much less attention in discrete optimization problems, possibly due to the difficulty of solving 0-1 nonlinear programs. For example, most off-the-shelf solvers cannot directly handle 0-1 SDPs, which often arise from discrete optimization problems with DRCCs. To the best of our knowledge, our results on chance constraints are most related to [@cheng2014distributionally] that studied DRCCs in the binary knapsack problem and derived 0-1 SDP reformulations. As compared to [@cheng2014distributionally], we investigate a different ambiguity set and derive a 0-1 *SOC* representation. Additionally, we solve the 0-1 SOC reformulation to *global* optimality instead of considering an SDP relaxation as in [@cheng2014distributionally].
In the seminal work [@nemhauser1978analysis], the authors identified submodularity in combinatorial and discrete optimization problems and proved a sufficient and necessary condition for 0-1 quadratic functions being submodular. We use this condition to exploit the submodularity of our 0-1 SOC reformulations. Indeed, submodular and supermodular knapsack sets (the discrete lower level set of a submodular function and discrete upper level set of a supermodular function, respectively) often arise when modeling utility, risk, and chance constraints on discrete variables. Extended polymatroid inequalities (see [@edmonds1970submodular; @Nemhauser1999CO]) can be efficiently obtained through greedy-based separation procedures to optimize submodular/supermodular functions. Recently, [@atamturk2009submodular; @atamturk2015supermodular; @bhardwaj2015binary] proposed cover and packing inequalities for efficiently solving submodular and supermodular knapsack sets with 0-1 variables. Our results on valid inequalities are most related to [@atamturk2018network; @bhardwaj2015binary] that identified a sufficient condition for the submodularity of 0-1 SOC constraints and strengthened their formulations by using the corresponding extended polymatroid inequalities (see Section 2.2 of [@atamturk2018network]). In contrast, we derive a different way to exploit the submodularity of general 0-1 SOC constraints. In particular, we apply the sufficient and necessary condition derived by [@nemhauser1978analysis] to search for “optimal” submodular approximations of the 0-1 SOC constraints (see Section \[subsec:submodular\]).
The main contributions of the paper are three-fold. First, using the general moment-based ambiguity set proposed by [@delage2010distributionally], we equivalently reformulate DRCCs as 0-1 SOC constraints that can readily be solved by solvers. Second, we exploit the (hidden) submodularity of the 0-1 SOC constraints and employ extended polymatroid valid inequalities to accelerate solving DCBP. In particular, we provide an efficient way of finding “optimal” submodular approximations of the 0-1 SOC constraints in the original variable space, and furthermore show that any 0-1 SOC constraint possesses submodularity in a lifted space. The valid inequalities in original and lifted spaces can both be efficiently separated via the well-known greedy algorithm (see, e.g., [@atamturk2017polymatroid; @atamturk2008polymatroids; @edmonds1970submodular]). Third, we conduct extensive numerical studies to demonstrate the computational efficacy of our solution approaches.
DCBP Models and Reformulations {#sec:model}
==============================
We study DRCCs under two alternatives of ambiguity set $\mathcal{D}$ based on the first two moments of $\tilde{t}_{i}, \ i \in [I]$. The first ambiguity set, denoted $\mathcal{D}_1$, exactly matches the mean and covariance matrix of each $\tilde{t}_{i}$. In contrast, the second ambiguity set, denoted $\mathcal{D}_2$, considers the estimation errors of sample mean and sample covariance matrix (see [@delage2010distributionally]). In Section \[subsec:as\], we introduce these two ambiguity sets and their calibration based on historical data. In Section \[subsec:soc-ref\], we derive SOC representations of DRCC under $\mathcal{D}_1$ and $\mathcal{D}_2$, respectively. While the former case (i.e., under $\mathcal{D}_1$) has been well studied (see, e.g., [@el2003worst; @CCP_calafiore-el-ghaoui-2006; @zymler2013distributionally]), to the best of our knowledge, this is the first work to show the SOC representation of the latter case [based on general covariance matrices]{} (i.e., under $\mathcal{D}_2$).
Ambiguity Sets {#subsec:as}
--------------
Suppose that a series of independent historical data samples $\{\tilde{t}^n_i\}_{n = 1}^N$ are drawn from the true probability distribution $\mathbb{P}$ of $\tilde{t}_{ij}$. Then, the first two moments of $\tilde{t}_i$ can be estimated by the sample mean and sample covariance matrix $$\mu_i \ = \ \frac{1}{N} \sum_{n=1}^N \tilde{t}^n_i, \qquad \Sigma_i \ = \ \frac{1}{N} \sum_{n=1}^N (\tilde{t}^n_i - \mu_i) (\tilde{t}^n_i - \mu_i)^{\top}.$$ Throughout this paper, we assume that both $\Sigma_i$ and the true covariance matrix of $\tilde{t}_i$ are [symmetric and]{} positive definite. As promised by the law of large numbers, as the data size $N$ grows, $\mu_i$ and $\Sigma_i$ converge to the true mean and true covariance matrix of $\tilde{t}_i$, respectively. Hence, when $N$ takes a large value, a natural choice of the ambiguity set consists of all probability distributions that match the sample moments $\mu_i$ and $\Sigma_i$, i.e., $$\mathcal{D}_1 \ = \ \left\{\mathbb{P} \in {\color{black}\mathcal{P}(\mathbb{R}^J)}: \
\begin{array}{l} \mathbb{E_P}[\tilde{t}_i] = \mu_i, \\[0.3cm]
\mathbb{E_P}[(\tilde{t}_i - \mu_i) (\tilde{t}_i - \mu_i)^{\top}] = \Sigma_i, \ \forall i \in [I]
\end{array}
\right\},$$ where [$\mathcal{P}(\mathbb{R}^J)$]{} represents the set of all probability distributions on $\mathbb{R}^J$. Under many circumstances, however, the historical data can be inadequate. With a small $N$, there may exist considerable estimation errors in $\mu_i$ and $\Sigma_i$, which brings a layer of “moment ambiguity” into $\mathcal{D}_1$ and adds to the existing distributional ambiguity of $\mathbb{P}$. To address the moment ambiguity and take into account the estimation errors, [@delage2010distributionally] proposed an alternative ambiguity set $$\mathcal{D}_2 \ = \ \left\{\mathbb{P} \in {\color{black}\mathcal{P}(\mathbb{R}^J)}: \
\begin{array}{l} (\mathbb{E_P}[\tilde{t}_i] - \mu_i)^{\top} \Sigma_i^{-1} (\mathbb{E_P}[\tilde{t}_i] - \mu_i) \ \leq \ \gamma_1, \\[0.3cm]
\mathbb{E_P}\bigl[(\tilde{t}_i - \mu_i) (\tilde{t}_i - \mu_i)^{\top}\bigr] \ \preceq \ \gamma_2 \Sigma_i, \ \ \forall i \in [I]
\end{array}
\right\},$$ where $\gamma_1 > 0$ and $\gamma_2 > \max\{\gamma_1, 1\}$ represent two given parameters. Set $\mathcal{D}_2$ designates that (i) the true mean of $\tilde{t}_i$ is within an ellipsoid centered at $\mu_i$, and (ii) the true covariance matrix of $\tilde{t}_i$ is bounded from above by $\gamma_2 \Sigma_i - (\mathbb{E_P}[\tilde{t}_i] - \mu_i)(\mathbb{E_P}[\tilde{t}_i] - \mu_i)^{\top}$ (note that $\mathbb{E_P}[(\tilde{t}_i - \mu_i) (\tilde{t}_i - \mu_i)^{\top}] = \mathbb{E_P}[(\tilde{t}_i - \mathbb{E_P}[\tilde{t}_i]) (\tilde{t}_i - \mathbb{E_P}[\tilde{t}_i])^{\top}] + (\mathbb{E_P}[\tilde{t}_i] - \mu_i)(\mathbb{E_P}[\tilde{t}_i] - \mu_i)^{\top}$).
[@delage2010distributionally] offered a rigorous guideline for selecting $\gamma_1$ and $\gamma_2$ values (see Theorem 2 in [@delage2010distributionally]) so that $\mathcal{D}_2$ includes the true distribution of $\tilde{t}_i$ with a high confidence level. In practice, we can select the values of $\gamma_1$ and $\gamma_2$ via cross validation. For example, we can divide the $N$ data points into two halves. We estimate sample moments ($\mu^1_i$, $\Sigma^1_i$) based on the first half of the data and ($\mu^2_i$, $\Sigma^2_i$) based on the second half. Then, we characterize $\mathcal{D}_2$ based on ($\mu^1_i$, $\Sigma^1_i$), and select $\gamma_1$ and $\gamma_2$ such that probability distributions with moments ($\mu^2_i$, $\Sigma^2_i$) belong to $\mathcal{D}_2$.
SOC Representations of the DRCC {#subsec:soc-ref}
-------------------------------
Now we derive SOC representations of DRCC for all $i \in [I]$. For notation brevity, we define vector $y := [y_{i1}, \ldots, y_{iJ}]$ and omit the subscript $i$ throughout this section. First, we review the celebrated SOC representation of DRCC under ambiguity set $\mathcal{D}_1$ in the following theorem.
\[thm:D-1\] (Adapted from [@el2003worst], also see [@wagner2008stochastic]) The DRCC with $\mathcal{D} = \mathcal{D}_1$ is equivalent to the following SOC constraint: $$\mu^{\top} y + \sqrt{\frac{1 - \alpha}{\alpha}} \sqrt{y^{\top} \Sigma y} \ \leq \ T. \label{ref-0}$$
Theorem \[thm:D-1\] shows that we can recapture the convexity of DRCC by employing ambiguity set $\mathcal{D}_1$ to model the $\tilde{t}$ uncertainty. Perhaps more surprisingly, in this case, the convex feasible region characterized by DRCC is SOC representable. It follows that the continuous relaxation of the DCBP model is an SOC program, [which can be solved very efficiently by standard nonlinear optimization solvers]{}. Next, we show that DRCC under the ambiguity set $\mathcal{D}_2$ is also SOC representable. This implies that the computational complexity of the DCBP remains the same even if we take the moment ambiguity into account. We present the main result of this section in the following theorem.
\[thm:D-2\] DRCC with $\mathcal{D} = \mathcal{D}_2$ is equivalent to
\[ref-1+2\] $$\mu^{\top} y + \Biggl(\sqrt{\gamma_1} + \sqrt{\Bigl(\frac{1 - \alpha}{\alpha}\Bigr) (\gamma_2 - \gamma_1)}\Biggr) \sqrt{y^{\top} \Sigma y} \ \leq \ T \label{ref-1}$$ if $\gamma_1/\gamma_2 \leq \alpha$, and is equivalent to $$\mu^{\top} y + \sqrt{\frac{\gamma_2}{\alpha}} \sqrt{y^{\top} \Sigma y} \ \leq \ T \label{ref-2}$$
if $\gamma_1/\gamma_2 > \alpha$.
[[@rujeerapaiboon2015robust] considered an ambiguity set similar to $\mathcal{D}_2$ and derived an SOC representation of DRCCs in portfolio optimization under an assumption of weak sense white noise, i.e., the uncertainty is stationary and mutually uncorrelated over time (see Definition 4 and Theorem 5 in [@rujeerapaiboon2015robust]). In contrast, the SOC representation in Theorem \[thm:D-2\] holds for general covariance matrices.]{} We prove Theorem \[thm:D-2\] in two steps. In the first step, we project the random vector $\tilde{t}$ and its ambiguity set $\mathcal{D}_2$ from $\mathbb{R}^J$ to the real line, i.e., $\mathbb{R}$. This simplifies DRCC as involving a one-dimensional random variable. In the second step, we derive optimal (i.e., worst-case) mean and covariance matrix in $\mathcal{D}_2$ that attain the worst-case probability bound in . We then [apply Cantelli’s inequality]{} to finish the representation. We present the first step of the proof in the following lemma.
\[lemma:projection\] Let $\tilde{s}$ be a random vector in $\mathbb{R}^J$ and $\tilde{\xi}$ be a random variable in $\mathbb{R}$. For a given $y \in \mathbb{R}^J$, define ambiguity sets $\mathcal{D}_{\tilde{s}}$ and $\mathcal{D}_{\tilde{\xi}}$ as
$$\mathcal{D}_{\tilde{s}} \ = \ \Bigl\{\mathbb{P} \in {\color{black}\mathcal{P}(\mathbb{R}^J)}: \ \mathbb{E_P}[\tilde{s}]^{\top}\Sigma^{-1}\mathbb{E_P}[\tilde{s}] \leq \gamma_1, \quad \mathbb{E_P}[\tilde{s} \tilde{s}^{\top}] \preceq \gamma_2 \Sigma\Bigr\} \label{D-s}$$
and $$\mathcal{D}_{\tilde{\xi}} \ = \ \Bigl\{\mathbb{P} \in {\color{black}\mathcal{P}(\mathbb{R})}: \ |\mathbb{E_P}[\tilde{\xi}]| \leq \sqrt{\gamma_1}\sqrt{y^{\top}\Sigma y}, \quad \mathbb{E_P}[\tilde{\xi}^2] \leq \gamma_2 (y^{\top}\Sigma y) \Bigr\}. \label{D-xi}$$
Then, for any Borel measurable function $f: \mathbb{R}^J \rightarrow \mathbb{R}$, we have $$\inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{f(y^{\top}\tilde{s}) \leq 0\} \ = \ \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{f(\tilde{\xi}) \leq 0\}.$$
[[*Proof:* ]{}]{}We first show that $\inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{f(y^{\top}\tilde{s}) \leq 0\} \geq \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{f(\tilde{\xi}) \leq 0\}$. Pick a $\mathbb{P} \in \mathcal{D}_{\tilde{s}}$, and let $\tilde{s}$ denote the corresponding random vector and $\tilde{\xi} = y^{\top}\tilde{s}$. It follows that
$$\begin{aligned}
\mathbb{E_P}[\tilde{\xi}] \ = & \ y^{\top} \mathbb{E_P}[\tilde{s}] \nonumber \\
\leq & \ \max_{s: \ s^{\top} \Sigma^{-1} s \ \leq \ \gamma_1} \ y^{\top} s \label{proj-note-1} \\
= & \ \max_{z: \ ||z||_2 \ \leq \sqrt{\gamma_1}} \ (\Sigma^{1/2}y)^{\top}z \ = \ \sqrt{\gamma_1} \sqrt{y^{\top} \Sigma y}, \nonumber\end{aligned}$$
where inequality is because $\mathbb{E_P}[\tilde{s}]^{\top}\Sigma^{-1}\mathbb{E_P}[\tilde{s}] \leq \gamma_1$. Similarly, we have $\mathbb{E_P}[\tilde{\xi}] \geq -\sqrt{\gamma_1} \sqrt{y^{\top} \Sigma y}$. Meanwhile, note that $$\begin{aligned}
\mathbb{E_P}[\tilde{\xi}^2] \ = & \ y^{\top} \mathbb{E_P}[\tilde{s}\tilde{s}^{\top}] y \nonumber \\
\leq & \ y^{\top} (\gamma_2 \Sigma) y \ = \ \gamma_2 (y^{\top} \Sigma y), \label{proj-note-2}\end{aligned}$$ where inequality is because $\mathbb{E_P}[\tilde{s}\tilde{s}^{\top}] \preceq \gamma_2 \Sigma$. Hence, the probability distribution of $\tilde{\xi}$ belongs to $\mathcal{D}_{\tilde{\xi}}$. It follows that $\inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{f(y^{\top}\tilde{s}) \leq 0\} \geq \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{f(\tilde{\xi}) \leq 0\}$.
Second, we show that $\inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{f(y^{\top}\tilde{s}) \leq 0\} \leq \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{f(\tilde{\xi}) \leq 0\}$. Pick a $\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}$, and let $\tilde{\xi}$ denote the corresponding random variable and $\tilde{s} = \bigl[\tilde{\xi}/(y^{\top}\Sigma y)\bigr] \Sigma y$. It follows that $$\begin{aligned}
\mathbb{E_P}[\tilde{s}]^{\top}\Sigma^{-1}\mathbb{E_P}[\tilde{s}] \ = & \ \mathbb{E_P}[\tilde{\xi}]^2 \frac{y^{\top} \Sigma}{y^{\top}\Sigma y}\Sigma^{-1} \frac{\Sigma y}{y^{\top}\Sigma y} \ = \ \frac{\mathbb{E_P}[\tilde{\xi}]^2}{y^{\top} \Sigma y} \nonumber \\
\leq & \ \frac{\gamma_1 y^{\top} \Sigma y}{y^{\top} \Sigma y} \ = \ \gamma_1, \label{proj-note-3}\end{aligned}$$ where inequality is because $|\mathbb{E_P}[\tilde{\xi}]| \leq \sqrt{\gamma_1} \sqrt{y^{\top} \Sigma y}$. Meanwhile, note that $$\begin{aligned}
\mathbb{E_P}[\tilde{s} \tilde{s}^{\top}] \ = & \ \mathbb{E_P}\left[\tilde{\xi}^2\frac{\Sigma y}{y^{\top} \Sigma y}\frac{y^{\top} \Sigma}{y^{\top} \Sigma y}\right] \nonumber \\
= & \ \mathbb{E_P}[\tilde{\xi}^2] \frac{(\Sigma y)(\Sigma y)^{\top}}{(y^{\top} \Sigma y)^2} \nonumber \\
\preceq & \ \gamma_2 (y^{\top} \Sigma y) \frac{(\Sigma y)(\Sigma y)^{\top}}{(y^{\top} \Sigma y)^2} \label{proj-note-4} \\
\preceq & \ \gamma_2 (y^{\top} \Sigma y) \frac{(y^{\top} \Sigma y)\Sigma}{(y^{\top} \Sigma y)^2} \ = \ \gamma_2 \Sigma, \label{proj-note-5}\end{aligned}$$ where inequality is because $\mathbb{E_P}[\tilde{\xi}^2] \leq \gamma_2 (y^{\top} \Sigma y)$ and inequality is because $(\Sigma y)(\Sigma y)^{\top} \preceq (y^{\top} \Sigma y) \Sigma$, which holds because for all $z \in \mathbb{R}^J$, $$\begin{aligned}
z^{\top} (\Sigma y)(\Sigma y)^{\top} z \ = & \ \left[ (\Sigma^{1/2}z)^{\top}(\Sigma^{1/2}y) \right]^2 \\
\leq & \ ||\Sigma^{1/2}z||^2 \ ||\Sigma^{1/2}y||^2 \ \ \ \ \ \ \ \ \mbox{(Cauchy-Schwarz inequality)} \\
= & \ (y^{\top} \Sigma y) (z^{\top} \Sigma z) \\
= & \ z^{\top} [(y^{\top} \Sigma y)\Sigma] z.\end{aligned}$$
Hence, the probability distribution of $\tilde{s}$ belongs to $\mathcal{D}_{\tilde{s}}$. It follows that $\inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{$ $f(y^{\top}\tilde{s}) \leq 0\} \leq \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{f(\tilde{\xi}) \leq 0\}$ because $\tilde{\xi} = y^{\top} \tilde{s}$, and the proof is completed. [$\Box$ ]{}
[@popescu2007robust] and [@yu2009general] showed a similar projection property for $\mathcal{D}_1$, i.e., when the first two moments of $\tilde{s}$ are exactly known. Lemma \[lemma:projection\] employs a different transformation approach to show the projection property for $\mathcal{D}_2$ when these moments are ambiguous in the sense of [@delage2010distributionally].
We are now ready to finish the proof of Theorem \[thm:D-2\].\
*Proof of Theorem \[thm:D-2\]*: First, we define random vector $\tilde{s} = \tilde{t} - \mu$, random variable $\tilde{\xi} = y^{\top}\tilde{s}$, constant $b = T - \mu^{\top} y$, and set $S$ such that $$S \ = \ \{(\mu_1, \sigma_1) \in \mathbb{R}\times\mathbb{R}_+: \ |\mu_1| \leq \sqrt{\gamma_1}\sqrt{y^{\top} \Sigma y}, \ \mu_1^2 + \sigma_1^2 \leq \gamma_2 y^{\top} \Sigma y\}.$$ It follows that
$$\begin{aligned}
\inf_{\mathbb{P} \in \mathcal{D}_2} \mathbb{P}\{\tilde{t}^{\top} y \leq T\} \ = & \ \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{s}}} \mathbb{P}\{y^{\top} \tilde{s} \leq b\} \nonumber \\
= & \inf_{\mathbb{P} \in \mathcal{D}_{\tilde{\xi}}} \mathbb{P}\{\tilde{\xi} \leq b\} \label{proj-note-6} \\
= & \inf_{(\mu_1, \sigma_1) \in S} \ \inf_{\mathbb{P} \in \mathcal{D}_1(\mu_1, \sigma_1^2)} \mathbb{P}\{\tilde{\xi} \leq b\}, \label{proj-note-7}\end{aligned}$$
where $\mathcal{D}_{\tilde{s}}$ and $\mathcal{D}_{\tilde{\xi}}$ are defined in and , respectively, equality follows from Lemma \[lemma:projection\], and equality decomposes the optimization problem in into two layers: the outer layer searches for the optimal (i.e., worst-case) mean and covariance, while the inner layer computes the worst-case probability bound under the given mean and covariance. For the inner layer, based on [Cantelli’s inequality]{}, we have $$\inf_{\mathbb{P} \in \mathcal{D}_1(\mu_1, \sigma_1^2)} \mathbb{P} \{\tilde{\xi} \leq b\} = \left\{
\begin{array}{ll}
\frac{(b-\mu_1)^2}{\sigma_1^2 + (b-\mu_1)^2}, & \mbox{if $b \geq \mu_1$,} \\[0.15cm]
0, & \mbox{o.w.}
\end{array}\right.$$ As DRCC states that $\inf_{\mathbb{P} \in \mathcal{D}_2} \mathbb{P}\{\tilde{t}^{\top} y \leq T\} \geq 1 - \alpha > 0$, we can assume $b \geq \mu_1$ for all $(\mu_1, \sigma_1) \in S$ without loss of generality. That is, $$b \ \geq \ \max_{(\mu_1, \sigma_1) \in S} \mu_1 \ = \ \sqrt{\gamma_1} \sqrt{y^{\top} \Sigma y}.$$ It follows that $$\begin{aligned}
\inf_{\mathbb{P} \in \mathcal{D}_2} \mathbb{P}\{\tilde{t}^{\top} y \leq T\} \ = & \ \inf_{(\mu_1, \sigma_1) \in S} \ \frac{(b-\mu_1)^2}{\sigma_1^2 + (b-\mu_1)^2} \nonumber \\
= & \ \inf_{(\mu_1, \sigma_1) \in S} \ \frac{1}{\left(\frac{\sigma_1}{b-\mu_1}\right)^2 + 1}. \label{proj-note-8}\end{aligned}$$ Note that the objective function value in decreases as $\sigma_1/(b-\mu_1)$ increases. Hence, shares optimal solutions with the following optimization problem: $$\inf_{(\mu_1, \sigma_1) \in S} \ - \Bigl(\frac{\sigma_1}{b-\mu_1}\Bigr). \label{proj-note-9}$$ The feasible region of problem is depicted in the shaded area of Figure \[fig:01\].
![Graphical Solution of Problem []{data-label="fig:01"}](circle.pdf){width=".9\textwidth"}
Furthermore, we note that the objective function of equals to the slope of the straight line connecting points $(b, 0)$ and $(\mu_1, \sigma_1)$ (see Figure \[fig:01\] for an example). It follows that an optimal solution ($\mu_1^*, \sigma_1^*$) to problem , and so to problem , lies in one of the following two cases:
Case 1.
: If $\sqrt{\gamma_1}\sqrt{y^{\top} \Sigma y} \leq b \leq (\gamma_2/\sqrt{\gamma_1}) \sqrt{y^{\top} \Sigma y}$, then $\mu_1^* = \sqrt{\gamma_1}\sqrt{y^{\top} \Sigma y}$ and $\sigma_1^* = \sqrt{\gamma_2 - \gamma_1}\sqrt{y^{\top} \Sigma y}$.
Case 2.
: If $b > (\gamma_2/\sqrt{\gamma_1}) \sqrt{y^{\top} \Sigma y}$, then $\mu_1^* = (\gamma_2 y^{\top} \Sigma y)/b$ and\
$\sigma_1^* = \sqrt{\gamma_2 y^{\top} \Sigma y - (\gamma_2 y^{\top} \Sigma y)^2/b^2}$.
Denoting $\kappa (b, y) = \frac{b}{\sqrt{y^{\top} \Sigma y}}$, we have $$\inf_{\mathbb{P} \in \mathcal{D}_2} \mathbb{P}\{\tilde{t}^{\top} y \leq T\} \ = \ \left\{
\begin{array}{ll} \frac{1}{\left( \frac{\sqrt{\gamma_2 - \gamma_1}}{\kappa(b, y) - \sqrt{\gamma_1}} \right)^2 + 1}, & \mbox{if } \sqrt{\gamma_1} \leq \kappa(b, y) \leq \frac{\gamma_2}{\sqrt{\gamma_1}}, \\[0.75cm]
{\color{black}\frac{\kappa(b,y)^2 - \gamma_2}{\kappa(b,y)^2}}, & \mbox{if } \kappa(b, y) > \frac{\gamma_2}{\sqrt{\gamma_1}}.
\end{array}
\right. \label{wc-prob-bound}$$
Second, based on , the DRCC $\inf_{\mathbb{P} \in \mathcal{D}_2} \mathbb{P}\{\tilde{t}^{\top} y \leq T\} \geq 1 - \alpha$ has the following representations: $$\mbox{DRCC} \ \Leftrightarrow \ \left\{
\begin{array}{ll} \frac{b}{\sqrt{y^{\top} \Sigma y}} \ \geq \ \sqrt{\gamma_1} + \sqrt{\Bigl(\frac{1 - \alpha}{\alpha}\Bigr)(\gamma_2 - \gamma_1)}, & \mbox{if } \sqrt{\gamma_1} \leq \frac{b}{\sqrt{y^{\top} \Sigma y}} \leq \frac{\gamma_2}{\sqrt{\gamma_1}}, \\[0.2cm]
\frac{b}{\sqrt{y^{\top} \Sigma y}} \ \geq \ \sqrt{\frac{\gamma_2}{\alpha}}, & \mbox{if } \frac{b}{\sqrt{y^{\top} \Sigma y}} > \frac{\gamma_2}{\sqrt{\gamma_1}}.
\end{array}
\right.$$
It follows that DRCC is equivalent to an SOC constraint by discussing the following two cases:
Case 1.
: If $\gamma_1 / \gamma_2 \leq \alpha$, then $\gamma_2/\sqrt{\gamma_1} \geq \sqrt{\gamma_1} + \sqrt{[(1 - \alpha)/\alpha](\gamma_2 - \gamma_1)}$ and $\gamma_2/\sqrt{\gamma_1} \geq \sqrt{\gamma_2/\alpha}$. It follows that [ (i) if $\sqrt{\gamma_1} \leq b/\sqrt{y^{\top} \Sigma y} \leq \gamma_2/\sqrt{\gamma_1}$, then DRCC is equivalent to $b/\sqrt{y^{\top} \Sigma y} \geq \sqrt{\gamma_1} + \sqrt{[(1 - \alpha)/\alpha](\gamma_2 - \gamma_1)}$ and (ii) if $b/\sqrt{y^{\top} \Sigma y}$ $> \gamma_2/\sqrt{\gamma_1}$, then DRCC always holds. Combining sub-cases (i) and (ii) yields that]{} DRCC is equivalent to $b/\sqrt{y^{\top} \Sigma y} \geq \sqrt{\gamma_1} + \sqrt{[(1 - \alpha)/\alpha](\gamma_2 - \gamma_1)}$.
Case 2.
: If $\gamma_1 / \gamma_2 > \alpha$, then $\gamma_2/\sqrt{\gamma_1} < \sqrt{\gamma_1} + \sqrt{[(1 - \alpha)/\alpha](\gamma_2 - \gamma_1)}$ and $\gamma_2/\sqrt{\gamma_1} < \sqrt{\gamma_2/\alpha}$. It follows that [ (i) if $\sqrt{\gamma_1} \leq b/\sqrt{y^{\top} \Sigma y} \leq \gamma_2/\sqrt{\gamma_1}$, then DRCC always fails and (ii) if $b/\sqrt{y^{\top} \Sigma y} > \gamma_2/\sqrt{\gamma_1}$, then DRCC is equivalent to $b/\sqrt{y^{\top} \Sigma y} \geq \sqrt{\gamma_2/\alpha}$. Combining sub-cases (i) and (ii) yields that]{} DRCC is equivalent to $b/\sqrt{y^{\top} \Sigma y} \geq \sqrt{\gamma_2/\alpha}$.
The proofs of the above two cases are both completed by the definition of $b$. [$\Box$ ]{}
To sum up, we have two exact 0-1 SOC constraint reformulations of DRCC under ambiguity sets $\mathcal{D}_1$ and $\mathcal{D}_2$, being constraints , /, respectively.
\[ex:threeOmega\] We consider a DRCC with $I = 1$ (the subscript $i$ is hence omitted), $J = 2$, $1- \alpha = 95\%$, mean vector $\mu = [0 \ 0]^\top$ and covariance matrix $
\Sigma \ = \ \begin{bmatrix}
2 & 1 \\
1 & 2
\end{bmatrix}.$ For the ambiguity set $\mathcal D_2$, we set $\gamma_1 = 1$ and $\gamma_2=2$. We note that $\Phi^{-1}(1 - \alpha) = 1.6449$ in the SOC reformulation , $\sqrt{(1-\alpha)/\alpha} = 4.3589$ in the SOC reformulation , and $\sqrt{\gamma_2/\alpha} = 6.3246$ in the SOC reformulation (since $\gamma_1/\gamma_2 > \alpha$). Without specifying the value of $T$, we depict the boundaries of the second-order cones associated with the three SOC reformulations in Figure \[fig:threeOmega\]. From this figure, we observe that the Gaussian approximation leads to the widest cone and so the largest SOC feasible region corresponding to DRCC . [$\Box$ ]{}
Valid Inequalities for DCBP {#sec:algorithm}
===========================
Although 0-1 SOC constraint reformulations can be directly handled by the off-the-shelf solvers, as we report in Section \[sec:result\], [ the resultant 0-1 SOC programs]{} are often time-consuming to solve, primarily because of the binary restrictions on variables. In this section, we derive valid inequalities for DRCC , with the objective of accelerating the branch-and-cut algorithm for solving DCBP with individual DRCCs and also general 0-1 SOC programs in commercial solvers. Specifically, we exploit the submodularity of 0-1 SOC constraints. In Section \[subsec:submodular\], we derive a sufficient condition for submodularity and two approximations of the 0-1 SOC constraints that satisfy this condition. Using the submodular approximations, we derive extended polymatroid inequalities. In Section \[subsec:bp\], we show the submodularity of the 0-1 SOC constraints in a lifted (i.e., higher-dimensional) space. While the extended polymatroid inequalities are well-known (see, e.g., [@atamturk2017polymatroid; @atamturk2008polymatroids; @edmonds1970submodular]), to the best of our knowledge, the two submodular approximations and the submodularity of 0-1 SOC constraints in the lifted space appear to be new and have not been studied in any literature.
Submodularity of the 0-1 SOC Constraints {#subsec:submodular}
----------------------------------------
We consider SOC constraints of the form $$\mu^{\top} y + \sqrt{y^{\top} \Lambda y} \ \leq \ T, \label{soc}$$ where $\Lambda$ represents a $J\times J$ positive definite matrix. Note that all SOC reformulations , , and derived in Section \[sec:model\], as well as the SOC reformulation of a chance constraint with Gaussian uncertainty, possess the form of . Before investigating the submodularity of , we review the definitions of submodular functions and extended polymatroids.
\[def:sub\] Define the collection of set $[J]$’s subsets $\mathcal{C} := \{R: R \subseteq [J]\}$. A set function $h$: $\mathcal{C} \rightarrow \mathbb R$ is called submodular if and only if $h(R \cup \{j\}) - h(R) \geq h(S \cup \{j\}) - h(S)$ for all $R \subseteq S \subseteq [J]$ and all $j \in [J]\setminus S$.
Throughout this section, we refer to a set function as $h(R)$ and $h(y)$ interchangeably, where $y \in \{0, 1\}^J$ represents the indicating vector for subset $R \subseteq [J]$, i.e., $y_j = 1$ if $j \in R$ and $y_j = 0$ otherwise.
For a submodular function $h(y)$, the polyhedron $$\mbox{EP}_h = \{\pi \in \mathbb R^J: \ \pi(R) \leq h(R), \quad \forall R \subseteq [J]\}$$ is called an extended polymatroid associated with $h(y)$, where $\pi(R) = \sum_{j \in R} \pi_j$.
For a submodular function $h(y)$ with $h(\emptyset) = 0$, inequality $$\pi^{\top} y \leq t, \label{ep}$$ termed an *extended polymatroid inequality*, is valid for the epigraph of $h$, i.e., $$\{(y, t) \in \{0, 1\}^J \times \mathbb{R}: t \geq h(y)\}, \mbox{ if and only if $\pi \in \mbox{EP}_h$ {\color{black} (see \cite{edmonds1970submodular})} }.$$ Furthermore, the separation of can be efficiently done by a greedy algorithm [@edmonds1970submodular], which we briefly describe in Algorithm \[algo:greedy\].
A point $(\hat{y}, \hat{t})$ with $\hat{y} \in [0, 1]^J$, $\hat{t} \in \mathbb{R}$, and a function $h$ that is submodular on $\{0,1\}^J$. Sort the entries in $\hat{y}$ such that $y_{(1)} \geq \cdots \geq y_{(J)}$. Obtain the permutation $\{(1), \ldots, (J)\}$ of $[J]$. Letting $R_{(j)} := \{(1), \ldots, (j)\}, \ \forall j \in \mathcal{J}$, compute $\hat{\pi}_{(1)} = h(R_{(1)})$, and $\hat{\pi}_{(j)} = h(R_{(j)}) - h(R_{(j-1)})$ for $j = 2, \ldots, J$. We generate a valid extended polymatroid inequality $\hat{\pi}^{\top} y \leq t$. The current solution $(\hat{y}, \hat{t})$ satisfies $h(\hat{y}) \leq \hat{t}$. either $(\hat{y}, \hat{t})$ is feasible, or a violated extended polymatroid inequality $\hat{\pi}^{\top} y \leq t$.
The strength and efficient separation of the extended polymatroid inequality motivate us to explore the submodularity of function $g(y) := \mu^{\top} y + \sqrt{y^{\top} \Lambda y}$. In a special case, $\Lambda$ is assumed to be a diagonal matrix and so the random coefficients $\tilde{t}_{ij}, j \in [J]$ for the same $i$ are uncorrelated. In this case, [@atamturk2008polymatroids] successfully show that $g(y)$ is submodular. As a result, we can strengthen a DCBP by incorporating extended polymatroid inequalities in the form $\pi^{\top} y \leq T$, where $\pi \in \mbox{EP}_g$. Unfortunately, the submodularity of $g(y)$ quickly fades when the off-diagonal entries of $\Lambda$ become non-zero, e.g., when $\Lambda$ is associated with a general covariance matrix. We present an example as follows.
\[ex:general-cov\] Suppose that $[J] = \{1, 2, 3\}$, $\mu = [0, 0, 0]^\top$, and $$\Lambda \ = \ \begin{bmatrix}
0.6 & -0.2 & 0.2 \\
-0.2 & 0.7 & 0.1 \\
0.2 & 0.1 & 0.6
\end{bmatrix}.$$ The three eigenvalues of $\Lambda$ are 0.2881, 0.7432, and 0.8687, and so $\Lambda \succ 0$. However, function $g(y) = \mu^{\top} y + \sqrt{y^{\top} \Lambda y}$ is not submodular because $g(R \cup \{j\}) - g(R) < g(S \cup \{j\}) - g(S)$, where $R = \{1\}$, $S = \{1, 2\}$, and $j = 3$. [$\Box$ ]{}
In this section, we provide a sufficient condition for function $g(y)$ being submodular for general $\Lambda$. To this end, we apply a necessary and sufficient condition derived in [@nemhauser1978analysis] for quadratic function $y^{\top}\Lambda y$ being submodular (see the second paragraph on Page 276 of [@nemhauser1978analysis], following Proposition 3.5). We summarize this condition in the following theorem.
\[thm:submodular\] ([@nemhauser1978analysis]) Define function $h: \{0, 1\}^J \rightarrow \mathbb{R}$ such that $h(y) := y^{\top}\Lambda y$, where $\Lambda \in \mathbb{R}^{J\times J}$ represents a symmetric matrix. Then, $h(y)$ is submodular if and only if $\Lambda_{rs} \leq 0$ for all $r, s \in [J]$ and $r \neq s$.
Note that Theorem \[thm:submodular\] does not assume $\Lambda \succeq 0$ and so can be applied to general (convex or non-convex) quadratic functions. Theorem \[thm:submodular\] leads to a sufficient condition for function $g(y)$ being submodular, [as presented in the following proposition.]{}
\[prop-sufficient\] Let $\Lambda \in \mathbb{R}^{J \times J}$ represent a symmetric and positive semidefinite matrix that satisfies (i) $2\sum_{s=1}^J \Lambda_{rs} \geq \Lambda_{rr}$ for all $r \in [J]$ and (ii) $\Lambda_{rs} \leq 0$ for all $r, s \in [J]$ and $r \neq s$. Then, function $g(y) = \mu^{\top} y + \sqrt{y^{\top} \Lambda y}$ is submodular.
[[*Proof:* ]{}]{}As $\mu^{\top}y$ is submodular in $y$, it suffices to prove that $\sqrt{y^{\top} \Lambda y}$ is submodular. Hence, we can assume $\mu = 0$ without loss of generality. We let $f(x) = \sqrt{x}$ and $h(y) = y^{\top} \Lambda y$. Then, $g(y) = f(h(y))$.
First, we note that $h(y \vee e_r) \geq h(y)$ for all $y \in \{0, 1\}^J$ and $r \in [J]$, where $a \vee b = [\max\{a_1, b_1\}, \ldots, \max\{a_J, b_J\}]^{\top}$ for $a, b \in \mathbb{R}^J$. Indeed, if $y_r = 1$ then $y \vee e_r = y$ and so $h(y \vee e_r) = h(y)$. Otherwise, if $y_r = 0$, then $y \vee e_r = y + e_r$ and so $$\begin{aligned}
h(y \vee e_r) \ = & \ y^{\top}\Lambda y + 2e_r^{\top}\Lambda y + e_r^{\top}\Lambda e_r \\
= & \ y^{\top}\Lambda y + 2\sum_{s: \ y_s = 1} \Lambda_{rs} + \Lambda_{rr}\\
\geq & \ y^{\top}\Lambda y + 2\sum_{\substack{s = 1 \\ s \neq r}}^J \Lambda_{rs} + \Lambda_{rr}\\
\geq & \ y^{\top}\Lambda y \ = \ h(y),\end{aligned}$$ where the first inequality is due to $y_r = 0$ and condition (ii), and the second inequality is due to condition (i). It follows that $h(y') \geq h(y)$ for all $y, y' \in \{0, 1\}^J$ such that $y' \geq y$. Hence, $h(y)$ is increasing.
Second, based on Theorem \[thm:submodular\], $h(y)$ is submodular due to condition (ii). It follows that $g(y) = f(h(y))$ is submodular because function $f$ is concave and nondecreasing and function $h$ is submodular and increasing (see, e.g., Proposition 2.2.5 in [@simchi2014logic])[^2]. [$\Box$ ]{}
Proposition \[prop-sufficient\] generalizes the sufficient condition in [@atamturk2008polymatroids] because conditions (i)–(ii) are automatically satisfied if $\Lambda$ is diagonal and positive definite.
For general $\Lambda \succ 0$ that does not satisfy sufficient conditions (i)–(ii), we can approximate SOC constraint by replacing $\Lambda$ with a matrix $\Delta$ that satisfies these conditions. We derive relaxed and conservative submodular approximations of constraint in the following [proposition]{}.
\[prop-approx\] Constraint implies the SOC constraint $$\mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny L}} y} \ \leq \ T, \label{soc-relaxed}$$ where function $g^{\mbox{\tiny L}}(y) := \mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny L}} y}$ is submodular and $\Delta^{\mbox{\tiny L}}$ is an optimal solution of SDP
\[eq:relaxSDP\] $$\begin{aligned}
\min_{\Delta} \ & \ ||\Delta - \Lambda||_2 \label{relax-obj} \\
\mbox{s.t.} \ & \ 0 \ \preceq \ \Delta \ \preceq \ \Lambda, \label{relax-con-1} \\
& \ 2\sum_{s=1}^J \Delta_{rs} \geq \Delta_{rr}, \ \ \forall r \in [J], \label{relax-con-2} \\
& \ \Delta_{rs} \leq 0, \ \ \forall r, s \in [J] \mbox{ and } r \neq s. \label{relax-con-3}\end{aligned}$$
Additionally, constraint is implied by the SOC constraint $$\mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny U}} y} \ \leq \ T, \label{soc-conservative}$$ where function $g^{\mbox{\tiny U}}(y) := \mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny U}} y}$ is submodular and $\Delta^{\mbox{\tiny U}}$ is an optimal solution of SDP
\[eq:conservativeSDP\] $$\begin{aligned}
\min_{\Delta} \ & \ ||\Delta - \Lambda||_2 \label{conservative-obj} \\
\mbox{s.t.} \ & \ \Delta \ \succeq \ \Lambda, \ \ \mbox{\eqref{relax-con-2}--\eqref{relax-con-3}}. \label{conservative-con-1}\end{aligned}$$
[[*Proof:* ]{}]{}By construction, $g^{\mbox{\tiny L}}(y)$ is submodular because $\Delta^{\mbox{\tiny L}}$ satisfies constraints – and so conditions (i)–(ii). Additionally, constraint implies because $\Delta^{\mbox{\tiny L}}$ satisfies constraint and so $\Delta^{\mbox{\tiny L}} \preceq \Lambda$. Similarly, we obtain that $g^{\mbox{\tiny U}}(y)$ is submodular and constraint is implied by . [$\Box$ ]{}
Note that there always exist matrices $\Delta^{\mbox{\tiny L}}$ and $\Delta^{\mbox{\tiny U}}$ that are feasible to SDPs – and –, respectively. For example, $\mbox{diag}(\lambda_{\mbox{\tiny min}}, \ldots, \lambda_{\mbox{\tiny min}}) \in \mathbb{R}^{J\times J}$ satisfy constraints –, where $\lambda_{\mbox{\tiny min}}$ represents the smallest eigenvalue of matrix $\Lambda$. Additionally, $\mbox{diag}(\lambda_{\mbox{\tiny max}}, \ldots, \lambda_{\mbox{\tiny max}}) \in \mathbb{R}^{J\times J}$ satisfy constraints , where $\lambda_{\mbox{\tiny max}}$ represents the largest eigenvalue of matrix $\Lambda$.
By minimizing the $\ell^2$ distance between $\Delta$ and $\Lambda$ in objective functions and , we find “optimal” approximations of $\Lambda$ that satisfies sufficient conditions (i)–(ii) in Proposition \[prop-sufficient\]. Accordingly, we obtain “optimal” submodular approximations of the 0-1 SOC constraint . There are many possible alternatives of the $\ell^2$ norm in and . For example, formulations – and – remain SDPs if the $\ell^2$ norm is replaced by the $\ell^1$ norm or the $\ell^{\infty}$ norm. [ We have empirically tested $\ell^1$, $\ell^2$, and $\ell^{\infty}$ norms based on a server allocation problem (see Section \[sec:comp\_setup\] for a brief description) and the $\ell^2$ norm leads to the largest improvement on CPU time.]{} In computation, we only need to solve these two SDPs once to obtain $\Delta^{\mbox{\tiny L}}$ and $\Delta^{\mbox{\tiny U}}$. Then, extended polymatroid inequalities can be obtained from the relaxed approximation . Additionally, the conservative approximation leads to an upper bound of the optimal objective value of the related DCBP.
\[ex:ApproxMatrices\] Recall the $3\times 3$ matrix $\Lambda$ in Example \[ex:general-cov\] and the corresponding function $g(y) = \mu^{\top} y + \sqrt{0.6 y^2_1 + 0.7 y^2_2 + 0.6 y^2_3 - 0.4 y_1 y_2 + 0.4 y_1 y_3 + 0.2 y_2 y_3}$ being not submodular. We set $\mu = [0, 0, 0]^\top$ and apply Proposition \[prop-approx\] to optimize the two SDPs and , yielding the following two positive semidefinite matrices: $$\Delta^{\mbox{\tiny L}} \ = \ \begin{bmatrix}
0.35 & -0.15 & 0\\
-0.15 & 0.37 & 0\\
0 & 0 & 0.38
\end{bmatrix} \mbox{ and}$$ $$\Delta^{\mbox{\tiny U}} \ = \ \begin{bmatrix}
0.83 & -0.22 & 0\\
-0.22 & 0.95 & 0\\
0 & 0 & 0.82
\end{bmatrix}.$$ Due to Proposition \[prop-sufficient\], $g^{\mbox{\tiny L}}(y) := \mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny L}} y } = \sqrt{0.35 y^2_1 + 0.37 y^2_2 + 0.38 y^2_3 - 0.3 y_1 y_2} $ and $g^{\mbox{\tiny U}}(y) := \mu^{\top} y + \sqrt{y^{\top} \Delta^{\mbox{\tiny U}} y }= $ $\sqrt{0.83y^2_1 + 0.95 y^2_2 + 0.82 y^2_3 - 0.44 y_1 y_2}$ are submodular.
Now suppose that $T = 0.8$ and we are given $\hat{y} = [\hat{y}_1, \hat{y}_2, \hat{y}_3]^{\top} = [1,0.5,0.9]^{\top}$ with $g(\hat{y}) = \mu^{\top} \hat{y} + \sqrt{\hat{y}^{\top}\Lambda \hat{y}} = 1.23 > 0.8 = T$. First, with respect to constraint $g^{\mbox{\tiny L}}(y) \leq T$, we follow Algorithm \[algo:greedy\] to find an extended polymatroid inequality and note that this inequality is also valid for constraint $g(y) \leq T$. Specifically, we sort the entries of $\hat{y}$ to obtain $\hat{y}_1 \geq \hat{y}_3 \geq \hat{y}_2$ and $\{1, 3, 2\}$, the corresponding permutation. It follows that $R_{(1)} = \{1\}$, $R_{(2)} = \{1,3\}$, and $R_{(3)} = \{1, 3, 2\}$. Hence, $\hat{\pi}_{(1)} = g^{\mbox{\tiny L}}([1, 0, 0]^{\top}) = 0.59$, $\hat{\pi}_{(2)} = g^{\mbox{\tiny L}}([1, 0, 1]^{\top}) - g^{\mbox{\tiny L}}([1, 0, 0]^{\top}) = 0.26$, and $\hat{\pi}_{(3)} = g^{\mbox{\tiny L}}([1, 1, 1]^{\top}) - g^{\mbox{\tiny L}}([1, 0, 1]^{\top}) = 0.04$. This generates the extended polymatroid inequality $0.59 y_1 + 0.26 y_3 + 0.04 y_2 \leq 0.8$ that cuts off $\hat{y}$. Second, we can replace constraint $g(y)\leq T$ with $g^{\mbox{\tiny U}}(y) \leq T$ in a DCBP to obtain a conservative approximation. [$\Box$ ]{}
Valid Inequalities in a Lifted Space {#subsec:bp}
------------------------------------
In Section \[subsec:submodular\], we derive extended polymatroid inequalities based on a relaxed approximation of SOC constraint . In this section, we show that the submodularity of holds for general $\Lambda$ in a lifted (i.e., higher-dimensional) space. Accordingly, we derive extended polymatroid inequalities in the lifted space. [We note that this approach was first proposed in [@bhardwaj2015binary] to test quadratic constrained problems (see Section 3.6.2 in [@bhardwaj2015binary]).]{}
To this end, we reformulate constraint as $\mu^{\top} y \leq T$ and $y^{\top} \Lambda y \leq (T - \mu^{\top} y)^2$, i.e., $y^{\top} (\Lambda - \mu\mu^{\top}) y + 2T\mu^{\top}y \leq T^2$. [Note that $\mu^{\top} y \leq T$ is because $T - \mu^{\top} y \geq \sqrt{y^{\top}\Lambda y} \geq 0$ by .]{} Then, we define $w_{jk} = y_j y_k$ for all $j, k \in [J]$ and augment vector $y$ to vector $v = [y_1, \ldots, y_J, w_{11}, \ldots, w_{1J}, w_{21}, \ldots, w_{JJ}]^{\top}$. We can incorporate the following McCormick inequalities to define each $w_{ij}$: $$w_{jk} \leq y_j, \quad w_{jk} \leq y_k, \quad w_{jk} \geq y_j + y_k - 1, \quad w_{ij} \geq 0. \label{mccormick}$$ Accordingly, we rewrite as $[\mu^{\top}, 0^{\top}] v \leq T$ and $$\label{eq:lifted:v}
a^{\top} v + v^{\top} B_{\mbox{\tiny N}} v \leq T^2,$$ where we decompose $(\Lambda - \mu\mu^{\top})$ to be the sum of two matrices, one containing all positive entries and the other containing all nonpositive entries. Accordingly, we define vector $a := [2T\mu; B_{\mbox{\tiny P}}]^{\top}$ with $B_{\mbox{\tiny P}} \in \mathbb{R}^{J^2}_+$ representing all the positive entries after vectorization, and matrix $B_{\mbox{\tiny N}} \in \mathbb{R}^{(J + J^2)\times (J + J^2)}_-$ collects all nonpositive entries, i.e., [$B_{\mbox{\tiny N}} = \begin{bmatrix} -(\Lambda - \mu\mu^{\top})_- & 0 \\ 0 & 0 \end{bmatrix}$]{}, where $(x)_- = - \min\{0, x\}$ for $x \in \mathbb{R}$.
As $a^{\top} v + v^{\top} B_{\mbox{\tiny N}} v$ is a submodular function of $v$ by Theorem \[thm:submodular\], we can incorporate extended polymatroid inequalities to strengthen the lifted SOC constraints . We summarize this result in the following [proposition]{}.
\[theorem:lifted\] [(See also Section 3.6.2 in [@bhardwaj2015binary])]{} Define function $h:\{0, 1\}^{J + J^2} \rightarrow \mathbb{R}$ such that $h(v) := a^{\top} v + v^{\top} B_{\mbox{\tiny N}} v$. Then, $h$ is submodular. Furthermore, inequality $\pi^{\top}v \leq T^2$ is valid for set $\{v \in \{0, 1\}^{J + J^2}: h(v) \leq T^2 \}$ for all $\pi \in \mbox{EP}_h$ and the separation of this inequality can be done by Algorithm \[algo:greedy\].
Note that this lifting procedure introduces $J^2$ additional variables $w_{ijk}$ for each $i$. However, $w_{ijk}$ can be treated as continuous variables when solving DCBP in view of the McCormick inequalities , and the number of $w_{ijk}$ variables can be reduced by half because $w_{ijk} = w_{ikj}$. In our numerical studies later, we derive more valid inequalities to strengthen the formulation in the lifted space for distributionally robust chance-constrained bin packing problems that involve DRCCs, using the bin packing structure.
\[ex:polymatroidCut\] Recall Example \[ex:general-cov\] and the corresponding function $g(y)$ being not submodular. We set $\mu = [0,0,0]^{\top}$ and $T = 0.8$, and rewrite the constraint $g(y) \leq T$ in the form of as $$0.6 v_4 + 0.2 v_6 + 0.7 v_8 + 0.1 v_9 + 0.2 v_{10} + 0.1 v_{11} + 0.6 v_{12} - 0.4 v_1 v_2 \leq 0.64,$$ where $v := [y_1, y_2, y_3, w_{11}, w_{12}, \ldots, w_{33}]^\top = [v_1,\ldots,v_{12}]^\top$. Now suppose that we are given $\hat{y} = [1, 0.5, 0.9]^{\top}$. The corresponding $\hat{v} = [1, 0.5, 0.9, 1, 0.5, 0.9, 0.5, 0.25, 0.45,$ $0.9, 0.45, 0.81]^\top$ violates the above inequality. We follow Algorithm \[algo:greedy\] to find an extended polymatroid inequality in the lifted space of $v$. Specifically, we sort the entries of $\hat{v}$ to obtain $\hat{v}_1 \geq \hat{v}_4 \geq \hat{v}_{3} \geq \hat{v}_6 \geq \hat{v}_{10} \geq \hat{v}_{12}\geq \hat{v}_2 \geq \hat{v}_5 \geq \hat{v}_7 \geq \hat{v}_9 \geq \hat{v}_{11} \geq \hat{v}_{8}$ and the corresponding permutation $\{1, 4, 3, 6, 10, 12, 2, 5, 7, 9, 11, 8\}$. It follows that $\hat{\pi}_{(2)} = 0.6$, $\hat{\pi}_{(4)} = 0.2$, $\hat{\pi}_{(5)} = 0.2$, $\hat{\pi}_{(6)} = 0.6$, $\hat{\pi}_{(7)} = -0.4$, $\hat{\pi}_{(10)} = 0.1$, $\hat{\pi}_{(11)} = 0.1$, $\hat{\pi}_{(12)} = 0.7$, and all other $\hat{\pi}_{(i)}$’s equal zero. This generates the following extended polymatroid inequality $$0.6 v_4 + 0.2 v_6 + 0.2 v_{10} + 0.6 v_{12} -0.4 v_{2} + 0.1 v_{9} +0.1 v_{11} + 0.7 v_{8} \leq 0.64$$ that cuts off $\hat{v}$.
[$\Box$ ]{}
Numerical Studies {#sec:result}
=================
We numerically evaluate the performance of our proposed models and solution approaches. In Section \[sec:ACCBP\], we present the formulation of a chance-constrained bin-packing problem with DRCCs and the related 0-1 SOC reformulations. We describe the solution methods and more valid inequalities based on the bin packing structure. In Section \[sec:comp\_setup\], we describe the experimental setup of the stochastic bin packing instances. Our results consist of two parts, which report the CPU time (Section \[sec:cpu\]) and the out-of-sample performance of solutions given by different models (Appendix ), respectively. More specifically, Section \[sec:cpu\] demonstrates the computational efficacy of the valid inequalities we derived in Section \[sec:algorithm\] for the original or lifted SOC constraints. Appendix demonstrates that DCBP solutions can well protect against the distributional ambiguity as opposed to the solutions obtained by following the Gaussian distribution assumption or by the SAA method.
Formulation of Ambiguous Chance-Constrained Bin Packing {#sec:ACCBP}
-------------------------------------------------------
For the classical bin packing problem, $[I]$ is the set of bins and $[J]$ is the set of items, where each bin $i$ has a weight capacity $T_i$ and each item $j$, if assigned to bin $i$, has a random weight $\tilde{t}_{ij}$. The deterministic bin packing problem aims to assign all $J$ items to a minimum number of bins, while respecting the capacity of each bin. If we consider a slightly more general setting by introducing a cost for each assignment, then the DCBP of bin packing with DRCCs is presented as:
\[bin-packing\] $$\begin{aligned}
\label{bp:0}\min_{\mathbf z, \mathbf y} \quad & & \sum_{i=1}^I c^{\text{z}}_i z_i + \sum_{i=1}^I \sum_{j=1}^J c^{\text{y}}_{ij} y_{ij} \\
\label{bp:1}\mbox{s.t.} \quad & & y_{ij} \leq \rho_{ij} z_i \quad \forall i \in [I], \ j \in [J] \\
\label{bp:2} & & \sum_{i=1}^I y_{ij} = 1 \quad \forall j \in [J] \\
\label{bp:4} & & y_{ij}, z_i \in \{0,1\} \quad \forall i \in [I], \ j \in [J], \\
\label{bp:3} & & \inf_{\mathbb{P} \in \mathcal{D}} \ \mathbb{P} \Biggl\{ \sum_{j=1}^J \tilde{t}_{ij} y_{ij} \leq T_i \Biggr\} \ \geq \ 1 - \alpha_i \quad \forall i \in [I], \end{aligned}$$
where $c^{\text{z}}_i$ represents the cost of opening bin $i$ and $c^{\text{y}}_{ij}$ represents the cost of assigning item $j$ to bin $i$. For each $i \in [I]$ and $j \in [J]$, we let binary variables $z_i$ represent if bin $i$ is open (i.e., $z_i = 1$ if open and $z_i = 0$ otherwise), binary variables $y_{ij}$ represent if item $j$ is assigned to bin $i$ (i.e., $y_{ij} = 1$ if assigned and $y_{ij} = 0$ otherwise), and parameters $\rho_{ij}$ represent if we can assign item $j$ to bin $i$ (i.e., $\rho_{ij} = 1$ if we can and $\rho_{ij} = 0$ otherwise). The objective function minimizes the total cost of opening bins and assigning items to bins. Constraints ensure that all items are assigned to open bins, constraints ensure that each item is assigned to one and only one bin, and constraints are the DRCCs. In our computational studies, we follow Section \[sec:model\] to derive 0-1 SOC reformulations of model and then follow Section \[sec:algorithm\] to derive valid inequalities in the original and lifted space for the 0-1 SOC reformulations. We strengthen the extended polymatroid inequalities, as well as derive valid inequalities in the lifted space containing variables $z_i$, $y_{ij}$, and $w_{ijk}$ to further strengthen the formulation as follows. We refer to Appendices and for the detailed proofs of the valid inequalities below, and will test their effectiveness later.
\[thm:poly-z\]
For all extended polymatroid inequalities $\pi^{\top} y_i \leq T$ with regard to bin $i$, $\forall i \in [I]$, inequality $$\pi^{\top} y_i \leq T z_i \label{submodular-1}$$ is valid for the DCBP formulation. Similarly, for all extended polymatroid inequalities $\pi^{\top} v_i \leq T^2$ with regard to bin $i$, $\forall i \in [I]$, inequality $$\pi^{\top} v_i \leq T^2 z_i \label{submodular-2}$$ is valid for the DCBP formulation.
\[thm:poly-lifted\] Consider set $$L = \Bigl\{(z, y, w) \in \{0, 1\}^{I \times (IJ) \times (IJ^2)}: \ \mbox{\eqref{bp:1}--\eqref{bp:4}}, \ w_{ijk} = y_{ij} y_{ik}, \ \forall j, k \in [J] \Bigr\}.$$ Without loss of optimality, the following inequalities are valid for $L$:
$$\begin{aligned}
w_{ijk} \ \geq & \ y_{ij} + y_{ik} + \sum_{\substack{\ell=1 \\ \ell \neq i}}^I w_{\ell jk} - 1 \quad \forall j, k \in [J] \label{lifted-1} \\
w_{ijk} \ \geq & \ y_{ij} + y_{ik} - z_i \quad \forall i \in [I], \ \forall j, k \in [J] \label{lifted-2} \\
\sum_{\substack{j=1\\ j \neq k}}^J w_{ijk} \ \leq & \ \sum_{j=1}^J y_{ij} - z_i \quad \forall i \in [I], \ \forall k \in [J] \label{lifted-3} \\
\sum_{j=1}^J \sum_{k = j+1}^J w_{ijk} \ \geq & \ \sum_{j=1}^J y_{ij} - z_i \quad \forall i \in [I]. \label{lifted-4}\end{aligned}$$
We note that valid inequalities – are *polynomially many* and all coefficients are in *closed-form*. Hence, we do not need any separation processes for these inequalities, and we can incorporate them in the DCBP formulation without dramatically increasing its size.
Computational Setup {#sec:comp_setup}
-------------------
We first consider $I= 6$ servers (i.e., bins) and $J = 32$ appointments (i.e., items) to test the DCBP model under various distributional assumptions and ambiguity sets. The daily operating time limit (i.e., capacity) $T_i$ of each server $i$ varies in between $[420,~540]$ minutes (i.e., 7–9 hours). We let the opening cost $c_i^{\mbox{\tiny z}}$ of each server $i$ be an increasing function of $T_i$ such that $c_i^{\textrm{z}} = T_i^2/3600+3T_i/60$, and let all assignment costs $c_{ij}^{\textrm{y}},~\forall i\in [I],~j\in [J]$ vary in between $[0,18]$, so that the total opening cost and the total assignment cost have similar magnitudes. The above problem size and parameter settings follow the literature of surgery block allocation (see, e.g., [@ORA_denton2010ORA_informs; @ORA_Shylo_CC; @yan2013ORP]).
To generate samples of random service time (i.e., random item weight), we consider “high mean (hM)" and “low mean ($\ell$M)" being $25$ minutes and $12.5$ minutes, respectively. We set the coefficient of variation (i.e., the ratio of the standard deviation to the mean) as $1.0$ for the “high variance (hV)" case and as $0.3$ for the “low variance ($\ell$V)" case. We equally mix all four types of appointments with “hMhV", “hM$\ell$V", “$\ell$MhV", “$\ell$M$\ell$V", and thus have eight appointments of each type. We sample 10,000 data points as the random service time of each appointment on each server, following a Gaussian distribution with the above settings of mean and standard deviation. We will hereafter call them the in-sample data. To formulate the 0-1 SOC models with diagonal covariance matrices, we use the empirical mean and standard deviation of each $\tilde{t}_{ij}$ obtained from the in-sample data and set $\alpha_i = 0.05, \ \forall i \in [I]$. Using the same $\alpha_i$-values, we formulate the 0-1 SOC models under general covariance matrices, for which we use the empirical mean and covariance matrix obtained from the in-sample data. The empirical covariance matrices we obtain have most of their off-diagonal entries being non-zero, and some being quite significant.
All the computation is performed on a Windows 7 machine with Intel(R) Core(TM) i7-2600 CPU 3.40 GHz and 8GB memory. We implement the optimization models and the branch-and-cut algorithm using commercial solver GUROBI 5.6.3 via Python 2.7.10. The GUROBI default settings are used for optimizing all the 0-1 programs, and we set the number of threads as one. When implementing the branch-and-cut algorithm, we add the violated extended polymatroid inequalities using GUROBI [callback]{} class by `Model.cbCut()` for both integer and fractional temporary solutions. For all the nodes in the branch-and-bound tree, we generate violated cuts at each node as long as any exists. The optimality gap tolerance is set as $0.01\%$. We also set the threshold for identifying violated cuts as $10^{-4}$, and set the time limit for computing each instance as 3600 seconds.
CPU Time Comparison {#sec:cpu}
-------------------
We solve 0-1 SOC reformulations or approximations of DCBP, and use either a diagonal or a general covariance matrix in each model. Our valid inequalities significantly reduce the solution time of directly solving the 0-1 SOC models in GUROBI, while the extended polymatroid inequalities generated based on the approximate and lifted SOC constraints perform differently depending on the problem size. The details are presented as follows.
### Computing 0-1 SOC models with diagonal matrices
We first optimize 0-1 SOC models with a diagonal matrix in constraint , of which the left-hand side function $g(y)$ is submodular, and thus we use extended polymatroid inequalities with $\pi \in EP_g$ in a branch-and-cut algorithm. Table \[tab1:diag\_cpu\] presents the CPU time (in seconds), optimal objective values, and other solution details (including “[**Server**]{}" as the number of open servers, “[**Node**]{}" as the total number of branching nodes, and “[**Cut**]{}" as the total number of cuts added) for solving the three 0-1 SOC models [**DCBP1**]{} (i.e., using ambiguity set $\mathcal D_1$), [**DCBP2**]{} (i.e., using ambiguity set $\mathcal D_2$ with $\gamma_1 = 1, \ \gamma_2 = 2$), and [**Gaussian**]{} (assuming Gaussian distributed service time). We also implement the SAA approach (i.e., row “[**SAA**]{}”) by optimizing the MILP reformulation of the chance-constrained bin packing model based on the 10,000 in-sample data points. We compare the branch-and-cut algorithm using our extended polymatroid inequalities (in rows “[**B&C**]{}”) with directly solving the 0-1 SOC models in GUROBI (in rows “[**w/o Cuts**]{}”).
\[tab1:diag\_cpu\]
In Table \[tab1:diag\_cpu\], the branch-and-cut algorithm quickly optimizes DCBP1 and DCBP2. Especially, if being directly solved by GUROBI, DCBP2 cannot be solved within the 3600-second time limit and ends with a $9.15\%$ optimality gap. Solving DCBP1 by using the branch-and-cut algorithm is much faster than solving the large-scale SAA-based MILP model, while the solution time of DCBP2 is similar to the latter. The two DCBP models also yield higher objective values, since they both provide more conservative solutions that open one more server than either the Gaussian or the SAA-based approach.
### Computing 0-1 SOC models with general covariance matrices
In this section, we focus on testing DCBP2 yielded by the ambiguity set $\mathcal D_2$ with parameters $\gamma_1=1$, $\gamma_2=2$, and $\alpha_i = 0.05, \ \forall i\in [I]$. We use empirical covariance matrices of the in-sample data. Note that these covariance matrices are general and non-diagonal. We compare the time of solving the 0-1 SOC reformulations of DCBP2 on ten independently generated instances. We examine two implementations of the branch-and-cut algorithm: one uses extended polymatroid inequalities with $\pi \in EP_{g^L}$ based on the relaxed 0-1 SOC constraint , and the other uses the extended polymatroid inequalities based on the lifted SOC constraint.
Table \[tab2:general\_cpu\] reports the CPU time (in seconds) and number of branching nodes (in column “**Node**") for various methods. First, we directly solve the 0-1 SOC models of DCPB2 in GUROBI, without or with the linear valid inequalities –, and report their results in columns “**w/o Cuts**” and “[**Ineq.**]{}”, respectively. We then implement the branch-and-cut algorithm, and examine the results of using extended polymatroid inequalities with $\pi \in EP_{g^L}$ (reported in columns “[**B&C-Relax**]{}”) and cuts based on lifted SOC constraints (reported in columns “[**B&C-Lifted**]{}”). [ (The latter does not involve valid inequalities used in the former.) The time reported under [**B&C-Relax**]{} also involves the time of solving SDPs for obtaining $\Delta^{\mbox{\tiny L}}$ and the relaxed 0-1 SOC constraint . The time of solving related SDPs are small (varying from 1 to 2 seconds for instances of different sizes) and negligible as compared to the total B&C time.]{} For both B&C methods, we also present the number of extended polymatroid inequalities (see column “**Cut**") added.
\[tab2:general\_cpu\]
In Table \[tab2:general\_cpu\], we highlight the solution time of the method that runs the fastest for each instance. Note that without the extended polymatroid inequalities or the valid inequalities, the GUROBI solver takes the longest time for solving all the instances except instance \#9. Adding the valid inequalities – to the solver reduces the solution time by 40% or more in almost all the instances, and drastically reduces the number of branching nodes. The extended polymatroid inequalities – further reduce the CPU time significantly (see columns “[**B&C-Relax**]{}” and “[**B&C-Lifted**]{}”). Moreover, for all the instances having 6 servers and 32 appointments, the algorithm using the extended polymatroid inequalities runs faster in eight out of ten instances than the algorithm using cuts based on relaxed SOC constraints without lifting. It indicates that the extended polymatroid inequalities generated by the lifted SOC constraints are more effective than those generated by the relaxed SOC constraints. This observation is overturned when we later increase the problem size.
In the following, we continue reporting the CPU time of solving DCBP2 with general covariance matrix. We vary the problem sizes (i.e., values of $I$ and $J$) in Section \[sec:time:size\], and vary the values of $\Lambda$ in the SOC constraint in Section \[sec:time:lambda\].
### Solving 0-1 SOC models under different problem sizes {#sec:time:size}
We use the same problem settings as in Section \[sec:comp\_setup\], and vary $I = 6, 8, 10$ and $J= 32, 40$ to test DCBP2 instances with different sizes. We still keep an equal mixture of all the four appointment types in each instance. Table \[tab3:size\] presents the computational time (in seconds), the total number of branching nodes (“**Node**"), and the total number of extended polymatroid inequalities generated (“**Cuts**"; if applicable) for solving the 0-1 SOC reformulation of DCBP2 by directly using GUROBI (“**w/o Cuts**”) and by using the two implementations of the extended polymatroid inequalities (“**B&C-Relax**” and “**B&C-Lifted**”).
\[tab3:size\]
In Table \[tab3:size\], we again highlight the solution time of the method that runs the fastest in each instance. We keep the first five instances we reported in Table \[tab2:general\_cpu\] for instances with $I=6, \ J =32$, and report five instances for other $(I, J)$ combinations. From Table \[tab3:size\], we observe that both implementations of the extended polymatroid inequalities run significantly faster than directly using GUROBI, especially when we increase the problem sizes (i.e., $I$ increased from 6 to 10, and $J$ increased from 32 to 40). In particular, the CPU time of directly using GUROBI is consistently 1 or 2 orders of magnitude larger than that of our approaches. For smaller $(I, J)$-values (e.g., $(I, J) = (6, 32)$ or $(I, J) =(8, 32)$), we see that [**B&C-Lifted**]{} sometimes runs faster than [**B&C-Relax**]{}, but for all the other $(I, J)$ combinations, the latter completely dominates the former. This is expected because the cuts are generated in a lifted space with $J^2$ additional variables for each server $i \in [I]$. Therefore, it makes sense that the scalability of [**B&C-Lifted**]{} is worse than that of [**B&C-Relax**]{}, which uses cuts without lifting.
### Solving 0-1 SOC models with different $\Lambda$-values in {#sec:time:lambda}
We again focus on instances with $I = 6$ and $J = 32$ under the same general covariance matrix $\Sigma$ obtained from the in-sample data points. We let $\Lambda := \Omega^2 \Sigma$ in the 0-1 SOC constraint and adjust the scalar $\Omega$ to obtain different $\Lambda$. We want to show how the computational time of directly using GUROBI increases as we increase $\Omega$, as compared to using the branch-and-cut algorithm with extended polymatroid inequalities. (The results of [**B&C-Lifted**]{} are used here and similar observations can be made if the results of [**B&C-Relax**]{} are used.)
Considering specific cases of the SOC constraint for modeling DCBP, we have $\Omega= \Phi^{-1}(1-\alpha)= 1.64$ for the Gaussian approximation model when $\alpha = 0.05$, and $\Omega=\sqrt{\gamma_2/\alpha} = 6.32$ for DCBP2 when $\gamma_2 = 2$ and $\alpha = 0.05$. We test four values of $\Omega$ equally distributed in between $[1.64,6.32]$ including the two end points.
Figure \[fig1:omega\] depicts the average CPU time and the average number of branching nodes of solving five independent DCBP2 instances for each $\Omega$-setting. Specifically, the four values $1.64$, $3.20$, $4.76$, and $6.32$ of $\Omega$ lead to $0.98$, $97.68$, $299.29$, and $363.56$ CPU seconds when directly using GUROBI, respectively, together with the significantly growing number of branching nodes $0$, $909.8$, $6662$, and $10418$, respectively. On the other hand, the branch-and-cut algorithm with the extended polymatroid inequalities respectively takes $1.03$, $35.19$, $16.09$, and $26.63$ CPU seconds on average for solving the same instances, and branches on average $0$, $188.6$, $247.2$, and $374.6$ nodes, respectively. This indicates that our approach is more scalable than directly using the off-the-shelf solvers.
Conclusions {#conclusion}
===========
In this paper, we considered distributionally robust individual chance constraints, where the true distributional information of the constraint coefficients is ambiguous and only the empirical first and second moments are given. The goal is to restrict the worst-case probability of violating a linear constraint under a given threshold. We provided 0-1 SOC representations of DRCCs under two types of ambiguity sets. In addition, we derived an efficient way of obtaining extended polymatroid inequalities for the 0-1 SOC constraints in both original and lifted spaces. Via extensive numerical studies, we demonstrated that our solution approaches significantly accelerate solving the DCBP model as compared to the state-of-the-art commercial solvers. In particular, a branch-and-cut algorithm with extended polymatroid inequalities in the original space scales very well as the problem size grows.
For future research, we plan to investigate DRCCs under other types of ambiguity sets, which could take into account not only the moment information but also density or structural information of their probability distribution. The connections between SOC program, SDP, and submodular optimization are also interesting to study.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported in part by the National Science Foundation under grants CMMI-1433066 and CMMI-1662774. The authors are grateful to Profs. Alper Atamt[ü]{}rk and Andr[é]{}s G[ó]{}mez for their comments on an earlier version of this paper and for bringing to our attention the works [@atamturk2018network; @atamturk2017polymatroid; @bhardwaj2015binary]. The authors are grateful to the two referees and the Associate Editor for their constructive comments and helpful suggestions.
Supplementary Materials {#supplementary-materials .unnumbered}
=======================
[**SM1. Proof of Proposition \[thm:poly-z\]**]{} \[sec:proofs\]\
[[*Proof:* ]{}]{}When $z_i = 1$, inequality reduces to the extended polymatroid inequality. When $z_i = 0$, we have $y_{ij} = 0$ for all $j \in [J]$ due to constraints . It follows that inequality holds valid.
When $z_i = 1$, inequality reduces to the extended polymatroid inequality. When $z_i = 0$, we have $y_{ij} = 0$ for all $j \in [J]$ due to constraints and so $w_{ijk} = 0$ for all $j, k \in [J]$. It follows that $v_i = 0$ by definition. Hence, inequality holds valid. [$\Box$ ]{}
[**SM2. Proof of Proposition \[thm:poly-lifted\]**]{} \[sec:proofs2\]\
[[*Proof:* ]{}]{}([**Validity of inequality** ]{}) If $j = k$, then $w_{ijk} = y_{ij}^2 = y_{ij}$. In this case, inequality reduces to $y_{ij} \geq 2y_{ij} + \sum_{\substack{\ell=1 \\ \ell \neq i}}^I y_{\ell j} - 1$, which clearly holds because $y_{ij} + \sum_{\substack{\ell=1 \\ \ell \neq i}}^I y_{\ell j} = \sum_{i=1}^I y_{ij} \leq 1$. If $j \neq k$, then we discuss the following two cases:
1. If $\max\{y_{ij}, y_{ik}\} = 1$, then we assume $y_{ij} = 1$ without loss of generality. It follows that $y_{\ell j} = 0$ due to constraints and so $w_{\ell jk} = y_{\ell j} y_{\ell k} = 0$ for all $\ell \neq i$. Hence, $\sum_{\substack{\ell=1 \\ \ell \neq i}}^I w_{\ell jk} = 0$ and inequality reduces to $w_{ijk} \geq y_{ij} + y_{ik} - 1$, which holds valid.
2. If $\max\{y_{ij}, y_{ik}\} = 0$, then $y_{ij} = y_{ik} = 0$ and $w_{ijk} = 0$. It remains to show $\sum_{\substack{\ell=1 \\ \ell \neq i}}^I w_{\ell jk} \leq 1$. Indeed, since $w_{\ell jk} \leq y_{\ell j}$, we have $\sum_{\substack{\ell=1 \\ \ell \neq i}}^I w_{\ell jk} \leq \sum_{\substack{\ell=1 \\ \ell \neq i}}^I y_{\ell j} \leq \sum_{\ell=1}^I y_{\ell j} = 1$, where the last equality is due to constraints .
([**Validity of inequality** ]{}) This inequality clearly holds valid when $z_i = 1$. When $z_i = 0$, we have $y_{ij} = y_{ik} = 0$ due to constraints . It follows that $w_{ijk} = y_{ij} y_{ik} = 0$ and so the inequality holds valid.
([**Validity of inequality** ]{}) This inequality holds valid when $z_i = 0$. When $z_i = 1$, this inequality is equivalent to $y_{ik}\sum_{\substack{j=1\\ j \neq k}}^J y_{ij} \leq \sum_{j=1}^J y_{ij} - 1$. We discuss the following two cases:
1. If $y_{ik} = 0$, then $\sum_{j=1}^J y_{ij} \geq 1$ without loss of optimality because $z_i = 1$. Inequality holds valid.
2. If $y_{ik} = 1$, then $y_{ik}\sum_{\substack{j=1\\ j \neq k}}^J y_{ij} = \sum_{\substack{j=1\\ j \neq k}}^J y_{ij}$. Meanwhile, $\sum_{j=1}^J y_{ij} - 1 = \sum_{\substack{j=1\\ j \neq k}}^J y_{ij} + y_{ik} - 1 = \sum_{\substack{j=1\\ j \neq k}}^J y_{ij}$. Inequality holds valid.
([**Validity of inequality** ]{}) This inequality holds valid when $z_i = 0$. When $z_i = 1$, we have $\sum_{j=1}^J y_{ij} \geq 1$ without loss of optimality. It follows that $\sum_{j=1}^J y_{ij} = 1$ or $\sum_{j=1}^J y_{ij} \geq 2$, and so $(\sum_{j=1}^J y_{ij} - 1)(\sum_{j=1}^J y_{ij} - 2) \geq 0$. Hence, $$\begin{aligned}
\Bigl(\sum_{j=1}^J y_{ij} - 1\Bigr)\Bigl(\sum_{j=1}^J y_{ij} - 2\Bigr) \ = & \ \sum_{j=1}^J y^2_{ij} + 2\sum_{j=1}^J \sum_{{\color{black}k = j+1}}^J y_{ij} y_{ik} - 3\Bigl(\sum_{j=1}^J y_{ij}\Bigr) + 2 \\
= & \ \sum_{j=1}^J y_{ij} + 2\sum_{j=1}^J \sum_{{\color{black}k = j+1}}^J w_{ijk} - 3\Bigl(\sum_{j=1}^J y_{ij}\Bigr) + 2 \\
= & \ 2\sum_{j=1}^J \sum_{{\color{black}k = j+1}}^J w_{ijk} - 2\biggl[\Bigl(\sum_{j=1}^J y_{ij}\Bigr) - 1\biggr] \ \geq \ 0.\end{aligned}$$ Inequality follows. [$\Box$ ]{}
[**SM3. Out-of-Sample Performance of DCBP**]{} \[sec:out-of-sample\]\
Through testing instances of chance-constrained bin packing, we show that DCBP solutions have very low probabilities of violating capacities in all the out-of-sample tests, even when the distributional information is misspecified. Specifically, we evaluate the out-of-sample performance of the optimal solutions to the DCBP1, DCBP2, Gaussian-based 0-1 SOC models, and the SAA-based MILP model. To generate the out-of-sample reference scenarios, we consider either misspecified distribution type or misspecified moment information as follows.
- [**Misspecified distribution type:**]{} We sample 10,000 out-of-sample data points from a two-point distribution having the same mean and standard deviation of each random variable $\tilde{t}_{ij}$ for $i\in [I]$ and $j\in [J]$ as the in-sample data. The service time is realized as $\mu_{ij} + \frac{(1-p)}{\sqrt{p(1-p)}} \sigma_{ij}$ with probability $p$ ($0 < p < 1$) and as $\mu_{ij} - \frac{\sqrt{p(1-p)}}{(1-p)} \sigma_{ij}$ with probability $1-p$, where $\mu_{ij}$ and $\sigma_{ij}$ are the sample mean and standard deviation of $\tilde{t}_{ij}$ obtained from the in-sample data. We set $p = 0.3$ so that we have smaller probability of having larger service time realizations.
- [**Misspecified moments:**]{} Alternatively, we sample 10,000 data points from the Gaussian distribution, but only consider the hM$\ell$V type of appointments, instead of an equal mixture of all the four types. In each sample and for each $i \in [I]$, we draw a standard-Gaussian random number $\rho_i$, and for each $j\in [J]$, generate a service time realization as $\mu_{ij} + \rho_i\sigma_{ij}$.
#### Performance of solutions under diagonal matrices
Under diagonal matrices, both of the two DCBP models open three servers (i.e., Servers 4, 5, 6 by DCBP1 and Servers 2, 4, 5 by DCBP2), while the Gaussian and SAA approaches only open Servers 4 and 6. We first use the $10,000$ out-of-sample data points given by misspecified distribution type, namely, the two-point distribution. Table \[tab4:discrete\] reports each solution’s probability of having the total time of assigned appointments not exceeding the capacity of the server to which they are assigned.
Model
---------- ------ ------ ------ ------
DCBP1 N/A 1.00 1.00 1.00
DCBP2 1.00 1.00 1.00 N/A
Gaussian N/A 0.69 N/A 0.91
SAA N/A 0.69 N/A 0.91
: Solution reliability in out-of-sample data following a misspecified distribution type
\[tab4:discrete\] \
“N/A”: the server is not opened by using the corresponding method.
Recall that $\alpha_i=0.05$ for all $i$ used in all four approaches. The reliability results of the Gaussian and SAA approaches are significantly lower than the desired probability threshold $1-\alpha_i=0.95$ on Server 4, and slightly lower than $0.95$ on Server 6. On the other hand, the optimal solutions of DCBP1 and DCBP2 do not exceed the capacity of any open servers.
Next, we use the $10,000$ out-of-sample data points given by misspecified moments. Table \[tab5:moment\_amb\] reports the reliability performance of each optimal solution.
Model
---------- ------ ------ ------ ------
DCBP1 N/A 0.94 1.00 1.00
DCBP2 0.98 1.00 0.99 N/A
Gaussian N/A 0.59 N/A 0.89
SAA N/A 0.59 N/A 0.89
: Solution reliability in out-of-sample scenarios with misspecified moments
\[tab5:moment\_amb\] \
“N/A”: the server is not opened by using the corresponding method.
The DCBP2 solution still outperforms solutions given by all the other approaches and achieves the desired reliability in all the three open servers. The Gaussian and SAA solutions perform poorly when the moment information is different from the empirical inputs. The DCBP1 solution respects the capacities of Servers 5 and 6 with sufficiently high probability (i.e., $> 0.95$), but yields a slightly lower reliability ($0.94$) than the threshold on Server 4.
#### Performance of solutions under general matrices
We optimize all the models under general matrices by using the empirical covariance matrices of the in-sample data, and report their corresponding solutions in Table \[tab:solution:general\]. Each entry illustrates the number of appointments assigned to an open server. Note that the Gaussian and SAA approaches yield the same solution of opening servers and assigning appointments.
Model
---------- ---- ----- ---- -----
DCBP1 12 N/A 13 7
DCBP2 12 11 9 N/A
Gaussian 15 N/A 17 N/A
SAA 15 N/A 17 N/A
: Optimal open servers and appointment-to-server assignments under general matrices
\[tab:solution:general\] \
“N/A”: the server is not opened by using the corresponding method.
We test the solutions shown in Table \[tab:solution:general\] in the out-of-sample scenarios under misspecified distribution type, and present their reliability performance in Table \[tab:out-of-sample:general\]. We again show that under general covariance matrices, the DCBP2 model yields the most conservative solution that does not exceed any open server’s capacity, while DCBP1 only ensures the desired reliability on Servers 3 and 6, but not on Server 5. The Gaussian and SAA approaches cannot produce solutions that can achieve the desired reliability threshold on any of their open servers.
Model
---------- ------ ------ ------ ------
DCBP1 1.00 N/A 0.91 1.00
DCBP2 1.00 1.00 1.00 N/A
Gaussian 0.91 N/A 0.91 N/A
SAA 0.91 N/A 0.91 N/A
: Solution reliability in out-of-sample scenarios with misspecified moments
\[tab:out-of-sample:general\] \
“N/A”: the server is not opened by using the corresponding method.
[10]{}
, [*Supermodular covering knapsack polytope*]{}, Discrete Optimization, 18 (2015), pp. 74–86.
, [*Network design with probabilistic capacities*]{}, Networks, 71 (2018), pp. 16–30.
, [*Polymatroid inequalities for p-order conic mixed 0-1 optimization*]{}. Available at arXiv e-prints <https://arxiv.org/abs/1705.05918>, 2017.
, [*Polymatroids and mean-risk minimization in discrete optimization*]{}, Operations Research Letters, 36 (2008), pp. 618–622.
, [*The submodular knapsack polytope*]{}, Discrete Optimization, 6 (2009), pp. 333–344.
, [*Models for minimax stochastic linear optimization problems with risk aversion*]{}, Mathematics of Operations Research, 35 (2010), pp. 580–602.
, [*Binary conic quadratic knapsacks*]{}, PhD thesis, University of California, Berkeley, 2015.
, [*On distributionally robust chance-constrained linear programs*]{}, Journal of Optimization Theory and Applications, 130 (2006), pp. 1–22.
, [*From [CVaR]{} to uncertainty set: [I]{}mplications in joint chance constrained optimization*]{}, Operations Research, 58 (2010), pp. 470–485.
, [*Distributionally robust stochastic knapsack problem*]{}, SIAM Journal on Optimization, 24 (2014), pp. 1485–1506.
, [*Distributionally robust optimization under moment uncertainty with application to data-driven problems*]{}, Operations Research, 58 (2010), pp. 595–612.
, [*Chance-constrained surgery planning under conditions of limited and ambiguous data*]{}. Available at SSRN: <http://dx.doi.org/10.2139/ssrn.2432375>, 2016.
, [*Optimal allocation of surgery blocks to operating rooms under uncertainty*]{}, Operations Research, 58 (2010), pp. 802–816.
, [*Submodular functions, matroids, and certain polyhedra*]{}, Combinatorial Structures and Their Applications, (1970), pp. 69–87.
, [*Worst-case [V]{}alue-at-[R]{}isk and robust portfolio optimization: A conic programming approach*]{}, Operations Research, 51 (2003), pp. 543–556.
, [*Data-driven chance constrained stochastic program*]{}, Mathematical Programming, Series A, 158 (2016), pp. 291–327.
, [*On mixing sets arising in chance-constrained programming*]{}, Mathematical Programming, 132 (2012), pp. 31–56.
, [*Decomposition algorithms for two-stage chance-constrained programs*]{}, Mathematical Programming, 157 (2016), pp. 219–243.
, [*A branch-and-cut decomposition algorithm for solving chance-constrained mathematical programs with finite support*]{}, Mathematical Programming, 146 (2014), pp. 219–244.
, [*A sample approximation approach for optimization with probabilistic constraints*]{}, SIAM Journal on Optimization, 19 (2008), pp. 674–699.
, [*An integer programming approach for linear programs with probabilistic constraints*]{}, Mathematical Programming, 122 (2010), pp. 247–272.
, [*Integer and Combinatorial Optimization*]{}, Wiley, 1999.
, [*An analysis of approximations for maximizing submodular set functions–[I]{}*]{}, Mathematical Programming, 14 (1978), pp. 265–294.
, [*Sample average approximation method for chance constrained programming: Theory and applications*]{}, Journal of Optimization Theory and Applications, 142 (2009), pp. 399–416.
, [*Robust mean-covariance solutions for stochastic optimization*]{}, Operations Research, 55 (2007), pp. 98–112.
, [*Probabilistic programming*]{}, in Stochastic Programming: Handbooks in Operations Research and Management Science, vol. 10, Elsevier, 2003, ch. 5, pp. 267–351.
, [*Robust growth-optimal portfolios*]{}, Management Science, 62 (2015), pp. 2090–2109.
, [*Stochastic modeling and approaches for managing energy footprints in cloud computing services*]{}, Service Science, 6 (2014), pp. 15–33.
, [*Stochastic operating room scheduling for high-volume specialties under block booking*]{}, INFORMS Journal on Computing, 25 (2012), pp. 682–692.
, [*The Logic of Logistics: [T]{}heory, Algorithms, and Applications for Logistics Management*]{}, Springer Verlag, 2014.
, [*The optimizer’s curse: Skepticism and postdecision surprise in decision analysis*]{}, Management Science, 52 (2006), pp. 311–322.
, [ *Chance-constrained binary packing problems*]{}, INFORMS Journal on Computing, 26 (2014), pp. 735–747.
, [*Stochastic 0–1 linear programming under limited distributional information*]{}, Operations Research Letters, 36 (2008), pp. 150–156.
, [*Distributionally robust convex optimization*]{}, Operations Research, 62 (2014), pp. 1358–1376.
, [*A general projection property for distribution families*]{}, in Advances in Neural Information Processing Systems, 2009, pp. 2232–2240.
, [*Distributionally robust joint chance constraints with second-order moment information*]{}, Mathematical Programming, 137 (2013), pp. 167–198.
[^1]: This paper has been accepted for publication at the SIAM Journal on Optimization. Copyright 2018 Society for Industrial and Applied Mathematics
[^2]: Proposition 2.2.5 in [@simchi2014logic] assumes that $y \in \mathbb{R}^n$ and provides a sufficient condition for $g$ being supermodular. It can be shown that a similar proof of this proposition applies to our case.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Although latent factor models (e.g., matrix factorization) achieve good accuracy in rating prediction, they suffer from several problems including cold-start, non-transparency, and suboptimal recommendation for local users or items. In this paper, we employ textual review information with ratings to tackle these limitations. Firstly, we apply a proposed aspect-aware topic model (ATM) on the review text to model user preferences and item features from different *aspects*, and estimate the *aspect importance* of a user towards an item. The aspect importance is then integrated into a novel aspect-aware latent factor model (ALFM), which learns user’s and item’s latent factors based on ratings. In particular, ALFM introduces a weighted matrix to associate those latent factors with the same set of aspects discovered by ATM, such that the latent factors could be used to estimate aspect ratings. Finally, the overall rating is computed via a linear combination of the aspect ratings, which are weighted by the corresponding aspect importance. To this end, our model could alleviate the data sparsity problem and gain good interpretability for recommendation. Besides, an aspect rating is weighted by an aspect importance, which is dependent on the targeted user’s preferences and targeted item’s features. Therefore, it is expected that the proposed method can model a user’s preferences on an item more accurately for each user-item pair locally. Comprehensive experimental studies have been conducted on 19 datasets from Amazon and Yelp 2017 Challenge dataset. Results show that our method achieves significant improvement compared with strong baseline methods, especially for users with only few ratings. Moreover, our model could interpret the recommendation results in depth.'
author:
- Zhiyong Cheng
- Ying Ding
- Lei Zhu
- Mohan Kankanhalli
bibliography:
- 'www\_long.bib'
title: |
Aspect-Aware Latent Factor Model:\
Rating Prediction with Ratings and Reviews
---
<ccs2012> <concept> <concept\_id>10002951.10003260.10003261.10003270</concept\_id> <concept\_desc>Information systems Social recommendation</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003260.10003261.10003271</concept\_id> <concept\_desc>Information systems Personalization</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003317.10003347.10003350</concept\_id> <concept\_desc>Information systems Recommender systems</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003260.10003261.10003269</concept\_id> <concept\_desc>Information systems Collaborative filtering</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10010147.10010257.10010258.10010260.10010268</concept\_id> <concept\_desc>Computing methodologies Topic modeling</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010147.10010257.10010293.10010309.10010311</concept\_id> <concept\_desc>Computing methodologies Factor analysis</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
Introduction
============
When making comments on an item (e.g., *product*, *movie*, and *restaurant*) in the online review/business websites, such as Yelp and Amazon, reviewers also provide an overall rating, which indicates their overall preference or satisfaction towards the corresponding items. Hence, predicting users’ overall ratings to unrated items or *personalized rating prediction* is an important research problem in recommender systems. Latent factor models (e.g., matrix factorization [@koren2009matrix; @zhang2016discrete; @cheng2017exploiting]) are the most widely used and successful techniques for rating prediction, as demonstrated by the Netflix Prize contest [@bell2007lessons]. These methods characterize user’s interests and item’s features using *latent factors* inferred from rating patterns in user-item rating records. As a typical collaborative filtering technique, the performance of MF suffers when the ratings of items or users are insufficient ( also known as the cold-start problem) [@he2015trirank]. Besides, a rating only indicates the overall satisfaction of a user towards an item, it cannot explain the underlying rationale. For example, a user could give a restaurant a high rating because of its delicious food or due to its nice ambience. Most existing MF models cannot provide such fine-grained analysis. Therefore, relying solely on ratings makes these methods hard to explicitly and accurately model users’ preferences [@wang2018rec; @he2015trirank; @ling2014ratings; @mcauley2013hidden; @wu2015flame].
Moreover, MF cannot achieve optimal rating prediction locally for each user-item pair, because it learns the latent factors of users ($\bm{p_u}$) and items ($\bm{q_i}$) via a global optimization strategy [@christakopoulou2016local]. In other words, $\bm{p_u}$ and $\bm{q_i}$ are optimized to achieve a global optimization over all the user-item ratings in the training dataset.[^1] As a result, the performance could be severely compromised locally for individual users or items. MF predicts an unknown rating by the dot product of the targeted user $u$’s and item $i$’s latent factors (e.g., $\bm{p_u}^T\bm{q_i}$). The overall rating of a user towards an item ($\hat{r}_{u,i}$) is decided by the importance/contribution of all factors. Take the $k$-th factor as an example, its contribution is $p_{u,k}*q_{i,k}$. For accurate prediction, it is important to accurately capture the importance of each latent factor for a user towards an item. It is well-known that different users may care about different *aspects* of an item. For example, in the domain of restaurants, some users care more about the taste of *food* while others pay more attention to the *ambience*. Even for the same aspect, the preference of users could be different from each other. For example, in the *food* aspect, some users like *Chinese cuisines* while some others favor *Italian cuisines*. Similarly, the characteristics of items on an aspect could also be different from each other. Thus, it is possible that “a user $u$ prefers item $i$ but dislikes item $j$ on a specific aspect", while “another user $u'$ favors item $j$ more than item $i$ on this aspect". Therefore, in MF, the importance of a latent factor for users towards an item should be treated differently. At first glance, MF achieves the goal as the influence of a factor (e.g., $k$-th factor) is dependent on both $p_{u,k}$ and $q_{i,k}$ (i.e., $p_{u,k}*q_{i,k}$). However, it is suboptimal to model the importance of a factor by a fixed value of an item or a user. In fact, MF treats each factor of an item with the same importance to all users (i.e., $q_{i,k}$); and similarly, each factor of a user is equally important to all items (i.e., $p_{u,k}$) in rating prediction. Take the previous example, “*a user $u$ prefers item $i$ but dislikes item $j$ on an aspect*", i.e., a factor (e..g, $k$) in MF), which means $p_{u,k}*q_{i,k}$ should be larger than $p_{u,k}*q_{j,k}$ (i.e., $p_{u,k}*q_{i,k}>p_{u,k}*q_{j,k}$). On the other hand, “*user $u'$ favors item $j$ more than item $i$ on this aspect*", thus $p_{u',k}*q_{j,k}$ should be larger than $p_{u',k}*q_{i,k}$ (i.e., $p_{u',k}*q_{i,k} <p_{u',k}*q_{j,k}$). Because the values of $p_{u,k}$ and $p_{u',k}$ are kept the same when predicting ratings, it is impossible for MF to satisfy the local requirements $p_{u,k}*q_{i,k}>p_{u,k}*q_{j,k}$ and $p_{u',k}*q_{i,k} <p_{u',k}*q_{j,k}$ simultaneously for these user-item pairs. A straightforward solution is to assign different weights (e.g., $w_{u,i,k}$) to different user-item pairs (e.g., $p_{u,k}*q_{i,k}$). However, how to compute a proper weight for each user-item pair is challenging.
A large amount of research effort has been devoted to deal with these weaknesses of MF methods. For example, various types of side information have been incorporated into MF to alleviate the cold-start problem, such as tags [@shi2013mining; @zhang2014attribute], social relations [@ma2011recommender; @wang2017item], reviews [@ling2014ratings; @mcauley2013hidden; @zhang2016integrating], and visual features [@he2016vbpr]. Among them, the accompanied review of a rating contains important complementary information. It not only encodes the information of user preferences and item features but also explains the underlying reasons for the rating. Therefore, in recent years, many models have been developed to exploit reviews with ratings to tackle the cold-start problem and also enhance the explainability of MF, such as HFT [@mcauley2013hidden], CTR [@wang2011collaborative], RMR [@ling2014ratings], and RBLT [@tan2016rating]. However, a limitation of these models is that they all assume an *one-to-one correspondence relationship* between latent topics (learned from reviews) and latent factors (learned from ratings), which not only limits their flexibility on modeling reviews and ratings but also may not be optimal. In addition, they cannot deal with the suboptimal recommendation for local users or items in MF. In fact, very few studies in literature have considered this problem.
In this paper, we focus on the problem of *personalized rating prediction* and attempt to tackle the above limitations together by utilizing reviews with ratings. Specifically, an *A*spect-aware *T*opic *M*odel (ATM) is proposed to extract *latent topics* from reviews, which are used to model users’ preferences and items’ features in different *aspects*. In particular, each *aspect* of users/items is represented as a probability distribution of latent topics. Based on the results, the relative importance of an aspect (i.e., *aspect importance*) for a user towards an item can be computed. Subsequently, the aspect importance is integrated into a developed *A*spect-aware *L*atent *F*actor *M*odel (ALFM) to estimate *aspect ratings*. In particular, a weight matrix is introduced in ALFM to associate the latent factors to the same set of aspects discovered by ATM. In this way, our model avoids referring to external sentiment analysis tools for aspect rating prediction as in [@zhang2014explicit; @diao2014jointly]. The overall rating is obtained by a linear combination of the *aspect ratings*, which are weighted by the importance of corresponding aspects (i.e., *aspect importance*). Note that the latent topics and latent factors in our model are not linked directly; instead, they are correlated via the *aspects* indirectly. Therefore, the number of latent topics and latent factors could be different and separately optimized to model reviews and ratings respectively, which is fundamentally different from the *one-to-one* mapping in previous models [@mcauley2013hidden; @wang2011collaborative; @ling2014ratings; @tan2016rating; @bao2014topicmf; @zhang2016integrating]. Besides, our model could learn an aspect importance for each user-item pair, namely, assigning a different weight to each $p_{u,k}*q_{i,k}$, and thus could alleviate the suboptimal local recommendation problem and achieve better performance.
A set of experimental studies has been conducted on 19 real-world datasets from Yelp and Amazon to validate the effectiveness of our proposed model. Experimental results show that our model significantly outperforms the state-of-the-art methods which also use both reviews and ratings for rating prediction. Besides, our model also obtains better results for users with few ratings, demonstrating the advantages of our model on alleviating the cold-start problem. Furthermore, we illustrate the interpretability of our model on recommendation results with examples. In summary, the main contributions of this work include:
- We propose a novel aspect-aware latent factor model, which could effectively combine reviews and ratings for rating prediction. Particularly, our model relaxes the constraint of one-to-one mappings between the latent topics and latent factors in previous models and thus could achieve better performance.
- Our model could automatically extract explainable aspects, and learn the aspect importance/weights for different user-item pairs. By associating latent factors with aspects, the aspect weights are integrated with latent factors for rating prediction. Thus, the proposed model could alleviate the suboptimal problem of MF for individual user-item pairs.
- We conduct comprehensive experimental studies to evaluate the effectiveness of our model. Results show that our model is significantly better than previous approaches on tasks of rating prediction, recommendation for sparse data, and recommendation interpretability.
Related Work {#sec:relwork}
============
A comprehensive review on the recommender system is beyond the scope of this work. We mainly discuss the works which utilize both reviews and ratings for rating prediction. Some works assume that the review is available when predicting the rating score, such as SUIT [@li2014suit], LARAM [@wang2011latent], and recent DeepCoNN [@zheng2017joint]. However, in real world recommendation settings, the task should be predicting ratings for the uncommented and unrated items. Therefore, the review is unavailable when predicting ratings. We broadly classify the approaches for the targeted task in three categories: (1) sentiment-based, (2) topic-based, and (3) deep learning-based. Our approach falls into the second category.
**Sentiment-based.** These works analyze user’s sentiments on items in reviews to boost the rating prediction performance, such as [@pappas2013sentiment; @pero2013opinion; @diao2014jointly; @zhang2014explicit]. For example, [@pappas2013sentiment] estimated a sentiment score for each review to build a user-item sentiment matrix, then a traditional collaborative filtering method was applied. Zhang et al. [@zhang2014explicit] analyzed the sentiment polarities of reviews and then jointly factorize the user–item rating matrix. These methods rely on the performance of external NLP tools for sentiment analysis and thus are not self-contained.
**Topic-based.** These approaches extract latent topics or aspects from reviews. An early work [@ganu2009beyond] in this direction relied on domain knowledge to manually label reviews into different aspects, which requires expensive domain knowledge and high labor cost. Later on, most works attempt to extract latent topics or aspects from reviews automatically [@mcauley2013hidden; @bao2014topicmf; @diao2014jointly; @he2015trirank; @ling2014ratings; @mcauley2013hidden; @wu2015flame; @zhang2016integrating; @tan2016rating]. A general approach of these methods is to extract latent topics from reviews using topic models [@wang2011collaborative; @mcauley2013hidden; @ling2014ratings; @zhang2016integrating; @tan2016rating] or non-negative MF [@bao2014topicmf; @qiu2016aspect] and learn latent factors from ratings using MF methods. HFT [@mcauley2013hidden] and TopicMF [@bao2014topicmf] link the latent topics and latent factors by using a defined transform function. ITLFM [@zhang2016integrating] and RBLT [@tan2016rating] assume that the latent topics and latent factors are in the same space, and linearly combine them to form the latent representations for users and items to model the ratings in MF. CTR [@wang2011collaborative] assumes that the latent factors of items depend on the latent topic distributions of their text, and adds a latent variable to offset the topic distributions of items when modeling the ratings. RMR [@ling2014ratings] also learns item’s features using topic models on reviews, while it models ratings using a mixture of Gaussian rather than MF methods. Diao et al. [@diao2014jointly] propose an integrated graphical model called JMARS to jointly model aspects, ratings and sentiments for movie rating prediction. Those models all assume an one-to-one mapping between the learned latent topics from reviews and latent factors from ratings. Although we adopt the same strategy to extract latent topics and learn latent factors, our model does not have the constraint of one-to-one mapping. Besides, Zhang et al. [@zhang2014explicit] extracted aspects by decomposing the user–item rating matrix into item–aspect and user–aspect matrices. He et al. [@he2015trirank] extracted latent topics from reviews by modeling the user-item-aspect relation with a tripartite graph. **Deep learning-based**. Recently, there has been a trend of applying deep learning techniques in recommendation [@he2017neural; @covington2016deep]. For example, He et al. generalized matrix factorization and factorization machines to neural collaborative filtering and achieved promising performance [@he2017neural; @he2017fm]. Textual reviews have also been used in deep learning models for recommendation [@zhang2016collaborative; @zheng2017joint; @catherine2017transnets; @zhang2017joint]. The most related works in this direction are DeepCoNN [@zheng2017joint] and TransNet [@catherine2017transnets], which apply deep techniques to reviews for rating prediction. In DeepCoNN, reviews are first processed by two CNNs to learn user’s and item’s representations, which are then concatenated and passed into a regression layer for rating prediction. A limitation of DeepCoNN is that it uses reviews in the testing phase. [@catherine2017transnets] shows that the performance of DeepCoNN decreases greatly when reviews are unavailable in the testing phase. To deal with the problem, TransNet [@catherine2017transnets] extends DeepCoNN by introducing an additional layer to simulate the review corresponding to the target user-item pair. The generated review is then used for rating prediction.
[ll]{} Notation & Definition\
$\mathcal{D}$ & corpus with reviews and ratings\
$d_{u,i}$ & review document of user $u$ to item $i$\
$s$ & a sentence in a review $d_{u,i}$\
$\mathcal{U}$, $\mathcal{I}$, $\mathcal{A}$ & user set, item set, and aspect set, respectively\
$M,N,A$ & number of users, items, and aspects, respectively\
$N_{w,s}$ & number of words in a sentence $s$\
$K$ & number of latent topics in ATM\
$y$ & an indicator variable in ATM\
$a_s$ & assigned aspect $a$ to sentence $s$\
$\pi_u$ & the parameter of Bernoulli distribution $P(y=0)$\
$\eta$ & Beta priors ($\eta=\{\eta_0, \eta_1\}$)\
$\bm{\alpha_u}, \bm{\alpha_i}$ & Dirichlet priors for aspect-topic distributions\
$\bm{\gamma_u}, \bm{\gamma_i}$ & Dirichlet priors for aspect distributions\
$\bm{\beta_w} $ & Dirichlet priors for topic-word distributions\
$\bm{\theta_{u,a}}$ & user’s aspect-topic distribution: denoting user’s preference on $a$\
$\bm{\psi_{i,a}}$ & item’s aspect-topic distribution: denoting item’s features on $a$\
$\bm{\lambda_u}, \bm{\lambda_v}$ & aspect distributions of user and item, respectively\
$\bm{\phi_w}$ & topic-text word distribution\
$f$ & number of latent factors in ALFM\
$\mu_\cdot$ & regularization coefficients\
$b_\cdot$ & bias terms, e.g., $b_u, b_i, b_0$\
$w_a$ & weight vector for aspect $a$\
$p_u$, $q_i$ & latent factors of user $u$ and item $i$, respectively\
$r_{u,i}$ & rating of user $u$ to item $i$\
$r_{u,i, a}$ & aspect rating of user $u$ towards item $i$ on aspect $a$\
$\rho_{u,i,a}$ & aspect importance of $a$ for $u$ with respect to $i$\
$s_{u,i,a}$ & the degree of item $i$’s attributes matching user $u$’s preference\
& on aspect $a$\
\[tab:notation\]
Proposed Model {#sec:ourmodel}
==============
Problem Setting
---------------
Let $\mathcal{D}$ be a collection of reviews of item set $\mathcal{I}$ from a specific category (e.g., restaurant) written by a set of users $\mathcal{U}$, and each review comes with an overall rating $r_{u,i}$ to indicate the overall satisfaction of user $u$ to item $i$. The primary goal is to predict the unknown ratings of items that the users have not reviewed yet. A review $d_{u,i}$ is a piece of text which describes opinions of user $u$ on different aspects $a \in \mathcal{A}$ towards item $i$, such as *food* for *restaurants*. In this paper, we only consider the case that all the items are from the same category, i.e., they share the same set of aspects $\mathcal{A}$. Aspects that users care for items are latent and learned from reviews by our proposed topic model, in which each aspect is represented as a distribution of the same set (e.g., $K$) of latent topics. Table \[tab:notation\] lists the key notations. Before introducing our method, we would like to first clarify the concepts of *aspects*, *latent topics*, and *latent factors*.
- **Aspect** - it is a high-level semantic concept, which represents the attribute of items that users commented on in reviews, such as *“food”* for *restaurant* and *“battery"* for *mobile phones*.
- **Latent topic & latent factor** - in our context, both concepts represent a more fine-grained concept than *“aspect"*. A latent topic or factor can be regarded as a *sub-aspect* of an item. For instance, for the “food" aspect, a related latent topic could be “*breakfast*" or “*Italian cuisine*". We adopt the terminology of *latent topic* in topic models and *latent factor* in matrix factorization. Accordingly, “latent topics" are discovered by topic model on reviews, and “latent factors" are learned by matrix factorization on ratings.
Aspect-aware Latent Factor Model
--------------------------------
Based on the observations that (1) different users may care for different aspects of an item and (2) users’ preferences may differ from each other for the same aspect, we claim that the overall satisfaction of a user $u$ towards an item $i$ (i.e., the overall rating $r_{u,i}$) depends on $u$’s satisfaction on each aspect $a$ of $i$ (i.e., *aspect rating* $r_{u,i,a}$) and the importance of each aspect (of $i$) to $u$ (i.e., *aspect importance* $\rho_{u,i,a}$). Based on the assumptions, the overall rating $r_{u,i}$ can be predicted as: $$\vspace{-2pt}
\hat{r}_{u,i} = \sum_{a\in\mathcal{A}} \overbrace{\rho_{u,i,a}}^{{ \mbox{\scriptsize\begin{tabular}{@{}c@{}}aspect importance\end{tabular}}}} \underbrace{r_{u,i,a}}_{{ \mbox{\scriptsize\begin{tabular}{@{}c@{}}aspect rating\end{tabular}}}}
\vspace{-2pt}$$
### Aspect rating estimation.
Aspect rating (i.e., $r_{u,i,a}$) reflects the satisfaction of a user $u$ towards an item $i$ on the aspect $a$. To receive a high aspect rating $r_{u,i,a}$, an item should at least possess the characteristics/attributes in which the user is interested in this aspect. Moreover, the item should satisfy user’s expectations on these attributes in this aspect. In other words, the item’s attributes on this aspect should be of high quality such that the user likes it. Take the “food" aspect as an example, for a user who likes Chinese cuisines, to receive a high rating on the *“food"* aspect from the user, a restaurant should provide Chinese dishes and the dishes should suit the user’s taste. Based on user’s text reviews, we can learn users’ preferences and items’ characteristics on each aspect and measure *how the attributes of an item $i$ on aspect “$a$" suit a user $u$’s requirements on this aspect*, denoted by $s_{u,i,a}$. We compute $s_{u,i,a}$ based on results of the proposed Aspect-aware Topic Model (ATM) (described in Sect. \[sec:matm\]), in which user’s preferences and item’s characteristics on each aspect are modeled as multinomial distributions of latent topics, denoted by $\bm{\theta_{u,a}}$ and $\bm{\psi_{i,a}}$, respectively. $s_{u,i,a} \in [0,1]$ is then computed as : $$\label{eq:jsd}
\vspace{-2pt}
s_{u,i,a} = 1-JSD(\bm{\theta_{u,a}}, \bm{\psi_{i,a}})
\vspace{-1pt}$$ where $JSD(\bm{\theta_{u,a}}, \bm{\psi_{i,a}})$ denotes the Jensen–Shannon divergence [@endres2003new] between $\bm{\theta_{u,a}}$ and $\bm{\psi_{i,a}}$. Notice that a large value of $s_{u,i,a}$ does not mean a high rating $r_{u,i,a}$ - an item providing all the features that a user $u$ requires does not mean that it satisfies $u$’s expectations, since the provided ones could be of low quality. For instance, a restaurant provides all the Chinese dishes the user $u$ likes (i.e., high score $s_{u,i,a}$), but these dishes taste bad from $u$’s opinion (i.e., low rating $r_{u,i,a}$). Therefore, we can expect that for this restaurant: users discuss its Chinese dishes in reviews with negative opinions and thus give low ratings. Instead of analyzing the review sentiments for aspect rating estimation by using external NLP tools (such as [@zhang2014explicit]), we refer to the matrix factorization (MF) [@koren2009matrix] technique.
MF maps users and items into a latent factor space and represents users’ preferences and items’ features by $f$-dim latent factor vectors (i.e., $\bm{p_u} \in \mathbb{R}^{f \times 1}$ and $\bm{q_i} \in \mathbb{R}^{f \times 1}$). The dot product of the user’s and item’s vectors ($\bm{p_u}^T\bm{q_i}$) characterizes the user’s overall interests on the item’s characteristics, and is thus used to predict the rating $r_{u,i}$. To extend MF for aspect rating prediction, we introduce a binary matrix $\bm{W} \in \mathbb{R}^{f \times A}$ to associate the latent factors to different aspects, where $A$ is the number of aspects considered. We call this model aspect-aware latent factor model (ALFM), in which the weight vector $\bm{w_a}$ in the $a$-th column of $\bm{W}$ indicates which factors are related to the aspect $a$. Thus, $\bm{p_{u,a}} = \bm{w_a} \odot \bm{p_u}$ denotes user’s interests in the aspect $a$, where $\odot$ represents element-wise product between vectors. Therefore, $(\bm{p_{u,a}})^T(\bm{q_{i,a}})$ represents the aspect rating of user $u$ to item $i$ on aspect $a$. Finally, we integrate the matching results of aspects (i.e., $s_{u,i,a}$) into ALFM to estimate the aspect ratings: $$\small
\vspace{-2pt}
r_{u,i,a} = s_{u,i,a} \cdot (\bm{w_a} \odot \bm{p_u})^T(\bm{w_a} \odot \bm{q_i})
\vspace{-1pt}$$ As a high aspect rating $r_{u,i,a}$ requires large values of both $s_{u,i,a}$ and $(\bm{w_a} \odot \bm{p_u})^T(\bm{w_a} \odot \bm{q_i})$, it is expected that the results learned from reviews could guide the learning of latent factors.
### Aspect importance estimation.
We rely on user reviews to estimate $\rho_{u,i,a}$, as users often discuss their interest topics of aspects in reviews, such as different *cuisines* in the *food* aspect. In general, the more a user comments on an aspect in reviews, the more important this aspect is (to this user). Thus, we estimate the importance of an aspect according to the possibility of a user writing review comments on this aspect. When writing a review, some users tend to write comments from the aspects according to their own preferences, while others like commenting on the most notable features of the targeted item. Based on this consideration, we introduce (1) $\pi_u$ to denote the probability of user $u$ commenting an item based on his own preference and (2) $\lambda_{u,a}$ ($\sum_{a\in\mathcal{A}}\lambda_{u,a}=1$) to denote the probability of user $u$ commenting on the aspect $a$ based on his own preference. Accordingly, $(1-\pi_{u})$ denotes the probability of the user commenting from the item $i$’s characteristics ($\sum_{a\in\mathcal{A}}\lambda_{i,a}=1$), and $\lambda_{i,a}$ is the probability of user $u$ commenting item $i$ from the item’s characteristics on the aspect $a$. Thus, the probability of a user $u$ commenting an item $i$ on an aspect $a$ (i.e., $\rho_{u,i,a}$) is: $$\label{eq:rho}
\small
\vspace{-2pt}
\rho_{u,i,a} = \pi_{u}\lambda_{u,a} + (1-\pi_{u})\lambda_{i,a}
\vspace{-1pt}$$ $\lambda_{u,a}$, $\lambda_{i,a}$, and $\pi_u$ are estimated by ATM, which simulates the process of a user writing a review, as detailed in the next subsection.
Aspect-aware Topic Model {#sec:matm}
------------------------
Given a corpus $\mathcal{D}$, which contains reviews of users towards items $\{d_{u,i}|d_{u,i} \in \mathcal{D}, u \in \mathcal{U}, i \in \mathcal{I}\}$, we assume that a set of latent topics (i.e., $K$ topics) covers all the topics that users discuss in the reviews. $\bm{\lambda_u}$ is a probability distribution of aspects in user $u$’s preferences, in which each value $\lambda_{u,a}$ denotes the relative importance of an aspect $a$ to the user $u$. Similarly, $\bm{\lambda_i}$ is the probability distribution of aspects in item $i$’s characteristics, in which each value $\lambda_{i,a}$ denotes the importance of an aspect $a$ to the item $i$. As the $K$ latent topics cover all the topics discussed in reviews, an aspect will only relate to some of the latent topics closely. For example, topic “*breakfast*" is closely related to aspect “food", while it is not related to aspects like “*service*" or “*price*". The relation between aspects and topics is also represented by a probabilistic distribution, i.e., $\bm{\theta_{u,a}}$ for users and $\bm{\psi_{i,a}}$ for items. More detailedly, the interests of a user $u$ in a specific aspect $a$ is represented by $\bm{\theta_{u,a}}$, which is a multinomial distribution of the latent topics; the characteristics of an item $i$ in a specific aspect $a$ is represented by $\bm{\psi_{i,a}}$, which is also a multinomial distribution of the same set of latent topics. $\bm{\theta_{u,a}}$ is determined based on all the reviews $\{d_{u,i} | i \in \mathcal{I}\}$ of user $u$ writing for items. $\bm{\psi_{i,a}}$ is learned from all the reviews $\{d_{u,i} | u \in \mathcal{U}\}$ of $i$ written by users. A latent topic is a multinomial distribution of text words in reviews. Based on these assumptions, we propose an aspect-aware topic model ATM to estimate the parameters $\{\bm{\lambda_i}$, $\bm{\lambda_i}$, $\bm{\theta_{u,a}}$, $\bm{\psi_{i,a}}$, $\pi_u\}$ by simulating the generation of the corpus $\mathcal{D}$.
The graphical representation of ATM is shown in Fig. \[fig:matm\]. In the figure, the shaded circles indicate observed variables, while the unshaded ones represent the latent variables. ATM mimics the processing of writing a review sentence by sentence. A sentence usually discusses the same topic $z$, which could be from user’s preferences or from item’s characteristics. To decide the topic $z_s$ for a sentence $s$, our model introduces an indicator variable $y \in \{0,1\}$ based on a Bernoulli distribution, which is parameterized by $\pi_u$. Specifically, when $y=0$, the sentence is generated from user’s preference; otherwise, it is generated according to item $i$’s characteristics. $\pi_u$ is user-dependent, indicating the tendency to comment from $u$’s personal preferences or from the item $i$’s characteristics is determined by $u$’s personality. The generation process of ATM is shown in Algorithm \[alg:geneproc\]. Let $a_s$ denote the aspect assigned to a sentence $s$. If $y=0$, $a_s$ is drawn from $\lambda_u$ and $z_s$ is then generated from $u$’s preferences on aspect $a_s$: $\bm{\theta_{u,a_s}}$; otherwise, if $y=1$, $a_s$ is drawn from $\lambda_i$ and $z_s$ is then generated from $i$’s characteristics on aspect $a_s$: $\bm{\psi_{i,a_s}}$. Then all the words $w$ in sentence $s$ is generated from $z_s$ according to the word distribution: $\bm{\phi_{z_s, w}}$.
\[alg:geneproc\]
In ATM, $\bm{\alpha_u}$, $\bm{\alpha_i}$, $\bm{\gamma_u}$, $\bm{\gamma_i}$, $\bm{\beta}$, and $\eta$ are pre-defined hyper-parameters and set to be symmetric for simplicity. Parameters need to be estimated including $\bm{\lambda_i}$, $\bm{\lambda_i}$, $\bm{\theta_{u,a}}$, $\bm{\psi_{i,a}}$, and $\pi_u$. Different approximate inference methods have been developed for parameter estimation in topic models, such as variation inference [@blei2003latent] and collapsed Gibbs sampling [@griffiths2004]. We apply collapsed Gibbs sampling to infer the parameters, since it has been successfully applied in many large scale applications of topic models [@cheng2016effective; @cheng2017sigir]. Due to the space limitation, we omit the detailed inference steps in this paper.
Model Inference
---------------
With the results of ATM, $\rho_{u,i,a}$ and $s_{u,i,a}$ can be computed using Eq. \[eq:rho\] and \[eq:jsd\], respectively. With the consideration of bias terms (i.e., $b_u, b_i, b_0$) in ALFM, the overall rating can be estimated as[^2], $$\label{eq:re2}
\hat{r}_{u,i} = \sum_{a\in\mathcal{A}}(\rho_{u,i,a}\cdot s_{u,i,a} \cdot (\bm{w_a} \odot \bm{p_u})^T(\bm{w_a} \odot \bm{q_i}) ) + b_u + b_i + b_0$$ where $b_0$ is the average rating, $b_u$ and $b_i$ are user and item biases, respectively. The estimation of parameters is to minimize the rating prediction error in the training dataset. The optimization objective function is $$\label{eq:ojf}
\begin{split}
\vspace{-2pt}
\underset{p*,q*}{\text{min}} \frac{1}{2}\sum_{u,i} (r_{u,i} & -\hat{r}_{u,i})^2 + \frac{\mu_u}{2} ||\bm{p_u}||_2^2 + \frac{\mu_i}{2} ||\bm{q_i}||_2^2 \\
& + \mu_w \sum_a||\bm{w_a}||_1 + \frac{\mu_b}{2} (||b_u||_2^2 + ||b_i||_2^2);
\vspace{-2pt}
\end{split}$$ where $||\cdot||_2$ denotes the $\ell_2$ norm for preventing model overfitting, and $||\cdot||_1$ denotes the $\ell_1$ norm. $\mu_u, \mu_i, \mu_w$, and $\mu_b$ are regularization parameters, which are tunable hyper-parameters. In practice, we relax the binary requirement of $\bm{w_a}$ by using $\ell_l$ norm. It is well known that $\ell_l$ regularization yields sparse solution of the weights [@mairal2010online]. The $\ell_2$ regularization of $\bm{p_u}$ and $\bm{q_i}$ prevents them to have arbitrarily large values, which would lead to arbitrarily small values of $\bm{w_a}$.
**Optimization.** We use the stochastic gradient descent (SGD) algorithm to learn the parameters by optimizing the objective function in Eq. \[eq:ojf\]. In each step of SGD, the localized optimization is performed on a rating $r_{u,i}$. Let $L$ denote the loss, and the gradients of parameters are given as follows: $$\begin{aligned}
\label{eq:user}
\small
\vspace{-2pt}
\frac{\partial L}{\partial p_u}&=\sum_{i=1}^{N}(\sum_{a}\rho_{u,i,a}s_{u,i,a}w_a^2)(\hat{r}_{u,i}-r_{u,i})q_i + \mu_u p_u \\
\frac{\partial L}{\partial q_i}&=\sum_{u=1}^{M}(\sum_{a}\rho_{u,i,a}s_{u,i,a}w_a^2)(\hat{r}_{u,i}-r_{u,i})p_u + \mu_i q_i \\
\frac{\partial L}{\partial w_a}&=\sum_{u=1}^{M}\sum_{i=1}^{N}\rho_{u,i,a}s_{u,i,a}(\hat{r}_{u,i}-r_{u,i})p_uq_iw_a + \frac{\mu_w w_a} {\sqrt{(w_a^2+\epsilon)}}
\vspace{-2pt}\end{aligned}$$ Here, we omit the gradients of $b_u$ and $b_i$, as they are the same as in the standard biased MF [@koren2009matrix]. $M$ and $N$ are the total number of users and items in the dataset. Notice that in the deriving of the gradient for $w_a$, we use $\sqrt{w_a^2+\epsilon}$ in place of $||w_a||_1$, because $\ell_1$ norm is not differentiable at 0. $\epsilon$ can be regarded as a “smoothing parameter" and is set to $10^{-6}$ in our implementation.
Experimental Study {#sec:expconfig}
==================
To validate the assumptions when designing the model and evaluate our proposed model, we conducted comprehensive experimental studies to answer the following questions:
- **RQ1:** How do the important parameters (e.g., the number of latent topics and latent factors) affect the performance of our model? More importantly, is the setting $f=K$ optimal, which is a default assumption for many previous models? (Sect. \[sec:modelanalysis\])
- **RQ2:** Can our ALFM model outperform the state-of-the-art recommendation methods, which consider both ratings and reviews, on rating prediction? (Sect. \[sec:comp\])
- **RQ3:** Compared to other methods which also use textual reviews and ratings, how does our ALFM model perform on the cold-start setting when users have only few ratings? (Sect. \[sec:coldstart\])
- **RQ4:** Can our model explicitly interpret the reasons for a high or low rating? (Sect. \[sec:interpret\])
Datasets \#users \#items \#ratings Sparsity
--------------------- --------- --------- ----------- ----------
Instant Video 4,902 1,683 36,486 0.9956
Automotive 2,788 1,835 20,218 0.9960
Baby 17,177 7,047 158,311 0.9987
Beauty 19,766 12,100 196,325 0.9992
Cell Phones 24,650 10,420 189,255 0.9993
Clothing 34,447 23,026 277,324 0.9997
Digital Music 5,426 3,568 64,475 0.9967
Grocery 13,979 8,711 149,434 0.9988
Health 34,850 18,533 342,262 0.9995
Home & Kitchen 58,901 28,231 544,239 0.9997
Musical Instruments 1,397 900 10,216 0.9919
Office Products 4,798 2,419 52,673 0.9955
Patio 1,672 962 13,077 0.9919
Pet Supplies 18,070 8,508 155,692 0.9990
Sports & Outdoors 31,176 18,355 293,306 0.9995
Tools & Home 15,438 10,214 133,414 0.9992
Toys & Games 17,692 11,924 166,180 0.9992
Video Games 22,348 10,672 228,164 0.9990
Yelp 2017 169,257 63,300 1,659,678 0.9998
: Statistics of the evaluation datasets
\[tab:dataset\]
Dataset Description {#sec:dataset}
-------------------
We conducted experiments on two publicly accessible datasets that provide user review and rating information. The first dataset is Amazon Product Review dataset collected by [@mcauley2013hidden][^3], which contains product reviews and metadata from Amazon. This dataset has been widely used for rating prediction with reviews and ratings in previous studies [@mcauley2013hidden; @ling2014ratings; @tan2016rating; @catherine2017transnets]. The dataset is organized into 24 product categories. In this paper, we used 18 categories (See Table \[tab:dataset\]) and focus on the 5-core version, with at least 5 reviews for each user or item. The other dataset is from Yelp Dataset Challenge 2017[^4], which includes reviews of local business in 12 metropolitan areas across 4 countries. For the Yelp 2017 dataset, we also processed it to keep users and items with at least 5 reviews. From each review in these datasets, we extract the corresponding “userID", “itemID", a rating score (from 1 to 5 rating stars), and a textual review for experiments. Notice that for all the datasets, we checked and removed the duplicates, and then filtered again to keep them as 5-core. Besides, we removed the infrequent terms in the reviews for each dataset.[^5] Some statistics of the datasets are shown in Table \[tab:dataset\].
Experimental Settings
---------------------
For each dataset, we randomly split it into training, validation, and testing set with ratio 80:10:10 for each user as in [@mcauley2013hidden; @ling2014ratings; @catherine2017transnets]. Because we take the 5-core dataset where each user has at least 5 interactions, we have at least 3 interactions per user for training, and at least 1 interaction per user for validation and testing. Note that we only used the review information in the training set, because the reviews in the validation or testing set are unavailable during the prediction process in real scenarios. The number of aspect is set to 5 in experiments.[^6]
**Baselines:** We compare the proposed **ALFM** model with the following baselines. It is worth noting that these methods are tuned on the validation dataset to obtain their optimal hyper-parameter settings for fair comparisons.
- **BMF [@koren2009matrix].** It is a standard MF method with the consideration of bias terms (i.e., user biases and item biases). This method only leverages ratings when modeling users’ and items’ latent factors. It is typically a strong baseline model in collaborative filtering [@koren2009matrix; @ling2014ratings].
- **HFT [@mcauley2013hidden].** It models ratings with MF and review text with latent topic model (e.g., LDA [@blei2003latent]). We use it as a representative of the methods which use an exponential transformation function to link the latent topics with latent factors, such as TopicMF [@bao2014topicmf]. The topic distribution can be modeled on either users or items. We use the topic distribution based on items, since it achieves better results. Note that in experiments, we add bias terms into HFT, which can achieve better performance.
- **CTR [@wang2011collaborative].** This method also utilizes both review and rating information. It uses a topic model to learn the topic distribution of items, which is then used as the latent factors of items in MF with an addition of a latent variable.
- **RMR [@ling2014ratings].** This method also uses both ratings and reviews. Different from HFT and CTR, which use MF to model rating, it uses a mixture of Gaussian to model the ratings.
- **RBLT [@tan2016rating].** This method is the most recent method, which also uses MF to model ratings and LDA to model review texts. Instead of using an exponential transformation function to link the latent topics and latent factors (as in HFT [@mcauley2013hidden]), this method linearly combines the latent factors and latent topics to represent users and items, with the assumption that the dimensions of topics and latent factors are equal and in the same latent space. The same strategy is also adopted by ITLFM [@zhang2016integrating]. Here, we use RBLT as a representative method for this strategy.
- **TransNet [@catherine2017transnets].** This method adopts neural network frameworks for rating prediction. In this model, the reviews of users and items are used as input to learn the latent representations of users and items. More descriptions about this method could be found in Section \[sec:relwork\]. We used the codes published by the authors in our experiments and tuned the parameters as described in [@catherine2017transnets].
The standard root-mean-square error (**RMSE**) is adopted in evaluation. A smaller RMSE value indicates better performance.
[|c|C[0.6cm]{}C[0.6cm]{}C[0.6cm]{}C[0.6cm]{}C[0.6cm]{}C[0.9cm]{}C[0.75cm]{}|C[0.87cm]{}C[0.89cm]{}C[0.87cm]{}C[0.89cm]{}C[0.95cm]{}C[0.85cm]{}|]{} & BMF & HFT & CTR & RMR & RBLT & TransNet & ALFM &\
& (a) & (b) & (c) & (d) & (e) & (f) & (g) & g vs. a & g vs. b & g vs. c & g vs. d & g vs. e & g vs. f\
Instant Video & 1.162 & 0.999 & 1.014 & 1.039 & 0.978 & 0.996 & **0.967** & 16.79 & 3.19 & 4.63\* & 6.94\* & 1.12\*\* & 2.88\
Automotive & 1.032 & 0.968 & 1.016 & 0.997 & 0.924 & 0.918 & **0.885** & 14.26\* & 8.58\*\* & 12.86\* & 11.19\* & 4.24\*\* & 3.56\*\
Baby & 1.359 & 1.112 & 1.144 & 1.178 & 1.122 & 1.110 & **1.076** & 20.83\*\* & 3.24 & 5.98\* & 8.66\*\* & 4.11 & 3.05\*\
Beauty & 1.342 & 1.132 & 1.171 & 1.190 & 1.117 & 1.123 & **1.082** & 19.39\*\* & 4.47 & 7.65\*\* & 9.12\*\* & 3.18\*\* & 3.65\*\*\
Phones & 1.432 & 1.216 & 1.271 & 1.289 & 1.220 & 1.207 & **1.167** & 18.47\*\* & 3.98\* & 8.18\* & 9.4\*\* & 4.33 & 3.27\*\*\
Clothing & 1.073 & 1.103 & 1.142 & 1.145 & 1.073 & 1.064 & **1.032** & 3.8\*\* & 6.47\*\* & 9.65 & 9.9\*\* & 3.86\*\* & 2.96\*\
Digital Music & 1.093 & **0.918** & 0.921 & 0.960 & **0.918** & 1.061 & 0.920 & 15.82 & -0.15 & 0.13\* & 4.49\*\* & -0.15\*\* & 4.13\*\*\
Grocery & 1.192 & 1.016 & 1.045 & 1.061 & 1.012 & 1.022 & **0.982** & 17.66\*\* & 3.36\*\* & 6.07 & 7.46\*\* & 3.01\*\* & 3.94\*\
Health & 1.263 & 1.073 & 1.105 & 1.135 & 1.070 & 1.114 & **1.042** & 17.48\* & 2.83 & 5.65\* & 8.20 & 2.56\*\* & 6.46\*\*\
Home & Kitchen & 1.297 & 1.083 & 1.123 & 1.149 & 1.086 & 1.123 & **1.049** & 19.16\*\* & 3.15\*\* & 6.62 & 8.7\*\* & 3.41\*\* & 6.61\*\*\
Musical Instruments & 1.004 & 0.972 & 0.979 & 0.983 & 0.946 & 0.901 & **0.893** & 11.08 & 8.17\*\* & 8.83\*\* & 9.2\*\* & 5.61 & 0.95\
Office Products & 1.025 & 0.879 & 0.898 & 0.934 & 0.872 & 0.898 & **0.848** & 17.29\*\* & 3.55\*\* & 5.61\* & 9.26\*\* & 2.77\*\* & 5.67\*\*\
Patio & 1.180 & 1.041 & 1.062 & 1.077 & 1.032 & 1.046 & **1.001** & 15.19\*\* & 3.84\* & 5.7\* & 7.07\* & 2.96 & 4.33\*\*\
Pet Supplies & 1.367 & 1.137 & 1.177 & 1.200 & 1.139 & 1.149 & **1.099** & 19.64\*\* & 3.41\* & 6.67\* & 8.41 & 3.54\*\* & 4.38\*\*\
Sports & Outdoors & 1.130 & 0.970 & 0.998 & 1.019 & 0.964 & 0.990 & **0.933** & 17.42\*\* & 3.8\* & 6.47 & 8.4\* & 3.2\*\* & 5.77\*\*\
Tools & Home & 1.168 & 1.013 & 1.047 & 1.090 & 1.011 & 1.041 & **0.974** & 16.63\*\* & 3.90 & 6.98 & 10.68\*\* & 3.7\*\* & 6.51\*\*\
Toys & Games & 1.072 & 0.926 & 0.948 & 0.974 & 0.923 & 0.951 & **0.902** & 15.81\*\* & 2.59\* & 4.82\*\* & 7.39\*\* & 2.3\*\* & 5.11\*\
Video Games & 1.321 & 1.096 & 1.115 & 1.150 & 1.094 & 1.123 & **1.070** & 19.02\* & 2.43 & 4.03\*\* & 6.97\* & 2.24\*\* & 4.77\*\
Yelp 2017 & 1.415 & 1.174 & 1.233 & 1.266 & 1.202 & 1.190 & **1.155** & 18.35\* & 1.60\*\* & 6.33\*\* & 8.74\* & 3.88\*\* & 2.92\*\
Average & 1.207 & 1.044 & 1.074 & 1.097 & 1.037 & 1.049 & **1.004** & 14.56\*\* & 2.84\* & 7.16\*\* & 8.31\* & 3.37\*\* & 4.26\*\*\
The improvements with \* are significant with $p-value < 0.05$, and the improvements with \*\* are significant with $p-value < 0.01$ with a two-tailed paired t-test.
\[tab:comp\]
Effect of Important Parameters (RQ1) {#sec:modelanalysis}
------------------------------------
In this subsection, we analyze the influence of *the number of latent factors* and *the number of latent topics* on the final performance of ALFM. As we know, in MF, more latent factors will lead to better performance unless overfitting occurs [@he2016fast; @koren2009matrix]; while the optimal number of latent topics in topic models (e.g., LDA) is dependent on the datasets [@blei2012probabilistic; @arun2010finding]. Accordingly, the optimal number of latent topics in topic model and the optimal number of latent factors in MF should be tuned separately. However, in the previous latent factor models (e.g., HFT, TopicMF [@bao2014topicmf], RMR, CTR, and RBLT), the number of factors (i.e., \#factors) and the number of topics (i.e., \#topics) are assumed to be the same, and thus cannot be optimized separately. Since our model does not have such constraint, we studied the effects of \#factors and \#topics individually. Fig. \[fig:effects\] show the performance variations with the change of \#factors and \#topics by setting the other one to 5. We only visualize the performance variations of three datasets, due to the space limitation and the similar performance variation behaviors of other datasets. From the figure, we can see that with the increase of \#factors, RMSE consistently decreases although the degree of decline is small. Notice that in our model, the rating prediction still relies on MF technique (Eq. \[eq:re2\]). Therefore, the increase of \#factors could lead to better representation capability and thus more accurate prediction. In contrast, the optimal number of latent topics is different from dataset to dataset.
To better visualize the impact of \#factors and \#topics, we also present 3D figures by varying the number of factors and topics in $\{5, 10, 15, 20, 25\}$, as shown in Fig. \[fig:factorvstopic\]. In this figure, we use the performance of three datasets as illustration. From the figure, we can see that the optimal numbers of topics and latent factors are varied across different datasets. In general, more latent factors usually lead to better performance, while the optimal number of latent topics is dependent on the reviews of different datasets. This also reveals that setting \#factors and \#topics to be the same may not be optimal.
Model Comparison (RQ2) {#sec:comp}
----------------------
We show the performance comparisons of our ALFM with all the baseline methods in Table \[tab:comp\], where the best prediction result on each dataset is in bold. For fair comparison, we set the number of latent factors ($f$) and the number of latent topics ($K$) to be the same as $f=K=5$. Notice that our model could obtain better performance when setting $f$ and $K$ differently. Still, ALFM achieves the best results on 18 out of the 19 datasets. Compared with BMF, which only uses ratings, we achieve much better prediction performance (16.49% relative improvement on average). More importantly, our model outperforms CTR and RMR with large margins - 6.28% and 8.18% relative improvements on average, respectively. Compared to the recently proposed RBLT and TransNet, ALFM can still achieve 3.37% and 4.26% relative improvement on average respectively with significance testing. It is worth mentioning that HFT achieves better performance than RMR and comparable performance with recent RBLT, because we added bias terms to the original HFT in [@mcauley2013hidden]. TransNet applies neural networks, which has exhibited strong capabilities on representation learning, in reviews to learn users’ preferences and items’ characteristics for rating prediction. However, it may suffer from (1) noisy information in reviews, which would deteriorate the performance; and (2) errors introduced when generating fake reviews for rating prediction, which will also cause bias in the final performance. Compared to those baselines, the advantage of ALFM is that it models users’ preferences on different aspects; and more importantly, it captures a user’s specific attention on each aspect of a targeted item. The substantial improvement of ALFM over those baselines demonstrates the benefits of modeling users’ specific preferences on each aspect of different items.
Cold-Start Setting (RQ3) {#sec:coldstart}
------------------------
As shown in Table \[tab:dataset\], the datasets are usually very sparse in practical systems. It is inherently difficult to provide satisfactory recommendation based on limited ratings. In the matrix factorization model, given a few ratings, the penalty function tends to push $q_u$ and $p_i$ towards zero. As a result, such users and items are modeled only with the bias terms [@ling2014ratings]. Therefore, matrix factorization easily suffers from the cold-start problem. By integrating reviews in users’ and items’ latent factor learning, our model could alleviate the problem of cold-start to a great extent, since reviews contain rich information about user preferences and item features.
To demonstrate the capability of our model in dealing with users with very limited ratings, we randomly split the datasets into training, validation, and testing sets in ratio 80:10:10 based on the number of ratings in each set. In this setting, it is not guaranteed that a user has at least 3 ratings in the training set. It is possible that a user has no rating in the training set. For the users without any ratings in the training set, we also removed them in the testing set. Then we evaluate the performance of users who have the number of ratings from 1 to 10 in the training set. In Fig. \[fig:coldstart\], we show the **Gain in RMSE** ($y$-axis) grouped by the number of ratings ($x$-axis) of users in the training set. The value of **Gain in RMSE** is equal to the average RMSE of baselines *minus* that of our model (e.g., “BMF-TALFM"). A positive value indicates that our model achieves better prediction. As we can see, our ALFM model substantially improves the prediction accuracy compared with the BMF model. More importantly, our model also outperforms all the other baselines which also utilize reviews. This demonstrates that our model is more effective in exploiting reviews and ratings, because it learns user’s preferences and item’s features in different aspects and is capable of estimating the aspect weights based on the targeted user’s preferences and targeted item’s features.
**Value** **Comfort** **Accessories** **Shoes** **Clothing**
------------- ------------- ----------------- ----------- --------------
price size ring socks shirt
color fit pretty foot back
quality wear dress boots bra
worth comfortable time comfort top
cute bra beautiful sandals feel
comfortable small gift walk soft
fits color earrings toe black
ring fits compliments pairs jeans
dress perfect chain hold pants
shirt material jewelry strap tight
material long shoes pockets material
: Top ten words of each aspect for a user (index 1511) from *Clothing*. Each column is corresponding to an aspect attached with an “interpretation” label.
\[tab:aspects\]
Aspects Value Comfort Accessories Shoes Clothing
---------------- ------- --------- ------------- ------- ----------
Importance (1) 0.621 0.042 0.241 0.001 0.095
Matching (1) 0.982 0.596 0.660 0.759 0.638
Polarity (1) **+** **-** **+** **-** **+**
Importance (2) 0.621 0.042 0.241 0.001 0.094
Matching (2) 0.920 0.303 0.362 1.000 0.638
Polarity (2) **-** **-** **-** **-** **+**
: Interpretation for why the “user 1511" rated “item 1" and “item 2" with 5 and 2, respectively, from *Clothing*.
\[tab:explaination\]
Interpretability (RQ4) {#sec:interpret}
----------------------
In our ALFM model, a user’s preference on an item is decomposed into user’s preference on different aspects and the importance of those aspects. An aspect is represented as a distribution of latent topics discovered based on reviews. A user’s attitude/sentiment on an aspect of the targeted item is decided by the latent factors (learned from ratings) associating with the aspect. Based on the topic distribution of an aspect ($\bm{\theta_{u,a_s}}$) and the word distribution of topics ($\bm{\phi_{w}}$), we can semantically represent an aspect by the top words in this aspect. Specifically, the probability of a word $w$ in an aspect $a_s$ of a user $u$ can be computed as $\sum_{k=1}^K\theta_{u,a_s,k}\phi_{k,w}$. The top 10 aspect words (\#aspect $= 5$) of “user 1511" from *Clothing* dataset discovered by our model are shown in Table \[tab:aspects\]. Notice that in order to obtain a better visualization of each aspect, we removed the “background” words that belong to more than 3 aspects. As shown in Table \[tab:aspects\], the five aspects can be semantically interpreted to “value"[^7], “comfort", “accessories", “shoes", and “clothing". Next, we illustrate the interpretability of our ALFM model on high or low ratings by examples from the same dataset. Table \[tab:explaination\] shows the aspect importance (i.e., $\rho_{u,i,a}$ in Eq. \[eq:rho\]) of the “user 1511" , the aspect matching scores (i.e., $s_{u,i,a}$ in Eq. \[eq:jsd\]) as well as sentiment polarity (obtained by Eq. \[eq:re2\]) on the five aspects with respect to “item 1" and “item 2" in *Clothing* dataset. From the results, we can see that “user 1511" pays more attention to “Value" and “Accessories" aspects. On the “Value" aspect, both “item 1" and “item 2" highly match her preference, however, she has a positive sentiment on “item 1" while a negative sentiment on “item 2".[^8] For the “Accessories" aspect, “item 1" has a higher matching score than “item 2"; and more importantly, the sentiment is positive on “item 1" while negative on “item 2". As a result, “user 1511" rated “item 1" with 5 while rated “item 2" with 2. From the examples, we can see that our model could provide explanations for the recommendations in depth with *aspect semantics*, *aspect matching score*, as well as *aspect ratings* (which shows sentiment polarity).
Conclusions {#sec:concl}
===========
In this paper, we proposed an aspect-aware latent factor model for rating prediction by effectively combining reviews and ratings. Our model correlates the latent topics learned from review text and the latent factors learned from ratings based on the same set of aspects, which are discovered from textual reviews. Accordingly, our model does not have the constraint of one-to-one mapping between latent factors and latent topics as previous models (e.g., HFT, RMR, RBLT, etc.), and thus could achieve better user preference and item feature modeling. Besides, our model is able to estimate aspect ratings and assign weights to different aspects. The aspect weight is dependent on each user-item pair, since it is estimated based on user’s personal preferences on the corresponding aspect towards an item. Experimental results on 19 real-world datasets show that our model greatly improves the rating prediction accuracy compared to the state-of-the-art methods, especially for users who have few ratings. With the extracted aspects from textual reviews, estimated aspect weights, and aspect ratings, our model could provide interpretation for recommendation results in great detail.
This research is supported by the National Research Foundation, Prime Minister’s Office, Singapore under its International Research Centre in Singapore Funding Initiative. The authors would like to thank Rose Catherine Kanjirathinkal (from CMU)’s great help on fine-tuning the results of TransNet on all datasets.
[^1]: In the paper, unless otherwise specified, notations in bold style denote matrices or vectors, and the ones in normal style denote scalars.
[^2]: In our experiments, we tried to normalize $\rho_{u,i,a}$ or $\rho_{u,i,a} \cdot s_{u,i,a}$ in Eq. \[eq:re2\], but no improvement has been observed.
[^3]: http://jmcauley.ucsd.edu/data/amazon/
[^4]: http://www.yelp.com/dataset\_challenge/
[^5]: The thresholds of infrequent terms varied across different datasets. For example, for the “Yelp 2017" dataset, which is relatively large, a term that appears less than 10 times in reviews is defined as an infrequent term; and the thresholds are smaller for relatively small datasets (e.g., the threshold is 5 for the “Music Instruments" dataset).)
[^6]: We tuned the number of aspects from 1 to 8 for all the datasets, and found that the performance does not change much unless setting the aspect number to 1 or 2.
[^7]: “Value" means value for money
[^8]: As a reminder, the aspect matching is based on the reviews. It is possible that both item 1 and item 2 contains comments on aspect “value". However, “item 1" has a high value while “item 2" has a low value.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An effective model is used to study the equation of state(EOS) of warm strange hadronic matter with nucleons, $\Lambda$-hyperons, $\Xi$-hyperons, $\sigma^{*}$ and $\phi$. In the calculation, a newest weak Y-Y interaction deduced from the recent observation[@Takahashi] of a $_{\Lambda\Lambda}^6$He double hypernucleus is adopted. Employing this effective model, the results with strong Y-Y interaction and weak Y-Y interaction are compared.'
author:
- 'W.L.Qian$^1$[^1], R.K. Su$^{2,1,4}$ [^2] and H.Q.Song$^{2,3,4}$[^3]'
title: 'Warm strange hadronic matter in an effective model with a weak Y-Y interaction'
---
ø
20.6pt
****
====
Normal nuclei are made of protons and neutrons. If hyperon carrying strangeness is added into a nucleus somehow, we then obtain hypernuclei. The first hypernucleus was seen in emulsion by Danysz and Pniewski[@Danysz1] in 1953. Since then, strangeness carried by s-quark opens a new dimension for the studies in nuclear physics. In recent years, exploring nuclear system with strangeness, i.e., strange matter has received increasing interest. Such system has many astrophysical and cosmological implications and is indeed interesting by itself. For instance, the core of neutron stars may contain a high fraction of hyperons[@Schaffner; @Yamamoto; @Sahu], resulting in a third family of compact stars which has a similar mass to neutron star but has a much smaller radius than the later[@Schaffner]. The strange matter may also be formed in relativistic heavy ion collisions. There are two kinds of strange matter: strange quark matter and strange hadronic matter. On one hand, it has been speculated[@Witten; @Bodmer; @Chin; @Jaffe] that lumps of quark matter (“strangelets”) with large strangeness per baryon might be more stable than the normal nuclei. The experimental work searching for strange quark matter has been going on in BNL-AGS and CERN-SPS [@Armstrong; @Appelquist; @Arsenescu; @Hill]. But unfortunately, the evidence for the production of strangelets has not yet been observed within the experimental limits. On the other hand, strange hadronic matter or hypernuclei have also been investigated [@Ikeda; @Barranco; @Schaffner1; @Schaffner2; @Schaffner3; @Schulze; @Vidana; @Zhang1; @Wang1; @Zhang2; @Wang2; @Wang3]. But the inclusion of multiple units of strangeness in nuclei remains rather largely unexplored yet. This is because the technical difficulty experimentally and the uncertainty of the interactions between baryons theoretically. Recently, Takahashi et al.[@Takahashi] reported their observation of a $_{\Lambda\Lambda}^6$He double hypernucleus, where the $\Lambda-\Lambda$ interaction energy $\Delta
B_{\Lambda\Lambda}=1.01\pm0.20^{+0.18}_{-0.11}$MeV is deduced from measured data. This value is much smaller than the previous estimation $\Delta B_{\Lambda\Lambda}\simeq 4-5$MeV from the early experiments[@Danysz2; @Prowse; @Aoki; @Song]. To incorporate the early strong Y-Y interaction, the $\Lambda$ well depth in “$\Lambda$ matter” at density $0.5\rho_0$ was estimated as $V_\Lambda^{(\Lambda)}\simeq 20$ MeV by using the Nijmegen model D.[@Schaffner3]. If the new value $\Delta
B_{\Lambda\Lambda}=1.01$MeV is used, the $V_\Lambda^{(\Lambda)}\simeq 5$ MeV is obtained. The measurement of $_{\Lambda\Lambda}^4H$ was also performed in BNL, but they did not obtain the interaction between the lambdas.
Although QCD is the fundamental theory for strong interaction, it is not available to describe strange hadronic matter directly because of its non-perturbative properties. Two kinds of effective models had been introduced. The first kind is focused on the chiral SU(3) symmetry[@Papazoglou1; @Papazoglou2]. For example, in ref. [@Papazoglou1], a generalized Lagrangian which is based on the linear realization of chiral SU(3) symmetry and the concept of broken scale invariance was proposed to describe the SHM. The second kind is based upon the successful models of nuclear matter, for example, Walecka model[@Schaffner2; @Schaffner3; @Schaffner4], quark-meson coupling(QMC) model[@Saito], or Furnstahl-Serot-Tang(FST) model[@FST], and adding hyperons into these models. In refs.[@Schaffner2] and [@Schaffner3], two models (denoted as model 1 and model 2) which are based on a SU(3) extension of the Walecka model are suggested to deal with the weak and strong Y-Y interaction. In particular, in the model 2, in order to incorporate the early strong Y-Y interaction data[@Danysz2], two additional mesons, namely, $\sigma^*$ and $\phi$ mesons are introduced to obtain the strong attraction between $\Lambda$ hyperons. The advantage of this treatment is that it can easily reduce to an effective model which can explain the nuclear systems very well when the hyperons are taken away, because the coupling constants between nucleons and mesons are determined by the experimental data of nuclear system. In this sense, the effect of hyperons can be exposed explicitly. Of course, part of the second kind of models suffer from the loss of chiral SU(3) symmetry. In fact, the treatment for adding hyperons to nuclear system to study the hypernuclei, for example, $\Lambda$ hypernuclei has been employed by many previous papers.
Recently, by using an extended modified quark meson coupling(MQMC) model we studied the implications[@Song] of the newest Y-Y interaction in strange hadronic matter. It is found that while the system with the strong $\Lambda-\Lambda$ interaction and in a quite large strangeness fraction region is more deeply bound than the ordinary nuclear matter due to the opening of the new degrees of freedom, the system with the weak $\Lambda-\Lambda$ interaction is rather loosely bound compared to the later. It is interesting to check if the above remarks depend on the model used and to see what happens if the system is at finite temperature. In a previous paper[@Zhang2] , we suggested an effective model, constructed by introducing hyperons in the Furnstahl-Serot-Tang(FST) model, to study the saturation properties and stabilities of strange hadronic matter at both zero[@Zhang2] and finite temperature[@Qian]. In this work, we will use the extended FST model to study warm strange hadronic matter with the newest weak Y-Y interactions and then compare the results to those[@Qian] with previous strong Y-Y interactions. The details of this model can be found in ref.[@Zhang2], here we only give a short description. Considering reactions, $\Lambda +\Lambda \rightarrow \Xi ^{-}+p$, $\Lambda +\Lambda \rightarrow \Xi ^0+n$ and their reverses, we take the mixture of the cascades $\Xi ^{-}$ and $\Xi ^0$ in the strange matter into our model, besides lambdas. For simplicity, we assume that $\Xi ^{-}$ and $\Xi ^0$ will appear in the strange matter with equal amount. This is similar to the protons and neutrons in symmetric nuclear matter. We used, therefore, a single symbol $\Xi $ for these particles. Furthermore, we have also include the $\sigma ^{*}$ and $\phi $ mesons in the model to describe the interaction between hyperons, as proposed by Schaffner et al.[@Schaffner3]. We will not consider the mixture of the $\Sigma$ hyperons in the same reason as mentioned in Ref.[@Zhang2]. The paper is organized as follows. In section II, we will gives a brief description of the model. The calculated results and some discussions will be presented in section III.
The extended FST model
======================
In our previous paper [@Zhang2; @Qian], the original FST model[@FST] was extended by including $\Lambda$ and $\Xi$ hyperons in the system and introducing a new hyperon-hyperon interaction mediated by two additional strange mesons $\sigma^*$ and $\phi$ which couple only to hyperons, just as we stated in the introduction section. Since we will study the unpolarized system, the $\pi$ meson has no influence on the system. Omitting the contributions from the $\pi$ meson, we have the following Lagrangian density for the FST model. $$\begin{aligned}
{\cal L}(x) &=&\bar{\psi}_N(i\gamma ^\mu \partial _\mu -g_{\omega
N}\gamma ^\mu V_\mu -M_N+g_{sN}\sigma )\psi _N \nonumber \\
&&+\bar{\psi}_\Lambda (i\gamma ^\mu {\partial }_\mu -g_{\omega
\Lambda }\gamma ^\mu V_\mu -g_{\phi \Lambda }\gamma ^\mu \phi _\mu
-M_\Lambda +g_{s\Lambda }\sigma +g_{\sigma ^{*}\Lambda }\sigma
^{*})\psi _\Lambda \nonumber \\ &&+\bar{\psi}_\Xi (i\gamma ^\mu
{\partial }_\mu -g_{\omega \Xi }\gamma ^\mu V_\mu -g_{\phi \Xi
}\gamma ^\mu \phi _\mu -M_\Xi +g_{s\Xi }\sigma +g_{\sigma ^{*}\Xi
}\sigma ^{*})\psi _\Xi \nonumber \\ &&-\frac 14G_{\mu \nu }G^{\nu
\mu }+\frac 12\left( 1+\eta \frac \sigma {S_0}\right) m_\omega
^2V_\mu V^\mu +\frac 1{4!}\zeta \left( g_\omega ^2V_\mu V^\mu
\right) ^2 \nonumber \\ &&+\frac 12\partial _\mu \sigma \partial
^\mu \sigma -H_q\left( \frac{S^2}{ S_0^2}\right) ^{2/d}\left(
\frac 1{2d}ln\frac{S^2}{S_0^2}-\frac 14\right) \nonumber \\
&&-\frac 14S_{\mu \nu }S^{\mu \nu }+\frac 12m_\phi ^2\phi _\mu
\phi ^\mu +\frac 12\left( \partial _\nu \sigma ^{*}\partial ^\nu
\sigma ^{*}-m_{\sigma ^{*}}^2\sigma ^{*^2}\right) .\end{aligned}$$ where $g_{ij}$ are the coupling constants of the baryons to the meson fields. $G_{\mu \nu }=\partial _\mu V_\nu -\partial _\nu
V_\mu $ and $S_{\mu \nu }=\partial _\mu \phi _\nu -\partial _\nu
\phi _\mu $ are the $\omega $ field and the $\phi $ field strength tensor, respectively. The scalar fluctuation field $\sigma $ is related to $S$ by $S(x)\equiv S_0-\sigma (x)$ . $H_q$ is linked to the mass of the light scalar $S$ by the relation $m_s^2=4H_q/(d^2S_0^2)$. For the symmetric matter considered here, there is no contribution from $\rho $ meson field. In mean field approximation, the Lagrangian density then takes the form $$\begin{aligned}
{\cal L}_{MFT} &=&\bar{\psi}_N(i\gamma ^\mu \partial _\mu
+g_{\omega N}\gamma ^0V_0-M_N+g_{sN}\sigma _0)\psi _N \nonumber
\\ &&+\bar{\psi}_\Lambda (i\gamma ^\mu {\partial }_\mu -g_{\omega
\Lambda }\gamma ^0V_0-g_{\phi \Lambda }\gamma ^0\phi _0-M_\Lambda
+g_{s\Lambda }\sigma _0+g_{\sigma ^{*}\Lambda }\sigma _0^{*})\psi
_\Lambda \nonumber \\ &&+\bar{\psi}_\Xi (i\gamma ^\mu {\partial
}_\mu -g_{\omega \Xi }\gamma ^0V_0-g_{\phi \Xi }\gamma ^0\phi
_0-M_\Xi +g_{s\Xi }\sigma _0+g_{\sigma ^{*}_0\Xi }\sigma
_0^{*})\psi _\Xi \nonumber \\ &&+\frac 12\left( 1+\eta
\frac{\sigma _0}{S_0}\right) m_\omega ^2V_0^2+\frac 1{4!}\zeta
\left( g_{\omega N}V_0\right) ^4+\frac 12m_\phi ^2\phi _0^2-\frac
12m_{\sigma ^{*}}^2\sigma_0^{*^2}. \nonumber \\
&&-H_q\left( 1-\frac{\sigma _0}{S_0}\right) ^{4/d}\left[ \frac 1dln\left( 1-\frac{\sigma _0}{S_0}\right) -\frac 14\right] ,\end{aligned}$$ where the meson field operators are replaced by their mean field values: $\phi _0$, $V_0$ , $\sigma _0$ and $\sigma _0^{*}$. To derive the equation of state at finite temperature, we calculate the thermodynamic potential $\Omega $ by using the standard technique in the field theory and statistical mechanics. The result reads $$\begin{aligned}
\Omega &=&V\{H_g[(1-\frac{\phi _0}{S_0})^{\frac 4d}(\frac 1d\ln
(1-\frac{ \phi _0}{S_0})-\frac 14)+\frac 14] \nonumber \\
&&-\frac 12(1+\eta \frac{\phi _o}{S_0})m_\omega^2V_0^2-\frac
1{4!}\zeta (g_{\omega N}V_0)^4-\frac 12m_\phi ^2\phi _0^2+\frac
12m_{\sigma ^{*}}\sigma_0 ^{*2}\} \nonumber \\
&&-2k_BT\{\sum_{i,{\bf k}}\ln {[1+e^{-\beta (E_i^{*}(k)-\nu
_i}]}+\sum_{{i, {\bf k}}}\ln {[1+e^{-\beta (E_i^{*}(k)+\nu
_i)}]}\},\end{aligned}$$ where $\beta =1/k_BT$ and $V$ is the volume of the system. $$E_i^{*}(k)=\sqrt{M_i^{*2}+k^2}$$ with $$M_i^{*}=M_i-g_{si}\sigma _0-g_{\sigma ^{*}i}\sigma
_0^{*}(i=\Lambda ,\Xi ),$$ $$M_i^{*}=M_i-g_{si}\sigma _0(i=N)$$ being the effective baryon masses.The chemical potential $\nu _i $ is calculated from baryon density $\rho_{Bi}$ by the subsidiary conditions $$\rho _{Bi}=\frac 2{(2\pi )^3}\int
d^3k[n_i(k)-\overline{n}_i(k)],\hspace{2cm} (i=n,p,\Lambda ,\Xi
^0,\Xi ^{-})$$ where $n_i(k)$ and $\overline{n}_i(k)$ are baryon and anti-baryon distributions and expressed as $$n_i(k)=\{exp[(E_i^{*}(k)-\nu _i)/k_BT]+1\}^{-1}$$ and $$\overline{n}_i(k)=\{exp[(E_i^{*}(k)+\nu
_i)/k_BT]+1\}^{-1}.\hspace{2cm}$$
Having obtained the thermodynamic potential, one can easily calculate all other thermodynamic quantities of the system. For example, the pressure $p$ is given by the relation $p=-\Omega /V$, the average energy density ${\cal E}$ by ${\cal E} V=\partial
(\beta \Omega )/\partial \beta +\mu \rho V $. The resulting expressions are as the follows. $$\begin{aligned}
{\cal E} &=&\frac 2{(2\pi )^3}\sum_i\int
d^3kE_i^{*}(k)[n_i(k)+\overline{n} _i(k)]+H_q\left\{ \left(
1-\frac{\phi _0}{S_0}\right) ^{\frac 4d}\left[ \frac 1dln\left(
1-\frac{\phi _0}{S_0}\right) -\frac 14\right] +\frac 14\right\}
\nonumber \\ &&+g_{\omega N}V_0\rho _{BN}+\left( g_{\omega \Lambda
}V_0+g_{\phi \Lambda }\phi _0\right) \rho _{B\Lambda }+\left(
g_{\omega \Xi }V_0+g_{\phi \Xi }\phi _0\right) \rho _{B\Xi }
\nonumber \\ &&-\frac 12\left( 1+\eta \frac{\phi _0}{S_0}\right)
m_\omega^2V_0^2-\frac 1{4!}\zeta g_{\omega N}^4V_0^4-\frac
12m_\phi ^2\phi _0^2+\frac 12m_{\sigma ^{*}}^2\sigma _0^{*^2}\end{aligned}$$ and $$\begin{aligned}
p &=&\frac 13\frac 2{(2\pi )^3}\sum_i\int
d^3k\frac{k^2}{E_i^{*}(k)}[n_i(k)+ \overline{n}_i(k)]-H_q\left\{
\left( 1-\frac{\phi _0}{S_0}\right) ^{\frac 4d}\left[ \frac
1dln\left( 1-\frac{\phi _0}{S_0}\right) -\frac 14\right] +\frac
14\right\} \nonumber \\ &&+\frac 12\left( 1+\eta \frac{\phi
_0}{S_0}\right) m_\omega^2V_0^2+\frac 1{4!}\zeta g_{\omega N}
^4V_0^4+\frac 12m_\phi ^2\phi _0^2-\frac 12m_{\sigma ^{*}}^2\sigma
_0^{*^2}.\end{aligned}$$
$g_{sN}^{2}$ $m_{s}$ $g_{\omega N}^{2}$ $S_{0}$ $\zeta$ $\eta$ d $g_{s\Lambda}^2$ $g_{s\Xi}^2$ $g_{\sigma^*\Lambda}^2$(S) $g_{\sigma^*\Xi}^2$(S) $g_{\sigma^*\Lambda}^2$(W) $g_{\sigma^*\Xi}^2$(W)
-------------- --------- -------------------- --------- --------- -------- ------ ------------------ -------------- ---------------------------- ------------------------ ---------------------------- ------------------------
99.3 509 154.5 90.6 0.0402 -0.496 2.70 37.32 9.99 48.31 154.62 28.73 129.06
: Parameter sets, where values for $S_{0}$, the scalar mass $m_{s}$ are in MeV.
Similarly, Helmholtz free energy $F$ is calculated from internal energy $E$ and entropy $S$ by formula $F=E-TS$. Entropy $S$ is obtained from thermodynamic potential $\Omega$ by using relation $S=-(\partial \Omega /\partial
T)_{V,\mu}$.
Now we come to discuss the chemical equilibrium condition for the reactions $\Lambda +\Lambda \rightleftharpoons n+\Xi ^0$ and $\Lambda
+\Lambda \rightleftharpoons p+\Xi ^{-}$. As pointed out in Sec.I, we will discuss the system with equal number of protons and neutrons as well as equal number of $
\Xi ^0$ and $\Xi ^{-}$. In this case, the chemical equilibrium condition reads $$2\nu _\Lambda -\nu _N-\nu _\Xi =0.$$ We defines a strangeness fraction $f_S$ as $$f_S\equiv \frac{\rho _{B\Lambda }+2\rho _{B\Xi }}{\rho _B},$$ where $$\rho _B=\rho _{BN}+\rho _{B\Lambda }+\rho _{B\Xi }.$$ Given $\rho _B$ and $f_S$, we determine $\rho _{BN}$, $\rho _{B\Lambda }$ and $\rho _{B\Xi }$ by above three equations.
Results and discussions
=======================
In numerical calculation, we adopt the same parameters as in Ref.[@Zhang2](quoted in Table 1). The symbol S(W) in the parentheses denotes these coupling constants are deduced from the strong(weak) Y-Y interaction. To determine the coupling constant $g_{\sigma^*\Lambda}$ and $g_{\sigma^*\Xi}$, we use the estimation made by Schaffner et. al.[@Schaffner3]. By using the Nijmegen model D and the method given by Millener[@Millener], they found $$U_\Lambda^{(\Xi)} \simeq U_\Xi^{(\Xi)} \simeq 2U_\Lambda^{(\Lambda)}$$ at densityes of $\rho_{\Xi}=\rho_0$ and $\rho_{\Lambda}=\rho_{0}/2$, where the notation $U_{Y}^{(Y')}$ stands for the potential depth for hyperon $Y$ in a “bath” of hyperon $Y'$, and $$\frac {U_\Lambda^{(\Lambda)}} {U_{N}^{(N)}} = \frac{1}{2} \frac{1/4 V_{\Lambda \Lambda}}{3/8 V_{NN}}$$ For strong Y-Y interactions, $V_{\Lambda \Lambda} \equiv \Delta B _{\Lambda\Lambda}\simeq 4-5 MeV$ and $V_{NN} \simeq 6-7 MeV$, we have $V_{\Lambda \Lambda}/V_{NN} \simeq 3/4$. In relativistic mean field, $U_{N}^{(N)} \simeq 80 MeV$, we obtain $U_\Lambda^{(\Xi)} \simeq U_\Xi^{(\Xi)} \simeq 40 MeV$. But for weak Y-Y interaction, $\Delta B _{\Lambda\Lambda}\simeq 1.01 MeV$, we have $U_\Lambda^{(\Xi)} \simeq U_\Xi^{(\Xi)} \simeq 10 MeV$, and $g_{\sigma^*\Lambda}(W) = 5.46$, $g_{\sigma^*\Xi}(W) = 11.39$.
Besides these, we set $g_{\omega\Lambda}/g_{ \omega N}=2/3$, $g_{\omega\Xi}/g_{\omega
N}=1/3$, according to the OZI rule [@Dover1] and used the quark model relationships $g_{\phi \Xi} = 2g_{\phi \Lambda}=
-2\sqrt{2}g_{\omega N}/3$. The bare masses of baryons and mesons are $M_{N}=939$ MeV ,$M_{\Lambda}=1116$ MeV ,$M_{\Xi}=1318.1$ MeV, $m_{\omega}=783$ MeV, $m_{\sigma^{*}}=975$ MeV and $m_{\phi}=1020$ MeV.
![Free energy per baryon versus baryon density in the strange hadronic matter with various values of strangeness fraction and at temperature $T=6MeV$, calculated with a strong $\Lambda-\Lambda$ interaction.\[fig1\] ](fig1.eps){width="7cm" height="10cm"}
We will discuss the Helmholtz free energy $F$ of the strange hadronic matter first. As usually, we subtract the baryon masses in the free energy per baryon of the strange matter. $${\frac{F }{B}}=({\frac{F
}{B}})_{tot}-M_N(1-Y_\Lambda-Y_\Xi)-M_\Lambda Y_\Lambda-M_\Xi
Y_\Xi.$$ where $Y_{\Lambda}=\rho_{B\Lambda}/\rho_{B}$ and $Y_{\Xi}=\rho_{B\Xi}/ \rho_{B}$ are the hyperon fractions in the matter. In Fig.1, we have plotted the free energy per baryon versus the baryon density of the matter with various strangeness fractions at temperature $T=6MeV$, calculated with the old strong Y-Y interactions. The outstanding feature is that with the increasing strangeness fraction, the saturation curves become deeper first and then shallower. The lowest minimum occurs around $f_{S}=1.3$. The corresponding saturation density increases from 0.148${fm}^{-3}$ for ordinary nuclear matter($fs=0$) to a maximum value $\sim 0.56{fm}^{-3}$ at $f_S \simeq 1.3$ and then decreases. We also note that the F/B curve for any value of $f_S$ has a negative minimum. It means that systems at $T=6MeV$ and with any combination satisfying constraint Eq.(12) will be stable against particle emission. The Fig.2 shows the same curves as in Fig.1 but calculated with the newest weak Y-Y interactions. Due to the weakness of the Y-Y interactions used in this case, the saturation curves become shallower and shallower with increasing strangeness fraction $fs$ except at very small $fs$ value around 0.1. For the $fs$ value larger than about 1.25, there is no negative minimum in the curve. It means that the system will no longer be stable in this region.
![The same curves as in Fig.1 but calculated with a weak $\Lambda-\Lambda$ interaction.\[fig2\] ](fig2.eps){width="7cm" height="10cm"}
![ (a)The minimized free energy per baryon in the strange hadronic matter;(b)The baryon density corresponding to the minimized free energy presented in (a), as a function of strangeness fraction $f_S$.\[fig3\] ](fig3a.eps "fig:"){width="7cm" height="10cm"} ![ (a)The minimized free energy per baryon in the strange hadronic matter;(b)The baryon density corresponding to the minimized free energy presented in (a), as a function of strangeness fraction $f_S$.\[fig3\] ](fig3b.eps "fig:"){width="7cm" height="10cm"}
To see the stability of the system against $f_S$, we minimize the $F/B$ with respect to $\rho_B$ at each strangeness fraction $f_S$ for both of strong and weak Y-Y interactions, leaving out the unstable points near zero density. As a function of the strangeness fraction $f_S$, we present in Fig.3(a) the minimized $F/B$ , in Fig.3(b) the corresponding baryon density $
\rho_B$. In order to examine the role of the strange mesons, we have also presented the results without $\sigma^*$ or without both of $\sigma^*$ and $\phi$ mesons. One can see that for the strong Y-Y interactions, the minimum free energy which is much deeper than the that of the ordinary nuclear matter($f_S=0$) appears around $f_S=1.3$, where the system has almost the highest density. It means that compared with the ordinal nuclear matter the system becomes more stable. On the contrary, for the weak Y-Y interactions, the minimum free energy which is only a little deeper than the ordinary nuclear matter case appears around $f_S=0.1$. After this point, the free energy increases monotonously as baryon density increases and becomes larger than the value for the ordinary nuclear matter when $f_S>0.45$. It means that the strange hadronic matter with weak Y-Y interactions has at most the comparable stability to the ordinary nuclear matter in small strangeness fraction $f_S$ region and becomes less stable than the ordinary nuclear matter in large $f_S$ region. If the $\sigma^*$ and $\phi$ meson fields are switched off, then the curve coincides with the weak case in small $f_S$ region and becomes deeper than the curve for weak case in large $f_S$ region. The curve without the $\sigma^*$ only grows very quickly with the strangeness fraction. It means that the $\sigma^*$($\phi$) meson gives rise the attractive(repulsive) interaction between hyperons. The above situation is quite similar to that in the MQMC model[@Song]. It means that the above consequence in not model dependent.
![Free energy per baryon at saturation point for the matter with different $f_S$ as a function of temperature, calculated with a strong Y-Y interaction.\[fig4\] ](fig4.eps){width="7cm" height="10cm"}
![The same curves as in Fig.4 but calculated with a weak Y-Y interaction.\[fig5\] ](fig5.eps){width="7cm" height="10cm"}
To see the situation of the strange hadronic matter at different temperature, we present in Fig.4 the free energy at saturation point verses temperature for the matter with the strong Y-Y interaction at different strangeness fractions. One can learn that the matter with different strangeness fraction has different limit temperature $T_l$, above which the system become unstable, because the $F/B - \rho$ curves become monotonous and have no minimum when $T>T_l$. The highest limit temperature appears around $f_S=1.3$ where the system is most stable. Fig.5 shows the same curves as in Fig.4 but with the weak Y-Y interactions. One can see that the highest limit temperature appears around $f_S=0.1$ where the system is most stable in this case.
![$p-\rho_B$ isotherms of the matter with strangeness fractions $f_S=0.5$ and at various temperature.\[fig6\] ](fig6.eps){width="7cm" height="10cm"}
![$p-\rho_B$ isotherms of the matter with strangeness fractions $
f_S=1.3$ and at various temperature.\[fig7\] ](fig7.eps){width="7cm" height="10cm"}
Finally, we discuss the EOS by plotting the pressure-density ($p-\rho_B)$ isotherms. We show in Fig.6 the $
p-\rho_B$ isotherms of the strange hadronic matter at $f_S=0.5$ and with various temperature. The curve for low temperature exhibits the same typical shape as given by the Van der Waals interaction, i.e., there is an unphysical region where the pressure decreases with increasing baryon density. An inflection point appears around $T=12MeV$. The situation is quite similar to the liquid-gas phase transition in ordinary nuclear matter[@Su]. Fig.7 shows the same curves as in Fig.6 but with $f_S=1.3$. In this case, the inflection point appears around $T=6MeV$. This again indicates the system with large strangeness fraction is less stable than the one with small strangeness fraction when the Y-Y interaction is weak. On contrary, in the strong Y-Y interaction, the inflection point appears at a temperature a little higher than 12 MeV for $f_S=0.5$ and around 25 MeV for $f_S=1.3$(see Ref.[@Qian]).
In summary, we have extended an effective model describing strange hadronic matter to finite temperature and then used it to discuss the properties of multi-hyperon nuclear matter at finite temperature. It is found the strange hadronic matter with different Y-Y interactions behaves very different. While the system with the strong Y-Y interactions and in a quite large strangeness fraction region is more stable than the ordinary nuclear matter, the system with the weak Y-Y interactions is less stable than the later. This conclusion is not model dependent and true for both zero and finite temperature. If the weak $\Lambda-\Lambda$ interaction is reliable, then the previous studies on strange hadronic matter and its consequences should be reexamined. In particular, if one hope to extend above discussions from infinite strange hardronic matter to finite hypernuclei, the Coulomb interaction must be considered[@Schaffner5]. It is therefore interesting to perform further precise measurements of double hypernuclei.
Acknowledgement {#acknowledgement .unnumbered}
===============
This work is supported in part by National Natural Science Foundation of China under No.10375013, 10347107, 10075071, 10247001, 10235030, National Basic Research Program of China 2003CB716300, the Foundation of Education Ministry of China 20030246005 and CAS Knowledge Innovation Project N0.KJCX2-N11. Also supported by the Major State Basic Research Development Program under contract number G200077400 and Exploration Project of Knowledge Innovation Program of Chinese Academy of Sciences.
[99]{} H. Takahashi et al., Phys. Rev. lett.[**87**]{}, 212502-1(2001). Danysz and Pniewski, Phil. Mag. [**44**]{}, 348 (1953). J. Schaffner-Bielich, M. Hanauske, H. Stöcker and W. Greiner, arXiv:astro-ph/0005490 24 May 2000. Y. Yamamoto, S. Nishizaki and T.Takatsuka, Nucl. Phys. [**A691**]{}, 432 (2001). P.K. Sahu and A. Ohnishi, Nucl. Phys. [**A691**]{}, 439(2001). E. Witten, Phys. Rev.[**D30**]{},272(1984). A.R. Bodmer, Phys. Rev. [**D4**]{},1601(1971). Y. Zhang and R. K. Su, Phys. Rev. [**C67**]{} 015202 (2003); Phys. Rev. [**C56**]{} 035202 (2002), Europe. Phys. Lett. [**56**]{} 361 (2001); Mod. Phys. Lett. [**A18**]{} 143 (2003) E. Gilson and R. Jaffe, Phys. Rev. Lett.[**71**]{},332(1993). T.A. Armstrong [*et al.*]{}\[E864 Collaboration\], Phys. Rev. [**63**]{},054903(2001); T.A. Armstrong [*et al.*]{}\[E864 Collaboration\],Nucl. Phys. [**A625**]{},494(1997); T.A. Armstrong [*et al.*]{}\[E864 Collaboration\],Phys. Rev. Lett. [**79**]{},3612(1997). G. Appelquist [*et al.*]{}\[NA52 Collaboration\], Phys. Rev. Lett. [**76**]{},3907(1996). R. Arsenescu [*et al.*]{}\[NA52 Collaboration\], J. Phys. [**G27**]{},487(2001). John C. Hill, Nucl. Phys. [**A675**]{}(2000)226c. K. Ikeda, H. Bando, and T. Motoba, Prog. Theor. Phys. Supplement[**81**]{} (1985)147. M. Barranco, R.J. Lombard, S. Marcos, and S.A. Moszkowski, Phys. Phys. [**C44**]{}(1991)178. J. Schaffner,H. Stocker and C. Greiner, Phys. Rev. [**C46**]{},322(1992). J. Schaffener, C.B. Dover, A. Gal, C. Greiner and H. Stocker, Phys. Rev. Lett.[**71**]{},1328(1993). J. Schaffner, C.B. Dover, A. Gal, C. Greiner, D.J. Millener, and H. Stöcher, Ann. Phys. [**235**]{}(1994)35. H.-J. Schulze, M. Baldo, U. Lombardo, J. Cugnon, and A. Lejeune, Phys. Rev. [**C57**]{}(1998)704. I. Vidana, A. Polls, A. Ramos, M. Hjorth-Jensen and V.G.J. Stoks, Phys. Rev. [**C61**]{}(2000)025802-1. L.L. Zhang, H.Q. Song, and R.K. Su, J. Phys. [**G 23**]{}(1997)557. P. Wang, R.K. Su, H.Q. Song and L.L. Zhang, Nucl. Phys. [A653]{} (1999)166. L.L. Zhang, H.Q. Song, P. Wang and R.K. Su, J. Phys. [**G 26**]{}(2000)1301. P. Wang, Z.Y. Zhang, Y.W. Yu, H. Guo, R.K. Su and H.Q. Song, Nucl. Phys. [**A705**]{},455(2002). P. Wang, Z.Y. Zhang, Y.W. Yu,R.K. Su and H.Q. Song, Nucl. Phys. [**A688**]{},791(2001). M. Danysz et al., Nycl. Phys. [**49**]{}, 121(1963). R.H. Dalitz, D.H. Davis, P.H. Fowler, A. Montwill, J. Pniewski and J.A. Zakrzewski, Proc. R. Soc. London [**A426**]{}), 1(1989). D.J. Prowse, Phys. Rev. Lett.[**17**]{}, 782(1966). S. Aoki et al., Prog. Theor. Phys. [**85**]{}, 1287(1991). H.Q. Song, R.K. Su, D.H. Lu and W.L. Qian, Phys. Rev. [**C68**]{}(2003)055201. P. Papazoglou, et. al. Phys. Rev. [**C57**]{} (1998) 2576. P. Papazoglou, et. al. Phys. Rev. [**C59**]{} (1999) 411. J. Schaffner and A. Gal, Phy. Rev. [**C62**]{} (2000) 034311. K. Saito and A.W. Thomas, Phys. Lett. [**B327**]{} (1994) 9. R.J. Furnstahl, H.B. Tang. and B.D. Serot, Phys. Rev. [**C52**]{} (1995) 1368. R.J. Furnstahl, B.D. Serot and H.B. Tang, Nucl. Phys. [**A598**]{} (1996) 539. R.J. Furnstahl, B.D. Serot and H.B. Tang, Nucl. Phys. [**A615**]{} (1997) 441. Wei-Liang Qian, Ru-Keng Su and Hong-Qiu Song, Commun. Theor. Phys. [**40**]{}(2003)466. Shmuel Balberg and Avraham Gal, Nucl. Phys. [**A625**]{} (1997) 435. C.B. Dover and A. Gal, Pro. Part. Nucl. Phys., ed. D. Wilkinson, Vol.12(Pergamon, Oxford, 1984); B.K. Jinnings, Phys. Lett. [**B246**]{}(1990)325. D.Zschiesche, P. Parazoglou, Ch. W. Beckmann, S. Schramm, J. Schaffner-Bielich, H. Stcker and W. Greiner, Nucl. Phys. [**A663**]{},737c(2000). R.K. Su, S.D. Yang and T.T.S. Kuo, Phys. Rev. [**C35**]{} (1987) 1539. D.J. Millener, A. Gal, C.B. Dover and R.H. Dalitz, Phys. Rev. [**C31**]{} (1985) 499.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The new model of a quantum ring (QR) defined inside a nanowire (NW) is proposed. The one-particle Hamiltonian for electron in \[111\]-oriented NW QR is constructed taking into account both Rashba and Dresselhaus spin-orbit coupling (SOC). The energy levels as a function of magnetic field are found using the exact numerical diagonalization. The persistent currents (both charge and spin) are calculated. The specificity of SOC and arising anticrossings in energy spectrum lead to unusual features in persistent current behavior. The variation of magnetic field or carrier concentration by means of gate can lead to pure spin persistent current with the charge current being zero.'
author:
- 'I.A. Kokurin'
title: Electronic states and persistent currents in nanowire quantum ring
---
Introduction
============
The recent progress in nanowire (NW) growth technology, in particular, the possibility of radial (core-shell) [@Funk2013] and axial [@Nylund2016] heterostructure growth, leads to the opportunity to fabricate various NW-based structures, e.g. so-called NW-quantum dots (quantum dot inside NW) and other more complex ones. We suppose that the quantum ring (QR) can be grown in NW by similar way (see Fig. \[fig1\]a), e.g. using well-known hetero-pair ${\rm
GaAs-Al_xGa_{1-x}As}$ with n-doped Al-rich barriers.
Alternatively, the tubular electron gas (TEG) formed close to InAs-NW surface [@Hernandez2010] can be electrostatically confined to form QR, by analogy with confinement of carriers in carbon nanotubes to ring geometry [@Bulaev2008]. Both mentioned structures significantly differ from known QRs defined in two-dimensional electron gas (2DEG) structures. Especially, the effects that are due to spin-orbit coupling (SOC) will be different.
QRs constitute the polygon for study such coherent effects as Aharonov-Bohm (AB) one [@Aharonov1959] and persistent currents (PCs) [@Buttiker1983]. Usually zero-temperature PC can be found by using the well-known equation
$$\label{PC} I=-c\sum_n\frac{\partial E_n}{\partial\Phi},$$
where $E_n$ is the energy levels in the ring, $\Phi$ is the magnetic flux through the ring and the summation is over all occupied states.
The presence of SOC leads to modification of energy spectrum and as a consequence to PC modification. The spectrum of thin (one-dimensional) QR defined in 2DEG structure with Rashba SOC [@Bychkov1984] is well known [@Chaplik1995; @Meijer2002]. However, the lack of inversion center in host semiconductor material leads to other type of spin-splitting [@Dresselhaus1955] (known as Dresselhaus SOC), that appear in a new light in low-dimensional structures [@Dyakonov1986] and particularly in QR-structures [@Sheng2006].
Here we will study only features of PCs that are due to specifics of spectrum and SOC, and neglect the disorder and interaction effects. We present the model of thin QR which is confined inside \[111\]-oriented NW (it is the usual growth direction for NWs of ${\rm
A_{III}B_V}$ materials with zinc-blende lattice). It should be noted, that SOC in NW QR sufficiently differs from SOC in planar QR, where effective Rashba field is constant at each point, and Dresselhaus SOC is different as well due to another crystallographic orientation. It is worth noted, that there is no difference between AB-flux and homogeneous magnetic field for model of thin one-dimensional QR if $g=0$. However, it seems to be unreal to realize AB-flux through the NWQR, and we discuss here only the case of homogeneous magnetic filed.
Model and Hamiltonian
=====================
Using one-particle Hamiltonian of NW with TEG [@Kokurin2014; @Kokurin2015], that takes into account both Rashba and $k$-linear Dresselhaus SOC, after dimension quantization along NW axis, we find the following effective-mass Hamiltonian for carriers in NWQR
$$\begin{aligned}
\label{Hamiltonian} \nonumber
H=\frac{\hbar^2K_\varphi^2}{2m}+\alpha\sigma_zK_\varphi-\frac{\beta}{2\sqrt
3}\left(\sigma_rK_\varphi-\frac{i}{2r_0}\sigma_\varphi\right)\\
+\sqrt\frac32\beta\sigma_z\left(\sin3\varphi
K_\varphi-\frac{3i}{2r_0}\cos
3\varphi\right)+\frac12g\mu_B\sigma_zB,\end{aligned}$$
where the total Hamiltonian consists of kinetic term, Rashba SOC, isotropic and anisotropic Dresselhaus SOC terms, and Zeeman splitting, respectively. Here $K_\varphi=r_0^{-1}(-i\partial/\partial\varphi+\Phi/\Phi_0)$ with $\Phi=\pi r_0^2B$ being the flux of the magnetic field ${\bf B}$, and $\Phi_0=2\pi\hbar c/|e|$ is the flux quantum, $m$, $\alpha$, $\beta$, $g$ and $r_0$ are the effective mass, Rashba and $k$-linear Dresselhaus SOC parameter, effective g-factor and QR radius, respectively. Here we use the polar Pauli matrices $\sigma_r=\cos\varphi\sigma_x+\sin\varphi\sigma_y$, $\sigma_\varphi=-\sin\varphi\sigma_x+\cos\varphi\sigma_y$, that connected with usual Cartesian Pauli matrices $\sigma_i$ ($i=x,y,z$). The constant energy shift that is due to dimension quantization is omitted in Eq. (\[Hamiltonian\]).
![\[fig1\] (a) Sketch of one possible realization of NWQR. (b) Energy spectrum of NWQR as a function of magnetic flux. $m=0.026m_0$, $2m\alpha r_0/\hbar^2=2.5$, $2m\beta r_0/\hbar^2=0.8$, $g=-3.0$. (c) Density of states at $\Phi=2\Phi_0$ marked by dashed line at spectrum, $\gamma=0.03E_0$ with $E_0=\hbar^2/2mr_0^2$.](fig01.eps){width="80mm"}
The spectral problem for Hamiltonian (\[Hamiltonian\]) can be solved only numerically, but any appropriate set of basis functions can be used for numerical diagonalization. One can see that the Hamiltonian without penultimate term commutes with $z$-projection of total angular momentum, $j_z=-i\hbar\partial/\partial\varphi+(\hbar/2)\sigma_z$ and therefore it can be diagonalized analytically. Thus, it is convenient to use the eigenfunction of such reduced Hamiltonian for diagonalization of Hamiltonian (\[Hamiltonian\]). The result of numerical diagonalization is depicted in Fig. \[fig1\]b, where we restrict ourselves by $40\times 40$ matrix that ensures the perfect precision for depicted levels. The numerical calculation was performed with material parameters that are typical for InAs-based structures for which SOC-effects are more pronounced. One can see the presence of ‘gap’ (anticrossings) at the spectrum that is due to penultimate term in Hamiltonian (\[Hamiltonian\]). It should be noted, that the energy levels will be $\Phi_0$-periodic at $g=0$.
Additionally, we can calculate the one-particle density of states (DoS) (see Fig. \[fig1\]c) using the Green’s function of Hamiltonian $${\rm DoS}(E)=-\frac{1}{\pi}{\rm Im Tr}(E-H+i\gamma)^{-1},$$ where $\gamma$ describes the level broadening.
Persistent currents
===================
Since we know the energy spectrum, then we can use Eq. (\[PC\]) to find PC numerically. However, it is convenient to use the equilibrium density matrix formalism, and to calculate PC by using equation $$\label{current}
I=\frac{e}{2\pi r_0}{\rm Tr}[v_\varphi f_0(H,\mu,T)],$$ where $f_0(E,\mu,T)$ is the Fermi distribution function, $E$, $\mu$ and $T$ are the energy, chemical potential and temperature, respectively. This approach is more general and permits to find current at finite temperature. In this case we have to know the matrix elements of velocity operator $v_\varphi$ that in turn can be found from Heisenberg equation of motion, $v_\varphi=-(i/\hbar)[r_0\varphi,H]$.
![\[fig2\] PC as a function of magnetic flux at constant chemical potential, $m=0.026m_0$, $2m\alpha r_0/\hbar^2=2.5$, $2m\beta r_0/\hbar^2=0.8$, $g=-3.0$. $I_0=|e|\hbar/4\pi mr_0^2$. (a) $\mu=4.0E_0$; (b) $\mu=0.0$. Full line corresponds to $T=0$, dashed line corresponds to $T=0.1E_0$.](fig02.eps){width="80mm"}
There are possible two different situations: (i) the constant particle number, $N={\rm const}$, that e.g. can be realized in QR defined in core-shell NW-structure and (ii) the constant chemical potential, $\mu={\rm const}$, that can be realized in electrostatically confined TEG in InAs-NW. In the first case one has to additionally find the dependence $\mu(\Phi)$ using well-known relation, $N={\rm Tr}[f_0(H,\mu,T)]$. In the case of $\mu={\rm
const}$ the abrupt change in PC occurs near the crossing of energy levels with the chemical potential, whereas in $N={\rm const}$ case it happens close to level crossing (anticrossing).
The numerical calculation of PC as a function of magnetic flux at $\mu={\rm const}$ is depicted in Fig. \[fig2\]. One can see the usual saw-toothed behavior for the high chemical potential (Fig. \[fig2\]a) at zero temperature, but there is no $\Phi_0$-periodicity due to non-zero g-factor. The increasing of the temperature leads to smoothing of $I(\Phi)$-dependence. When the chemical potential lies in a ‘gap’ the PC behavior is significantly different. In this case $I(\Phi)$-dependence is smoothly oscillatory even at $T=0$ (see Fig. \[fig2\]b).
It is interesting to study persistent spin current (PSC) as well. Usually, the spin current is the pseudo-tensor $I^i_j$ ($i,j=x,y,z$), that components describe the spin component $s_i$ carrying in $j$-th spatial direction. In our case of thin QR there is only one coordinate direction, tangential to QR circumference, i.e. in our case we deal with pseudo-vector $I^i_\varphi\equiv I^i$ ($i=r,\varphi,z$). Here we use the simple definition of spin-current operator and PSC can be found from Eq. (\[current\]) with replacement
$$\label{spin_current}
ev_\varphi\rightarrow\frac{\hbar}{4}(\sigma_iv_\varphi+v_\varphi\sigma_i).$$
Numerical calculations show that at specific values of magnetic field and chemical potential there is possibility to observe a pure PSC, i.e. in this case charge current is zero (see for instance Ref. [@Splettstoesser2003]). The realization of pure PSC is convenient with help of electric gates, that can control not only electron concentration, but the SOC parameter $\alpha$ and in a less degree $\beta$.
Conclusion
==========
In conclusion, we proposed two models of QR defined in zinc-blende NWs and constructed the one-particle Hamiltonian for electron in NWQR taking into account both Rashba and $k$-linear Dresselhaus SOC. The specificity of spin-orbit terms in \[111\]-oriented QR manifests in energy levels and PCs. The energy levels as a function of magnetic field are found revealing several anticrossings that are due to anisotropic part of Dresselhaus SOC. Using the equilibrium density matrix formalism PC and PSC are calculated. Charge and spin PC behavior differs from conventional one, especially for carrier concentration corresponding to the Fermi level position close to anticrossings. The variation of magnetic field or carrier concentration and SOC strength by means of gate can lead to pure spin PC when the charge current is zero.
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
S. Funk, M. Royo, I. Zardo, D. Rudolph, S. Morkötter, B. Mayer, J. Becker, A. Bechtold, S. Matich, M. Döblinger, M. Bichler, G. Koblmüller, J.J. Finley, A. Bertoni, G. Goldoni, G. Abstreiter, High mobility one- and two-dimensional electron systems in nanowire-based quantum heterostructures. Nano Lett. [**13**]{}, 6189 (2013). [](http://dx.doi.org/10.1021/nl403561w).
G. Nylund, K. Storm, S. Lehmann, F. Capasso, L. Samuelson, Designed quasi-1D potential structures realized in compositionally graded ${\rm InAs_{1-x}P_x}$ nanowires. Nano Lett. [**16**]{}, 1017 (2016). [](http://dx.doi.org/10.1021/acs.nanolett.5b04067).
S. Estévez Hernández, M. Akabori, K. Sladek, C. Volk, S. Alagha, H. Hardtdegen, M.G. Pala, N. Demarina, D. Grützmacher, T. Schäpers, Spin-orbit coupling and phase coherence in InAs nanowires. Phys. Rev. B 82, 235303 (2010). [](http://dx.doi.org/10.1103/PhysRevB.82.235303).
D.V. Bulaev, B. Trauzettel, D. Loss, Spin-orbit interaction and anomalous spin relaxation in carbon nanotube quantum dots. Phys. Rev. B [**77**]{}, 235301 (2008). [](http://dx.doi.org/10.1103/PhysRevB.77.235301).
Y. Aharonov, D. Bohm, Significance of electromagnetic potentials in the quantum theory. Phys. Rev. [**115**]{}, 485 (1959). [](http://dx.doi.org/10.1103/PhysRev.115.485).
M. Buttiker, Y. Imry, R. Landauer, Josephson behavior in small normal one-dimensional rings. Phys. Lett. A [**96**]{}, 365 (1983). [](http://dx.doi.org/10.1016/0375-9601(83)90011-7).
Yu.A. Bychkov, E.I. Rashba, Properties of a 2D electron gas with lifted spectral degeneracy. JETP Lett. [**39**]{}, 78 (1984).
A.V. Chaplik, L.I. Magarill, Effect of the spin-orbit interaction on persistent currents in quantum rings. Superlatt. Microstruct. [**18**]{}, 321 (1995). [](http://dx.doi.org/10.1006/spmi.1995.1118).
F.E. Meijer, A.F. Morpurgo, T.M. Klapwijk, One-dimensional ring in the presence of Rashba spin-orbit interaction: Derivation of the correct Hamiltonian. Phys. Rev. B [**66**]{}, 033107 (2002). [](http://dx.doi.org/10.1103/PhysRevB.66.033107).
G. Dresselhaus, Spin-orbit coupling effects in zinc blende structures. Phys. Rev. [**100**]{}, 580 (1955). [](http://dx.doi.org/10.1103/PhysRev.100.580).
M.I. Dyakonov, V.Yu. Kachorovskii, Spin relaxation of two-dimensional electrons in noncentrosymmetric semiconductors. Sov. Phys. Semicond. [**20**]{}, 110 (1986).
J.S. Sheng, K. Chang, Spin states and persistent currents in mesoscopic rings: Spin-orbit interactions. Phys. Rev. B [**74**]{}, 235315 (2006). [](http://dx.doi.org/10.1103/PhysRevB.74.235315).
I.A. Kokurin, Determination of Rashba-coupling strength for surface two-dimensional electron gas in InAs nanowires. Solid State Commun. [**195**]{}, 49 (2014). [](http://dx.doi.org/10.1016/j.ssc.2014.07.002).
I.A. Kokurin, Effect of spin-orbit coupling on spectral and transport properties of tubular electron gas in InAs nanowires. Physica E [**74**]{}, 264 (2015). [](http://dx.doi.org/10.1016/j.physe.2015.07.003).
J. Splettstoesser, M. Governale, U. Zülicke, Persistent current in ballistic mesoscopic rings with Rashba spin-orbit coupling. Phys. Rev. B [**68**]{}, 165341 (2003). [](http://dx.doi.org/10.1103/PhysRevB.68.165341).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the *Solar Dynamics Observatory* observations of two flux ropes respectively tracked out by material from a surge and a failed filament eruption on 2012 July 29 and August 04. For the first event, the interaction between the erupting surge and a loop-shaped filament in the east seems to “peel off" the filament and add bright mass into the flux rope body. The second event is associated with a C-class flare that occurs several minutes before the filament activation. The two flux ropes are respectively composed of 85$\pm$12 and 102$\pm$15 fine-scale structures, with an average width of about 1$\arcsec$.6. Our observations show that two extreme ends of the flux rope are rooted in the opposite polarity fields and each end is composed of multiple footpoints (FPs) of the fine-scale structures. The FPs of the fine-scale structures are located at network magnetic fields, with magnetic fluxes from 5.6$\times$10$^{18}$ Mx to 8.6$\times$10$^{19}$ Mx. Moreover, almost half of the FPs show converging motion of smaller magnetic structures over 10 hr before the appearance of the flux rope. By calculating the magnetic fields of the FPs, we deduce that the two flux ropes occupy at least 4.3$\times$10$^{20}$ Mx and 7.6$\times$10$^{20}$ Mx magnetic fluxes, respectively.'
author:
- 'Ting Li & Jun Zhang'
title: 'Fine-scale Structures of Flux Ropes Tracked by Erupting Material'
---
Introduction
============
The flux rope is thought to be closely related to the coronal mass ejection (CME), which generally has a three$-$part structure: the bright core, the dark cavity and the leading edge (see e.g., Illing & Hundhausen 1986). It is often believed that the twisted flux rope is the dark cavity which accumulates magnetic energy and mass within it (Chen 1996; Hudson & Schwenn 2000; Gibson et al. 2006). The ideal magnetohydrodynamic (MHD) instabilities of the flux rope such as the kink instability and the torus instability are thought to be one type of mechanism that triggers the CME and associated activities (T[ö]{}r[ö]{}k & Kliem 2003, 2005; Fan 2005; Kliem & T[ö]{}r[ö]{}k 2006; Olmedo & Zhang 2010). Thus, a detailed study of the flux rope is important for a clear understanding of CMEs, and this leads to a good ability to forecast CMEs and associated space weather.
The flux rope has been reconstructed in a complex magnetic topology from observed vector magnetograms by using non linear force-free field models (Guo et al. 2010; Canou & Amari 2010; Jing et al. 2010). Moreover, the formation and dynamic behavior of flux ropes have been simulated successfully by many authors (Amari & Luciani 1999; Amari et al. 2010; Fan & Gibson 2004). In the simulation of Aulanier et al. (2010), the flux rope is progressively formed by flux-cancellation-driven photospheric reconnection in a bald-patch separatrix.
Recently, the direct observations of flux ropes during the eruption process have been reported by using the data from the Atmospheric Imaging Assembly (AIA; Lemen et al. 2012) onboard the *Solar Dynamics Observatory* (*SDO*; Pesnell et al. 2012). Cheng et al. (2011) presented the observation of a flux rope as a bright blob of hot plasma in the channel of 131 [Å]{}. The hot flux rope rapidly moved outward and stretched the surrounding magnetic field upward, similar to the classical magnetic reconnection scenario in eruptive flares. Detailed thermal property of the flux rope based on differential emission measure was investigated by Cheng et al. (2012). However, there is to our knowledge no observations of fine-scale structures of flux ropes tracked by erupting material. In this letter, we present two events of flux ropes and investigate their fine-scale structures and magnetic properties by using the data from *SDO*/AIA and the Helioseismic and Magnetic Imager (HMI; Schou & Larson 2011).
Observations and Data Analysis
==============================
On 2012 July 29 and August 04, *SDO*/AIA observed two flux ropes that are respectively tracked by material from a surge and a failed filament eruption. Both of them are composed of thread-like structures, which warp and interweave together.
The *SDO*/AIA takes full-disk images in 10 (E)UV channels at 1$\arcsec$.5 resolution and high cadence of 12 s. The flux ropes could be observed in all the 7 EUV channels. The 171 channel best shows the flux rope and we focus on this channel in this study. We also present the observations of flux ropes in different channels such as 304, 193, 335 and 131 [Å]{}.
The 5 EUV channels correspond to different temperatures: 171 [Å]{} (Fe IX) at 0.6 MK, 304 [Å]{} (He II) at 0.05 MK, 193 [Å]{} (Fe XII) at 1.5 MK (with a hot contribution of Fe XXIV at 20 MK and cooler O V at 0.2 MK), 335 [Å]{} (Fe XVI) at 2.5 MK and 131 [Å]{} (Fe VIII, Fe XXI) at 11 MK (O’Dwyer et al. 2010; Boerner et al. 2012; Parenti et al. 2012). We also use the full-disk line-of-sight magnetic field data from the HMI onboard *SDO*, with a cadence of $\sim$ 45 s and a sampling of 0$\arcsec$.5 pixel$^{-1}$.
Results
=======
Overview of the Two Flux Ropes
------------------------------
At about 00:30 UT on 2012 July 29, a surge occurred in NOAA AR 11530 (S19W00) and plenty of material was ejected northward (see Figure 1*b* and Animation 1 in the online journal). By examining the *SDO*/HMI line-of-sight magnetograms, we found that the magnetic flux cancellation took place several hours prior to the surge (Figure 1*d*). The obvious brightening was observed at 171 [Å]{} at the location of the cancelled flux. This cancellation could be what led to the occurrence of the surge. When the surge first appears and starts to ascend, there is evidence of interaction between it and a loop-shaped filament in the east side of the erupting surge (see Figure 1*e* and Animation 2 in the online journal). This interaction seems to “peel off" the filament and to add mass into the flux rope body. Simultaneously, brightenings at the interaction location and the footpoint of the surge at 304 [Å]{} are also observed (see Figure 1*e* and Animation 2 in the online journal).
At about 00:52 UT, the erupting material seemed to be confined and moved toward the west (see Animation 1 in the online journal). Meanwhile, the bright fine-scale structures are clearly observed. About 52 min later (01:44 UT), the erupting material arrived at the west end of the flux rope and the entire flux rope was tracked out by the ejected material (see Figure 1*c*). It seems that the surge occurs within the flux rope and the footpoint of the surge is located at one end of the flux rope, and thus the material from the surge flows along the flux rope. The approximate length of the flux rope is 596 Mm, and the apparent flow speed of the material along the flux rope body is about 150 km s$^{-1}$. The flux ropes observed at 304, 193, 335 and 131 [Å]{} are roughly the same on the whole. However, they are different in some details. Taking the area denoted by white rectangles in Figure 1 for example, the 193 [Å]{} observations are similar to those of 171 [Å]{}, and the 304, 335 and 131 [Å]{} observations are different from 171 [Å]{} observations as the fine-scale structures (pointed by white arrows) are not clearly identified in these three channels.
On 2012 August 04, the second flux rope was observed in NOAA AR 11539 (S23E32). Before the flux rope was tracked, there existed a filament at the east part of the flux rope (see Figure 2*a*). At 11:04 UT, a C2.9 flare occurred at the east of the filament (see Figure 2*e* and Animation 4 in the online journal), which peaked at 11:47 UT and ended at 12:49 UT. At about 11:14 UT, the filament started to turn over and brighten (Figure 2*b*). At about 11:40 UT, the ejected material from the filament successively moved toward the southwest (see Animation 3 in the online journal). Then the arch-shaped flux rope with helical fine-scale structures was observed clearly (Figure 2*c*). The observed length of the flux rope is about 546 Mm, and it takes the filament material about 51 min to flow from the east to the west end. The apparent flow velocity along the flux rope body is approximately 180 km s$^{-1}$. Similar to the first flux rope on 2012 July 29, the fine-scale structures denoted by the white arrow in white rectangles of Figure 2 are identified clearly at 171 and 193 [Å]{} and seems obscure at 304, 335 and 131 [Å]{}. At about 16:00 UT on August 6, the east part of the second flux rope erupted, and this eruption resulted in a faint CME with a speed of about 260 km s$^{-1}$. The whole flux rope erupted at about 03:00 UT on August 8, associated with a halo CME with a speed of about 230 km s$^{-1}$.
Fine-scale Structures of the Two Flux Ropes
-------------------------------------------
As the erupting material arrived at the west extreme end of the first flux rope on 2012 July 29, the west end showed obvious brightening (Figures 1*c* and 3*a*). Then partial material went back toward the east and brightened the east end at about 03:02 UT (see Figure 3*d* and Animation 1 in the online journal). The brightening at the ends makes it possible to determine the ends location. Thus we select two areas (denoted by red and blue rectangles in Figure 1*c*) to investigate the ends and fine-scale structures.
By counting the number of fine-scale structures one by one, we notice that the first flux rope is composed of 85$\pm$12 fine-scale structures, and 15 well identified fine-scale structures are selected to measure their widths. Two examples are shown in Figures 3*c*, *f* and *g*. Firstly, the intensity-location profiles (black curves in Figures 3*f* and *g*) along slices perpendicular to the fine-scale structures (Slices “S1" and “S2" in Figure 3*c*) are obtained. Secondly, we use Gaussian function to fit the intensity-location profiles and two Gaussian fitting profiles (blue and red ones) are shown in Figures 3*f* and *g*. The full width at half maximum (FWHW) of the Gaussian fitting profile is thought to be the width of fine-scale structure. The average width of these fine-scale structures is 1$\arcsec$.8, with the maximum value of 2$\arcsec$.0 and the minimum value of 1$\arcsec$.4.
For the first flux rope, there are 12 western footpoints (FPs) of the fine-scale structures that form the west end of the flux rope (white circles in Figure 3*a*). By comparing the 171 [Å]{} observations with line-of-sight magnetograms, we find that all the western FPs are rooted in negative polarity fields (Figure 3*b*). The net magnetic fluxes of these FPs are in the range of 8.6$\times$10$^{18}$$-$8.6$\times$10$^{19}$ Mx. The magnetic flux of the west end of the flux rope is $-$4.3$\times$10$^{20}$ Mx. This is the lower limit of the magnetic flux since some FPs are not identified and calculated for they are not accompanied by brightening. The 8 eastern FPs are rooted in positive polarity fields (Figures 3*d* and *e*). The net magnetic fluxes of eastern FPs are from 5.6$\times$10$^{18}$ to 2.8$\times$10$^{19}$ Mx. The magnetic flux of the east end of the flux rope is 1.3$\times$10$^{20}$ Mx, which is less than that of the western end. Not all the eastern FPs of the fine-scale structures are identified and thus there exists the discrepancy between the two sides.
For the second flux rope on 2012 August 04, we similarly select two areas where the brightening occurs at the ends (Figures 2*c* and 4). This arch-shaped flux rope seems more complex than the first one and has two main western ends (one in Figure 4*a* and the other one in Figure 4*d*) which are separated apart. The flux rope is composed of 102$\pm$15 fine-scale structures. The average width of 20 clearly identified fine-scale structures is 1$\arcsec$.5. The thickest one has a width of 1$\arcsec$.7, and the thinnest one has that of 1$\arcsec$.1 (Figures 4*c*, *f* and *g*). The line-of-sight magnetograms show that 22 western FPs (15 ones in Figures 4*a*$-$*b* and 7 ones in Figures 4*d*$-$*e*) of the fine-scale structures are anchored at positive polarity fields (Figures 4*a*$-$*b*, *d*$-$*e*). The net magnetic fluxes of these western FPs are in the range of 1.1$-$8.1$\times$10$^{19}$ Mx. The total magnetic flux of western ends of the flux rope is 7.6$\times$10$^{20}$ Mx. The eastern end of the flux rope is close to the solar limb and not accompanied by EUV enhancements, thus it could not be identified.
By examining the magnetic field evolution at the FPs of the fine-scale structures over 10 hr before the appearance of the flux rope, we find almost half of these FPs show converging motion of smaller magnetic structures for both the flux ropes. Figure 5 presents one example of the eastern FPs of the first flux rope. As seen in the stack plot along Slice “A$-$B", the west magnetic structure obviously moved toward the east one with a velocity of 0.2 km s$^{-1}$ and the motion of the east one was slower than the west one, with a velocity of 0.1 km s$^{-1}$ (Figure 5*e*).
Summary and Discussion
======================
We present the *SDO*/AIA observations of two flux ropes on 2012 July 29 and August 04 which are tracked out by material from a surge and a failed filament eruption. For the two flux ropes, the apparent speeds of filling of the flux rope structure with chromospheric and coronal plasma are respectively 150 and 180 km s$^{-1}$, which are comparable to the typical sound speed for the corona of about 100$-$200 km s$^{-1}$. For the first flux rope, the approximate length of the flux rope is 596 Mm. By examining the fine-scale structures which are observed more clearly, we roughly estimate the twist is about $\pi$. For event 2, the observed length of the flux rope is about 546 Mm, and the average twist is about 2$\pi$. Seen from the *Solar-Terrestrial Relations Observatory* (*STEREO*; Kaiser et al. 2008) B viewpoint, the two flux ropes are both located at the west limb. By using three-dimensional reconstructions, we obtain the heights of the two flux ropes, which are respectively 90 and 140 Mm above the solar surface. The two flux ropes analyzed here are respectively composed of 85$\pm$12 and 102$\pm$15 fine-scale structures, which probably outline the magnetic field structures of flux ropes (Martin et al. 2008; Lin 2011). The width of the fine-scale structures ranges from 1$\arcsec$.1 to 2$\arcsec$.0, with an average of about 1$\arcsec$.6. It is comparable to the resolution limit of the AIA telescope of about 1$\arcsec$.2, which suggests that even thinner structures may exist. Moreover, the width of the fine-scale structures is an order of magnitude larger than the ultrafine magnetic loop structures observed by Ji et al. (2012).
Before the flux ropes are tracked out by erupting material, part of magnetic flux rope structures may exist in the space filled in by the flux ropes. For event 1, there are several similar events in two days before the appearance of the first flux rope on 2012 July 29. Moreover, the 304 [Å]{} images shortly before the surge injection and flux rope appearance reveal the presence of long and thin absorbing threads along the first flux rope. For event 2, the filament at the east location of the flux rope may indicate part of the pre-existing flux rope structures.
When the surge in event 1 appears and the filament in event 2 is activated, the obvious brightenings and flare activities are observed simultaneously at the east of the two flux ropes. This implies that heating takes place and may illuminate the flux rope body by filling it with hot and dense plasma emitting in the EUV channels. This is similar to the observations of Raouafi (2009), who suggested that a C-class flare near one footpoint of the flux rope led to the brightening of the magnetic structure showing its fine structure. While the material arrives at the FPs of the fine-scale structures, the FPs are consequently brightened. The brightening at the FPs may be caused by the conversion from the kinetic energy to the thermal energy.
Our observations show that there exist 7$-$15 FPs of the fine-scale structures for each end of the flux rope. By comparing the EUV observations with the HMI magnetograms, we find that the FPs at one end of the flux rope on July 29 are rooted in the same polarity fields and the FPs at the other end are anchored at the opposite polarity fields (Figure 3). For the flux rope on August 04, the eastern end could not be identified and only the western end is analyzed here. The FPs of the fine-scale structures are located at network magnetic fields and their magnetic fluxes are in the range of 5.6$\times$10$^{18}$$-$8.6$\times$10$^{19}$ Mx. The magnetic fluxes of the two flux ropes are at least 4.3$\times$10$^{20}$ and 7.6$\times$10$^{20}$ Mx. According to the statistical study of Sung et al. (2009), the magnetic flux of 34 magnetic clouds (MC) for in-situ observations varies from 1.25$\times$10$^{18}$ to 4.69$\times$10$^{21}$ Mx with the average of 1.1$\times$10$^{21}$Mx. The magnetic flux of the flux ropes in our observations is comparable to that of the MC. Moreover, almost half of the FPs of the fine-scale structures show converging motion of smaller magnetic structures over 10 hr before the appearance of the flux rope. The network magnetic field is often thought to be the converging center (Zhang et al. 1998), which is consistent with our observations. As small-scale magnetic fields located at the converging centers always exist tens of hours (Liu et al. 1994), implying that the flux ropes have a relatively long lifetime.
The flux ropes are observed in all the 7 EUV channels (304, 171, 193, 211, 335, 94 and 131 [Å]{}) of the *SDO*/AIA that cover the temperature from 0.05 MK to 11 MK. This is consistent with recent observations of Patsourakos et al. (2013) and Li & Zhang (2013), who reported the hot and cool components of flux ropes. However, there exist the flux ropes that could only be observed in hot channels such as 94 and 131 [Å]{} (Zhang et al. 2012; Cheng et al. 2011, 2012). The comprehensive characteristics of flux ropes need to be analyzed in further studies.
Amari, T., Aly, J.-J., Mikic, Z., & Linker, J. 2010, , 717, L26 Amari, T., & Luciani, J. F. 1999, , 515, L81 Aulanier, G., T[ö]{}r[ö]{}k, T., D[é]{}moulin, P., & DeLuca, E. E.2010, , 708, 314 Boerner, P., Edwards, C., Lemen, J., et al. 2012, , 275, 41 Canou, A., & Amari, T. 2010, , 715, 1566 Chen, J. 1996, , 101, 27499 Cheng, X., Zhang, J., Liu, Y., & Ding, M. D. 2011, , 732, L25 Cheng, X., Zhang, J., Saar, S. H., & Ding, M. D. 2012, , 761, 62 Fan, Y. 2005, , 630, 543 Fan, Y., & Gibson, S. E. 2004, , 609, 1123 Gibson, S. E., Foster, D., Burkepile, J., de Toma, G., & Stanger, A. 2006, , 641, 590 Guo, Y., Schmieder, B., D[é]{}moulin, P., et al. 2010, , 714, 343 Hudson, H., & Schwenn, R. 2000, Advances in Space Research, 25, 1859 Illing, R. M. E., & Hundhausen, A. J. 1986, , 91, 1095 Ji, H., Cao, W., & Goode, P. R. 2012, , 750, L25 Jing, J., Yuan, Y., Wiegelmann, T., et al. 2010, , 719, L56 Kaiser, M. L., Kucera, T. A., Davila, J. M., St. Cyr, O. C., Guhathakurta, M., & Christian, E. 2008, Space Sci. Rev., 136, 5 Kliem, B., & T[ö]{}r[ö]{}k, T. 2006, Physical Review Letters, 96, 255002 Li, L. & Zhang, J. 2013, , 552, L11 Lin, Y. 2011, , 158, 237 Liu, Y., Zhang, H., Ai, G., Wang, H., & Zirin, H. 1994, , 283, 215 Lemen, J. R., Title, A. M., Akin, D. J., et al. 2012, Sol. Phys., 275, 17 Martin, S. F., Lin, Y., & Engvold, O. 2008, , 250, 31 O’Dwyer, B., Del Zanna, G., Mason, H. E., Weber, M. A., & Tripathi, D. 2010, , 521, A21 Olmedo, O., & Zhang, J. 2010, , 718, 433 Patsourakos, S., Vourlidas, A., & Stenborg, G. 2013, , 764, 125 Parenti, S., Schmieder, B., Heinzel, P., & Golub, L. 2012, , 754, 66 Pesnell, W. D., Thompson, B. J., & Chamberlin, P. C. 2012, Sol. Phys., 275, 3 Raouafi, N.-E. 2009, , 691, L128 Schou, J., & Larson, T. P. 2011, Bulletin of the American Astronomical Society, 1605 Sung, S.-K., Marubashi, K., Cho, K.-S., et al. 2009, , 699, 298 T[ö]{}r[ö]{}k, T., & Kliem, B. 2003, , 406, 1043 T[ö]{}r[ö]{}k, T., & Kliem, B. 2005, , 630, L97 Zhang, J., Cheng, X., & Ding, M.-D. 2012, Nature Communications, 3, 747 Zhang, J., Wang, J., Wang, H., & Zirin, H. 1998, , 335, 341
![ [*SDO*]{}$/$AIA multi-wavelength images and HMI line-of-sight magnetogram showing the evolution of the flux rope on 2012 July 29 (see Animations 1 and 2, available in the online edition of the journal). The red solid rectangle in panel *c* denotes the FOV of Figure 3*a* and the blue one denotes the FOV of Figure 3*d*. \[fig1\]](fig1.eps)
![[*SDO*]{}$/$AIA multi-wavelength images and HMI line-of-sight magnetogram showing the evolution of the flux rope on 2012 August 04 (see Animations 3 and 4, available in the online edition of the journal). The red solid rectangle in panel *c* denotes the FOV of Figure 4*a* and the blue one denotes the FOV of Figure 4*d*. \[fig2\]](fig2.eps)
![[*SDO*]{}$/$AIA 171 [Å]{} images and HMI magnetograms showing the western (panels *a* and *b*) and eastern ends (panels *d* and *e*) of the flux rope on 2012 July 29, and the Gaussian fitting profiles (panels *f* and *g*) showing the widths of fine-scale structures. The white rectangle in panel *a* denotes the FOV of panel *c* and white rectangles in panels *d* and *e* denote the FOV of Figure 5. The blue and red curves in panels *f* and *g* are respectively the Gaussian fitting profiles of the intensity-location curves (black ones) along the blue and red slices (Slices “S1" and “S2") in panel *c*. \[fig3\]](fig3.eps)
![[*SDO*]{}$/$AIA 171 [Å]{} images and HMI magnetograms showing two main western ends of the flux rope (one in panels *a*$-$*b* and the other in panels *d*$-$*e*) on 2012 August 4, and the Gaussian fitting profiles (panels *f* and *g*) showing the widths of fine-scale structures. The white rectangle in panel *a* denotes the FOV of panel *c*.\[fig4\]](fig4.eps)
![[*SDO*]{}$/$HMI magnetograms showing the converging motion of smaller magnetic structures at the FPs of the fine-scale structures. Blue circles denote one of the eastern FPs (“FP 1" in Figure 3*e*) of the fine-scale structures for the first flux rope on 2012 July 29. The stack plot along Slice “A$-$B" (red dashed line in panel *a*) is shown in panel *e*. \[fig5\]](fig5.eps)
|
{
"pile_set_name": "ArXiv"
}
|
---
address: 'Department of Mathematics, Columbia University, 2990 Broadway, New York, NY 10027, USA'
author:
- 'Chiu-Chu Melissa Liu'
date: 'May 10, 2019'
title: A Lecture on Holomorphic Anomaly Equations and Extended Holomorphic Anomaly Equations
---
[^1]
Mirror Symmetry for Compact Calabi-Yau threefolds
=================================================
Mirror symmetry relates the A-model on a compact Calabi-Yau threefold $X$, defined in terms of the symplectic structure on $X$, to the B-model on a mirror compact Calabi-Yau threefold $Y$, defined in terms of the complex structure on $Y$, $$\textup{A-model }(X,\omega) \ \stackrel{\textup{Mirror Symmetry}}{\longleftrightarrow} \ \textup{B-model }(Y,\Omega),$$ where $\omega$ is a Ricci flat Kähler form on $X$ and $\Omega$ is a holomorphic volume form (i.e. a nowhere vanishing holomorphic 3-form) on $Y$.
The Hodge diamond of $X$ is of the form $$\begin{array}{ccccccc}
&&& 1 &&& \\
&& 0 && 0 && \\
& 0 && h^{2,2}(X) && 0 \\
1 \quad & & h^{2,1}(X) && h^{1,2}(X) && \quad 1 \\
& 0 && h^{1,1}(X) && 0\\
&& 0 && 0 && \\
&&& 1. &&&
\end{array}$$ The Hodge diamond of $Y$ is of the same form. Let $$\kappa:=h^{1,1}(X)=h^{2,2}(X) = \dim_{{\mathbb{C}}}H^2(X;{\mathbb{C}}) =\dim_{{\mathbb{C}}}{\mathcal{M}}_A,$$ where ${\mathcal{M}}_A$ is the complexified Kähler moduli of $X$. Then $$\kappa= h^{2,1}(Y)=h^{1,2}(Y)=\dim_{\mathbb{C}}H^1(Y,T_Y) =\dim_{\mathbb{C}}{\mathcal{M}}_B,$$ where ${\mathcal{M}}_B$ is the moduli of complex structures on $Y$.
A-model on compact Calabi-Yau threefolds
========================================
A-model topological closed strings: Gromov-Witten invariants
------------------------------------------------------------
Given a non-negative integer $g$ and an effective curve class $\beta\in H_2(X;{\mathbb{Z}})$, let ${\overline{{\mathcal{M}}}}_g(X,\beta)$ be the moduli space of genus $g$ degree $\beta$ stable maps to $X$ (with no marked points). Note that ${\overline{{\mathcal{M}}}}_g(X,\beta)$ is empty if $(g,\beta) = (0,0)$ or $(1,0)$. The moduli ${\overline{{\mathcal{M}}}}_g(X,\beta)$ is a proper Deligne-Mumford stack equipped with a virtual fundamental class $$[{\overline{{\mathcal{M}}}}_g(X,\beta)]^{ {\mathrm{vir}} }\in H_0({\overline{{\mathcal{M}}}}_g(X,\beta);{\mathbb{Q}}).$$ The genus $g$, degree $\beta$ Gromov-Witten invariant of $X$ is defined by $$N_{g,\beta}^X:= \int_{[{\overline{{\mathcal{M}}}}_g(X,\beta)]^{ {\mathrm{vir}} }} 1\in {\mathbb{Q}}$$ when $(g,\beta)\neq (0,0),(1,0)$. The integral sign in the above equation stands for the natural pairing between $[{\overline{{\mathcal{M}}}}_g(X,\beta)]^{ {\mathrm{vir}} }\in H_0({\overline{{\mathcal{M}}}}_g(X,\beta);{\mathbb{Q}})$ and $1\in H^0({\overline{{\mathcal{M}}}}_g(X,\beta);{\mathbb{Q}})$. For $g\geq 2$, we have $${\overline{{\mathcal{M}}}}_g(X,0) ={\overline{{\mathcal{M}}}}_g\times X,$$ and $$N_{g,0}^X = \frac{(-1)^g}{2} \int_{{\overline{{\mathcal{M}}}}_g}\lambda_{g-1}^3\int_X c_3(X)
= \frac{(-1)^g |B_{2g}|\cdot |B_{2g-2}|}{4g(2g-2)\cdot (2g-2)!} \int_X c_3(X),$$ where $B_{2m}$ are Bernoulli numbers.
We choose a basis $H_1,\ldots,H_k$ of $H^{1,1}(X)=H^2(X;{\mathbb{C}})$ such that each $H_i$ is in $H^2(X;{\mathbb{Z}})$ and also in the nef cone (which is the closure of the Kähler cone of $X$). A complexified Kähler class is of the form $$t = \sum_{i=1}^\kappa t^i H_i,$$ where $t^1,\ldots,t^\kappa\in {\mathbb{C}}$ are complexified Kähler parameters. The genus $g$ Gromov-Witten potential of $X$, $F_g^X$, is a generating function of genus $g$ Gromov-Witten invariants of $X$: $$F_g^X(t) := \begin{cases}
\displaystyle{\frac{1}{6}\int_X t^3 + \sum_{\beta\neq 0} N_{0,\beta}^X \exp(\int_\beta t)}, & g=0;\\
\displaystyle{-\frac{1}{24}\int_X c_2(X)t+ \sum_{\beta\neq 0} N_{1,\beta}^X \exp(\int_\beta t) }, & g=1;\\
\displaystyle{ N_{g,0}^X + \sum_{\beta\neq 0} N_{g,\beta}^X \exp(\int_\beta t)}, & g\geq 2.
\end{cases}$$ The sum $\displaystyle{\sum_{\beta\neq 0}}$ is over all non-zero effective classes. We may write $$F_g^X = F_g^{X,\textup{classical}} +F_g^{X,\textup{quantum}},$$ where $$F_g^{X,\textup{quantum}} =\sum_{\beta\neq 0} N_{g,\beta}^X \exp(\int_\beta t)$$ is the contribution from non-constant genus $g$ stable maps to $X$. We have $$\exp(\int_\beta t) = Q_1^{d_1}\cdots Q_\kappa^{d_\kappa},$$ where $Q_i = \exp(t^i)$ and $d_i = \int_\beta H_i\in {\mathbb{Z}}_{\geq 0}$. So $F_g^{X,\mathrm{quantum}} $ is a formal power series in $Q_1,\ldots, Q_\kappa$ with rational coefficients; it tends to zero at the large radius limit $Q_i\to 0$: $$\lim_{Q_i\to 0} F_g^{X,\mathrm{quantum}}(t) =0.$$ We have $$\begin{aligned}
&& \frac{\partial ^3 F_0^X}{\partial t^i\partial t^j\partial t^k} = (H_i\star H_j, H_k)\\
&=& \int_X H_i\cup H_j\cup H_k + \sum_{\beta\neq 0} \Big(\int_\beta H_i \int_\beta H_j \int_\beta H_k \Big)N_{0,\beta}^X \exp(\int_\beta t),\end{aligned}$$ where $\star$ is the quantum product, where $\cup$ is the classical cup product, and where $(\ ,\ )$ is the Poincaré pairing on $H^*(X)$.
A-model topological open strings: open Gromov-Witten invariants
---------------------------------------------------------------
Let $L\subset X$ be a closed oriented Lagrangian submanifold. Then the tangent bundle of $L$ is trivial – recall that the tangent bundle of any orientable 3-manifold is trivial. In this paper, we further assume that $L$ is a rational homology 3-sphere: $$H_*(L;{\mathbb{Q}})= H_*(S^3;{\mathbb{Q}}).$$ Then the map $H_2(X;{\mathbb{Q}})\to H_2(X,L;{\mathbb{Q}})$ is an isomorphism, so $H_2(X;{\mathbb{Z}})\to H_2(X,L;{\mathbb{Z}})$ has finite kernel and cokernel. Since $H_1(L;{\mathbb{Z}})$ is torsion and $H_0(L;{\mathbb{Z}})={\mathbb{Z}}$, we have $H^1(L;{\mathbb{Z}})=0$ by the universal coefficient theorem. In particular, the Maslov class $\mu(L)\in H^1(L;{\mathbb{Z}})$ (which is defined for any Lagrangian submanifold in a Calabi-Yau manifold) is zero. Given a pair $(g,h)$, where $g$ is a nonnegative integer and $h$ is a positive integer, let ${\overline{{\mathcal{M}}}}_{(g,h)}(X,L,\beta)$ be the stable compactification of the moduli ${\mathcal{M}}_{g,h}(X,L,\beta)$, parametrizing holomorphic maps $u: ({ {\Sigma} },\partial { {\Sigma} })\to (X,L)$, where ${ {\Sigma} }$ is a bordered Riemann surface with $g$ handles and $h$ holes, and $u_*[\Sigma]= \beta \in H_2(X,L;{\mathbb{Z}})$, where the domain ${ {\Sigma} }$ is oriented by its complex structure. Then ${\overline{{\mathcal{M}}}}_{g,h}(X,L,\beta)$ is a (usually singular) orbifold whose virtual dimension is zero. In some cases it is possible to define a virtual number $N_{(g,h),\beta}^{X,L}$ of points in ${\overline{{\mathcal{M}}}}_{g,h}(X,L,\beta)$; in general, $N_{(g,h),\beta}^{X,L}$ is a rational number (instead of an integer) due to the existence of orbifold points. We define generating functions of open Gromov-Witten invariants of the pair $(X,L)$ by $$F_{(g,h)}^{X,L}(t) = \sum_{\beta\neq 0} N^{X,L}_{(g,h),\beta} \exp(\int_\beta t),$$ where the sum is over nonzero relative homology classes $\beta\in H_2(X,L;{\mathbb{Z}})$. By assumption, for any $\beta\in H_2(X,L;{\mathbb{Z}})$ there exists a positive integer $r$ such that $r\beta$ lies in the image of $H_2(X;{\mathbb{Z}})\to H_2(X,L;{\mathbb{Z}})$, so $F_{(g,h)}^{X,L}(t)$ is a formal power series in $Q_1^{1/r_1},\ldots, Q_\kappa ^{1/r_\kappa}$ for some positive integers $r_1,\ldots, r_\kappa$, and it tends to zero at the large radius limit: $$\lim_{Q_i\to 0} F_{(g,h)}^{X,L}(t) =0.$$
Let $X$ be a quintic Calabi-Yau threefold with real coefficients, and let $L$ be the real quintic. Then $L$ is a Lagrangian submanifold of $X$, which is diffeomorphic to ${\mathbb{R}}{\mathbb{P}}^3$, so it is orientable and is a rational homology sphere. The group homomorphism $$H_2(X;{\mathbb{Z}})= {\mathbb{Z}}\longrightarrow H_2(X,L;{\mathbb{Z}})={\mathbb{Z}}$$ is injective with cokernel ${\mathbb{Z}}/2{\mathbb{Z}}$. In this case $\kappa=1$, and $$F_g^{X,\textup{quantum}} \in {\mathbb{Q}}[\![ Q ]\!],\quad F_{(g,h)}^{X,L}\in {\mathbb{Q}}[\![Q^{1/2}]\!],$$ where $Q=Q_1$.
Preliminaries on moduli of complex structures
=============================================
The complex moduli and the vacuum line bundle {#sec:complex}
---------------------------------------------
Recall that ${\mathcal{M}}_B$ is the moduli space of complex structures on $Y$, and that $\dim_{\mathbb{C}}{\mathcal{M}}_B =\kappa$. Let $q=(q_1,\ldots,q_\kappa)$ be the local holomorphic coordinates on ${\mathcal{M}}_B$ such that $q = 0$ corresponds to a maximal unipotent monodromy point in the boundary of a (partial) compactification ${\overline{{\mathcal{M}}}}_B$ of ${\mathcal{M}}_B$. (In particular, $q=0$ is in ${\overline{{\mathcal{M}}}}_B\setminus {\mathcal{M}}_B$.) Let ${\mathcal{L}}^\vee \to {\mathcal{M}}_B$ be the complex line bundle over ${\mathcal{M}}_B$ whose fiber over $q\in {\mathcal{M}}_B$ is ${\mathcal{L}}^\vee_q= H^0(Y_q,\Omega^3_{Y_q})\cong {\mathbb{C}}$; then ${\mathcal{L}}^\vee$ is a holomorphic line bundle over ${\mathcal{M}}_B$, and its dual ${\mathcal{L}}$ is the vacuum line bundle in the physics literature such as [@S] and [@BCOV]. The extended moduli space ${\widetilde{{\mathcal{M}}}}_B$ is the total space of the frame bundle of ${\mathcal{L}}^\vee$; it parametrizes pairs $(Y_q,\Omega)$, where $Y_q$ corresponds to a point $q\in {\mathcal{M}}_B$ and $\Omega$ is a nonzero holomorphic 3-form on $Y_q$. So $p: {\widetilde{{\mathcal{M}}}}_B \to {\mathcal{M}}_B$ is a principal ${\mathbb{C}}^*$-bundle, and $\dim_{\mathbb{C}}{\widetilde{{\mathcal{M}}}}_B = \kappa+1$.
The Torelli space
-----------------
Let $H$ denote the rank 2 lattice ${\mathbb{Z}}\oplus {\mathbb{Z}}$ equipped with the symplectic form $\Big(\begin{array}{cc} 0&1\\ -1& 0\end{array}\Big)$. The Torelli space of $Y$ is the moduli of the marked Calabi-Yau threefold $(Y_q,\gamma)$, where $Y_q$ corresponds to a point $q\in {\mathcal{M}}_B$ and the marking $\gamma$ is an isometry from $H^{\oplus (\kappa+1)}$ to $H^3(X_q;{\mathbb{Z}})/\mathrm{Tor}$. Forgetting the marking $\gamma$ defines a covering map $\pi: {\mathcal{T}}\to {\mathcal{M}}_B$, which is a principal $Sp(2\kappa+2;{\mathbb{Z}})$-bundle. Let ${\widetilde{{\mathcal{T}}}}$ be the fiber product: $$\begin{CD}
{\widetilde{{\mathcal{T}}}}@>>> {\widetilde{{\mathcal{M}}}}_B \\
@VVV @VV{p}V\\
{\mathcal{T}}@>{\pi}>> {\mathcal{M}}_B.
\end{CD}$$ Then ${\widetilde{{\mathcal{T}}}}\to {\mathcal{T}}$ is a principal ${\mathbb{C}}^*$-bundle that is the frame bundle of $\pi^*{\mathcal{L}}^\vee$, and ${\widetilde{{\mathcal{T}}}}\to{\widetilde{{\mathcal{M}}}}_B$ is a covering map that is a principal $Sp(2\kappa+2;{\mathbb{Z}})$ bundle.
The Hodge bundle ${\mathbb{H}}_{\mathbb{C}}$ and the Gauss-Manin connection
---------------------------------------------------------------------------
Let ${\mathbb{H}}_{\mathbb{Z}}$ be the local system of lattices on ${\mathcal{M}}_B$ whose fiber over $q\in {\mathcal{M}}_B$ is $H^3(Y_q;{\mathbb{Z}})\cong {\mathbb{Z}}^{2k+2}$. Then ${\mathbb{H}}_{\mathbb{R}}= {\mathbb{H}}_{\mathbb{Z}}\otimes_{{\mathbb{Z}}}{\mathbb{R}}$ (resp. ${\mathbb{H}}_{\mathbb{C}}= {\mathbb{H}}_{\mathbb{Z}}\otimes_{\mathbb{Z}}{\mathbb{C}}$) is a flat real (resp. complex) vector bundle of rank $2\kappa+2$ whose fiber at $q\in {\mathcal{M}}_B$ is $H^3(Y_q;{\mathbb{R}})$ (resp. $H^3(Y_q;{\mathbb{C}})$); the flat connection is known as the Gauss-Manin connection.
More explicitly, let $$\nabla: \Omega^0({\mathcal{M}}_B, {\mathbb{H}}_{\mathbb{C}}) = C^\infty({\mathcal{M}}_B, {\mathbb{H}}_{\mathbb{C}}) \longrightarrow \Omega^1({\mathcal{M}}_B, {\mathbb{H}}_{\mathbb{C}}) =
C^\infty({\mathcal{M}}_B, T^*M_B\otimes {\mathbb{H}}_{\mathbb{C}})$$ be the Gauss-Manin connection. Let $U$ be an open subset on ${\mathcal{M}}_B$ such that ${\mathbb{H}}_{\mathbb{Z}}|_U$ is trivial; we choose a trivialization of ${\mathbb{H}}_{\mathbb{Z}}|_U$, or equivalently, a symplectic basis $\{ \alpha_i,\beta^i: i=0,1,\ldots, \kappa\}$ of $H^3(Y_q;{\mathbb{Z}}) \cong H^{\oplus (\kappa+1)}$ for $q\in U$: $$\int_{Y_q} \alpha_i \cup \alpha_j = \int_{Y_q} \beta^i\cup \beta^j=0,\quad
\int_{Y_q} \alpha_i\cup \beta^j = -\int_{Y_q} \beta^j \cup \alpha_i =\delta_{ij}.$$ Then $\{\alpha_i, \beta^i\}$ is a frame of ${\mathbb{H}}_{\mathbb{C}}|_U$ and $$\nabla \alpha_i = \nabla \beta^i =0,\quad i=0,1,\ldots, \kappa.$$ Any $C^\infty$ section $s \in C^\infty(U,{\mathbb{H}}_{\mathbb{C}})$ is of the form $$\label{eqn:s}
s =\sum_{i=0}^\kappa (a^i \alpha_i + b_i \beta^i),$$ where $a^i, b_i$ are complex-valued $C^\infty$ functions on $U$. Then $$\nabla s = \sum_{i=0}^\kappa (da^i \alpha_i + db_i \beta^i),$$ where $da^i, db_i \in \Omega^1(U,{\mathbb{C}})$ are $C^\infty$ 1-forms on $U$.
We may write $$\nabla = \nabla' + \nabla'',$$ where $\nabla':\Omega^0({\mathcal{M}}_B,{\mathbb{H}}_{\mathbb{C}}) \to \Omega^{1,0}({\mathcal{M}}_B,{\mathbb{H}}_{\mathbb{C}})$ is a $(1,0)$-connection on the $C^\infty$ complex vector bundle ${\mathbb{H}}_{\mathbb{C}}$ and where $\nabla'':\Omega^0({\mathcal{M}}_B,{\mathbb{H}}_{\mathbb{C}})\to \Omega^{0,1}({\mathcal{M}}_B,{\mathbb{H}}_{\mathbb{C}})$ is a $(0,1)$-connection on ${\mathbb{H}}_{\mathbb{C}}$. The $(0,1)$-connection $\nabla''$ defines a holomorphic structure on ${\mathbb{H}}_{\mathbb{C}}$: a section $s \in C^\infty(U,{\mathbb{H}}_{\mathbb{C}})$ is holomorphic iff $\nabla'' s =0$ iff $$s = \sum_{i=0}^\kappa (a^i \alpha_i + b_i \beta^i),$$ where $a^i, b_i$ are holomorphic functions on $U$.
Hodge filtration and the holomorphic polarization
-------------------------------------------------
We have the Hodge filtration $$0\subset {\mathbb{F}}^3\subset {\mathbb{F}}^2 \subset {\mathbb{F}}^1\subset {\mathbb{F}}^0 ={\mathbb{H}}_{\mathbb{C}},$$ where $${\mathbb{F}}^i_q = \bigoplus_{p\geq i} H^{p,3-p}(Y_q;{\mathbb{C}}).$$ The complex vector bundles ${\mathbb{F}}^1, {\mathbb{F}}^2, {\mathbb{F}}^3$ are (non-flat) holomorphic subbundles of ${\mathbb{F}}^0={\mathbb{H}}_{\mathbb{C}}$ of ranks $2\kappa+1$, $\kappa+1$, $1$, respectively. In particular, ${\mathbb{F}}^3={\mathcal{L}}^\vee$ is the dual of the vacuum line bundle. We also have $${\mathbb{H}}_{\mathbb{C}}= {\mathbb{F}}^2 \oplus \overline{{\mathbb{F}}^2},$$ where $${\mathbb{F}}^2_q = H^{3,0}(Y_q;{\mathbb{C}})\oplus H^{2,1}(Y_q;{\mathbb{C}}),\quad \overline{{\mathbb{F}}^2_q} = H^{1,2}(Y_q;{\mathbb{C}}) \oplus H^{0,3}(Y_q;{\mathbb{C}}).$$ For each $q\in U$, $$H^3(Y_q;{\mathbb{C}}) = {\mathbb{F}}^2_q\oplus \overline{{\mathbb{F}}^2_q}$$ is the holomorphic polarization of the complex symplectic space $H^3(Y_q;{\mathbb{C}})$.
Real polarization
-----------------
We choose a symplectic basis $\{ A^i, B_i: i=0,1,\ldots,\kappa\}$ of $(H_3(Y;{\mathbb{Z}}),\cap)$, where $\cap$ is the intersection form, and let $\{\alpha_i, \beta^i: i=0,1,\ldots, \kappa\}$ be the dual symplectic basis of $(H^3(Y;{\mathbb{Z}}), Q)$, where $$Q(\alpha, \beta) = \int_Y \alpha\cup \beta \in {\mathbb{Z}}\quad \textup{ for } \alpha,\beta \in H^3(Y;{\mathbb{Z}}).$$ We have $$\delta^i_j = A^i\cap B_j = - B_j \cap A^i = \int_Y \alpha_i \cup \beta^j= -\int_Y \beta ^j\cup \alpha_i = \int_{A^i} \alpha_j =
\int_{B_j}\beta^i.$$ $$0 =A^i\cap A^j = B_i \cup B_j = \int_Y \alpha_i\cup \alpha_j =\int_Y \beta^i \cup \beta^j =\int_{A^i}\beta^j = \int_{B_j} \alpha_i.$$ For any $\phi \in H^3(X;{\mathbb{R}})$ we have $$\phi =\sum_{i=0}^\kappa x^i \alpha_i +\sum_{i=0}^\kappa p_i \beta^i,$$ where $$x^i=\int_{A^i}\phi,\quad p_i =\int_{B_i}\phi.$$ Then $\{ x^i, p_i: i=0,1,\ldots, k\}$ are Darboux coordinates of the real linear symplectic space $(H^3(Y;{\mathbb{R}}),Q)$. The linear symplectic form $$\sum_{i=0}^k dx^i\wedge dp_i$$ on $H^3(Y;{\mathbb{R}})$ is independent of the choice of a symplectic basis. The integral symplectic basis $\{\alpha_i ,\beta^i : i=0,1,\ldots,\kappa\}$ extends to a flat frame of ${\mathbb{H}}_{\mathbb{R}}$ on an open neighborhood $U$ of $q_0$ in ${\mathcal{M}}_B$, and $\{ x^i, p_i : i=0,1,\ldots, \kappa\}$ are flat fiber coordinates of ${\mathbb{H}}_{\mathbb{R}}|_U \cong U\times {\mathbb{R}}^{2\kappa +2}$.
The Gauss-Manin connection is compatible with the symplectic structure $Q$ on ${\mathbb{H}}_{\mathbb{C}}$: given two $C^\infty$ sections $s_1, s_2$ of ${\mathbb{H}}_{\mathbb{C}}$, and a $C^\infty$ vector field $\xi$ on ${\mathcal{M}}_B$, we have $$L_\xi \left( Q(s_1,s_2) \right) = Q(\nabla_\xi s_1, s_2) + Q(s_1, \nabla_\xi s_2),$$ where $$L_\xi : C^\infty({\mathcal{M}}_B;{\mathbb{C}})\to C^\infty({\mathcal{M}}_B;{\mathbb{C}})$$ is the Lie derivative on $C^\infty$ functions, and where $$\nabla_\xi: C^\infty({\mathcal{M}}_B,{\mathbb{H}}_{\mathbb{C}})\to C^\infty({\mathcal{M}}_B, {\mathbb{H}}_{\mathbb{C}})$$ is the covariant derivative defined by the Gauss-Manin connection.
Special homogeneous coordinates and the period map
--------------------------------------------------
Let $s:{\widetilde{{\mathcal{M}}}}_B\to p^*{\mathcal{L}}^\vee$ be the tautological section. Write $$s(\xi) = \sum_{i=0}^k (x^i(\xi) \alpha_i + p_i(\xi) \beta^i).$$ Then $x^0(\xi),\ldots, x^\kappa(\xi)$ are the local holomorphic coordinates on the extended complex moduli ${\widetilde{{\mathcal{M}}}}_B$ and the local homogeneous coordinates on the complex moduli ${\mathcal{M}}_B$; they are called “special homogeneous coordinates” in [@S].
Let $V = \left(H^{\oplus (\kappa+1)}\right)\otimes_{\mathbb{Z}}{\mathbb{C}}$, which is a complex symplectic vector space of dimension $2(\kappa+1)$. The isometry $\gamma: H^{\oplus(\kappa+1)} \to H^3(X_q;{\mathbb{Z}})/\mathrm{Tor}$ extends to an isomorphism $\gamma: V\to H^3(X_q;{\mathbb{C}})$ of complex symplectic vector spaces. There is a period map $$P_{\mathcal{T}}: {\mathcal{T}}\to {\mathbb{P}}(V)\cong {\mathbb{P}}^{2\kappa+1}, \quad (X_q,\gamma) \mapsto \gamma^{-1}(H^{3,0}(X_q)).$$ More explicitly, $$P_{\mathcal{T}}(q)= [x^0(\xi), x^1(\xi),\ldots, x^\kappa(\xi), p_0(\xi), p_1(\xi),\ldots, p_{\kappa}(\xi)],$$ where $\xi \in {\widetilde{{\mathcal{T}}}}$ is any point in the fiber of ${\widetilde{{\mathcal{T}}}}\to {\mathcal{T}}$ over $q\in {\mathcal{T}}$.
Let ${\mathcal{O}}_{{\mathbb{P}}(V)}(-1)$ be the tautological line bundle over ${\mathbb{P}}(V)$. Then $$P_{\mathcal{T}}^* {\mathcal{O}}_{{\mathbb{P}}(V)}(-1) = \pi^*{\mathcal{L}}^\vee.$$
Recall the Euler sequence: $$0\to {\mathcal{O}}_{{\mathbb{P}}(V)} \to V\otimes {\mathcal{O}}_{{\mathbb{P}}(V)}(1) \to T_{{\mathbb{P}}(V)} \to 0.$$ Pulling back the above sequence under $P_{\mathcal{T}}$, we obtain $$0\to {\mathcal{O}}_{\mathcal{T}}\to V \otimes \pi^*{\mathcal{L}}\to P_{\mathcal{T}}^*T_{{\mathbb{P}}(V)} \to 0.$$ Here, $\pi: {\mathcal{T}}\to {\mathcal{M}}_B$ is the projection from the Torelli space to the complex moduli as before, and $$V\otimes \pi^*{\mathcal{L}}= \pi^*({\mathbb{H}}_{\mathbb{C}}\otimes {\mathcal{L}}).$$
The Hodge metric
----------------
The symplectic form $Q$ on $H^3(Y;{\mathbb{R}})$ extends to $V= H^3(Y;{\mathbb{C}})$. For $\alpha,\beta\in H^3(Y,{\mathbb{C}})$, define $$(\alpha,\beta):= \sqrt{-1}Q(\alpha,\beta).$$
By the Hodge-Riemann bilinear relation,
- $(\alpha,\alpha) =0$ if $\alpha\in H^{3,0}(Y)\oplus H^{2,1}(Y)$;
- $(\alpha,\bar{\alpha})>0$ if $\alpha\in H^{3,0}(Y)$ and $\alpha\neq 0$;
- $(\alpha,\bar{\alpha})<0$ if $\alpha\in H^{2,1}(Y)$ and $\alpha\neq 0$.
Define a Hermitian metric on the holomorphic line bundle ${\mathcal{L}}^\vee={\mathbb{F}}^3$ by $$\parallel \Omega \parallel^2 = (\Omega,\bar{\Omega}).$$ This is known as the Hodge metric on ${\mathbb{F}}^3$.
Let ${\mathbb{H}}^{2,1}$ be the complex vector bundle over ${\mathcal{M}}_B$ whose fiber over $q\in {\mathcal{M}}_B$ is $H^{2,1}(Y_q)$. Then ${\mathbb{H}}^{2,1}$ is a $C^\infty$ complex subbundle of ${\mathbb{H}}_{\mathbb{C}}$ but not a holomorphic subbundle of ${\mathbb{H}}_{\mathbb{C}}$. Define the Hodge metric $h$ on ${\mathbb{H}}^{2,1}$ by $$h(s_1,s_2) = -(s_1,\bar{s}_2).$$ Then $({\mathbb{H}}^{2,1}, h)$ is a Hermitian vector bundle of rank $\kappa$ over ${\mathcal{M}}_B$.
The Chern connection on ${\mathcal{L}}^\vee = {\mathbb{F}}^3$ {#sec:chern}
-------------------------------------------------------------
Let $$D:\Omega^0({\mathcal{M}}_B, {\mathcal{L}}^\vee) \to \Omega^1({\mathcal{M}}_B, {\mathcal{L}}^\vee) = \Omega^{1,0}({\mathcal{M}}_B,{\mathcal{L}}^\vee) \oplus \Omega^{0,1}({\mathcal{M}}_B,{\mathcal{L}}^\vee)$$ be the Chern connection determined by the holomorphic structure and the Hodge metric on ${\mathcal{L}}^\vee={\mathbb{F}}^3$, and let $D'$ and $D''$ be the $(1,0)$ and $(0,1)$ parts of $D$, so that $$D = D' + D'',$$ where $D''$ depends on the holomorphic structure. Any $C^\infty$ section $s$ of ${\mathcal{L}}^\vee$ is also a $C^\infty$ section of ${\mathbb{H}}_{\mathbb{C}}$, and we have $$D''s = \nabla''s$$ since ${\mathcal{L}}^\vee$ is a holomorphic subbundle of ${\mathbb{H}}_{\mathbb{C}}$.
If $(Y_q,\Omega_q)$ is a local holomorphic frame of ${\mathcal{L}}^\vee$ over an open neighborhood $U$ then $D''\Omega =0$, and for any tangent vector $v\in T_q {\mathcal{M}}_B$, where $q\in U$, we have $$\label{eqn:Dv}
D'_v \Omega = D_v \Omega = \frac{(\nabla_v \Omega, \bar{\Omega})}{(\Omega, \bar{\Omega})} \Omega.$$ The connection 1-form is $$A= \partial \log(\Omega,\bar{\Omega}) \in \Omega^{1,0}(U).$$ The curvature form 2-form is $$F = d A = -\partial \bar{\partial} \log \left( \parallel \Omega \parallel^2\right) \in \Omega^{1,1}(U).$$ The right-hand side is independent of the choice of the local holomorphic frame, so $F$ is a global $(1,1)$ form on ${\mathcal{M}}_B$. More explicitly, we define $${\mathbb{D}}: \Omega^0({\mathcal{M}}_B,{\mathcal{L}}^\vee)\to \Omega^{1,0}({\mathcal{M}}_B, {\mathbb{H}}^{2,1})$$ by $${\mathbb{D}}s = \nabla' s - D's = \nabla s - D s.$$ Write $$F = F_{i\bar{j}} dz^i \wedge d\bar{z}^j,$$ where $(z^1,\ldots, z^\kappa)$ are the local holomorphic coordinates on $U\subset {\mathcal{M}}_B$. Define $$\nabla_i:= \nabla_{\frac{\partial}{\partial z^i}}, \quad D_i:= D_{\frac{\partial}{\partial z^i}}, \quad
{\mathbb{D}}_i = \nabla_i - D_i.$$ Then $$\begin{aligned}
F_{i\bar j} &=& -\frac{\partial^2}{\partial z^i \partial \bar{z}^j} \log(\Omega,\bar{\Omega})\\
&=& -\frac{\partial}{\partial z^i} \frac{(\Omega, \overline{\nabla_j\Omega})}{(\Omega,\bar{\Omega})} \\
&=& -\frac{(\nabla_i \Omega, \overline{\nabla_j\Omega})}{(\Omega,\bar{\Omega})} +
\frac{ (\nabla_i\Omega, \bar{\Omega}) (\Omega, \overline{\nabla_j\Omega}) }{(\Omega,\bar{\Omega})^2} \\
&=& \frac{-(\nabla_i\Omega,\overline{\nabla_j\Omega} ) + (D_i\Omega, \overline{D_j\Omega} )}{(\Omega,\bar{\Omega})} \\
&=& \frac{ -({\mathbb{D}}_i \Omega,\overline{{\mathbb{D}}_j\Omega}) }{ (\Omega,\bar{\Omega}) }\end{aligned}$$ where the fourth equality follows from Equation , and the fifth (and last) equality follows from the identity $\nabla_k \Omega = D_k \Omega + {\mathbb{D}}_k\Omega$ where $D_k\Omega$ is a (3,0)-form and ${\mathbb{D}}_k \Omega$ is a (2,1)-form.
The above computation shows that the first Chern form $$c_1({\mathcal{L}}^\vee, D) = \frac{\sqrt{-1}}{2\pi} F$$ is a positive $(1,1)$-form.
The Weil-Petersson metric
-------------------------
The Weil-Petersson metric on ${\mathcal{M}}_B$ is defined by the Kähler form $$\omega_{\mathrm{WP}} := 2\pi c_1({\mathcal{L}}^\vee,D) = \sqrt{-1} F.$$ In local holomorphic coordinates, the Weil-Petersson metric is given by $$G_{i\bar{j}} := \langle \frac{\partial}{\partial z_i}, \frac{\partial}{\partial z_j} \rangle
= \frac{-({\mathbb{D}}_i\Omega,\overline{{\mathbb{D}}_j\Omega})}{(\Omega,\bar{\Omega})} = \frac{g_{i\bar{j}}}{g_{0\bar{0}}},$$ where $g_{i\bar{j}}$ is the Hodge metric on ${\mathbb{H}}^{2,1}$ and $g_{0\bar{0}}$ is the Hodge metric on ${\mathcal{L}}^\vee={\mathbb{F}}^3$.
We have an isomorphism of Hermitian vector bundles: $$\label{eqn:tangent-hodge}
T{\mathcal{M}}_B \otimes {\mathcal{L}}^\vee \cong {\mathbb{H}}^{2,1},$$ where the tangent bundle $T{\mathcal{M}}_B$ is equipped with the Weil-Petersson metric $G_{i\bar{j}}$, while ${\mathbb{H}}^{2,1}$ is equipped with the Hodge metric $g_{i\bar{j}}$, and ${\mathcal{L}}^\vee$ is equipped with the Hodge metric $g_{0\bar{0}}$. The restriction of to a point $q\in {\mathcal{M}}_B$ can be identified with $$H^1(Y_q,T_Y)\otimes H^0(Y_q,\Omega^3_{Y_q}) \cong H^1(Y_q,\Omega^2_{Y_q}).$$
B-model on compact Calabi-Yau threefold
=======================================
Genus zero free energy
-----------------------
$$x^0, x^1,\ldots, x^\kappa, p_0, p_1,\ldots, p_\kappa \in \Gamma({\mathcal{T}}, p^*{\mathcal{L}})$$ are holomorphic sections of the line bundle $$p^*{\mathcal{L}}= {\mathcal{P}}_{{\mathcal{T}}}^*{\mathcal{O}}_{{\mathbb{P}}(V)}(1)$$ over the Torelli space ${\mathcal{T}}$. Note that $${\mathcal{F}}_0 := \frac{1}{2} \sum_{i=0}^\kappa x^i p_i$$ is a holomorphic section on $p^*{\mathcal{L}}^2 \to {\mathcal{T}}$ and is a multi-valued holomorphic section of ${\mathcal{L}}^2\to {\mathcal{M}}_B$. Define the local holomorphic functions $$\label{eqn:Fi}
z^i:= \frac{x^i}{x^0},\quad F_i:= \frac{p_i}{x^0},\quad g := \frac{p_0}{x^0},\quad \check{F}_0 = (x^0)^{-2} {\mathcal{F}}_0,$$ where $i=1,\ldots,\kappa$. In particular, $$g = 2\check{F}_0 -\sum_{i=1}^\kappa z^i F_i.$$ The functions $(z^1,\ldots, z^\kappa)$ are known as special coordinates (or the B-model flat coordinates), which are local holomorphic coordinates on ${\mathcal{T}}$, defined in an open neighborhood of large complex structure with $A^0$ a vanishing cycle. The function $\check{F}_0$ is the B-model genus zero free energy.
We may write $\Omega = x^0 \Omega_0$, where $$\Omega_0 = \alpha_0 + \sum_{j=1}^\kappa z^j \alpha_j + \sum_{j=1}^\kappa F_j \beta^j
+ (2\check{F}_0 - \sum_{j=1}^\kappa z^j F_j)\beta^0.$$ For $i=1,\ldots,\kappa$, $$\nabla_i \Omega_0 = \alpha_i + \sum_{j=1}^\kappa \frac{\partial F_j}{\partial z^i}\beta^j
+ \left(2\frac{\partial \check{F}_0}{\partial z^i} - F_i - \sum z^j\frac{\partial F_j}{\partial z^i} \right)\beta^0.$$ Here, $\Omega_0$ is a local holomorphic section of ${\mathbb{F}}^3$, and $\nabla_i \Omega_0$ is a local holomorphic section of ${\mathbb{F}}^2$ by Griffiths transversality, so Hodge-Riemann bilinear relations imply $$0 = Q(\Omega_0, \nabla_i \Omega_0) = 2\frac{\partial \check{F}_0}{\partial z^i} - 2 F_i.$$
To summarize:
- The B-model flat coordinates are given by $$z^i = \frac{x^i}{x^0} =\frac{\int_{A^i}\Omega}{\int_{A^0}\Omega},\quad i=1,\ldots, \kappa.$$
- The B-model genus zero free energy $\check{F}_0$ is defined by $$\check{F}_0 =\frac{1}{2} (x^0)^{-2} \cdot \sum_{i=0}^{\kappa} x^ip_i,$$ which satisfies $$\frac{\partial \check{F}_0}{\partial z^i} = \frac{p_i}{x^0}=
\frac{\int_{B_i}\Omega}{\int_{A^0} \Omega},\quad i=1,\ldots, \kappa.$$
Yukawa coupling
---------------
$$\nabla_i \Omega_0 = \alpha_i + \sum_{j=1}^\kappa \frac{\partial^2 \check{F}_0}{\partial z^i\partial z^j} \beta^j \mod \beta^0.$$ $$\nabla_j \nabla_k \Omega_0 = \sum_{j,k,\ell=1}^\kappa
\frac{\partial^3 \check{F}_0}{\partial z^j\partial z^k \partial z^\ell} \beta^\ell
\mod \beta^0.$$ $$Q(\nabla_i \Omega_0, \nabla_j \nabla_k \Omega_0) =
\frac{\partial^3 \check{F}_0}{\partial z^i \partial z^j\partial z^k}.$$ We also have $$Q(\nabla_i \Omega_0, \nabla_j \nabla_k \Omega_0) = \frac{\partial}{\partial z^i} Q(\Omega_0, \nabla_j\nabla_k \Omega_0)
-Q(\Omega_0, \nabla_i \nabla_j \nabla_k \Omega_0) = -Q(\Omega_0, \nabla_i \nabla_j \nabla_k \Omega_0).$$ So $$\frac{\partial^3 \check{F}_0}{\partial z^i \partial z^j\partial z^k} = -Q(\Omega_0, \nabla_i \nabla_j \nabla_k \Omega_0)
=-\int_Y \Omega_0 \wedge \nabla_i \nabla_j \nabla_k \Omega_0.$$ Define $C\in \Gamma((T^*M)^3\otimes {\mathcal{L}}^2 )$ by $$\begin{aligned}
C_{ijk} = C(\frac{\partial}{\partial z^i}, \frac{\partial}{\partial z^j},
\frac{\partial}{\partial z^k}) &:=& -\int_Y \Omega\wedge \nabla_i \nabla_j \nabla_k \Omega\\
&=& -(x^0)^2 \int_Y \Omega_0 \wedge \nabla_i \nabla_j \nabla_k \Omega_0
= (x^0)^2 \frac{\partial^3 \check{F}_0}{\partial z^i \partial z^j\partial z^k}.\end{aligned}$$ Then $C_{ijk}$ is symmetric in $i,j,k$, so $$C \in \Gamma({ \mathrm{Sym} }^3(T^*M)\otimes {\mathcal{L}}^2 ).$$
Define $$C_{\bar{i}}^{jk} := \overline{C_{ilm}}g^{j\bar{l}} g^{k\bar{m}}.$$ Then $$\label{eqn:C}
{\mathcal{C}}:= C_{\bar{i}}^{jk}d\bar{z}^i \otimes \frac{\partial}{\partial z_j} \otimes
\frac{\partial}{\partial z_k} \in \Omega^{0,1}({\mathcal{M}}_B, (T{\mathcal{M}}_B)^{\otimes 2} \otimes {\mathcal{L}}^{-2}).$$
Genus one free energy
---------------------
The genus one free energy $\check{F}_1$ is a linear combination of Ray-Singer torsions. A mathematical definition of $\check{F}_1$ is given in [@FLY].
Genus $g\geq 2$ free energies and the Holomorphic Anomaly Equations
-------------------------------------------------------------------
The Weil-Petersson metric is Kähler, so the Chern connection on $T{\mathcal{M}}_B$ defined by the holomorphic structure and the Weil-Petersson metric is also torsion free. We equip ${\mathcal{L}}^\vee={\mathbb{F}}^3$ with the connection $D$ in Section \[sec:chern\]. These two connections induce connections on tensor bundles $$(T^*{\mathcal{M}}_B)^{\otimes m} \otimes {\mathcal{L}}^k$$ for any integers $m,k$. Let $D$ be the covariant derivative defined by these connections.
The special homogeneous coordinates $x^0,x^1,\ldots, x^{\kappa}$ are local holomorphic sections of the vacuum line bundle ${\mathcal{L}}$. For $g\geq 2$, $${\mathcal{F}}_g(z,\bar{z}) \in \Gamma({\mathcal{M}}_B, {\mathcal{L}}^{2-2g}).$$ The limit $$\check{F}_g(z)=\lim_{\bar{z}\to \sqrt{-1}\infty} (x^0)^{2g-2}{\mathcal{F}}_g(z,\bar{z})$$ is a holomorphic function on $U$. The non-holomorphic section ${\mathcal{F}}_g(z, \bar{z})$ satisfies the following Holomorphic Anomaly Equation (BCOV [@BCOV]): $$\label{eqn:BCOV}
{\bar{\partial}}_i {\mathcal{F}}_g = \frac{1}{2} C_{\bar{i}}^{jk} \Big(D_j D_k {\mathcal{F}}_{g-1} + \sum_{\substack{g_1,g_2>0\\ g_1+g_2=g} } D_j {\mathcal{F}}_{g_1}D_k {\mathcal{F}}_{g_2} \Big).$$
More precisely, we have $$DD{\mathcal{F}}_{g-1} + \sum_{\substack{g_1,g_2>0\\ g_1+g_2=g} } D{\mathcal{F}}_{g_1} \otimes D{\mathcal{F}}_{g_2} \in \Gamma( {\mathcal{M}}_B, (T^*{\mathcal{M}}_B)^{\otimes 2}\otimes {\mathcal{L}}^{4-2g}),$$ $${\mathcal{C}}\in \Omega^{0,1}({\mathcal{M}}_B, (T{\mathcal{M}}_B)^{\otimes 2} \otimes {\mathcal{L}}^{-2}),$$ where ${\mathcal{C}}$ is defined by Equation . Using the natural pairing between $T{\mathcal{M}}_B$ and $T^*{\mathcal{M}}_B$, we obtain $${\mathcal{C}}\diamond \Big( DD{\mathcal{F}}_{g-1} + \sum_{\substack{g_1,g_2>0\\ g_1+g_2=g} } D{\mathcal{F}}_{g_1} \otimes D{\mathcal{F}}_{g_2}\Big) \in \Omega^{0,1}({\mathcal{M}}_B, {\mathcal{L}}^{2-2g}).$$ With the above notation, the Holomorphic Anomaly Equation can be rewritten in the following coordinate-free form: $$\label{eqn:BCOV-tensor}
D''{\mathcal{F}}_g = \frac{1}{2} {\mathcal{C}}\diamond \Big( DD{\mathcal{F}}_{g-1} + \sum_{\substack{g_1,g_2>0\\ g_1+g_2=g} } D{\mathcal{F}}_{g_1} \otimes D{\mathcal{F}}_{g_2}\Big).$$
B-model topological open string and the Extended Holomorphic Anomaly Equation
-----------------------------------------------------------------------------
Let $$\Delta_{ij} := D_jD_i {\mathcal{F}}_{(0,1)}$$ be the second covariant derivatives of the disk potential, and define $$\Delta_{\bar{i}}^j := g^{j\bar{k}} \overline{\Delta_{ki}}.$$ Then $$\Delta_{\bar{i}}^j d\bar{z}^i \otimes \frac{\partial}{\partial z_j} \in \Omega^{0,1}({\mathcal{M}}_B, T{\mathcal{M}}_B \otimes {\mathcal{L}}^{-1}).$$ If $h>0$ and $2g-2+h>0$, then $${\mathcal{F}}_{(g,h)}(z,\bar{z}) \in \Gamma({\mathcal{L}}^{2-2g-h}).$$ The limit $$\check{F}_{(g,h)}(z)=\lim_{\bar{z}\to \infty} (x^0)^{2g-2+h}{\mathcal{F}}_{(g,h)}(z,\bar{z})$$ is a holomorphic function on $U$. The non-holomorphic section ${\mathcal{F}}_{(g,h)}(z,\bar{z})$ satisfies the following Extended Holomorphic Anomaly Equation (Walcher [@W]) $$\label{eqn:Walcher}
\bar{\partial}_i {\mathcal{F}}_{(g,h)} = \frac{1}{2} C_{\bar{i}}^{jk} \Big(D_j D_k {\mathcal{F}}_{g-1,h} +
\sum'_{\substack{g_1+ g_2 =g\\ h_1+h_2=h} } D_j {\mathcal{F}}_{g_1,h_1}D_k {\mathcal{F}}_{g_2,h_2} \Big) -\Delta_{\bar{i}}^j D_j F_{g,h-1},$$ where the sum $\displaystyle{\sum'}$ excludes the unstable case $(g_i,h_i)=(0,0), (0,1)$, and the last term on the RHS corresponds to the cases $(g_1,h_1)=(0,1)$ or $(g_2,h_2)=(0,1)$.
Mirror symmetry
---------------
Under the mirror map $$(t^1,\ldots, t^\kappa) = (z^1(q),\ldots, z^\kappa(q))$$ we have the following mirror conjectures:
- For any $i,j,k\in \{1,\ldots, \kappa\}$, $$\frac{\partial^3\check{F}_0}{\partial z^i \partial z^j \partial z^k}(z)
=\frac{\partial^3 F_0^X}{\partial t^i \partial t^j \partial t^k}(t).$$
- For $i=1,\ldots, \kappa$, $$\frac{\partial \check{F}_1}{\partial z^i}(z) = \frac{\partial F_1^X}{\partial t^i}(t),$$ or equivalently, $$d\check{F}_1 = d F_1^X.$$
- For $g\geq 2$, $$\check{F}_g(z) = F_g^X(t).$$
- If $h>0$ and $2g-2+h>0$ then $$\check{F}_{(g,h)}(z) = F_{(g,h)}^{X,L}(t).$$
Acknowledgments {#acknowledgments .unnumbered}
---------------
This note is based on the author’s talks at the 6th Workshop on Combinatorics of Moduli Spaces, Cluster Algebras, and Topological Recursion in Moscow on June 4–9, 2018, and the conference on Crossing the Walls in Enumerative Geometry in Snowbird, Utah on May 21–June 1, 2018. The author sincerely thanks the organizers of these events for the invitation to participate as a speaker. The author also wishes to thank Bohan Fang, Sheldon Katz, Zhengyu Zong, and the anonymous referee for their helpful comments on an earlier version of this note.
[AA]{}
M. Aganagic, V. Bouchard, A. Klemm, “Topological strings and (almost) modular forms,” Comm. Math. Phys. [**277**]{} (2008), no. 3, 771–819.
M. Alim and J. D. Länge, “Polynomial Structure of the (Open) Topological String Partition Function,” JEHP [**0710**]{} (2007), no. 045.
M. Alim, E. Scheidegger, S.-T. Yau, J. Zhou, “Special polynomial rings, quasi modular forms and duality of topological strings,” Adv. Theor. Math. Phys. [**18**]{} (2014), 401–467.
M. Bershadsky, S. Cecotti, H. Ooguri, C. Vafa, “Kodaira-Spencer theory of gravity and exact results for quantum string amplitudes,” Comm. Math. Phys. [**165**]{} (1994), no. 2, 311–428.
G. Bonelli and A. Tanzini, “The holomorphic anomaly for open string moduli,” J. High Energy Phys. 2007, no. 10, 060, 17 pp.
D. Cox and S. Katz, “Mirror symmetry and algebraic geometry,” Mathematical Surveys and Monographs, [**68**]{}. American Mathematical Society, Providence, RI, 1999. xxii+469 pp.
H. Fang, Z. Lu, “Generalized Hodge metrics and BCOV torsion on Calabi-Yau moduli,” J. Reine Angew. Math. [**588**]{} (2005), 49–69.
H. Fang, Z. Lu, K.I. Yoshikawa, “Analytic torsion for Calabi-Yau threefolds,” J. Differential Geom. [**80**]{} (2008), no. 2, 175–259.
K. Hori, S. Katz, A. Klemm, R. Pandharipande, R. Thomas, C. Vafa, R. Vakil, E. Zaslow, [*Mirror symmetry*]{}. With a preface by Vafa. Clay Mathematics Monographs, [**1**]{}. American Mathematical Society, Providence, RI; Clay Mathematics Institute, Cambridge, MA, 2003. xx+929 pp.
Y. Konishi and S. Minabe, “On solutions to Walcher’s extended holomorphic anomaly equation,” Commun. Number Theory Phys. [**1**]{} (2007), no. 3, 579–603.
A. Kanazawa and J. Zhou, “Lectures on BCOV Holomorphic Anomaly Equations,” Calabi-Yau varieties: arithmetic, geometry and physics, 445–-473, Fields Inst. Monogr., [**34**]{}, Fields Inst. Res. Math. Sci., Toronto, ON, 2015.
R. Pandharipande, “Three questions in Gromov-Witten theory,” Proceedings of the International Congress of Mathematicians, Vol. II (Beijing, 2002), 503–512, Higher Ed. Press, Beijing, 2002.
A. Strominger, “Special geometry," Comm. Math. Phys. [**133**]{} (1990), no. 1, 163–180.
G. Tian, “Smoothness of the universal deformation space of compact Calabi-Yau manifolds and its Petersson-Weil metric,” [*Mathematical aspects of string theory*]{} (San Diego, Calif., 1986), 629–646, Adv. Ser. Math. Phys., [**1**]{}, World Sci. Publishing, Singapore, 1987.
J. Walcher, “Extended holomorphic anomaly and loop amplitudes in open topological string,” Nuclear Phys. B [**817**]{} (2009), no. 3, 167–207.
S. Yamaguchi and S.-T. Yau, “Topological string partition functions as polynomials,” JEHP [**0407**]{} (2004), no. 047.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
B. Ahmmed$^{1,2,*}$, M. K. Mudunuru$^1$, S. Karra$^1$, S. C. James$^{2,3}$, and V. V. Vesselinov$^1$\
[$^1$Computational Earth Science Group, Earth and Environmental Sciences Division, Los Alamos National Laboratory, Los Alamos, NM 87545]{}\
$^2$Department of Geosciences, Baylor University, Waco, TX 76706\
$^3$Departments of Geosciences and Mechanical Engineering, Baylor University, Waco, TX 76706\
bibliography:
- 'Master\_References/References.bib'
title: A Comparative Study of Machine Learning Models for Predicting the State of Reactive Mixing
---
Abstract {#abstract .unnumbered}
========
Mixing phenomena are important mechanisms controlling flow, species transport, and reaction processes in fluids and porous media. Accurate predictions of reactive mixing are critical for many Earth and environmental science problems such as contaminant fate and remediation, macroalgae growth, and plankton biomass evolution. To investigate mixing dynamics over time under different scenarios (e.g., anisotropy, fluctuating velocity fields), a high-fidelity, finite-element-based numerical model is built to solve the fast, irreversible bimolecular reaction-diffusion equations to simulate a range of reactive-mixing scenarios. A total of 2,315 simulations are performed using different sets of model input parameters comprising various spatial scales of vortex structures in the velocity field, time-scales associated with velocity oscillations, the perturbation parameter for the vortex-based velocity, anisotropic dispersion contrast (i.e., ratio of longitudinal-to-transverse dispersion), and molecular diffusion. Outputs comprise concentration profiles of the reactants and products. The inputs and outputs of these simulations are concatenated into feature and label matrices, respectively, to train 20 different machine learning (ML) emulators to approximate system behavior. The 20 ML emulators based on linear methods, Bayesian methods, ensemble learning methods, and multilayer perceptron (MLP), are compared to assess these models. The ML emulators are specifically trained to classify the state of mixing and predict three quantities of interest (QoIs) characterizing species production, decay (i.e., average concentration, square of average concentration), and degree of mixing (i.e., variances of species concentration). Linear classifiers and regressors fail to reproduce the QoIs; however, ensemble methods (classifiers and regressors) and the MLP accurately classify the state of reactive mixing and the QoIs. Among ensemble methods, random forest and decision-tree-based AdaBoost faithfully predict the QoIs. At run time, trained ML emulators are $\approx10^5$ times faster than the high-fidelity numerical simulations. Speed and accuracy of the ensemble and MLP models facilitate uncertainty quantification, which usually requires 1,000s of model run, to estimate the uncertainty bounds on the QoIs. **Keywords:** Surrogate modeling, machine learning, reaction-diffusion equations, random forests, ensemble methods, artificial neural networks.
Acknowledgments {#acknowledgments .unnumbered}
===============
Bulbul Ahmmed thanks the support from Mickey Leland Energy Fellowship (MLEF) awarded by U.S. Department of Energy’s (DOE) Office of Fossil Energy (FE). MKM and SK also thank the support of the LANL Laboratory Directed Research and Development (LDRD) Early Career Award 20150693ECR. VVV thanks the support of LANL LDRD-DR Grant 20190020DR. Los Alamos National Laboratory is operated by Triad National Security, LLC, for the National Nuclear Security Administration of U.S. Department of Energy (Contract No. 89233218CNA000001). Additional information regarding the simulation datasets and codes can be obtained from Bulbul Ahmmed (Email: `[email protected]`) and Maruti Mudunuru (Email: `[email protected]`).
Conflict of Interest {#conflict-of-interest .unnumbered}
====================
The authors declare that they do not have conflict of interest.
Computer Code Availability {#computer-code-availability .unnumbered}
==========================
Codes for machine learning implementation are available in the public Github repository <https://github.com/bulbulahmmed/ML-to-reactive-mixing-data>. Additional information regarding the simulation datasets can be obtained from Bulbul Ahmmed (Email: `[email protected]`) and Maruti Kumar Mudunuru (Email: `[email protected]`).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'R.G. Paremuzyan'
title: Timelike Compton Scattering at JLab
---
Timelike Compton Scattering at JLab\
R.G. Paremuzyan$^{1, 2}$
[ 1. Institut de Physique Nucléaire, 15 rue Georges Clémenceau, 91406 - ORSAY Cedex - FRANCE\
2. Yerevan Physics Institute, 2 Alikhanyan Brothers St.,Yerevan 0036, Armenia. ]{} 0.5cm
It is demonstrated, that with exclusive final state, data from electron scattering experiments that are recorded with loose trigger requirements can be used to analyze photoproduction reactions. A preliminary results on Timelike Compton Scattering using the electroproduction data from the CLAS detector at Jefferson Lab are presented. In particular, using final state ($pe^-e^+$) photoproduction of vector mesons and timelike photon is studied. Angular asymmetries in Timelike Compton Scattering region is compared with model predictions in the framework of Generalized Parton Distribution.
Introduction
============
In the description of the nucleon structure an important roll plays a formalism of Generalized Parton Distributions (GPD)s [@GPDs]. GPDs provide 3 dimensional description of the quark-gluon structure of the nucleon. QCD factorization theorem for exclusive processes [@Factorize] allows to access GPDs through exclusive processes in a certain kinematic domain ($Q^{2}>>$, $t/Q^{2} << 1$, $s > 4\;GeV^{2}$). Theoretically and experimentally best studied reaction in GPD framework is Deeply Virtual Compton Scattering (DVCS) i.e. $\gamma^{*}p\rightarrow \gamma p$. where incoming photon has large spacelike virtuality, whereas the outgoing photon is on shell. GPDs enter into Real part of Compton Form Factors (CFFs) as integral over $x$ (quark internal loop momentum), or into imaginary part at $x = \pm\xi$ point. The variable $\xi$ (skewdness) is defined as $\Delta^{+} = -2\xi \overline{P}^{+}$, where $\Delta = p - p^{\prime}$ is the momentum transfer in the process and $\overline{P} = (p + p^{\prime})/2$ is the average nucleon momentum. Until now most of the published DVCS observables are sensitive to $Im$ part of CFF, whereas real part is accessible through Beam Charge Asymmetries (BCA)[@BCA], cross-section measurements [@DVCS_CRS], or double spin asymmetries [@DSA]. Measurement of the $Re$ part of CFFs is also important for constraining GPDs, and the $Re$ part of CFFs have strong model sensitivity, whereas this is less the case for the $Im$ part. This is
shown in Fig.\[fig:CFF\_Model\_Dep\] where for two models of GPDs, double distribution [@DD] and dual parametrization [@Dual] the $Im$ and $Re$ parts of the CFF $\cal{H}$ are presented.
Recently a lot of theoretical work has been devoted to the evaluation of the next to leading order (NLO) corrections to CFFs, e.g. [@NLO1] and [@TCS_DVCS_NLO]. As in [@TCS] shown, the $Re$ and $Im$ parts of CCFs can be accessed also through angular asymmetries in the inverse DVCS process, called Timelike Compton Scattering (TCS), using unpolarized and circularly polarized photon beams, respectively. In TCS, $\gamma p \rightarrow \gamma^{*}(\rightarrow l^{-}l^{+})p$ the incoming photon is on shell, and the outgoing photon has large timelike virtuality and decays into lepton pairs. TCS also is an important reaction for testing universalities of GPDs, like in the same way Drell-Yan was used for checking universalities of PDFs.
Current status of experimental measurements
===========================================
Experimental data used in this analysis are from two high energy electroproduction experiments with the CLAS detector at Jefferson Lab [@CLAS] at beam energies $5.76\;GeV$ and $5.479\;GeV$. In electroproduction experiments, when beam electron scatters at small $\sim 0 \degree$ angle, the interacting intermediate photon will have $Q^{2}\sim 0$, and the reaction $ep\rarr e^{-}e^{+}p(e^{\prime})$ (Fig.\[fig:quasireal\_diagram\]), where $e^{\prime}$ is the scattered electron, can be interpreted as quasireal photoproduction of lepton pairs. The exclusivity of the event is ensured through the $Q^{2} < 0.01\;GeV^{2}$ and $|M_{x}|^{2} < 0.1\;GeV^{2}$ cuts, where $|M_{x}|^{2}$ is the missing mass squared of the $e^{-}e^{+}p$ system and the $Q^{2}$ is calculated from missing momentum. The invariant mass distribution of
$e^{-}e^{+}$ pairs is shown in Fig.\[fig:Minv\]. One can clearly see peaks for $\omega(782)$ and $\varphi(1020)$ mesons. The sholder on the left side of the $\omega(782)$ corresponds $\rho(770)$. For the TCS analysis we have used data above $\varphi(1020)$, $M(e^{-}e^{+}) > 1.01\;GeV$. The proposed observable in [@TCS] is the normalized ratio R, which is directly related to the scattering amplitude, and is defined as: $$R = \frac{\dstl 2 \int_{0}^{2\pi} d\phi cos(\phi) \frac{dS}{dQ^{2}dtd\phi}}{\dstl \int_{0}^{2\pi} d\phi \frac{dS}{dQ^{2}dtd\phi}}
\label{eq:R_Def}$$ where $$\frac{\dstl dS}{\dstl dQ^{2}dtd\phi} = \int_{\pi/2 - \Delta}^{\pi/2 + \Delta}\frac{\dstl L(\theta, \phi)}{\dstl L_{0}\theta}
\frac{\dstl d\sigma}{\dstl dQ^{2}dtd\phi}
\label{eq:dS_Def}$$ CLAS acceptance is not symmetric w.r.t. $\theta = \pi/2$. Therefore in order to compare experimental data with theoretical predictions the integration limits over $\theta$ in eq.(\[eq:dS\_Def\]) was chosen to fit CLAS acceptance.
In Fig.\[fig:R\_Data\_Theor\] preliminary result for the extracted ratio $R$ along with theoretical predictions are shown. While $\phi$-dependent $\theta$ integration creates artificial asymmetry, and the ratio $R$ doesn’t reflect the scattering amplitude, it is still usefull information for model comparison and as shown data favor to Dual parametrization model.
Currently work is in progress to to extract the ratio $R$ in the same way as proposed in [@TCS], through the extrapolation of experimental data to outside of the CLAS acceptance region.
Future plans
============
The 12 GeV upgrade of Jefferson lab will provide better conditions for TCS studies. The expected beam energy will be $11\;GeV$, which will allow to reach $M(e^{-}e^{+}) < 3.7\;GeV$ region. The expected luminosity for CLAS12 detector in Hall-B is $\approx 10^{35}cm^{-2}s^{-1}$, which is one order of magnitude higher than the CLAS maximum luminosity at $6\;GeV$. Proposal for studying TCS and $J/\Psi$ photoproduction near threshold on the proton target with $11\;GeV$ electron beam is already approved by JLAb Program Advisory Committee (PAC). It is proposed to study TCS in the mass range $2\; GeV < M(e^{-}e^{+})<3\;GeV$ where there is no contribution from meson resonances and pQCD descibes the ratio $R(s) = \frac{\dstl \sigma(e^{-}+e^{+}\rarr hadrons, s)}{\dstl \sigma(e^{-}+e^{+}\rarr \mu^{+}\mu^{-}, s)}$ [@PDG]. With $11\;GeV$ high luminosity beams $J/\Psi$ production near threshold can be thoroughly studied. Currently there are no published data for the $J/\Psi$ photoproduction near threshold, and the $J/\Psi$ production mechanism is not well understood. Measurement of $J/\Psi$ through it’s $J/\Psi\rarr e^{-}e^{+}$ decay channel can be studied with the same technique as TCS, and will be an important input for understanding the gluonic form factors of the nucleon.
Summary
=======
TCS offers a complementary way of studying GPDs, and is an important reaction for testing universalities of GPDs. Analysis of $6\;GeV$ CLAS data showed feasibility of studying TCS using data from electroproduction experiments. The same analysis technique will be employed for the TCS analysis with $11\;GeV$ electron beam with CLAS12 detector. 0.5cm This work was supported by the French P2IO laboratory of excellence, and by the Department of Education and Science of Republic of Armenia, Grant-11-1C015
[50]{} M. Diehl. Physics Reports, 388(24):41 – 277, 2003. John C. Collins, Leonid Frankfurt, and Mark Strikman. Phys. Rev. D, 56:2982–3006, Sep 1997. A. Airapetian et al. Phys. Rev. D 75, 011103(R) (2007) C. Muñoz Camacho et al. Phys. Rev. Lett. 97, 262002 (2006) JHEP 1006 (2010) 019. arXiv:1004.0177 A.V. Radyushkin Phys. Rev., D59:014030, 1999 V. Guzey and T. Teckentrup. Phys. Rev., D74:054027, 2006. B. Pire, L. Szymanowski and J. Wagner, Phys. Rev. D [83]{}, 034009 (2011) arXiv:1101.0555 \[hep-ph\]. H. Moutarde, B. Pire, F. Sabatie, L. Szymanowski and J. Wagner, arXiv:1301.3819 \[hep-ph\]. E. R. Berger, M. Diehl, B. Pire Eur. Phys. J. C [23]{} (2002) 675 B.A. Mecking et al. Nucl.Instrum.Meth. A503 (2003) J. Beringer et al. Phys.Rev. D86 010001 (2012)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, algorithms for calculation of 3-loop propagator diagrams in HQET and on-shell QCD with a heavy quark have been constructed and implemented. These algorithms (based on integration by parts recurrence relations) reduce an arbitrary diagram to a combination of a finite number of basis integrals. Here I discuss various ways to calculate non-trivial bases integrals, either exactly or as expansions in $\varepsilon$. Some integrals of these two classes are related to each other by inversion, which provides a useful cross-check.'
address: 'Budker Institute of Nuclear Physics, Novosibirsk'
author:
- 'A.G. Grozin'
title: Multiloop calculations in HQET
---
I presented a review talk about multiloop calculations in HQET at this conference in Pisa in 1995 [@BG:95]. Methods of calculation of two-loop propagator diagrams in HQET [@BG:91] and on-shell massive QCD [@B:92; @FT:92], based on integration by parts [@CT:81], were discussed there. Recently, three-loop HQET [@G:00] and on-shell [@MR:00] algorithms have been constructed. Here I discuss this substantial progress.
Three-loop massless diagrams {#QCD}
============================
First, I briefly remind you the classic method of calculation of 3-loop massless propagator diagrams. There are 3 generic topologies of such diagrams. They can be reduces, using integration by parts, to 6 basis integrals [@CT:81]. This algorithm is implemented in the package Mincer [@Mincer] (first written in SCHOONSCHIP [@Sch] and later rewritten in FORM [@Form]), and in the package Slicer [@Sli] written in REDUCE [@H:99; @G:97].
Four basis integrals are trivial. One is a two-loop diagram with a non-integer power of the middle line. It can be found as a particular case of a more general expression [@K:96; @BGK:97] for the the two-loop diagram with three non-integer powers via a hypergeometric ${}_3F_2$ function of the unit argument, with indices tending to integers at $\varepsilon\to0$. There is a rather straightforward algorithm for expanding such functions in $\varepsilon$, with coefficients expressed via multiple $\zeta$-values. I have implemented it in REDUCE in the summer of 2000, some results produced by this program are published in [@G:01]. It is clearly presented as Algorithm A in [@MUW:02]; this paper also contains other, more complicated, algorithms. The algorithms of [@MUW:02] are implemented in the C++ library nestedsums [@W:02] based on the computer-algebra library GiNaC [@BFK:00]. This implementation is very convenient; unfortunately, it requires one to install an outdated version of GiNaC. The Algorithm A seems to be also implemented in FORM [@V:99], but I could not understand how to use it. Using my REDUCE procedure or nestedsums [@W:02], one can quickly find as many terms of expansion of this basis integral in $\varepsilon$ as needed, in terms of multiple $\zeta$-values. They can be expressed, up to weight 9, via a minimum set of independent $\zeta$-values, using the results of [@B:96; @V:99]. The two-loop diagram with a non-integer power of the middle line can also be expressed [@K:85] via an ${}_3F_2$ function of the argument $-1$. Expanding this expression in $\varepsilon$ (say, using nestedsums [@W:02]), we encounter more general Euler–Zagier sums, which were also considered in [@B:96]. Reducing them to the minimal basis, we obtain, of course, the same $\varepsilon$-expansion of our basis integral.
Using this expansion and integration-by-parts relations, it is easy to recover the well-known result for the 3-loop ladder diagram, which is finite $\varepsilon=0$: $20\zeta(5)+\mathcal{O}(\varepsilon)$. The last and most difficult basis diagram is non-planar. It is also finite at $\varepsilon=0$. Using gluing of its external vertices [@CT:81], one can easily understand that it has the same value $20\zeta(5)$ at $\varepsilon=0$. There is no easy way to find further terms of its $\varepsilon$- expansion.
(92,50) (46,25)[(0,0)[]{}]{} (13.5,26)[(0,0)\[b\][$n_1$]{}]{} (28.5,26)[(0,0)\[b\][$n_2$]{}]{} (63.5,26)[(0,0)\[b\][$n_1$]{}]{} (78.5,26)[(0,0)\[b\][$n_2$]{}]{} (8.5,37.5)[(0,0)\[r\][$n_3$]{}]{} (33,37.5)[(0,0)\[l\][$n_4$]{}]{} (58.5,39)[(0,0)\[r\][$d-n_1-n_3$]{}]{} (57,36)[(0,0)\[r\][${}-n_5-n_7$]{}]{} (83,39)[(0,0)\[l\][$d-n_2-n_4$]{}]{} (84,36)[(0,0)\[l\][${}-n_5-n_8$]{}]{} (22,33.75)[(0,0)\[l\][$n_5$]{}]{} (72,33.75)[(0,0)\[l\][$n_5$]{}]{} (17.7,39)[(0,0)\[r\][$n_7$]{}]{} (25,39)[(0,0)\[l\][$n_8$]{}]{} (67.7,39)[(0,0)\[r\][$n_7$]{}]{} (75,39)[(0,0)\[l\][$n_8$]{}]{} (21,46)[(0,0)\[b\][$n_6$]{}]{} (71,46)[(0,0)\[b\][$d-n_6-n_7-n_8$]{}]{} (46,30)[(0,0)[$=$]{}]{} (11,1)[(0,0)\[b\][$n_1$]{}]{} (21,1)[(0,0)\[b\][$n_3$]{}]{} (31,1)[(0,0)\[b\][$n_2$]{}]{} (61,1)[(0,0)\[b\][$n_1$]{}]{} (71,1)[(0,0)\[b\][$n_3$]{}]{} (81,1)[(0,0)\[b\][$n_2$]{}]{} (9.5,10)[(0,0)\[r\][$n_4$]{}]{} (32,10)[(0,0)\[l\][$n_5$]{}]{} (59.5,12)[(0,0)\[r\][$d-n_1-n_4$]{}]{} (57.5,9)[(0,0)\[r\][${}-n_6$]{}]{} (82,12)[(0,0)\[l\][$d-n_2-n_5$]{}]{} (83,9)[(0,0)\[l\][${}-n_7$]{}]{} (17,10)[(0,0)\[l\][$n_6$]{}]{} (25,10)[(0,0)\[r\][$n_7$]{}]{} (67,10)[(0,0)\[l\][$n_6$]{}]{} (75,10)[(0,0)\[r\][$n_7$]{}]{} (21,16)[(0,0)\[b\][$n_8$]{}]{} (71,16)[(0,0)\[b\][$d-n_3-n_6-n_7-n_8$]{}]{} (46,5)[(0,0)[$=$]{}]{}
Three-loop HQET diagrams {#HQET}
========================
There are 10 generic topologies of 3-loop HQET propagator diagrams. They can be reduced, using integration by parts, to 8 basis integrals [@G:00]. This algorithm is implemented in the REDUCE package Grinder [@G:00], available at http://www-ttp.physik.uni-karlsruhe.de/Progdata/ttp00/ttp00-01/. Five basis integrals are trivial. Two can be expressed via ${}_3F_2$ hypergeometric functions of the unit argument [@BB:94; @G:00]. Their expansions in $\varepsilon$ can be obtained in the same way as in the massless case, the results are presented in [@G:01]. The last and most difficult basis integral was found in [@G:01] up to the finite term in $\varepsilon$, using direct integration in the coordinate space. More terms of its $\varepsilon$-expansion were recently obtained in [@CM:02] using inversion, as explained in the next Section.
Three-loop on-shell diagrams {#OS}
============================
Calculations of on-shell diagrams with massive quarks in QCD are necessary for obtaining coefficients in the HQET Lagrangian and $1/m$ HQET expansions of QCD operators by matching. There are 2 generic topologies of 2-loop on-shell propagator diagrams with a single non-zero mass. They can be reduced, using integration by parts, to 3 basis integrals. This algorithm is implemented in the REDUCE package RECURSOR [@B:92] and the FORM package SHELL2 [@FT:92]. Two basis integrals are trivial, and the third one is expressed via two ${}_3F_2$ hypergeometric functions of the unit argument. However, some of their indices tend to half-integers at $\varepsilon\to0$, and the algorithm of expansion in $\varepsilon$ discussed in Sect. \[QCD\] is not applicable. This approach was used for QCD/HQET matching of heavy-light quark currents [@BG:95a] and chromomagnetic interaction [@CG:97].
The case when there is another non-zero mass was systematically studied in [@DG:99]. There are 4 basis integrals, 2 of them trivial, and 2 are expressed via ${}_3F_2$ hypergeometric functions of the mass ratio squared. Finite parts at $\varepsilon\to0$ are expressed via dilogarithms. More terms of expansions of the general results [@DG:99] in $\varepsilon$ were recently obtained [@AMR:02]. The REDUCE package [@DG:99] is available at http://wwwthep.physik.uni-mainz.de/Publications/progdata/mzth9838/ Mm.red.
There are 11 generic topologies of 3-loop on-shell propagator diagrams with a single non-zero mass (10 of them are the same as in HQET, and one involves a heavy-quark loop). They can be reduced, using integration by parts, to 18 basis integrals [@MR:00]. This algorithm is implemented as the FORM package SHELL3 [@MR:00]. The basis integrals are mostly known from QED [@LR:96].
Some on-shell diagrams are related to HQET ones by inversion of Euclidean integration momenta. One- and two-loop relations were presented in [@BG:95]. Three-loop relations are shown in Fig. \[Fig\]. The second of them was used in [@CM:02] to relate the the convergent ladder HQET diagram at $\varepsilon=0$ to the known on-shell ladder diagram.
[99]{}
D.J. Broadhurst and A.G. Grozin, in *New computing techniques in physics research IV*, ed. B. Denby and D. Perret-Gallix, World Scientific (1996) 217.
D.J. Broadhurst and A.G. Grozin, Phys. Lett. B267 (1991) 105.
N. Gray, D.J. Broadhurst, W. Grafe, K. Schilcher, Z. Phys. C48 (1990) 673;\
D.J. Broadhurst, N. Gray, K. Schilcher, Z. Phys. C52 (1991) 111;\
D.J. Broadhurst, Z. Phys. C54 (1992) 599.
J. Fleischer and O.V. Tarasov, Phys. Lett. B283 (1992) 129; Comput. Phys. Commun. 71 (1992) 193.
F.V. Tkachov, Phys. Lett. B100 (1981) 65;\
K.G. Chetyrkin and F.V. Tkachov, Nucl. Phys. B192 (1981) 159.
A.G. Grozin, J. High Energy Physics 03 (2000) 013; hep-ph/0002266.
K. Melnikov and T. van Ritbergen, Phys. Lett. B482 (2000) 99; Nucl. Phys. B591 (2000) 515.
S.G. Gorishny, S.A. Larin, F.V. Tkachev, Preprint INR P-0330, Moscow (1984);\
S.G. Gorishny, S.A. Larin, L.R. Surguladze, F.V. Tkachev, Comput. Phys. Commun. 55 (1989) 381;\
S.A. Larin, F.V. Tkachev, J.A.M. Vermaseren, Preprint NIKHEF-H/91-18, Amsterdam (1991).
M. Veltman, SCHOONSCHIP, CERN (1967);\
H. Strubbe, Comput. Phys. Commun. 8 (1974) 1.
J.A.M. Vermaseren, Symbolic manipulations with FORM, Amsterdam (1991); math-ph/0010025 (2002).
D.J. Broadhurst, Preprint OUT-4102-41 (1992), see in: D.J. Broadhurst, A.L. Kataev, O.V. Tarasov, Phys. Lett. B298 (1993) 445, and in: D.J. Broadhurst, hep-th/9909185.
A.C. Hearn, REDUCE User’s Manual, Version 3.7 (1999).
A.G. Grozin, Using REDUCE in High Energy Physics, Cambridge University Press (1997).
A.V. Kotikov, Phys. Lett. B375 (1996) 240.
D.J. Broadhurst, J.A. Gracey, D. Kreimer, Z. Phys. C75 (1997) 559.
A.G. Grozin, in *QCD: theory and experiment*, AIP conference proceedings 602 (2001) 271.
S. Moch, P. Uwer, S. Weinzierl, J. Math. Phys. 43 (2002) 3363.
S. Weinzierl, Comput. Phys. Commun. 145 (2002) 357.
C. Bauer, A. Frink, R. Kreckel, Preprint MZ-TH-00-17 (2000), cs.sc/0004015.
J.A.M. Vermaseren, Int. J. Mod. Phys. A14 (1999) 2037.
D.I. Kazakov, Theor. Math. Phys. 62 (1985) 84.
D.J. Broadhurst, Preprint OUT-4102-62 (1996), hep-th/9604128;\
J.M. Borwein, D.M. Bradley, D.J. Broadhurst, Electronic J. Combinatorics 4 No. 2 (1997) \#R5, hep-th/9611004;\
J.M. Borwein, D.M. Bradley, D.J. Broadhurst, P. Lisonek, Electronic J. Combinatorics 5 (1998) \#R38, math.NT/9812020.
M. Beneke and V.M. Braun, Nucl. Phys. B426 (1994) 301.
A. Czarnecki and K. Melnikov, Phys. Rev. D66 (2002) 011502.
D.J. Broadhurst and A.G. Grozin, Phys. Rev. D52 (1995) 4082.
A. Czarnecki and A.G. Grozin, Phys. Lett. B405 (1997) 142.
A.I. Davydychev and A.G. Grozin, Phys. Rev. D59 (1999) 054023.
M. Argeri, P. Mastrolia, E. Remiddi, Nucl. Phys. B631 (2002) 388.
S. Laporta and E. Remiddi, Phys. Lett. B379 (1996) 283;\
K. Melnikov and T. van Ritbergen, Phys. Rev. Lett. 84 (2000) 1673.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the quantum criticality of the phase transition between Dirac semimetal and excitonic insulator in two dimensions. Even though the system has a semimetallic ground state, there are observable effects of excitonic pairing at finite temperatures and/or finite energies, provided that the system is in proximity to excitonic insulating transition. To determine the quantum critical behavior, we consider three potentially important interactions, including the Yukawa coupling between Dirac fermions and excitonic order parameter fluctuation, the long-range Coulomb interaction, and the disorder scattering. We employ the renormalization group technique to study how these interactions affect quantum criticality and also how they influence each other. We first investigate the Yukawa coupling in the clean limit, and show that it gives rise to typical non-Fermi liquid behavior. Adding random scalar potential to the system always turns such a non-Fermi liquid into a compressible diffusive metal. In comparison, the non-Fermi liquid behavior is further enhanced by random vector potential, but is nearly unaffected by random mass. Incorporating the Coulomb interaction may change the results qualitatively. In particular, the non-Fermi liquid state is protected by the Coulomb interaction for weak random scalar potential, and it becomes a diffusive metal only when random scalar potential becomes sufficiently strong. When random vector potential or random mass coexists with Yukawa coupling and Coulomb interaction, the system is stable non-Fermi liquid state, with fermion velocities flowing to constants in the former case and being singularly renormalized in the latter case. These quantum critical phenomena can be probed by measuring observable quantities. We also find that, while the fermion velocity anisotropy is not altered by the excitonic quantum fluctuation, it may be driven by the Coulomb interaction to flow to the isotropic limit.'
author:
- 'Xiao-Yin Pan'
- 'Jing-Rong Wang'
- 'Guo-Zhu Liu'
title: Quantum critical phenomena of excitonic insulating transition in two dimensions
---
Introduction
============
In the past decade, the unconventional properties of various Dirac/Weyl semimetal (SM) materials [@Sarma11; @Kotov12; @Vafek14; @Wehling14; @Wan11; @Weng16; @FangChen16; @Yan17; @Hasan17; @Armitage18] have been investigated extensively. Many of the unconventional properties are related to the existence of isolated Dirac/Weyl points, at which the conduction and valence bands touch. When the chemical potential is tuned to exactly the Dirac points, the fermion density of states (DOS) vanishes at the Fermi level. As a result, the Coulomb interaction is long-ranged due to the absence of static screening. Extensive previous studies [@Kotov12; @Gonzalez99; @Hofmann14; @Sharma16; @Goswami11; @Hosur12; @Hofmann15; @Throckmorton15; @Sharma18; @Yang14; @Abrikosov72; @Abrikosov71; @Abrikosov74; @Moon13; @Herbut14; @Janssen15; @Dumitrescu15; @Janssen17; @Huh16; @Cho16; @Isobe16B; @Lai15; @Jian15; @Zhang17; @WangLiuZhang17MWSM] have revealed that the Coulomb interaction leads to a variety of unconventional low-energy behaviors.
Among all the known SM materials, two-dimensional Dirac SM, abbreviated as 2D DSM hereafter, has been studied most extensively, usually in the context of graphene. Renormalization group (RG) analysis [@Shankar94; @Kotov12] has revealed that the long-range Coulomb interaction is marginally irrelevant in the weak-coupling regime. When the Coulomb interaction is strong enough, the originally massless fermions can acquire a dynamical mass gap via the formation of stable particle-hole pairs [@CastroNetoPhysics09; @Khveshchenko01; @Gorbar02; @Khveshchenko04; @Liu09; @Khveshchenko09; @Gamayun10; @Sabio10; @Zhang11; @Liu11; @WangLiu11A; @WangLiu11B; @WangLiu12; @Popovici13; @WangLiu14; @Gonzalez15; @Carrington16; @Sharma17; @Xiao17; @Carrington18; @Gamayun09; @WangJianhui11; @Katanin16; @Gonzalez10; @Gonzalez12; @Drut09A; @Drut09B; @Drut09C; @Armour10; @Armour11; @Buividovich12; @Ulybyshev13; @Smith14; @Juan12; @Kotikov16; @Gonzalez14; @Braguta16; @Xiao18; @Janssen16; @WangLiuZhangSemiDSM]. This gap generating scenario is non-perturbative, and has the same picture as excitonic pairing, a notion proposed decades ago [@Keldysh64; @Jerome67]. In the special case of 2D DSM, such an excitonic gap dynamically breaks a continuous chiral (sublattice) symmetry, which can be regarded as a condensed-matter realization of the dynamical chiral symmetry breaking [@Nambu61; @Miransky94]. The finite gap opened at the Dirac point drives the SM to undergo a quantum phase transition (QPT) into an excitonic insulator (EI). The EI is induced only when the effective interaction strength, denoted by $\alpha$, exceeds some critical value $\alpha_c$, which defines SM-EI quantum critical point (QCP).
In recent years, the possibility of SM-EI transition in graphene has been investigated by means of various analytical and numerical techniques. Early calculations [@CastroNetoPhysics09; @Khveshchenko01; @Gorbar02; @Khveshchenko04; @Liu09; @Khveshchenko09; @Gamayun10; @Drut09A; @Drut09B; @Drut09C] predicted that the Coulomb interaction in suspended graphene is strong enough to open an excitonic gap at zero temperature. Specifically, the critical value $\alpha_c$ was claimed to be smaller than the physical value $\alpha
= 2.16$. However, no visible experimental evidence for excitonic gap has been observed at low temperatures [@Elias11; @Mayorov12]. More careful numerical calculations [@WangLiu12; @Gonzalez15; @Carrington16; @Ulybyshev13; @Smith14; @Kotikov16] revealed that the critical value $\alpha_c$ is actually larger than $2.16$, which implies that the Coulomb interaction cannot generate a finite excitonic gap. Owing to the conceptual importance and also the potential technical applications, theorists are still searching for possible approaches to promote excitonic pairing in various SM materials. For instance, it was proposed that excitonic pairing may be promoted by an additional short-range repulsive interaction [@Liu09; @Gamayun10; @WangLiu12] or by certain extrinsic effects, such as strain [@Tang15].
![Global phase diagram of 2D DSM on the $\alpha$-$T$ or $\alpha$-$\omega$ plane. Here, $\omega$ stands for the fermion energy. Deep in the insulating phase, the fermions are suppressed at low energies. Deep in the semimetallic phase, the Coulomb interaction is too weak to form excitonic pairs. The excitonic insulating transition occurs as $\alpha$ increases up to $\alpha_c$ at $T=0$. This point is broadened into a finite quantum critical regime at finite $T$ and/or finite $\omega$. The excitonic quantum fluctuation has observable effects in the whole quantum critical regime.[]{data-label="Fig:PhaseEIQCP"}](PhaseEIQCP.eps){width="2.88in"}
Most previous works on SM-EI QPT have focused on the precise calculation of $\alpha_c$ at zero temperature ($T=0$) by means of various techniques [@CastroNetoPhysics09; @Khveshchenko01; @Gorbar02; @Khveshchenko04; @Liu09; @Khveshchenko09; @Gamayun10; @Sabio10; @Zhang11; @Liu11; @WangLiu11A; @WangLiu11B; @WangLiu12; @Popovici13; @WangLiu14; @Gonzalez15; @Carrington16; @Sharma17; @Xiao17; @Carrington18; @Gamayun09; @WangJianhui11; @Katanin16; @Gonzalez10; @Gonzalez12; @Drut09A; @Drut09B; @Drut09C; @Armour10; @Armour11; @Buividovich12; @Ulybyshev13; @Smith14; @Juan12; @Kotikov16; @Gonzalez14; @Braguta16; @Xiao18; @Janssen16; @WangLiuZhangSemiDSM]. In this paper, we propose to explore the signatures of excitonic pairing at finite $T$ and/or finite energy $\omega$. Here is our logic: even though the exact zero-$T$ ground state of suspended graphene (or other 2D DSMs) is gapless, the quantum fluctuation of excitonic pairs still have observable effects at finite $T$ and/or $\omega$ if the system is in the quantum critical regime around the putative SM-EI QCP. Recent Monte Carlo simulations [@Ulybyshev13] and Dyson-Schwinger equation study [@Carrington18] both suggest that the value $\alpha_c$ is not far from the physical value of suspended graphene. As illustrated in the schematic phase diagram Fig. \[Fig:PhaseEIQCP\], if $\alpha$ is slightly smaller than $\alpha_c$, no excitonic gap is opened at $T=0$ and the excitonic order parameter has a vanishing mean-value. However, the quantum fluctuation of excitonic order parameter is not negligible at finite $T$ and/or $\omega$ and may lead to considerable corrections to observable quantities. For instance, the nuclear-magnetic-resonance measurements performed by Hirata *et al.* [@Hirata17] indicate that the compound $\alpha$-(BEDT-TTF)$_{2}$I$_{3}$ is close to an SM-EI QCP and that the excitonic fluctuation results in singular corrections to the nuclear magnetic resonance relaxation rate.
We study the quantum critical phenomena emerging in the broad quantum critical regime around SM-EI QCP, with the aim to explore observable effects of excitonic pairing. For this purpose, we take suspended graphene (typical 2D DSM) as our starting model, and calculate the interaction corrections to some observable quantities of Dirac fermions. In this regime, the gapless fermions interact with the quantum critical fluctuation of excitonic order parameter, which is described by a Yukawa coupling term. The long-range Coulomb interaction is still present and needs to be properly taken into account. Moreover, there is always certain amount of quenched disorder [@Kotov12] in realistic materials, and the fermion-disorder coupling might play a vital role. The actual quantum critical phenomena cannot be accurately determined if one or more of these interactions are naively ignored or improperly treated. We emphasize that, these three kinds of interaction may have very complicated mutual influence. To make a generic analysis, we will treat all the three kinds of interaction on equal footing and study their interplay carefully.
As the first step, we treat the Yukawa coupling in the clean limit, and demonstrate that this coupling leads to strong violation of Fermi liquid (FL) theory. Indeed, the quasiparticle residue $Z_f$ vanishes at low energies, and the fermion DOS $\rho(\omega)$ receives power-law corrections from the excitonic fluctuation. Both of these two features are typical non-Fermi-liquid (NFL) behaviors. If the fermion dispersion is originally anisotropic, the ratio between two fermion velocities is unrenormalized.
The next step is to incorporate quenched disorder and analyze its interplay with the Yukawa coupling. We find that the resultant low-energy properties depend sensitively on the nature of the disorder. Adding random scalar potential (RSP) to the system always turns the NFL caused by Yukawa coupling in the clean limit into a compressible diffusive metal (CDM). The CDM state is characterized by the generation of a finite zero-energy fermion DOS and a constant zero-$T$ disorder scattering rate. Different from RSP, random vector potential (RVP) tends to further enhance the NFL behavior, whereas random mass (RM) has negligible effects on the system.
We finally incorporate the Coulomb interaction, and find that it changes the above results qualitatively. In the case of weak RSP, the Coulomb interaction suppresses disorder scattering and as such renders the stability of the NFL state caused by Yukawa coupling. However, such a NFL is converted into CDM once RSP becomes sufficiently strong. The combination of Yukawa coupling, Coulomb interaction, and RVP produces a stable NFL state in which the two fermion velocities flow to constant values in the zero energy limit. When the Yukawa coupling, Coulomb interaction, and RM are considered simultaneously, we show that the Coulomb interaction is marginally irrelevant and RM is irrelevant. These results indicate that the true quantum critical phenomena are determined by a delicate interplay of excitonic fluctuation, Coulomb interaction, and disorder scattering.
Our results might be applied to understand some 2D DSM materials, such as uniaxially strained graphene [@Sharma17; @Xiao17] and organic compound $\alpha$-(BEDT-TTF)$_{2}$I$_{3}$ [@Hirata17]. In these systems, the fermion velocities along different directions may be unequal. It is thus necessary to examine how interactions change the anisotropy. According to our RG analysis, the fermion velocity anisotropy is unaffected by the excitonic fluctuation, but could be significantly suppressed by the Coulomb interaction.
The rest of the paper will be arranged as following. The model is presented in Sec. \[Sec:Model\]. The RG equations for the corresponding parameters are shown in Sec. \[Sec:RGEquations\]. The numerical results for different conditions are given and analyzed in Sec. \[Sec:NumResults\]. The mains results are summarized in Sec. \[Sec:SummaryDiscuss\]. The detailed derivation of the RG equations can be found in Appendices.
The model\[Sec:Model\]
======================
The fermion energy dispersion in intrinsic graphene is isotropic. It becomes anisotropic when graphene is deformed. Generically, the action of free 2D Dirac fermions with anisotropic dispersion is given by $$\begin{aligned}
S_{f} = \sum_{\sigma=1}^{N}\int_{\tau,\mathbf{x}}
\bar{\Psi}_{\sigma}(\tau,\mathbf{x})\left[\partial_{\tau}\gamma_{0}
+ \mathcal{H}_{f}\right] \Psi_{\sigma}(\tau,\mathbf{x}),\end{aligned}$$ where $\int_{\tau,\mathbf{x}}\equiv\int d\tau\int d^2\mathbf{x}$ and $\mathcal{H}_{f} = -iv_{1}\nabla_{1}\gamma_{1} -iv_{2}\nabla_{2}
\gamma_{2}$. Here, $\Psi$ is a four-component spinor, and $\bar{\Psi}=\Psi^{\dag}\gamma_{0}$. The matrices $\gamma_{0,1,2}$ are defined as $\gamma_{0,1,2} =
\left(\tau_{3},-i\tau_{2},i\tau_{1}\right)\otimes\tau_{3}$ in terms of Pauli matrices $\tau_{i}$ with $i = 1, 2, 3$. The gamma matrices satisfy the anti-commutative rule $\left\{\gamma_{\mu},\gamma_{\nu}\right\} =
2\mathrm{diag}(1,-1,-1)$. The fermion species is denoted by $\sigma$, which sums from $1$ to $N$. Fermion flavor $N$ is assumed to be a general large integer. We use $v_{1}$ and $v_{2}$ to represent the fermion velocities along two orthogonal directions.
The action of the quantum fluctuation of excitonic order parameter can be written as $$\begin{aligned}
S_{b} = \int_{\tau,\mathbf{x}}\left[\frac{1}{2}
\left(\partial_{\tau}\phi\right)^{2}+\frac{c^{2}}{2}
\left(\mathbf{\nabla}\phi\right)^{2} + \frac{r}{2}\phi^{2} +
\frac{u}{24}\phi^{4}\right],\end{aligned}$$ where $c$ is the boson velocity. Varying boson mass $r$ tunes the QPT between SM and EI phases. At the QCP, the mass vanishes, i.e., $r=0$, and the boson field $\phi$ describes the quantum critical fluctuation of excitonic order parameter. The quartic self-interacting term has a coupling constant $u$. The Yukakwa coupling between fermions and excitonic order parameter is given by $$\begin{aligned}
S_{fb} = \lambda\sum_{\sigma=1}^{N}\int_{\tau,\mathbf{x}}
\phi\bar{\Psi}_{\sigma}\Psi_{\sigma},\end{aligned}$$ where $\lambda$ is the corresponding coupling constant.
The excitonic pairing originates from the Coulomb interaction between fermions and their anti-fermions (holes). Inside the EI phase, a finite gap is opened at the Fermi level and strongly suppresses the low-energy fermion DOS. In this case, the Coulomb interaction and even the fermionic degrees of freedom can be neglected, and the low-energy properties of the EI phase is mainly governed by the dynamics of neutral excitons. In contrast, the fermions remain gapless at the SM-EI QCP. The Coulomb interaction between gapless fermions may play an important role at low energies. The action for Coulomb interaction is described by $$\begin{aligned}
S_{ee}=\frac{1}{4\pi}\sum_{\sigma,\sigma'=1}^{N}
\int_{\tau,\mathbf{x},\mathbf{x}'}\rho_{\sigma}(\tau,\mathbf{x})
\frac{e^2/\epsilon}{|\mathbf{x}-\mathbf{x}'|}
\rho_{\sigma}(\tau,\mathbf{x}'),\end{aligned}$$ where $\int_{\tau,\mathbf{x},\mathbf{x}'}\equiv\int d\tau\int
d^2\mathbf{x}\int d^2\mathbf{x}'$. The fermion density operator is defined as $\rho_{\sigma}(\tau,\mathbf{x}) =
\bar{\Psi}_{\sigma}(\tau,\mathbf{x})\gamma_{0}
\Psi_{\sigma}(\tau,\mathbf{x})$. In addition, $e$ is electric charge and $\epsilon$ dielectric constant.
Disorder exists in almost all realistic materials. Many of the low-energy behaviors of fermions are heavily affected by disorder scattering, especially at low $T$. The fermion-disorder coupling is formally described by $$\begin{aligned}
S_{\mathrm{dis}} = v_{\Gamma}\int d\tau d^2\mathbf{x}
\bar{\Psi}_{\sigma}(\mathbf{x})\Gamma\Psi_{\sigma}(\mathbf{x})
A(\mathbf{x}).\end{aligned}$$ The random field $A(\mathbf{x})$ is assumed to be a Gaussian white noise, i.e., $\langle A(\mathbf{x})\rangle = 0$ and $\langle
A(\mathbf{x})A(\mathbf{x}')\rangle = \Delta
\delta^{2}(\mathbf{x}-\mathbf{x}')$. Here, $\Delta$ is the impurity concentration, and $v_{\Gamma}$ measures the strength of a single impurity. The disorders are classified by the expression of $\Gamma$ matrix [@Ludwig94; @Nersesyan95; @Altland02]. For $\Gamma_{0}=\gamma_{0}$, $A(\mathbf{x})$ is a RSP. For $\Gamma_{j}=\mathbbm{1}_{4\time4}$, $A(\mathbf{x})$ serves as a RM. In comparison, RVP has two components $A_{1,2}(\mathbf{x})$, characterized by $\Gamma=(\gamma_{1},\gamma_{2})$ and $v_{\Gamma}=(v_{\Gamma1}, v_{\Gamma2})$.
The free fermion propagator has the form $$\begin{aligned}
G_{0}(\omega,\mathbf{k}) = \frac{1}{-i\omega\gamma_{0} +
v_{1}k_{1}\gamma_{1}+v_{1}k_{2}\gamma_{2}}.
\label{Eq:FermionPropagator}\end{aligned}$$ The Yukawa coupling can be treated by the RG method in combination with the $1/N$ expansion. Following the scheme developed by Huh and Sachdev [@Huh08], we re-scale $\phi$ and $r$ as follows: $\phi\rightarrow\phi/\lambda$ and $r\rightarrow Nr\lambda^{2}$. Accordingly, the bare propagator of $\phi$ is expressed as $$\begin{aligned}
D_{0}^{A}(\Omega,\mathbf{q})=\frac{1}{\frac{\Omega^{2} +
c^2\mathbf{q}^{2}}{\lambda^{2}}+Nr}.\end{aligned}$$ Near the QCP, we take $r=0$ and then get $$\begin{aligned}
D_{0}^{A}(\Omega,\mathbf{q}) = \frac{\lambda^{2}}{\Omega^{2} +
c^2\mathbf{q}^{2}}.\end{aligned}$$ The free boson propagator is drastically altered by the polarization function, which, to the leading order of $1/N$ expansion, is $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) &=& N \int\frac{d\omega}{2\pi}\frac{d^2
\mathbf{k}}{(2\pi)^2} \nonumber \\
&& \times \mathrm{Tr} \left[G_{0}(\omega,\mathbf{k})
G_{0}(\omega+\Omega,\mathbf{k}+\mathbf{q})\right]
\nonumber \\
&=& \frac{N}{4v_1 v_2}\sqrt{\Omega^2 + v_{1}^{2}q_{1}^{2} +
v_{2}^{2} q_{2}^{2}}.\end{aligned}$$ Now the dressed boson propagator becomes $$\begin{aligned}
D^{A}(\Omega,\mathbf{q})=\frac{1}{\frac{\Omega^{2} + c^2
\mathbf{q}^{2}}{\lambda^{2}}+\Pi^{A}(\Omega,\mathbf{q})}.\end{aligned}$$ It is obvious that $\Pi^{A}$ dominates over the free term in the low-energy regime. Thus, the above expression can be further simplified to $$\begin{aligned}
D^{A}(\Omega,\mathbf{q}) \approx \frac{1}{\Pi^{A}(\Omega,\mathbf{q})}.\end{aligned}$$
The bare Coulomb interaction is described by $$\begin{aligned}
D_{0}^{B}(\mathbf{q})=\frac{2\pi e^{2}}{\epsilon|\mathbf{q}|}.\end{aligned}$$ The dynamical screening is encoded in the polarization $\Pi^{B}(\Omega,\mathbf{q})$, whose leading order expression is given by $$\begin{aligned}
\Pi^{B}(\Omega,\mathbf{q}) &=& -N \int \frac{d\omega}{2\pi}
\frac{d^2 \mathbf{k}}{(2\pi)^2} \mathrm{Tr}
\left[\gamma_{0}G_{0}(\omega,\mathbf{k})\gamma_{0}\right.\nonumber
\\
&&\left.\times G_{0}(\omega+\Omega,\mathbf{k}+\mathbf{q})\right]
\nonumber \\
&=& \frac{N}{8v_1v_2}\frac{v_{1}^2q_{1}^{2} + v_{2}^2
q_{2}^{2}}{\sqrt{\Omega^2+v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}}.\end{aligned}$$ The dressed Coulomb interaction can be written as $$\begin{aligned}
D^{B}(\Omega,\mathbf{q}) = \frac{1}{\frac{\epsilon|\mathbf{q}|}{2\pi
e^{2}} + \Pi^{B}(\Omega,\mathbf{q})}.\end{aligned}$$
In previous works on the quantum criticality of SM-EI transition, the interplay of Yukawa coupling, Coulomb interaction, and disorder has never been systematically studied. Here, we emphasize that all the three interactions could be very important at low energies and thus should be treated equally.
Renormalization group equations \[Sec:RGEquations\]
===================================================
The interplay of distinct interactions can be handled by means of perturbative RG approach. The detailed RG calculations are presented in the Appendices. In this section, we only list the coupled RG equations of a number of model parameters and then analyze their low-energy properties. The effective model contains several independent parameters, such as $v_1$, $v_2$, and $v_\Gamma$. These parameters are renormalized by interactions. To specify how the interactions alter the fermion dispersion anisotropy, we need to determine the flow of the ratio $v_2/v_1$. Moreover, to judge whether FL theory is applicable, we should compute the flow equation of the residue $Z_f$.
After incorporating three types of interaction in a self-consistent way, we find that the coupled RG equations for $Z_f$, $v_{1}$, $v_{2}$, and $v_{2}/v_{1}$ are given by $$\begin{aligned}
\frac{dZ_{f}}{d\ell} &=& \left(C_{0}^{A}+C_{0}^{B} -
C_{g}\right)Z_{f}, \label{Eq:VRGZf} \\
\frac{dv_{1}}{d\ell} &=& \left(C_{0}^{A} + C_{0}^{B} - C_{1}^{A} -
C_{1}^{B} - C_{g}\right)v_{1}, \label{Eq:VRGV1} \\
\frac{dv_{2}}{d\ell} &=& \left(C_{0}^{A}+C_{0}^{B} - C_{2}^{A} -
C_{2}^{B} - C_{g}\right)v_{2}, \label{Eq:VRGV2}
\\
\frac{d\left(v_{2}/v_{1}\right)}{d\ell} &=& \left(C_{1}^{A} -
C_{2}^{A} + C_{1}^{B} - C_{2}^{B}\right)\frac{v_{2}}{v_{1}}.
\label{Eq:VRGVRatio}\end{aligned}$$ RG analysis is performed by integrating out the modes defined within the momentum shell $e^{-\ell}\Lambda < |\mathbf{k}| < \Lambda$, where $\Lambda$ is an UV cutoff and $\ell$ is a running parameter [@Shankar94]. The lowest energy limit is reached as $\ell
\rightarrow \infty$. For RSP, the flow equation of $v_{\Gamma}$ takes the form $$\begin{aligned}
\frac{d v_{\Gamma}}{d\ell}=0. \label{Eq:VRGVGammaRCP}\end{aligned}$$ For the two components of RVP, the flow equations for $v_{\Gamma1}$ and $v_{\Gamma2}$ are $$\begin{aligned}
\frac{dv_{\Gamma1}}{d\ell}=\left(C_{0}^{A}+C_{0}^{B}-C_{1}^{A}
-C_{1}^{B}-C_{g}\right)v_{\Gamma1}, \label{Eq:VRGVGammaRVP1}
\\
\frac{dv_{\Gamma2}}{d\ell}=\left(C_{0}^{A}+C_{0}^{B}-C_{2}^{A}
-C_{2}^{B}-C_{g}\right)v_{\Gamma2}. \label{Eq:VRGVGammaRVP2}\end{aligned}$$ For RM, the flow equation of $v_{\Gamma}$ is $$\begin{aligned}
\frac{dv_{\Gamma}}{d\ell}&=&\left(2C_{0}^{A}+C_{1}^{A}+C_{2}^{A} +
2C_{0}^{B}-C_{1}^{B}-C_{2}^{B}\right.\nonumber
\\
&&\left.-2C_{g}\right)v_{\Gamma}. \label{Eq:VRGVGammaRM}\end{aligned}$$ Here, we introduce a new parameter $C_{g}$ to characterize the effective strength of disorder. For RSP and RM, it is $$\begin{aligned}
C_g =\frac{v_{\Gamma}^{2}\Delta}{2\pi v_{1}v_{2}}.\end{aligned}$$ For RVP, we have $$\begin{aligned}
C_g = \frac{\left(v_{\Gamma1}^{2}+v_{\Gamma2}^{2}\right)
\Delta}{2\pi v_{1}v_{2}}.\end{aligned}$$ The three coefficients $C_{0}^{A}$, $C_{1}^{A}$, and $C_{2}^{A}$ appearing in the coupled RG equations are $$\begin{aligned}
C_{0}^{A}&=&\frac{1}{8\pi^3}\int_{-\infty}^{+\infty}dx
\int_{0}^{2\pi}d\theta\nonumber \\
&&\times\frac{x^2-\cos^2\theta-(v_{2}/v_{1})^{2}
\sin^2\theta}{\left(x^2+\cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta\right)^{2}} \mathcal{G}^{A}(x,\theta), \label{Eq:C0A}
\\
C_{1}^{A}&=&\frac{1}{8\pi^3}\int_{-\infty}^{+\infty} dx
\int_{0}^{2\pi}d\theta \nonumber
\\
&&\times\frac{-x^2+\cos^2\theta-(v_{2}/v_{1})^{2}\sin^2\theta}
{\left(x^2+\cos^2\theta+(v_{2}/v_{1})^{2}\sin^2\theta\right)^{2}}
\mathcal{G}^{A}(x,\theta), \label{Eq:C1A}
\\
C_{2}^{A}&=&\frac{1}{8\pi^3}\int_{-\infty}^{+\infty} dx
\int_{0}^{2\pi}d\theta\nonumber
\\
&&\times\frac{-x^2-\cos^2\theta+(v_{2}/v_{1})^{2}\sin^2\theta}
{\left(x^2+\cos^2\theta+(v_{2}/v_{1})^{2}\sin^2\theta\right)^{2}}
\mathcal{G}^{A}(x,\theta),\label{Eq:C2A}\end{aligned}$$ where $$\begin{aligned}
\mathcal{G}^{A}(x,\theta) = \frac{1}{\frac{N}{4v_2/v_{1}}
\sqrt{x^{2}+\cos^{2}\theta + \left(v_{2}/v_{1}\right)^{2}
\sin^{2}\theta}}. \label{Eq:GExpressionA}\end{aligned}$$ The coefficients $C_{0}^{B}$, $C_{1}^{B}$, and $C_{2}^{B}$ are $$\begin{aligned}
C_{0}^{B} &=& \frac{1}{8\pi^3}\int_{-\infty}^{+\infty}dx
\int_{0}^{2\pi}d\theta \nonumber \\
&&\times\frac{-x^2+\cos^2\theta + (v_{2}/v_{1})^{2}
\sin^2\theta}{\left(x^2+\cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta\right)^{2}}\mathcal{G}^{B}(x,\theta),
\label{Eq:C0B} \\
C_{1}^{B}&=&\frac{1}{8\pi^3}\int_{-\infty}^{+\infty}
dx\int_{0}^{2\pi}d\theta \nonumber \\
&&\times\frac{-x^2+\cos^2\theta-(v_{2}/v_{1})^{2}
\sin^2\theta}{\left(x^2 + \cos^2\theta + (v_{2}/v_{1})^{2}
\sin^2\theta\right)^{2}}\mathcal{G}^{B}(x,\theta), \label{Eq:C1B}
\\
C_{2}^{B} &=& \frac{1}{8\pi^3}\int_{-\infty}^{+\infty}dx
\int_{0}^{2\pi}d\theta \nonumber \\
&& \times \frac{-x^2-\cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta}{\left(x^2 + \cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta\right)^{2}} \mathcal{G}^{B}(x,\theta), \label{Eq:C2B}\end{aligned}$$ with $$\begin{aligned}
\mathcal{G}^{B}(x,\theta) = \frac{1}{\frac{1}{2\pi\alpha_{1}} +
\frac{N}{8v_2/v_{1}}\frac{\cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta}{\sqrt{x^2+\cos^2\theta+(v_{2}/v_{1})^{2}
\sin^2\theta}}}. \label{Eq:GExpressionB}\end{aligned}$$ An effective parameter $$\begin{aligned}
\alpha_{1} = \frac{e^2}{\epsilon v_{1}}\end{aligned}$$ is defined to represent the Coulomb interaction strength. The electric charge $e$ is not renormalized due to the absence of logarithmic term in the polarization $\Pi^{B}$ [@Kotov12], and $\epsilon$ takes a constant value in any given sample. The value of $\alpha_1$ is determined by the renormalization of velocity $v_1$.
The coupled flow equations can be simplified. According to Eq. (\[Eq:VRGVGammaRCP\]), we know that $$\begin{aligned}
v_{\Gamma} = v_{\Gamma 0}\end{aligned}$$ is independent of $\ell$ for RSP. Thus we re-write $C_{g}$ as $$\begin{aligned}
C_g = \frac{v_{\Gamma 0}^{2}\Delta}{2\pi v_{1}v_{2}}.\end{aligned}$$ The flow equation for $C_{g}$ is given by $$\begin{aligned}
\frac{dC_{g}}{d\ell} &=& \left(-2C_{0}^{A}-2C_{0}^{B} +
C_{1}^{A}+C_{1}^{B}+C_{2}^{A}+C_{2}^{B}\right.\nonumber \\
&&\left.+2C_{g}\right)C_{g}.\end{aligned}$$ For RVP, from Eqs. (\[Eq:VRGV1\]), (\[Eq:VRGV2\]), (\[Eq:VRGVGammaRVP1\]), and (\[Eq:VRGVGammaRVP2\]), one gets $$\begin{aligned}
\frac{d(v_{\Gamma1}/v_{1})}{d\ell}=0,\qquad
\frac{d(v_{\Gamma2}/v_{2})}{d\ell}=0,\end{aligned}$$ which indicate that $$\begin{aligned}
\frac{v_{\Gamma1}}{v_{1}}=\frac{v_{\Gamma10}}{v_{10}},\qquad
\frac{v_{\Gamma2}}{v_{2}}=\frac{v_{\Gamma20}}{v_{20}}.\end{aligned}$$ Accordingly, $C_g$ now can be written as $$\begin{aligned}
C_g = \frac{\Delta}{2\pi} \left(\frac{v_{\Gamma10}^{2}}{v_{10}^{2}}
\frac{v_{1}}{v_{2}}+\frac{v_{\Gamma20}^{2}}{v_{20}^{2}}
\frac{v_{2}}{v_{1}}\right).\end{aligned}$$ The corresponding RG equation is $$\begin{aligned}
\frac{dC_g}{d\ell} = \frac{\left(v_{\Gamma1}^{2} -
v_{\Gamma2}^{2}\right)\Delta}{2\pi v_{1}v_{2}}\left(-C_{1}^{A} -
C_{1}^{B}+C_{2}^{A} +C_{2}^{B}\right).\end{aligned}$$
For RM, through Eqs. (\[Eq:VRGV1\]), (\[Eq:VRGV2\]), and (\[Eq:VRGVGammaRM\]), we obtain the following flow equation $$\begin{aligned}
\frac{d C_{g}}{d\ell} &=& \left(2C_{0}^{A} + 3C_{1}^{A} + 3C_{2}^{A}
+ 2C_{0}^{B} - C_{1}^{B} - C_{2}^{B}\right.\nonumber
\\
&&\left.-2C_{g}\right)C_{g}.\end{aligned}$$
Quantum critical phenomena \[Sec:NumResults\]
=============================================
In this section, we will solve the RG equations and then apply the solutions to analyze the quantum critical phenomena. We adopt the following steps: first, examine the low-energy behaviors induced solely by the quantum critical fluctuation of excitonic order parameter; second, introduce quenched disorder into the system and study its interplay with the Yukawa coupling; finally, investigate the impact of Coulomb interaction on the results.
Although the RG calculations are carried out at $T=0$, it is possible to extract the $T$-dependence of observable quantities from RG results. We can regard $k_B T$, where $k_B$ is Boltzmann constant, as a free parameter that tunes the energy scale: increasing (decreasing) $T$ amounts to increasing (decreasing) the energy $\omega$. The dependence of observable quantities on $\omega$ and/or $T$ can be computed from the solutions of RG equations as follows. One solves the flow equations at $T=0$ and gets the $\ell$-dependence of model parameters, such as fermion velocities, which leads to the $\ell$-dependence of various observable quantities. On the basis of these results, one converts the $\ell$-dependence of an observable quantity into the $\omega$-dependence of the same quantity at $T=0$ by using the transformation $\omega = \omega_0 e^{-\ell}$, where $\omega_0$ is some high energy, or into the $T$-dependence of the same quantity by using the transformation $T = T_0 e^{-\ell}$, where $T_0$ takes a large value. For examples, the low-energy DOS $\rho(\omega)$ can be directly obtained from $\rho(\ell)$, and the $T$-dependent specific heat $C_{v}(T)$ can be obtained from $C_{v}(\ell)$. This approach has been extensively employed to calculate the $\omega$- and/or $T$-dependence of many observable quantities of Dirac/Weyl fermions subject to the Coulomb interaction [@Goswami11; @Hosur12; @Moon13; @WangLiu14; @Lai15; @Jian15; @Isobe16B; @Cho16; @WangLiuZhang17MWSM; @WangLiuZhangSemiDSM; @Zhang17] and gapless nodal fermions coupled to the nematic quantum fluctuation [@Huh08; @Xu08; @Fritz09; @Wang11; @Liu12; @She15; @WangLiuZhang16NJP].
Non-Fermi liquid behavior induced by excitonic fluctuation \[SubSec:NumResOnlyEIFL\]
------------------------------------------------------------------------------------
If 2D DSM is far from SM-EI transition, the ground state is a robust SM. While the Coulomb interaction is long-ranged, it can only produce normal FL behavior [@Kotov12; @Gonzalez99; @Hofmann14; @Hofmann15]. As the system approaches to the SM-EI QCP, the excitonic fluctuation becomes stronger and eventually invalidates the FL description at $T=0$. Now we illustrate how FL theory breaks down at the QCP by analyzing the solutions of RG equations.
In the clean limit, the excitonic fluctuation leads to the following RG equations $$\begin{aligned}
\frac{dZ_{f}}{d\ell}&=&C_{0}^{A}Z_{f},
\\
\frac{dv_{1}}{d\ell}&=&\left(C_{0}^{A}-C_{1}^{A}\right)v_{1},
\\
\frac{dv_{2}}{d\ell}&=&\left(C_{0}^{A}-C_{2}^{A}\right)v_{2},
\\
\frac{d\left(v_{2}/v_{1}\right)}{d\ell}
&=&\left(C_{1}^{A}-C_{2}^{A}\right)\frac{v_{2}}{v_{1}}.\end{aligned}$$ These equations will be solved in the isotropic and anisotropic cases respectively.
### Isotropic limit
We first consider the isotropic limit, i.e., $v_{1}=v_{2}=v$. In this case, we have $$\begin{aligned}
C_{0}^{A}=C_{1}^{A}=C_{2}^{A} = -\frac{2}{3\pi^{2}N}=-\eta^{A}.
\label{Eq:C1AC2AC3AIsotropic}\end{aligned}$$ Accordingly, the RG equations can be simplified to $$\begin{aligned}
\frac{dZ_{f}}{d\ell} &=& -\eta^{A}Z_{f}, \\
\frac{dv}{d\ell} &=& 0.\end{aligned}$$ The velocity is a constant, i.e., $v = v_{0}$. Thus, the fermion dispersion is unrenormalized, and the dynamical exponents is $z=1$ [@Herbut09]. The specific heat behaves as [@Herbut09] $$\begin{aligned}
C_{v}(T)\sim T^{d/z}\sim T^{2}.\end{aligned}$$ The residue is given by [@Herbut09] $$\begin{aligned}
Z_{f} = Z_{f0}e^{-\eta^{A}\ell}=e^{-\eta^{A}\ell},\end{aligned}$$ which flows to zero quickly in the limit $\ell\rightarrow \infty$. $Z_{f}$ is connected to the real part of retarded self-energy $\mathrm{Re}\Sigma^{R}(\omega)$ via the definition $$\begin{aligned}
Z_{f} = \frac{1}{\left|1-\frac{\partial}{\partial\omega}
\mathrm{Re}\Sigma^{R}(\omega)\right|}.\end{aligned}$$ Employing the transformation $\omega=\omega_{0}e^{-\ell}$, we get the following expression $$\begin{aligned}
\mathrm{Re}\Sigma^{R}(\omega)\sim\omega^{1-\eta^{A}}.\end{aligned}$$ Using the Kramers-Kronig relation, we can easily obtain the imaginary part $$\begin{aligned}
\mathrm{Im}\Sigma^{R}(\omega)\sim\omega^{1-\eta^{A}},\end{aligned}$$ which exhibits typical NFL behavior. The renormalized DOS depends on $\omega$ as follows $$\begin{aligned}
\rho(\omega)\sim\omega^{1+\eta^{A}}.\end{aligned}$$
### Anisotropic case
In the generic anisotropic case, namely $v_{1}\neq v_{2}$, we integrate over variable $x$ in Eqs. (\[Eq:C0A\])-(\[Eq:C2A\]) and find $$\begin{aligned}
C_{0}^{A} &=&-\frac{v_{2}/v_{1}}{3\pi^3N}\int_{0}^{2\pi}d\theta
\frac{1}{\left(\cos^2\theta +
(v_{2}/v_{1})^{2}\sin^2\theta\right)}\nonumber \\
&=&-\frac{v_{2}/v_{1}}{3\pi^3N}\frac{2\pi}{v_{2}/v_{1}} = -\eta^{A},
\\
C_{1}^{A} &=& \frac{v_{2}/v_{1}}{3\pi^3 N}\int_{0}^{2\pi}d\theta
\frac{\cos^2\theta-3(v_{2}/v_{1})^{2}\sin^2\theta}{\left(\cos^2\theta
+ (v_{2}/v_{1})^{2}\sin^2\theta\right)^{2}}\nonumber
\\
&=&\frac{v_{2}/v_{1}}{3\pi^3 N}\left(-\frac{2 \pi}{v_{2}/v_{1}}
\right) = -\eta^{A}, \\
C_{2}^{A} &=&\frac{v_{2}/v_{1}}{3\pi^3 N}\int_{0}^{2\pi}d\theta
\frac{-3\cos^2\theta+(v_{2}/v_{1})^{2}\sin^2\theta}{\left(\cos^2
\theta + (v_{2}/v_{1})^{2}\sin^2\theta\right)^{2}} \nonumber \\
&=&\frac{v_{2}/v_{1}}{3\pi^3 N}\left(-\frac{2 \pi}{v_{2}/v_{1}}
\right)=-\eta^{A},\end{aligned}$$ which are exactly the same as the isotropic case. Accordingly, the RG equations for $v_{1}$ and $v_{2}$ are $$\begin{aligned}
\frac{dv_{1}}{d\ell} = \frac{dv_{2}}{d\ell} = 0,\end{aligned}$$ which implies that $$\begin{aligned}
v_{1}=v_{10},\quad v_{2}=v_{20}.\end{aligned}$$ Thus, the fermion velocities are not renormalized, and the anisotropy is not changed by the Yukawa coupling. The low-energy properties of specific heat $C_{v}(T)$, residue $Z_{f}$, fermion damping rate $|\mathrm{Im}\Sigma^{R}(\omega)|$, and DOS $\rho(\omega)$ are the same as those obtained in the isotropic case.
Excitonic fluctuation and disorder
----------------------------------
We then include disorder and examine how it affects the above results. Now the coupled RG equations of $Z_{f}$, $v_{1}$, $v_{2}$, and $v_{2}/v_{1}$ are $$\begin{aligned}
\frac{dZ_{f}}{d\ell} &=& \left(C_{0}^{A}-C_{g}\right)Z_{f} =
-\left(\eta^{A}+C_{g}\right)Z_{f}, \label{Eq:VRGZfQFDis}
\\
\frac{dv_{1}}{d\ell} &=& \left(C_{0}^{A}-C_{1}^{A}-C_{g}\right)v_{1}
= -C_{g}v_{1}, \label{Eq:VRGV1QFDis}
\\
\frac{dv_{2}}{d\ell} &=& \left(C_{0}^{A}-C_{2}^{A} -
C_{g}\right)v_{2}=-C_{g}v_{2}, \label{Eq:VRGV2QFDis}
\\
\frac{d\left(v_{2}/v_{1}\right)}{d\ell} &=& \left(C_{1}^{A} -
C_{2}^{A}\right)\frac{v_{2}}{v_{1}}=0. \label{Eq:VRGVRatioQFDis}\end{aligned}$$
For RSP, $C_{g}$ satisfies $$\begin{aligned}
\frac{dC_{g}}{d\ell} = 2C_{g}^{2}, \label{Eq:VRGCgRCPQFDis}\end{aligned}$$ whose solution is $$\begin{aligned}
C_{g} = \frac{C_{g0}}{1-2C_{g0}\ell}. \label{Eq:SolutionCgRCPQFDis}\end{aligned}$$ It is clear that this $C_{g}$ diverges as $\ell \rightarrow
\ell_{c}$, where $\ell_{c} = 1/2C_{g0}$. Substituting Eq. (\[Eq:SolutionCgRCPQFDis\]) into Eqs. (\[Eq:VRGZfQFDis\])-(\[Eq:VRGV2QFDis\]), we obtain $$\begin{aligned}
Z_{f}&=&e^{-\eta^{A}\ell}\sqrt{1-2C_{g0}\ell},
\\
v_{1}&=&v_{10}\sqrt{1-2C_{g0}\ell},
\\
v_{2}&=&v_{20}\sqrt{1-2C_{g0}\ell}.\end{aligned}$$ We can see that, $Z_{f}$, $v_1$, and $v_2$ all flow to zero as $\ell
\rightarrow \ell_{c}$. Such singular behaviors are generally believed to indicate the instability of the system: RSP drives the system into a disorder-dominated CDM. The characteristic feature of CDM is that, the fermions acquire a finite disorder scattering rate $$\begin{aligned}
\gamma_{\mathrm{imp}} = \left|\mathrm{Im}\Sigma^{R}(0)\right|.\end{aligned}$$ In the meantime, the zero-energy DOS $\rho(0)$ also becomes finite, being a function of $\gamma_{\mathrm{imp}}$. According to the calculations given in Refs.[@WangLiu14; @WangLiuZhang16NJP], the specific heat displays a linear-in-$T$ behavior, namely $$\begin{aligned}
C_{v}(T)\sim T.\end{aligned}$$ The NFL quantum critical state realized in the clean limit is turned into a CDM once RSP is added to the system, even when RSP is very weak. The fermion damping effect, the low-energy DOS, and the specific heat of CDM phase are all distinct from those of the NFL phase.
For RVP, the RG equation for $C_{g}$ is $$\begin{aligned}
\frac{dC_g}{d\ell}=\frac{\left(v_{\Gamma1}^{2}-v_{\Gamma2}^{2}
\right)\Delta}{2\pi v_{1}v_{2}} \left(-C_{1}^{A}+C_{2}^{A}\right) =
0, \label{Eq:VRGCgRGPQFDis}\end{aligned}$$ implying that $$\begin{aligned}
C_{g} = C_{g0}. \label{Eq:SolutionCgRGPQFDis}\end{aligned}$$ Substituting Eq. (\[Eq:SolutionCgRGPQFDis\]) into Eqs. (\[Eq:VRGZfQFDis\])-(\[Eq:VRGV2QFDis\]) yields $$\begin{aligned}
Z_{f}&=&e^{-\left(\eta^{A}+C_{g0}\right)\ell}, \\
v_{1}&=&v_{10}e^{-C_{g0}\ell}, \\
v_{2}&=&v_{20}e^{-C_{g0}\ell}.\end{aligned}$$ The real and imaginary parts of retarded fermion self-energy are $$\begin{aligned}
\mathrm{Re}\Sigma^{R}(\omega)\sim \omega^{1 -
\left(\eta^{A}+C_{g0}\right)}, \\
\mathrm{Im}\Sigma^{R}(\omega)\sim \omega^{1 -
\left(\eta^{A}+C_{g0}\right)},\end{aligned}$$ which are still NFL-like behaviors. Comparing to the clean limit, $Z_{f}$ approaches to zero more quickly and the fermion damping becomes stronger. The velocity $v$ goes to zero rapidly with growing $\ell$, thus the fermion dispersion is substantially altered. In addition, the dynamical exponent $z$ becomes $z = 1+C_{g0}$. It is easy to find that, the specific heat is $$\begin{aligned}
C_{v}(T)\sim T^{d/z}\sim T^{2/(1+C_{g0})},\end{aligned}$$ and the low-energy DOS is $$\begin{aligned}
\rho(\omega)\sim \omega^{(1-C_{g0})/(1+C_{g0})+\eta^{A}}.\end{aligned}$$ An apparent conclusion is that both DOS and specific heat are enhanced by RVP at low energies.
For RM, the RG equation for $C_{g}$ becomes $$\begin{aligned}
\frac{dC_{g}}{d\ell} = -8\eta^{A}C_{g}-2C_{g}^{2}.
\label{Eq:VRGCgRMQFDis}\end{aligned}$$ Its solution is $$\begin{aligned}
C_{g}(\ell) = \frac{4\eta^{A}C_{g0}}{\left(C_{g0} + 4\eta^{A}\right)
e^{8\eta^{A}\ell}-C_{g0}}, \label{Eq:SolutionCgRMQFDis}\end{aligned}$$ which vanishes in the limit $\ell \rightarrow \infty$. Substituting Eq. (\[Eq:SolutionCgRMQFDis\]) into Eqs. (\[Eq:VRGZfQFDis\])-(\[Eq:VRGV2QFDis\]), we get $$\begin{aligned}
Z_{f} &=& e^{-\eta^{A}\ell}\sqrt{\frac{4\eta^{A}}{C_{g0} + 4\eta^{A}
- C_{g0}e^{-8\eta^{A}\ell}}}, \label{Eq:VRGZfCgRMQFDis}
\\
v_{1} &=& v_{10}\sqrt{\frac{4\eta^{A}}{C_{g0} + 4\eta^{A} -
C_{g0}e^{-8\eta^{A}\ell}}}, \label{Eq:VRGV1CgRMQFDis}
\\
v_{2} &=& v_{20}\sqrt{\frac{4\eta^{A}}{C_{g0} + 4\eta^{A} -
C_{g0}e^{-8\eta^{A}\ell}}}. \label{Eq:VRGV2CgRMQFDis}\end{aligned}$$ In the low-energy regime, the residue still behaves as $Z_{f} \sim
e^{-\eta^{A}\ell}$. From the $\ell$-dependence of $Z_{f}$, we obtain $$\begin{aligned}
\mathrm{Re}\Sigma^{R}(\omega) &\sim& \omega^{1-\eta^{A}},
\\
\mathrm{Im}\Sigma^{R}(\omega)&\sim&\omega^{1-\eta^{A}},\end{aligned}$$ which are the same as the clean case. As shown by Eqs. (\[Eq:VRGV1CgRMQFDis\]) and (\[Eq:VRGV2CgRMQFDis\]), $v_{1}$ and $v_{2}$ approach to finite values in the lowest energy limit. Accordingly, the fermion DOS still exhibits the behavior $\rho(\omega)\sim\omega^{1+\eta^{A}}$, and the specific heat is still of the form $C_{v}(T)\sim T^{2}$. We thus see that RM does not qualitatively change the low-energy properties of observable quantities.
The above RG results indicate that, the low-energy properties of the SM-EI QCP depend heavily on the disorder type. Such properties can be experimentally probed by measuring observable quantities, such as DOS and specific heat. However, we should remember that the long-range Coulomb interaction is entirely ignored in the above RG analysis. This might miss important quantum many-body effects. In the next subsection, we will study whether or not the above results are substantially altered when the Coulomb interaction is incorporated.
Interplay of three kinds of interaction
---------------------------------------
We now analyze the physical consequence of the interplay of all the three kinds of interaction, first in the isotropic limit and then in the more generic anisotropic case. We will see that the Coulomb interaction tends to suppress the fermion velocity anisotropy.
![Flowing behavior of $Z_{f}$ and $v$ caused by excitonic fluctuation and Coulomb interaction. In this and all the subsequent figures, we assume $N=2$ in numerical calculations.[]{data-label="Fig:VRGIsoClean"}](VRGIsoClean.eps){width="3.38in"}
![Flowing behavior of $Z_{f}$, $v$, $\alpha$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RSP. Blue, red, green, black, and magenta lines correspond to $C_{g0} = 0.08,
0.1, 0.12, 0.14, 0.16$. Here, $\alpha_{10}=1.0$. \[Fig:VRGIsoRSP\]](VRGIsoRSP.eps){width="3.38in"}
### Isotropic limit
In the isotropic limit with $v_{1} = v_{2} = v$, the RG equations for $Z_{f}$ and $v$ are $$\begin{aligned}
\frac{dZ_{f}}{d\ell} &=& \left(-\eta^{A}+C_{0}^{B}-C_{g}\right)
Z_{f}, \label{Eq:VRGZfQFDisCoulombIso}
\\
\frac{dv}{d\ell}&=&\left(C^{B}-C_{g}\right)v.
\label{Eq:VRGVQFDisCoulombIso}\end{aligned}$$ Here, $C^{B} = C_{0}^{B}-C_{1}^{B} = C_{0}-C_{2}^{B}$, in which $$\begin{aligned}
C_{0}^{B} &=& \frac{4}{N\pi^2}\left[2-\frac{1}{\lambda}\pi +
\frac{2-\lambda^{2}}{\lambda}f(\lambda)\right], \\
C_{1,2}^{B} &=& \frac{4}{N\pi^2}\left[1-\frac{1}{\lambda}
\frac{\pi}{2}+\frac{1-\lambda^{2}}{\lambda}f(\lambda)\right].\end{aligned}$$ The variable $\lambda$ is $\lambda = N\pi\alpha/4$, and the function $f(\lambda)$ is $$\begin{aligned}
f(\lambda)=\left\{
\begin{array}{ll}
\frac{1}{\sqrt{1-\lambda^2}}\arccos\left(\lambda\right) & \lambda<1
\\
\\
\frac{1}{\sqrt{\lambda^2-1}}\mathrm{arccosh}\left(\lambda\right) &
\lambda>1
\\
\\
1 & \lambda=1.
\end{array}
\right.\end{aligned}$$
In the clean limit, $Z_{f}$ and $v$ flow as follows $$\begin{aligned}
\frac{dZ_{f}}{d\ell} &=& \left(-\eta^{A}+C_{0}^{B}\right)Z_{f},
\label{Eq:VRGZfQFCoulombIso} \\
\frac{dv}{d\ell} &=& C^{B}v. \label{Eq:VRGVQFCoulombIso}\end{aligned}$$ The numerical solutions are shown in Fig. \[Fig:VRGIsoClean\]. The velocity $v$ increases as the energy is lowered. The Coulomb interaction is marginally irrelevant since its strength parameter $\alpha = e^{2}/v\epsilon$ flows to zero slowly in the lowest energy limit. Both $C_{0}^{B}$ and $C^{B}$ vanish as $\alpha \rightarrow
0$. The velocity renormalization produces logarithmic-like correction to the temperature or energy dependence of some observable quantities, including specific heat and compressibility [@Kotov12]. The singular renormalization of fermion velocities has been observed by various experimental tools [@Elias11; @Siegel11; @Chae12; @Yu13]. At low energies, $C_{0}^{B}$ is much smaller than $\eta^{A}$. Thus, the Coulomb interaction only slightly alters the low-energy behavior of $Z_{f}$ induced by the excitonic fluctuation.
For RSP, the RG equation of $C_{g}$ is $$\begin{aligned}
\frac{dC_{g}}{d\ell} = \left(-2C^{B}+2C_{g}\right)C_{g}.\end{aligned}$$ For a given $\alpha_0$, there exists a critical value $C^B(\alpha_0)$. The system exhibits entirely different low-energy properties when $C_{g0}$ is greater and smaller than $C^{B}(\alpha_{0})$. To illustrate this, we show the $\ell$-dependence of $Z_f$, $v$, $\alpha$, and $C_g$ in Fig. \[Fig:VRGIsoRSP\]. If $C_{g0} < C^{B}(\alpha_{0})$, $Z_f$, $\alpha$, and $C_{g}$ all flow to zero as $\ell \rightarrow \infty$, but $v$ increases with growing $\ell$. These results indicate that weak RSP is suppressed by the Coulomb interaction. If $C_{g0}>C^{B}(\alpha_{0})$, both $C_{g}$ and $\alpha$ formally diverge at some finite energy scale, whereas both $Z_f$ and $v$ decrease rapidly down to zero at the same energy scale. Thus, strong RSP still drives a NFL-to-CDM transition. As can be seen from the flow diagram presented in Fig. \[Fig:FlowDiagrams\](a), the $(\alpha, C_{g})$ plane is divided by the critical line $C_{g0} =
C^{B}(\alpha_{0})$ into two distinct phases: the NFL phase and the CDM phase.
![Flowing diagrams on the $\alpha$-$C_{g}$ plane. Result for RSP is in (a), RVP in (b), and RM in (c). \[Fig:FlowDiagrams\]](FlowIsoRSP.eps "fig:"){width="2.38in"} ![Flowing diagrams on the $\alpha$-$C_{g}$ plane. Result for RSP is in (a), RVP in (b), and RM in (c). \[Fig:FlowDiagrams\]](FlowIsoRVP.eps "fig:"){width="2.38in"} ![Flowing diagrams on the $\alpha$-$C_{g}$ plane. Result for RSP is in (a), RVP in (b), and RM in (c). \[Fig:FlowDiagrams\]](FlowIsoRM.eps "fig:"){width="2.38in"}
For RVP, the $\ell$-dependence of $Z_f$, $v$, $\alpha$, and $C_g$ are shown in Fig. \[Fig:VRGIsoRVP\]. The parameter $C_{g}$ does not flow at all, namely $$\begin{aligned}
\frac{dC_{g}}{d\ell} = 0.\end{aligned}$$ We fix $C_g$ at a constant: $C_{g} = C_{g0}$. For a given $C_{g0}$, $v$ approaches to a constant value $v^{*}$ in the zero energy limit. The value of $v^{*}$ is obtained from $$\begin{aligned}
C^{B}(\alpha^{*}) = C_{g0},\end{aligned}$$ where $\alpha^{*} = e^{2}/v^{*}\epsilon$. RG analysis indicates that the system always flows to a stable infrared fixed point for any two given initial values of $\alpha$ and $C_g$. Connecting all of these fixed points forms a critical line on the $\alpha$-$C_g$ plane, as shown in Fig. \[Fig:FlowDiagrams\](b). Near the critical line, the specific heat behaves as $$\begin{aligned}
C_{v}(T)\sim \frac{1}{{v^{*}}^{2}}T^{2} \sim T^{2}.\end{aligned}$$ The residue is $$\begin{aligned}
Z_{f} &\sim& e^{\left(-\eta^{A} + C_{0}^{B}(\alpha^{*}) -
C_{g0}\right)\ell}\nonumber \\
&\sim& e^{\left(-\eta^{A}+C_{1}^{B}(\alpha^{*})\right)\ell},\end{aligned}$$ where $C_{1}^{B}(\alpha^{*})$ is negative. This $Z_{f}$ flows to zero more quickly than that induced purely by excitonic fluctuation. The retarded fermion self-energy is $$\begin{aligned}
\mathrm{Re}\Sigma^{R}(\omega)&\sim& \omega^{1-\left(\eta^{A}
- C_{1}^{B}(\alpha^{*})\right)}, \\
\mathrm{Im}\Sigma^{R}(\omega)&\sim& \omega^{1 -
\left(\eta^{A}-C_{1}^{B}(\alpha^{*})\right)}.\end{aligned}$$ The DOS takes the form $$\begin{aligned}
\rho(\omega) \sim \omega^{1 + \eta^{A} - C_{1}^{B}(\alpha^{*})}.\end{aligned}$$
![Flowing behavior of $Z_{f}$, $v$, $\alpha$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RVP. Blue, red, green, black, and magenta lines correspond to $C_{g0} = 0.08,
0.1, 0.12, 0.14, 0.16$. Here, $\alpha_{10}=1.0$. \[Fig:VRGIsoRVP\]](VRGIsoRVP.eps){width="3.3in"}
![Flowing behavior of $Z_{f}$, $v$, $\alpha$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RM. Blue, red, green, black, and magenta lines correspond to $C_{g0}=0.08,
0.15, 0.2, 0.25, 0.3$. Here, $\alpha_{10}=1.0$. \[Fig:VRGIsoRM\]](VRGIsoRM.eps){width="3.3in"}
For RM, the RG equation for $C_{g}$ is given by $$\begin{aligned}
\frac{d C_{g}}{d\ell} = \left(-8\eta^{A}+2C^{B}-2C_{g}\right)C_{g}.\end{aligned}$$ The numerical results are plotted in Fig. \[Fig:VRGIsoRM\]. We observe that $C_{g}$ always approaches to zero quickly, which indicates that RM is irrelevant in the low-energy regime. The Coulomb interaction is marginally irrelevant and leads to singular renormalization of fermion velocity. Accordingly, the DOS and specific heat are $$\begin{aligned}
\rho(\omega) &\sim& \frac{\omega^{1 + \eta^{A}
}}{\ln^{2}(\omega_{0}/\omega)}, \\
C_{v}(T) &\sim& \frac{T^{2}}{\ln^{2}(T_{0}/T)}.\end{aligned}$$ In the presence of RM, the two parameters $(\alpha, C_{g})$ always flow to the stable infrared fixed point $(0,0)$.
[|c|c|c|c|c|c|]{} & $Z_{f}(\ell)$ & $\mathrm{Im}\Sigma^{R}(\omega)$ & $\rho(\omega)$ & $C_{v}(T)$\
&
---------------------
$e^{-\eta^{A}\ell}$
[@Herbut09]
---------------------
&
-----------------------
$\omega^{1-\eta^{A}}$
[@Herbut09]
-----------------------
& $\omega^{1+\eta^{A}}$ &
-------------
$T^{2}$
[@Herbut09]
-------------
\
& $e^{-\eta^{A}\ell}\sqrt{1-2C_{g0}\ell}$ & $\gamma_{\mathrm{imp}}$ & $ \gamma_{\mathrm{imp}}\ln(v\Lambda/\gamma_{\mathrm{imp}})$ & $\rho(0)T$\
& $e^{-\left(\eta^{A}+C_{g0}\right)\ell}$ & $\omega^{1-\left(\eta^{A}+C_{g0}\right)}$ & $
\;\omega^{(1-C_{g0})/(1+C_{g0})+\eta^{A}}\;$ & $T^{2/(1+C_{g0})}$\
& $e^{-\eta^{A}\ell}$ & $\omega^{1-\eta^{A}}$ & $\omega^{1+\eta^{A}}$ & $T^{2}$\
& $e^{-\eta^{A}\ell} $ & $\omega^{1-\eta^{A}}$ & $\omega^{1+\eta^{A}}/\ln^2(\omega_{0}/\omega)$ & $\;
T^{2}/\ln^{2}\left(T_{0}/T\right)\; $\
& $C_{g0}<
C_{B}(\alpha_{0})$ & $e^{-\eta^{A}\ell}$ & $\omega^{1-\eta^{A}}$ & $\omega^{1+\eta^{A}}/\ln^{2}(\omega_{0}/\omega)$ & $T^{2}/\ln^{2}(T_{0}/T)$\
& $C_{g0}> C_{B}(\alpha_{0})$ & $\lim_{\ell\rightarrow
l_{c}}Z_{f}(\ell)\rightarrow0$ & $\gamma_{\mathrm{imp}}$ & $
\gamma_{\mathrm{imp}}\ln(v\Lambda/\gamma_{\mathrm{imp}})$ & $\rho(0)T$\
& $e^{\left(-\eta^{A}+C_{1}^{B}(\alpha^{*})\right)\ell}$ & $\omega^{1-\left(\eta^{A}-C_{1}^{B}(\alpha^{*})\right)}$ & $\omega^{1+\eta^{A}-C_{1}^{B}(\alpha^{*})}$ & $T^{2}$\
& $e^{-\eta^{A}\ell}$ & $\omega^{1-\eta^{A}}$ & $\omega^{1+\eta^{A}}/\ln^{2}(\omega_{0}/\omega)$ & $T^{2}/\ln^{2}\left(T_{0}/T\right)$\
&
-------------------------------------------------------
$\;\lim_{\ell\rightarrow\infty}Z_{f}(\ell)\rightarrow
\mathrm{Const.}\;$
[@Kotov12; @Gonzalez99; @Hofmann14; @WangLiu14]
-------------------------------------------------------
&
-------------------------------------
$\omega/\ln^{2}(\omega_{0}/\omega)$
[@Kotov12; @WangLiu14]
-------------------------------------
&
-------------------------------------
$\omega/\ln^{2}(\omega_{0}/\omega)$
[@Kotov12; @WangLiu14]
-------------------------------------
&
--------------------------
$T^{2}/\ln^{2}(T_{0}/T)$
[@Kotov12; @WangLiu14]
--------------------------
\
& $C_{g0}< C_{B}(\alpha_{0})$ &
----------------------------------------------------------------------
$\lim_{\ell\rightarrow\infty}Z_{f}(\ell)\rightarrow \mathrm{Const.}$
[@WangLiu14]
----------------------------------------------------------------------
&
-------------------------------------
$\omega/\ln^{2}(\omega_{0}/\omega)$
[@WangLiu14]
-------------------------------------
&
-------------------------------------
$\omega/\ln^{2}(\omega_{0}/\omega)$
[@WangLiu14; @Stauber05]
-------------------------------------
&
--------------------------
$T^{2}/\ln^{2}(T_{0}/T)$
[@WangLiu14; @Stauber05]
--------------------------
\
& $C_{g0} > C_{B}(\alpha_{0})$ &
-------------------------------------------------------
$\lim_{\ell\rightarrow l_{c}}Z_{f}(\ell)\rightarrow0$
[@WangLiu14]
-------------------------------------------------------
&
-------------------------
$\gamma_{\mathrm{imp}}$
[@Sarma11]
-------------------------
&
------------------------------------------------------------
$\gamma_{\mathrm{imp}}\ln(v\Lambda/\gamma_{\mathrm{imp}})$
[@Sarma11]
------------------------------------------------------------
&
------------
$\rho(0)T$
[@Sarma11]
------------
\
&
---------------------------------
$e^{C_{1}^{B}(\alpha^{*})\ell}$
[@WangLiu14]
---------------------------------
&
------------------------------------
$\omega^{1+C_{1}^{B}(\alpha^{*})}$
[@WangLiu14]
------------------------------------
&
------------------------------------
$\omega^{1-C_{1}^{B}(\alpha^{*})}$
[@WangLiu14]
------------------------------------
&
-----------------------------------------------
$T^{2}$
[@WangLiu14; @Stauber05; @Herbut08; @Vafek08]
-----------------------------------------------
\
&
---------------------------------
$e^{C_{1}^{B}(\alpha^{*})\ell}$
[@WangLiu14]
---------------------------------
&
------------------------------------
$\omega^{1+C_{1}^{B}(\alpha^{*})}$
[@WangLiu14]
------------------------------------
&
------------------------------------
$\omega^{1-C_{1}^{B}(\alpha^{*})}$
[@WangLiu14]
------------------------------------
&
-----------------------------------------------
$T^{2}$
[@WangLiu14; @Stauber05; @Herbut08; @Vafek08]
-----------------------------------------------
\
We now compare the quantum critical phenomena to the physical properties of the SM phase. Deep in the SM phase, the excitonic fluctuation can be completely ignored. The low-energy behavior is governed by the interplay of Coulomb interaction and disorder, which has already been extensively investigated [@WangLiu14; @Ye98; @Ye99; @Stauber05; @Herbut08; @Vafek08; @Foster08]. When the Coulomb interaction and RSP are both present, the system is a normal FL if RSP is weak, but is turned into a CDM phase by strong RSP. Thus, increasing the effective strength of RSP drives a FL-CDM phase transition. In the SM-EI quantum critical regime, increasing the effective strength of RSP leads to a NFL-CDM transition. If RM is added to the system, it is irrelevant around the SM-EI QCP, but is marginal and results in a stable critical line on the $\alpha$-$C_g$ plane deep in the SM phase. In contrast, RVP produces the same qualitative low-energy behaviors in the SM phase and around the SM-EI QCP.
We learn from the above analysis that, even if 2D DSM has a gapless SM ground state, the fluctuation of excitonic order parameter gives rise to observable effects at finite $T$ and/or $\omega$. The quantum critical regime can be distinguished from the pure SM phase by measuring the $\omega$-dependence of fermion damping rate and/or the $T$-dependence of specific heat.
To provide a complete analysis of the quantum critical phenomena, we summarize in Table \[Table:SummaryObQuant\] the low-energy properties induced by all the possible combinations of three types of interaction. The quantities presented in Table \[Table:SummaryObQuant\] include the residue $Z_f$, damping rate $\mathrm{Im}\Sigma^{R}(\omega)$, fermion DOS $\rho(\omega)$, and specific heat $C_{v}(T)$. We can see that distinct interactions affect each other significantly. The critical phenomena cannot be reliably determined if their mutual influence is not carefully handled.
### Anisotropic case
For different values of fermion velocity ratio, the running behaviors of $Z_{f}$, $v_{1}$, $v_{2}$, and $v_{2}/v_{1}$ obtained in the clean limit are plotted in Figs. \[Fig:VRGAniClean\](a)-(d), respectively. Firstly, $Z_{f}$ flows to zero very quickly, implying the violation of FL description. This is essentially induced by the excitonic quantum fluctuation, because the Coulomb interaction by itself would yield a finite $Z_f$. Secondly, the two fermion velocities $v_1$ and $v_2$ both increase as the energy is lowered, whereas the velocity ratio $v_2/v_1$ flows to unity in the lowest energy limit. Remember that the excitonic quantum fluctuation does not renormalize fermion velocities at all, as illustrated in Sec. \[SubSec:NumResOnlyEIFL\]. It is clear that the renormalization of $v_{1}$ and $v_{2}$ are mainly determined by the Coulomb interaction. These results indicate that both excitonic fluctuation and Coulomb interaction are important in the low-energy region.
![Flowing behavior of $Z_{f}$, $v_{1}$, $v_{2}$, and $v_{2}/v_{1}$ caused by excitonic fluctuation and Coulomb interaction in the anisotropic case. Blue, red, green, black, and magenta lines correspond to $v_{20}/v_{10}=10, 5, 1, 0.5, 0.1$. We choose $\alpha_{10} = 1.0$. As $\ell \rightarrow \infty$, the system flows to the isotropic limit. \[Fig:VRGAniClean\]](VRGAniClean.eps){width="3.3in"}
![Flowing behavior of $Z_{f}$, $v_{1}$, $v_{2}/v_{1}$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RSP. Blue, red, green, black, and magenta lines correspond to $v_{20}/v_{10} = 10, 5, 1, 0.5, 0.1$. Here, $\alpha_{10}=1.0$ and $C_{g0}=0.1$. \[Fig:VRGAniRSP\]](VRGAniRSP.eps){width="3.3in"}
After including three types of disorder, we find that the system still flows to the isotropic limit in the zero energy limit. The numerical results obtained in the cases of RSP, RVP, and RM are presented in Fig. \[Fig:VRGAniRSP\], Fig. \[Fig:VRGAniRVP\], and Fig. \[Fig:VRGAniRM\], respectively.
First, we consider the case of RSP. As shown in Fig. \[Fig:VRGAniRSP\], for given values of $\alpha_{10}$ and $C_{g0}$, $C_{g}$ becomes divergent at some finite energy scale if the bare velocity ratio $v_{20}/v_{10}$ exceeds a critical value. Both $Z_{f}$ and fermion velocities flow to zero at the same energy scale. The anisotropy is suppressed, but the ratio does not flow to the isotropic limit. If the bare value $v_{20}/v_{10}$ is small, $C_{g0}$ flows to zero quickly as the energy is lowered. Meanwhile, the fermion velocities increase, and the ratio $v_{2}/v_{1}\rightarrow 1$. Apparently, the isotropic limit is mainly driven by the Coulomb interaction. The residue $Z_{f}$ still vanishes, owing to the excitonic fluctuation. For given values of $\alpha_{10}$ and $C_{g0}$, varying the velocity ratio $v_{20}/v_{10}$ leads to QPT between CDM phase and NFL phase.
![Flowing behavior of $Z_{f}$, $v_{1}$, $v_{2}/v_{1}$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RVP. Blue, red, green, black, and magenta lines correspond to $v_{20}/v_{10}=5, 2, 1, 0.5, 0.2$. Here, $\alpha_{10}=1.0$, $\Delta/2\pi = 0.05$, $v_{\Gamma10}/v_{10} = 1$, and $v_{\Gamma20}/v_{20} = 1$. \[Fig:VRGAniRVP\]](VRGAniRVP.eps){width="3.3in"}
![Flowing behavior of $Z_{f}$, $v_{1}$, $v_{2}/v_{1}$, and $C_{g}$ caused by excitonic fluctuation, Coulomb interaction, and RM. Blue, red, green, black, and magenta lines correspond to $v_{20}/v_{10}=10, 5, 1, 0.5, 0.1$. Here, $\alpha_{10}=1.0$ and $C_{g0}=0.1$. \[Fig:VRGAniRM\]](VRGAniRM.eps){width="3.3in"}
In the case of RVP, we show the evolution of $Z_{f}$, $v_{1}$, $v_{2}/v_{1}$, and $C_{g}$ in Fig. \[Fig:VRGAniRVP\]. Comparing to the clean limit, the ratio $v_{2}/v_{1}$ approaches to unity more quickly. This should be attributed to the fact that the Coulomb interaction strength $\alpha$ flows to certain finite value in the presence of RVP but vanishes in the clean limit. Therefore, the suppression of velocity anisotropy is more significant once RVP is introduced.
We finally turn to the impact of RM. According to Fig. \[Fig:VRGAniRM\], the disorder parameter $C_{g}$ of RM always flows to zero quickly with decreasing energy. The low-energy behaviors of $Z_{f}$ and $v_{1}$ are nearly the same as those obtained in the clean limit, and the velocity ratio $v_{2}/v_{1}
\rightarrow 1$ as the energy is lowered down to zero.
Summary and Discussion \[Sec:SummaryDiscuss\]
=============================================
In summary, we have presented a systematic study of the quantum critical phenomena around the SM-EI QCP in 2D DSM. The Yukawa coupling between Dirac fermions and excitonic quantum fluctuation, the long-range Coulomb interaction, and the disorder scattering are treated on equal footing, focusing on their mutual influence and the consequent low-energy properties of the quantum critical regime. We first studied the influence of quantum critical fluctuation of excitonic order parameter, and showed that it invalidates the FL description. We further demonstrated that, adding RSP always drives a NFL-to-CDM transition, and adding RVP further reinforces the NFL behaviors. Nevertheless, adding RM does not change the qualitative results obtained in the clean limit. Once Coulomb interaction is also incorporated, the above results are altered. In particular, the NFL state is protected by the Coulomb interaction for weak RSP, but is eventually replaced by CDM state if RSP is strong enough. When RVP or RM coexist with excitonic fluctuation and Coulomb interaction, the system is in a NFL state. To characterize the NFL and CDM phases, we have calculated several quantities, including the residue, damping rate, fermion DOS, and specific heat. The predicted quantum critical phenomena can be directly probed by experiments.
The results obtained in this paper might be applied to judge whether or not a 2D DSM is close to the SM-EI QCP. Deep in the gapless SM phase, the properties of the system are determined by the combination of Coulomb interaction and disorder. As the system approaches the SM-EI QCP, i.e., $\alpha \rightarrow \alpha_c$, the excitonic quantum fluctuation becomes progressively more important, driving the system to enter into the quantum critical regime. Even when the zero-$T$ ground state is gapless, the system could exhibit nontrivial quantum critical behaviors in the $\omega$- and/or $T$-dependence of observable quantities, as illustrated in Fig. \[Fig:PhaseEIQCP\] and Table \[Table:SummaryObQuant\].
We finally give a brief remark on the existence of the excitonic QCP in realistic graphene. For a 2D DSM, all the previous analytical and numerical calculations [@CastroNetoPhysics09; @Khveshchenko01; @Gorbar02; @Khveshchenko04; @Liu09; @Khveshchenko09; @Gamayun10; @Sabio10; @Zhang11; @Liu11; @WangLiu11A; @WangLiu11B; @WangLiu12; @Popovici13; @WangLiu14; @Gonzalez15; @Carrington16; @Sharma17; @Xiao17; @Carrington18; @Gamayun09; @WangJianhui11; @Katanin16; @Gonzalez10; @Gonzalez12; @Drut09A; @Drut09B; @Drut09C; @Armour10; @Armour11; @Buividovich12; @Ulybyshev13; @Smith14; @Juan12; @Kotikov16; @Gonzalez14; @Braguta16; @Xiao18] have confirmed that an excitonic gap is generated only when $\alpha > \alpha_c$, where $\alpha_{c}$ is a nonzero critical value. Recent theoretical studies revealed that the physical value of $\alpha$ in suspendend graphene is not far from the critical value $\alpha_{c}$ [@Ulybyshev13; @Carrington18]. The system would become even closer to the excitonic QCP when strain is applied [@Tang15; @Sharma17; @Xiao17]. The organic conductor $\alpha$-(BEDT-TTF)$_{2}$I$_{3}$, an anisotropic 2D DSM, may also be close to the excitonic QCP [@Hirata17]. The theoretical results obtained in this work could be utilized to explore the quantum critical phenomena around the putative excitonic QCP in 2D DSM materials.
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
We thank Jing Wang and Peng-Lu Zhao for helpful discussions. We acknowledge the financial support by the National Natural Science Foundation of China under Grants 11275100, 11504379, and 11574285. X.Y.P. also acknowledges the support by the K. C. Wong Magna Foundation in Ningbo University. J.R.W. is partly supported by the Natural Science Foundation of Anhui Province under Grant 1608085MA19.
Polarization functions
======================
We now calculate the polarization functions caused by the particle-hole collective excitations. There are two polarization functions, corresponding to the dynamical screening effects of the quantum critical fluctuation of excitonic order parameter and the long-range Coulomb interaction, respectively.
Polarization function for excitonic fluctuation
-----------------------------------------------
For the quantum excitonic fluctuation, the polarization function is defined as $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) &=& N\int \frac{d\omega}{2\pi} \frac{d^2
\mathbf{k}}{(2\pi)^2}\mathrm{Tr}
\left[G_{0}(\omega,\mathbf{k})\right.\nonumber
\\
&&\left.\times G_{0}(\omega+\Omega,\mathbf{k}+\mathbf{q})\right].
\label{Eq:PolarizationADef}\end{aligned}$$ Substituting the free fermion propagator into Eq. (\[Eq:PolarizationADef\]), we obtain $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q})=-\frac{4N}{v_{1}v_{2}}\int \frac{d^3
k}{(2\pi)^3}\frac{k\cdot(k+q)}{k^2(k+q)^2},\end{aligned}$$ where $k=(\omega,\mathbf{k})$. Here, we have employed the following transformations $$\begin{aligned}
v_{1}k_{1}\rightarrow k_{1}, \quad v_{2}k_{2}\rightarrow k_{2},
\quad v_{1}q_{1}\rightarrow q_{1}, \quad v_{2}q_{2}\rightarrow
q_{2}. \nonumber \\\end{aligned}$$ Using the Feynman parametrization formula $$\begin{aligned}
\frac{1}{AB}=\int_{0}^{1}dx \frac{1}{[Ax+(1-xB)]^2},
\label{Eq:FeynmanParaFormula}\end{aligned}$$ one gets $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) &=& -\frac{4N}{v_{1}v_{2}}
\int_{0}^{1}dx\int\frac{d^3 k}{(2\pi)^3}\nonumber
\\
&&\times\frac{k\cdot(k+q)}{\left[(k+xq)^{2}+x(1-x)q^{2}\right]}.\end{aligned}$$ Let $k+xq\rightarrow k$, $\Pi_{A}$ can be further written as $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) &=& -\frac{4N}{v_{1}v_{2}} \int^1_0
dx\left\{\int \frac{d^3 k}{(2\pi)^3} \frac{k^{2}}{[k^2 +
x(1-x)q^2]^2}\right.\nonumber \\
&& \left. -\int \frac{d^3 k}{(2\pi)^3}
\frac{x(1-x)q^2}{[k^2+x(1-x)q^2]^2}\right\}.\end{aligned}$$ Performing integration over $k$ by using the standard formula of dimensional regularization $$\begin{aligned}
\int \frac{d^dk}{(2\pi)^d}\frac{1}{(k^2+\Delta)^n} &=&
\frac{1}{(4\pi)^{d/2}}\frac{\Gamma(n-\frac{d}{2})}{\Gamma(n)}
\frac{1}{\Delta^{n-\frac{d}{2}}}, \label{Eq:FormulaDimReGA}
\\
\int \frac{d^dk}{(2\pi)^d}\frac{k^2}{(k^2+\Delta)^n} &=&
\frac{1}{(4\pi)^{d/2}}\frac{d}{2}\frac{\Gamma(n -
\frac{d}{2}-1)}{\Gamma(n)}\frac{1}{\Delta^{n-\frac{d}{2}-1}},
\label{Eq:FormulaDimReGB} \nonumber \\\end{aligned}$$ we find that $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) &=& \frac{2N}{v_{1}v_{2}\pi}
\sqrt{q^{2}}\int^1_0 dx \sqrt{x(1-x)}\nonumber \\
&=&\frac{N}{4v_{1}v_{2}}\sqrt{\Omega^2+q_{1}^{2}+q_{2}^{2}}.\end{aligned}$$ By taking $q_{1}\rightarrow v_{1}q_{1}$ and $q_{2}\rightarrow
v_{2}q_{2}$, we get $$\begin{aligned}
\Pi^{A}(\Omega,\mathbf{q}) = \frac{N}{4v_1 v_2} \sqrt{\Omega^2 +
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}. \label{Eq:PolarizationAResultApp}\end{aligned}$$
Polarization function for Coulomb interaction
---------------------------------------------
For the Coulomb interaction, the polarization function is given by $$\begin{aligned}
\Pi^{B}(\Omega,\mathbf{q}) &=& -N\int\frac{d\omega}{2\pi}
\frac{d^2\mathbf{k}}{(2\pi)^2}\mathrm{Tr} \left[\gamma_{0}
G_{0}(\omega,\mathbf{k})\gamma_{0}\right.\nonumber
\\
&&\left.\times G_{0}(\omega+\Omega,\mathbf{k}+\mathbf{q})\right].
\label{Eq:PolarizationBDef}\end{aligned}$$ Substituting Eq. (\[Eq:FermionPropagator\]) into Eq. (\[Eq:PolarizationBDef\]) leads to $$\begin{aligned}
\Pi^{B}(\Omega,\mathbf{q})=\frac{4N}{v_{1}v_{2}}\int \frac{d^3
k}{(2\pi)^3}\frac{2k_{0}(k_{0}+q_{0}) -
k\cdot(k+q)}{k^2(k+q)^2}.\nonumber \\\end{aligned}$$ Making use of the Feynman parametrization formula Eq. (\[Eq:FeynmanParaFormula\]), along with the transformation $k+xq \rightarrow k$, we recast the above expression as $$\begin{aligned}
\Pi^{B}(\Omega,\mathbf{q}) &=& \frac{4N}{v_{1}v_{2}} \int^1_0
dx\left\{\int \frac{d^3 k}{(2\pi)^3} \frac{-k^{2}/3}{[k^2 +
x(1-x)q^2]^2} \right.\nonumber \\
&&\left. +\int\frac{d^3 k}{(2\pi)^3}\frac{x(1-x) \left(q^2 -
2q_{0}^{2}\right)}{[k^2+x(1-x)q^2]^2}\right\}.\end{aligned}$$ Repeating the calculational steps that lead to Eq. (\[Eq:PolarizationAResultApp\]), we finally obtain $$\begin{aligned}
\Pi^{B}(\Omega,\mathbf{q}) = \frac{N}{8v_1 v_2} \frac{v_{1}^2
q_{1}^{2} + v_{2}^2q_{2}^{2}}{\sqrt{\Omega^2 + v_{1}^{2}q_{1}^{2} +
v_{2}^{2}q_{2}^{2}}}.\end{aligned}$$
Fermion self-energy
===================
The fermion self-energy corrections come from three sorts of interaction, namely the Yukawa coupling, Coulomb interaction, and disorder scattering. The former two interactions are inelastic, and the third one is elastic. We now calculate them in order.
Contribution from Yukawa coupling
---------------------------------
The fermion self-energy induced by the Yukawa coupling takes the form $$\begin{aligned}
\Sigma^{A}(\omega,\mathbf{k}) &=& \int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}} G_{0}(\Omega+\omega,\mathbf{q} +
\mathbf{k})D^{A}(\Omega,\mathbf{q})\nonumber
\\
&=& -\int'\frac{d\Omega}{2\pi}\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}
\frac{\left[-i(\Omega+\omega)\gamma_{0}+v_{1}(q_{1}+k_{1})
\gamma_{1}+v_{2}(q_{2}+k_{2})\gamma_{2}\right]} {\left[(\Omega +
\omega)^{2}+v_{1}^{2}(q_{1}+k_{1})^{2} + v_{2}^{2}(q_{2} +
k_{2})^{2}\right]} D^{A}(\Omega,\mathbf{q}).\end{aligned}$$ This self-energy can be expanded in powers of $i\omega$, $v_{1}k_{1}$, and $v_{2}k_{2}$. To the leading order, we get $$\begin{aligned}
\Sigma^{A}(\omega,\mathbf{k}) &=& i\omega\gamma_{0}\int'
\frac{d\Omega}{2\pi}\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}
\frac{-\Omega^{2}+v_{1}^{2}q_{1}^{2}+v_{2}^{2}
q_{2}^{2}}{\left(\Omega^{2}+v_{1}^{2}q_{1}^{2} +
v_{2}^{2}q_{2}^{2}\right)^{2}}\frac{1}{\frac{N}{4v_1v_2}
\sqrt{\Omega^2+v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}}
\\
&&-v_{1}k_{1}\gamma_{1}\int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}\frac{\Omega^{2} -
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}{\left(\Omega^{2} +
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}\right)^{2}}
\frac{1}{\frac{N}{4v_1v_2}
\sqrt{\Omega^2+v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}}
\\
&&-v_{2}k_{2}\gamma_{2}\int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}\frac{\Omega^{2} +
v_{1}^{2}q_{1}^{2}-v_{2}^{2}q_{2}^{2}}{\left(\Omega^{2} +
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}\right)^{2}}
\frac{1}{\frac{N}{4v_1v_2}\sqrt{\Omega^2 +
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}}.\end{aligned}$$ To carry out RG calculation, we choose to integrate over the integral variables within the range $$\begin{aligned}
\int'\frac{d\Omega}{2\pi}\frac{d^2{\mathbf{q}}}{(2\pi)^{2}} =
\frac{1}{8\pi^{3}}\int_{-\infty}^{+\infty}d\Omega\int_{0}^{2\pi}
d\theta\int_{b\Lambda}^{\Lambda}d|\mathbf{q}|\left|\mathbf{q}\right|,
\label{Eq:IntegrationRange}\end{aligned}$$ where $b=e^{-\ell}$. It is then easy to obtain $$\begin{aligned}
\Sigma^{A}(\omega,\mathbf{k}) = \left(-i\omega\gamma_{0}C_{0}^{A} +
v_{1}k_{1}\gamma_{1}C_{1}^{A} + v_{2}k_{2}\gamma_{2}
C_{2}^{A}\right)\ell.\end{aligned}$$ The expressions of $C_{i}^{A}$ are given by Eqs. (\[Eq:C0A\])-(\[Eq:GExpressionA\]).
Contribution from Coulomb interaction
-------------------------------------
The fermion self-energy induced by the Coulomb interaction is $$\begin{aligned}
\Sigma^{B}(\omega,\mathbf{k}) &=& -\int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}\gamma_{0} G(\Omega +
\omega,\mathbf{q}+\mathbf{k})\gamma_{0}D^{B}(\Omega,\mathbf{q})
\nonumber \\
&=&\int'\frac{d\Omega}{2\pi}\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}
\gamma_{0}\frac{\left[-i(\Omega+\omega)\gamma_{0}+v_{1}(q_{1} +
k_{1})\gamma_{1}+v_{2}(q_{2}+k_{2})\gamma_{2}\right]}{\left[(\Omega+\omega)^{2}
+ v_{1}^{2}(q_{1}+k_{1})^{2} +v_{2}^{2}(q_{2}+k_{2})^{2}\right]}
\gamma_{0}D^{B}(\Omega,\mathbf{q}).\end{aligned}$$ To the leading order of small energy/momenta expansion, $\Sigma^{B}$ can be approximately written as $$\begin{aligned}
\Sigma^{B}(\omega,\mathbf{k}) &=& -i\omega\gamma_{0}\int'
\frac{d\Omega}{2\pi}\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}
\frac{-\Omega^{2}+v_{1}^{2}q_{1}^{2}+v_{2}^{2}
q_{2}^{2}}{\left(\Omega^{2}+v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}
\right)^{2}}\frac{1}{\frac{|\mathbf{q}|}{\frac{2\pi
e^{2}}{\epsilon}}+\frac{N}{8v_1 v_2} \frac{v_{1}^{2}q_{1}^{2} +
v_{2}^{2}q_{2}^{2}}{\sqrt{\Omega^2+v_{1}^{2}q_{1}^{2} +
v_{2}^{2}q_{2}^{2}}}} \nonumber \\
&& -v_{1}k_{1}\gamma_{1}\int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}\frac{\Omega^{2}-v_{1}^{2}
q_{1}^{2} + v_{2}^{2}q_{2}^{2}}{\left(\Omega^{2}+v_{1}^{2} q_{1}^{2}
+ v_{2}^{2}q_{2}^{2}\right)^{2}}\frac{1}{
\frac{|\mathbf{q}|}{\frac{2\pi e^{2}}{\epsilon}} + \frac{N}{8v_1
v_2}\frac{v_{1}^{2}q_{1}^{2} + v_{2}^{2}q_{2}^{2}}{\sqrt{\Omega^2 +
v_{1}^{2}q_{1}^{2} + v_{2}^{2}q_{2}^{2}}}}\nonumber \\
&&-v_{2}k_{2}\gamma_{2}\int'\frac{d\Omega}{2\pi}
\frac{d^2{\mathbf{q}}}{(2\pi)^{2}}\frac{\Omega^{2} +
v_{1}^{2}q_{1}^{2}-v_{2}^{2}q_{2}^{2}}{\left(\Omega^{2} +
v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}\right)^{2}}
\frac{1}{\frac{|\mathbf{q}|}{\frac{2\pi e^{2}}{\epsilon}} +
\frac{N}{8v_1 v_2}\frac{v_{1}^{2}q_{1}^{2}+v_{2}^{2}
q_{2}^{2}}{\sqrt{\Omega^2+v_{1}^{2}q_{1}^{2}+v_{2}^{2}q_{2}^{2}}}}.\end{aligned}$$ Performing integrations according to Eq. (\[Eq:IntegrationRange\]), we obtain $$\begin{aligned}
\Sigma^{B}(\omega,\mathbf{k}) = \left(-i\omega\gamma_{0}C_{0}^{B} +
v_{1}k_{1}\gamma_{1}C_{1}^{B}+v_{2}k_{2}\gamma_{2}C_{2}^{B}\right)\ell.\end{aligned}$$ The expressions of $C_{i}^{B}$ can be found in Eqs. (\[Eq:C0B\])-(\[Eq:GExpressionB\]).
Contribution from disorder scattering
-------------------------------------
The fermion self-energy generated by disorder is $$\begin{aligned}
\Sigma_{\mathrm{dis}}(\omega) &=& \Delta
v_{\Gamma}^{2}\int'\frac{d^2 \mathbf{k}}{(2\pi)^2}\Gamma
G_{0}(\omega,\mathbf{k})\Gamma \nonumber
\\
&=&i\omega v_{\Gamma}^{2}\Delta\int'\frac{d^2\mathbf{k}}{(2\pi)^2}
\frac{\Gamma\gamma_{0}\Gamma}{\left(\omega^{2}+v_{1}^{2}k_{1}^{2} +
v_{2}^{2}k_{2}^{2}\right)}\nonumber \\
&\approx&i\omega\gamma_{0}C_{g}\ell,\end{aligned}$$ where $$\begin{aligned}
C_g = \frac{v_{\Gamma}^{2}\Delta}{2\pi v_{1}v_{2}}\end{aligned}$$ for both RSP and RM, and $$\begin{aligned}
C_g = \frac{\left(v_{\Gamma1}^{2} + v_{\Gamma2}^{2}\right)
\Delta}{2\pi v_{1}v_{2}}\end{aligned}$$ for RVP.
Corrections to fermion-disorder coupling
========================================
The fermion-disorder coupling receives vertex corrections from three sorts of interaction, including the Yukawa coupling, the Coulomb interaction, and the fermion-disorder interaction, which will be studied below.
Vertex correction due to Yukawa coupling
----------------------------------------
The vertex correction due to Yukawa coupling is $$\begin{aligned}
V^{A} = -\int'\frac{d\Omega}{2\pi}\frac{d^2\mathbf{q}}{(2\pi)^2}
G_{0}(\Omega,\mathbf{q}) v_\Gamma \Gamma
G_{0}(\Omega,\mathbf{q})D^{A}(\Omega,\mathbf{q}). \nonumber \\\end{aligned}$$ For RSP, $\Gamma=\gamma_{0}$ and we get $$\begin{aligned}
V^{A} = v_\Gamma\gamma_{0}\left(-C_{0}^{A}\right)\ell.\end{aligned}$$ For the two components of RVP defined by $\Gamma=\gamma_{1}$ and $\Gamma=\gamma_{2}$, $V_{A}$ is given by $$\begin{aligned}
V^{A} = v_\Gamma\gamma_{1}\left(-C_{1}^{A}\right)\ell,\end{aligned}$$ and $$\begin{aligned}
V^{A} = v_{\Gamma}\gamma_{2}\left(-C_{2}^{A}\right)\ell,\end{aligned}$$ respectively. For RM with $\Gamma=\mathbbm{1}$, $V_{A}$ is $$\begin{aligned}
V^{A} = v_\Gamma\mathbbm{1}\left(C_{0}^{A} + C_{1}^{A} +
C_{2}^{A}\right)\ell.\end{aligned}$$
Vertex correction due to Coulomb interaction
--------------------------------------------
The vertex correction due to Coulomb interaction is $$\begin{aligned}
V^{B} &=& -\int'\frac{d\Omega}{2\pi}\frac{d^2\mathbf{q}}{(2\pi)^2}
\gamma_{0}G_{0}(\Omega,\mathbf{q}) v_\Gamma \Gamma
G_{0}(\Omega,\mathbf{q})\gamma_{0} \nonumber \\
&&\times D^{B}(\Omega,\mathbf{q}).\end{aligned}$$ For RSP with $\Gamma=\gamma_{0}$, $V_{B}$ is $$\begin{aligned}
V^{B} = v_\Gamma\gamma_{0}\left(-C_{0}^{B}\right)\ell.\end{aligned}$$ For the two components of RVP defined by $\Gamma=\gamma_{1}$ and $\Gamma=\gamma_{2}$, we obtain $$\begin{aligned}
V^{B} = v_\Gamma\gamma_{1}\left(-C_{1}^{B}\right)\ell,\end{aligned}$$ and $$\begin{aligned}
V^{B} = v_{\Gamma}\gamma_{2}\left(-C_{2}^{B}\right)\ell,\end{aligned}$$ respectively. For RM with $\Gamma=\mathbbm{1}$, we find $$\begin{aligned}
V^{B} = v_\Gamma \mathbbm{1}\left(C_{0}^{B} - C_{1}^{B} -
C_{2}^{B}\right)\ell.\end{aligned}$$
Vertex correction from disorder
-------------------------------
The vertex correction due to disorder has the form $$\begin{aligned}
V_{\mathrm{dis}} &=& \Delta v_{\Gamma}^{2}\int'
\frac{d^2\mathbf{p}}{(2\pi)^2} \Gamma G_0(0,\mathbf{k})
v_\Gamma\Gamma G_0(0,\mathbf{k})\Gamma\nonumber
\\
&=& v_{\Gamma}\Delta v_{\Gamma}^{2}\int
\frac{d^2\mathbf{p}}{(2\pi)^2}\frac{1}{\left(v_{1}^{2}k_{1}^{2} +
v_{2}^{2}k_{2}^{2}\right)^{2}}\nonumber \\
&& \times \Gamma\left(v_{1}k_{1} \gamma_{1} + v_{2}k_{2}
\gamma_{2}\right)\Gamma \left(v_{1}k_{1} \gamma_{1} +
v_{2}k_{2}\gamma_{2}\right)\Gamma. \nonumber \\\end{aligned}$$ For RSP with $\gamma=\gamma_{0}$, $V_{dis}$ is $$\begin{aligned}
V_{\mathrm{dis}} = v_{\Gamma}\gamma_{0}C_{g}\ell.\end{aligned}$$ For the two components of RVP defined by $\gamma=\gamma_{1}$ and $\gamma_{2}$, $V_{\mathrm{dis}}$ is $$\begin{aligned}
V_{\mathrm{dis}} = 0.\end{aligned}$$ For RM with $\Gamma=\mathbbm{1}$, $V_{\mathrm{dis}}$ is $$\begin{aligned}
V_{\mathrm{dis}} = -v_{\Gamma}\mathbbm{1}C_{g}\ell.\end{aligned}$$
Derivation of the coupled RG equations
======================================
The action for the free fermions is given by $$\begin{aligned}
S_{\Psi} &=& \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \bar{\Psi}_{\sigma}
(\omega,\mathbf{k})\left(-i\omega \gamma_{0} +
v_{1}k_{1}\gamma_{1}+v_{2}k_{2}\gamma_{2}\right)
\Psi_{\sigma}(\omega,\mathbf{k}).\end{aligned}$$ Including the fermion self-energies induced by excitonic quantum fluctuation, Coulomb interaction, and disorder scattering, the action of fermions becomes $$\begin{aligned}
S_{\Psi} &=& \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k}) \left[-i\omega\gamma_{0} +
v_{1}k_{1}\gamma_{1} + v_{2}k_{2}\gamma_{2} -
\Sigma^{A}(\omega,\mathbf{k})-\Sigma^{B}(\omega,\mathbf{k}) -
\Sigma_{\mathrm{dis}}(\omega)\right]\Psi_{\sigma}(\omega,\mathbf{k})\nonumber
\\
&\approx&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})\left[-i\omega\gamma_{0}
e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{g}\right)\ell} + v_{1}k_{1}
\gamma_{1}e^{-\left(C_{1}^{A}+C_{1}^{B}\right)\ell} + v_{2}k_{2}
\gamma_{2}e^{-\left(C_{2}^{B}+C_{2}^{B}\right)\ell}\right]
\Psi_{\sigma}(\omega,\mathbf{k}).\end{aligned}$$ Making the following re-scaling transformations: $$\begin{aligned}
\omega&=&\omega'e^{-\ell}, \label{Eq:Scalingomega}
\\
k_{1}&=&k'_{1}e^{-\ell}, \label{Eq:Scalingk1}
\\
k_{2}&=&k'_{2}e^{-\ell}, \label{Eq:Scalingk2}
\\
\Psi &=& \Psi' e^{\left(2+\frac{C_{0}^{A}}{2} + \frac{C_{0}^{B}}{2}
- \frac{C_{g}}{2}\right)\ell}, \label{Eq:ScalingPsi}
\\
v_{1}&=&v_{1}'e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{1}^{A} +
C_{1}^{B}+C_{g}\right)\ell}, \label{Eq:Scalingv1}
\\
v_{2}&=&v_{2}'e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{2}^{A} +
C_{2}^{B}+C_{g}\right)\ell}, \label{Eq:Scalingv2}\end{aligned}$$ the fermion action is re-written as $$\begin{aligned}
S_{\Psi'} = \sum_{\sigma=1}^{N}\int\frac{d\omega'}{2\pi}
\frac{d^{2}\mathbf{k}'}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}'(\omega',\mathbf{k}') \left[-i\omega'\gamma_{0}
+ v_{1}'k_{1}'\gamma_{1}+v_{2}'k_{2}'\gamma_{2}\right]
\Psi_{\sigma}'(\omega',\mathbf{k}'),\end{aligned}$$ which recovers the form of the original action.
The action for the fermion-disorder coupling is $$\begin{aligned}
S_{\mathrm{dis}} = \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\int\frac{d^{2}\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})v_{\Gamma}\Gamma
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1})\end{aligned}$$ After taking into account the quantum corrections, it becomes $$\begin{aligned}
S_{\mathrm{dis}} = \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}}{(2\pi)^{2}} \bar{\Psi}_{\sigma}(\omega,\mathbf{k})
\left(v_{\Gamma}\Gamma+V^{A}+V^{B}+V_{\mathrm{dis}}\right)
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1}).\end{aligned}$$ In the case of RSP, we obtain $$\begin{aligned}
S_{\mathrm{dis}} &=& \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}}{(2\pi)^{2}} \bar{\Psi}_{\sigma}(\omega,\mathbf{k})
\left[v_{\Gamma}\gamma_{0}+v_\Gamma\gamma_{0}\left(-C_{0}^{A}\right)\ell
+v_\Gamma\gamma_{0}\left(-C_{0}^{B}\right)\ell+v_{\Gamma}\gamma_{0}C_{g}\ell\right]
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1})
\nonumber \\
&\approx&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}}{(2\pi)^{2}} \bar{\Psi}_{\sigma}(\omega,\mathbf{k})
v_{\Gamma}\gamma_{0}e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{g}\right)\ell}
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1}).
\label{Eq:SdisCorrectRCP}\end{aligned}$$ For the two components of RVP, $S_{\mathrm{dis}}$ is expressed as $$\begin{aligned}
S_{\mathrm{dis}} &=& \sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})\left[v_{\Gamma}\gamma_{1}
+v_\Gamma\gamma_{1}\left(-C_{1}^{A}\right)\ell+v_\Gamma\gamma_{1}
\left(-C_{1}^{B}\right)\ell\right]
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1})
\nonumber \\
&\approx&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\int\frac{d^{2}\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})v_{\Gamma}\gamma_{1}
e^{-\left(C_{1}^{A}+C_{1}^{B}\right)\ell}
\Psi_{\sigma}(\omega,\mathbf{k}_{1})A(\mathbf{k}-\mathbf{k}_{1}),
\label{Eq:SdisCorrectRGP1}\end{aligned}$$ and $$\begin{aligned}
S_{\mathrm{dis}} &=& \sum_{\sigma=1}^{N}\int
\frac{d\omega}{2\pi}\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\int\frac{d^{2}\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})\left[v_{\Gamma}\gamma_{2} +
v_\Gamma\gamma_{2}\left(-C_{2}^{A}\right)\ell +
v_\Gamma\gamma_{2}\left(-C_{2}^{B}\right)\ell\right]
\Psi_{\sigma}(\omega,\mathbf{k}_{1}) A(\mathbf{k}-\mathbf{k}_{1})
\nonumber \\
&\approx&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\int\frac{d^{2}\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})v_{\Gamma}\gamma_{2}
e^{-\left(C_{2}^{A}+C_{2}^{B}\right)\ell}
\Psi_{\sigma}(\omega,\mathbf{k}_{1})A(\mathbf{k}-\mathbf{k}_{1}),
\label{Eq:SdisCorrectRGP2}\end{aligned}$$ respectively. For RM, $S_{\mathrm{dis}}$ is cast in the form $$\begin{aligned}
S_{\mathrm{dis}}&=&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}}
\int\frac{d^{2}\mathbf{k}_{1}}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}(\omega,\mathbf{k})\left[v_{\Gamma}
\mathbbm{1}+v_\Gamma\mathbbm{1}\left(C_{0}^{A}+C_{1}^{A} +
C_{2}^{A}\right)\ell + v_\Gamma\mathbbm{1} \left(C_{0}^{B} -
C_{1}^{B}-C_{2}^{B}\right)\ell-v_{\Gamma}\mathbbm{1}C_{g}\ell
\right]\nonumber \\
&&\times\Psi_{\sigma}(\omega,\mathbf{k}_{1})
A(\mathbf{k}-\mathbf{k}_{1})\nonumber \\
&=&\sum_{\sigma=1}^{N}\int\frac{d\omega}{2\pi}
\frac{d^{2}\mathbf{k}}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}}{(2\pi)^{2}} \bar{\Psi}_{\sigma}(\omega,\mathbf{k})
v_{\Gamma}\mathbbm{1}e^{\left(C_{0}^{A}+C_{1}^{A}+C_{2}^{A}+C_{0}^{B}
- C_{1}^{B}-C_{2}^{B} -C_{g}\right)\ell}
\Psi_{\sigma}(\omega,\mathbf{k}_{1})A(\mathbf{k}-\mathbf{k}_{1}).
\label{Eq:SdisCorrectRM}\end{aligned}$$ We then employ the re-scaling transformations given by Eqs. (\[Eq:Scalingomega\])-(\[Eq:ScalingPsi\]). The random potential $A(\mathbf{k})$ should be re-scaled as follows $$\begin{aligned}
A(\mathbf{k}) = A'(\mathbf{k'})e^{\ell}.\end{aligned}$$ The parameter $v_{\Gamma}$ is re-scaled as $$\begin{aligned}
v_{\Gamma}=v_{\Gamma}' \label{Eq:ScalingvGammaRCP}\end{aligned}$$ for Eq. (\[Eq:SdisCorrectRCP\]), $$\begin{aligned}
v_{\Gamma}=v_{\Gamma}'e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{1}^{A} +
C_{1}^{B}+C_{g}\right)\ell} \label{Eq:ScalingvGammaRVP1}\end{aligned}$$ for Eq. (\[Eq:SdisCorrectRGP1\]), $$\begin{aligned}
v_{\Gamma}=v_{\Gamma}'e^{\left(-C_{0}^{A}-C_{0}^{B}+C_{2}^{A} +
C_{2}^{B}+C_{g}\right)\ell} \label{Eq:ScalingvGammaRVP2}\end{aligned}$$ for Eq. (\[Eq:SdisCorrectRGP2\]), and $$\begin{aligned}
v_{\Gamma}=v_{\Gamma}'e^{\left(-2C_{0}^{A}-C_{1}^{A}-C_{2}^{A} -
2C_{0}^{B}+C_{1}^{B}+C_{2}^{B} +2C_{g}\right)\ell}
\label{Eq:ScalingvGammaRM}\end{aligned}$$ for Eq. (\[Eq:SdisCorrectRM\]). After carrying out the above manipulations, we re-write the action for fermion-disorder coupling as follows $$\begin{aligned}
S_{\mathrm{dis}} = \sum_{\sigma=1}^{N}\int\frac{d\omega'}{2\pi}
\frac{d^{2}\mathbf{k}'}{(2\pi)^{2}} \int\frac{d^{2}
\mathbf{k}_{1}'}{(2\pi)^{2}}
\bar{\Psi}_{\sigma}'(\omega',\mathbf{k}')v_{\Gamma}'\mathbbm{1}
\Psi_{\sigma}'(\omega',\mathbf{k}_{1}')A'(\mathbf{k}'-\mathbf{k}_{1}'),\end{aligned}$$ which restores the form of the original action.
From Eqs. (\[Eq:ScalingPsi\]), we obtain the RG equation for the quasiparticle $Z_{f}$ $$\begin{aligned}
\frac{dZ_{f}}{d\ell} = \left(C_{0}^{A} + C_{0}^{B} -
C_{g}\right)Z_{f}.\end{aligned}$$ According to Eqs. (\[Eq:Scalingv1\]) and (\[Eq:Scalingv2\]), the RG equations for $v_{1}$ and $v_{2}$ are given by $$\begin{aligned}
\frac{dv_{1}}{d\ell} &=& \left(C_{0}^{A}+C_{0}^{B}-C_{1}^{A} -
C_{1}^{B}-C_{g}\right)v_{1},
\\
\frac{dv_{2}}{d\ell} &=& \left(C_{0}^{A}+C_{0}^{B}-C_{2}^{A} -
C_{2}^{B}-C_{g}\right)v_{2}.\end{aligned}$$ The RG equation for the velocity ratio $v_{2}/v_{1}$ can be readily derived: $$\begin{aligned}
\frac{d\left(v_{2}/v_{1}\right)}{d\ell} &=&
\frac{\frac{dv_{2}}{dl}v_{1}-v_{2}\frac{dv_{1}}{dl}}{v_{1}^{2}}
\nonumber \\
&=& \left(C_{1}^{A}-C_{2}^{A}+C_{1}^{B}-C_{2}^{B}\right)
\frac{v_{2}}{v_{1}}.\end{aligned}$$ Based on Eqs. (\[Eq:ScalingvGammaRCP\])-(\[Eq:ScalingvGammaRM\]), we obtain the RG equation for the parameter $v_{\Gamma}$ $$\begin{aligned}
\left\{\begin{array}{ll} \frac{d v_{\Gamma}}{d\ell}=0 &\texttt{RSP},
\\
\\
\frac{dv_{\Gamma}}{d\ell}=\left(C_{0}^{A}+C_{0}^{B}-C_{1}^{A} -
C_{1}^{B}-C_{g}\right)v_{\Gamma} & \gamma_{1} \ \texttt{component of
RVP}
\\
\\
\frac{dv_{\Gamma}}{d\ell}=\left(C_{0}^{A}+C_{0}^{B}-C_{2}^{A} -
C_{2}^{B}-C_{g}\right)v_{\Gamma} & \gamma_{2} \ \texttt{component of
RVP}
\\
\\
\frac{dv_{\Gamma}}{d\ell}=\left(2C_{0}^{A}+C_{1}^{A}+C_{2}^{A} +
2C_{0}^{B}-C_{1}^{B}-C_{2}^{B} -2C_{g}\right)v_{\Gamma} &\texttt{RM}
\end{array}\right.\end{aligned}$$
[99]{}
S. Das Sarma, S. Adam, E. H. Hwang, and E. Rossi, Rev. Mod. Phys. [**83**]{}, 407 (2011).
V. N. Kotov, B. Uchoa, V. M. Pereira, F. Guinea, and A. H. Castro Neto, Rev. Mod. Phys. [**84**]{}, 1067 (2012).
O. Vafek and A. Vishwanath, Annu. Rev. Condens. Matter Phys. [**5**]{}, 83 (2014).
T. O. Wehling, A. M. Black-Schaffer, and A. V. Balatsky, Adv. Phys. [**63**]{}, 1 (2014).
X. Wan, A. M. Turner, A. Vishwanath, and S. Y. Savrasov, Phys. Rev. B [**83**]{}, 205101 (2011).
H. Weng, X. Dai, and Z. Fang, J. Phys.: Condens. Matter [**28**]{}, 303001 (2016).
C. Fang, H. Weng, X. Dai, and Z. Fang, Chin. Phys. B [**25**]{}, 117106 (2016).
B. Yan and C. Felser, Annu. Rev. Condens. Matter Phys. [**8**]{}, 337 (2017).
M. Z. Hasan, S.-Y. Xu, I. Belopolski, and S.-M. Huang, Annu. Rev. Condens. Matter Phys. [**8**]{}, 289 (2017).
N. P. Armitage, E. J. Mele, and A. Vishwanath, Rev. Mod. Phys. [**90**]{}, 015001 (2018).
J. González, F. Guinea, and M. A. H. Vozmediano, Phys. Rev. B [**59**]{}, R2474(R) (1999).
J. Hofmann, E. Barnes, and S. Das Sarma, Phys. Rev. Lett. [**113**]{}, 105502 (2014).
A. Sharma and P. Kopietz, Phys. Rev. B [**93**]{}, 235425 (2016).
P. Goswami and S. Chakravarty, Phys. Rev. Lett. [**107**]{}, 196803 (2011).
P. Hosur, S. A. Parameswaran, and A. Vishwanath, Phys. Rev. Lett. [**108**]{}, 046602 (2012).
J. Hofmann, E. Barnes, and S. Das Sarma, Phys. Rev. B [**92**]{}, 045104 (2015).
R. E. Throckmorton, J. Hofmann, E. Barnes, and S. Das Sarma, Phys. Rev. B [**92**]{}, 115101 (2015).
A. Sharma, A. Scammell, J. Krieg, and P. Kopietz, Phys. Rev. B [**97**]{}, 125113 (2018).
B.-J. Yang, E.-G. Moon, H. Isobe, and N. Nagaosa, Nat. Phys. [**10**]{}, 774 (2014).
A. A. Abrikosov, J. Low. Temp. Phys. [**8**]{}, 315 (1972).
A. A. Abrikosov and S. D. Beneslavskii, Sov. Phys. JETP [**32**]{}, 699 (1971).
A. A. Abrikosov, Sov. Phys. JETP [**39**]{}, 709 (1974).
E.-G. Moon, C. Xu, Y. B. Kim, and L. Balents, Phys. Rev. Lett. [**111**]{}, 206401 (2013).
I. F. Herbut and L. Janssen, Phys. Rev. Lett. [**113**]{}, 106401 (2014).
L. Janssen and I. F. Herbut, Phys. Rev. B [**92**]{}, 045117 (2015).
P. T. Dumitrescu, Phys. Rev. B [**92**]{}, 121102(R) (2015).
L. Janssen and I. F. Herbut, Phys. Rev. B [**95**]{}, 075101 (2017).
Y. Huh, E.-G. Moon, and Y. B. Kim, Phys. Rev. B [**93**]{}, 035138 (2016).
G. Y. Cho and E.-G. Moon, Sci. Rep. [**6**]{}, 19198 (2016).
H. Isobe and L. Fu, Phys. Rev. B [**93**]{}, 241113 (2016).
H.-H. Lai, Phys. Rev. B [**91**]{}, 235131 (2015).
S.-K. Jian and H. Yao, Phys. Rev. B [**92**]{}, 045121 (2015).
S.-X. Zhang, S.-K. Jian, and H. Yao, Phys. Rev. B [**96**]{}, 241111 (2017).
J.-R. Wang, G.-Z. Liu, and C.-J. Zhang, Phys. Rev. B [**96**]{}, 165142 (2017).
R. Shankar, Rev. Mod, Phys. [**66**]{}, 129 (1994).
A. H. Castro Neto, Physics [**2**]{}, 30 (2009).
D. V. Khveshchenko, Phys. Rev. Lett. [**87**]{}, 246802 (2001).
E. V. Gorbar, V. P. Gusynin, V. A. Miransky, and I. A. Shovkovy, Phys. Rev. B [**66**]{}, 045108 (2002).
D. V. Khveshchenko and H. Leal, Nucl. Phys. B [**687**]{}, 323 (2004).
G.-Z. Liu, W. Li, and G. Cheng, Phys. Rev. B [**79**]{}, 205429 (2009).
D. V. Khveshchenko, J. Phys.:Condens. Matter [**21**]{}, 075303 (2009).
O. V. Gamayun, E. V. Gorbar, and V. P. Gusynin, Phys. Rev. B [**81**]{}, 075429 (2010).
J. Sabio, F. Sols, and F. Guinea, Phys. Rev. B [**82**]{}, 121413(R) (2010).
C.-X. Zhang, G.-Z. Liu, and M.-Q. Huang, Phys. Rev. B [**83**]{}, 115438 (2011).
G.-Z. Liu and J.-R. Wang, New J. Phys. [**13**]{}, 033022 (2011).
J.-R. Wang and G.-Z. Liu, J. Phys. Condens. Matter [**23**]{}, 155602 (2011).
J.-R. Wang and G.-Z. Liu, J. Phys. Condens. Matter [**23**]{}, 345601 (2011).
J.-R. Wang and G.-Z. Liu, New J. Phys. [**14**]{}, 043036 (2012).
C. Popovici, C. S. Fischer, and L. von Smekal, Phys. Rev. B [**88**]{}, 205429 (2013).
J.-R. Wang and G.-Z. Liu, Phys. Rev. B [**89**]{}, 195404 (2014).
J. González, Phys. Rev. B [**92**]{}, 125115 (2015).
M. E. Carrington, C. S. Fischer, L. von Smekal, and M. H. Thoma, Phys. Rev. B [**94**]{}, 125102 (2016).
A. Sharma, V. N. Kotov, and A. H. Castro Neto, Phys. Rev. B [**95**]{}, 235124 (2017).
H.-X. Xiao, J.-R. Wang, H.-T. Feng, P.-L. Yin, and H.-S. Zong, Phys. Rev. B [**96**]{}, 155114 (2017).
M. E. Carrington, C. S. Fischer, L. von Smekal, and M. H. Thoma, Phys. Rev. B [**97**]{}, 115411 (2018).
O. V. Gamayun, E. V. Gorbar, and V. P. Gusynin, Phys. Rev. B [**80**]{}, 165429 (2009).
J. Wang, H. A. Fertig, G. Murthy, and L. Brey, Phys. Rev. B [**83**]{}, 035404 (2011).
A. Katanin, Phys. Rev. B [**93**]{}, 035132 (2016).
J. González, Phys. Rev. B [**82**]{}, 155404 (2010).
J. González, Phys. Rev. B [**85**]{}, 085420 (2012).
J. E. Drut and T. A. Lähde, Phys. Rev. Lett. [**102**]{}, 026802 (2009).
J. E. Drut and T. A. Lähde, Phys. Rev. B [**79**]{}, 165425 (2009).
J. E. Drut and T. A. Lähde, Phys. Rev. B [**79**]{}, 241405(R) (2009).
W. Armour, S. Hands, and C. Strouthos, Phys. Rev. B [**81**]{}, 125105 (2010).
W. Armour, S. Hands, and C. Strouthos, Phys. Rev. B [**84**]{}, 075123 (2011).
P. V. Buividovich and M. I. Polikarpov, Phys. Rev. B [**86**]{}, 245117 (2012).
F. de Juan and H. A. Fertig, Solid State Commun. [**152**]{}, 1460 (2012).
M. V. Ulybyshev, P. V. Buividovich, M. I. Katsnelson, and M. I. Polikarpov, Phys. Rev. Lett. [**111**]{}, 056801 (2013).
D. Smith and L. von Smekal, Phys. Rev. B [**89**]{}, 195429 (2014).
A. V. Kotikov and S. Teber, Phys. Rev. D [**94**]{}, 114010 (2016).
J. González, Phys. Rev. B [**90**]{}, 121107(R) (2014).
V. V. Braguta, M. I. Katsnelson, A. Y. Kotov, and A. A. Nikolaev, Phys. Rev. B [**94**]{}, 205147 (2016).
H.-X. Xiao, J.-R. Wang, G.-Z. Liu, and H.-S. Zong, Phys. Rev. B [**97**]{}, 155122 (2018).
L. Janssen and I. F. Herbut, Phys. Rev. B [**93**]{}, 165109 (2016).
J.-R. Wang, G.-Z. Liu, and C.-J. Zhang, Phys. Rev. B [**95**]{}, 075129 (2017).
L. V. Keldysh and Y. V. Kopaev, Fiz. Tverd. Tela. [**6**]{}, 2791 (1964).
D. Jerome, T. M. Rice, and W. Kohn, Phys. Rev. [**158**]{}, 462 (1967).
Y. Nambu and G. Jona-Lasinio, Phys. Rev. [**122**]{}, 345 (1961).
V. A. Miransky, *Dynamical Symmetry Breaking in Quantum Field Theories*, (World Scientific, 1994).
D. C. Elias, R. V. Gorbachev, A. S. Mayorov, S. V. Morozov, A. A. Zhukov, P. Blake, L. A. Ponomarenko, I. V. Grigorieva, K. S. Novoselov, F. Guinea, and A. K. Geim, Nat. Phys. [**7**]{}, 701 (2011).
A. S. Mayorov, D. C. Elias, I. S. Mukhin, S. V. Morozov, L. A. Ponomarenko, K. S. Novoselov, A. K. Geim, and R. V. Gorbachev, Nano. Lett. [**12**]{}, 4629 (2012).
H.-K. Tang, E. Laksono, J. N. B. Rodrigues, P. Sengupta, F. F. Assaad, and S. Adam, Phys. Rev. Lett. [**115**]{}, 186602 (2015).
M. Hirata, K. Ishikawa, G. Matsuno, A. Kobayashi, K. Miyagawa, M. Tamura, C. Berthier, and K. Kanoda, Science [**358**]{}, 1403 (2017).
A. W. W. Ludwig, M. P. A. Fisher, R. Shankar, and G. Grinstein, Phys. Rev. B [**50**]{}, 7526 (1994).
A. A. Nersesyan, A. M. Tsvelik, and F. Wenger, Nucl. Phys. B [**438**]{}, 561 (1995).
A. Altland, B. D. Simons, and M. R. Zirnbauer, Phys. Rep. [**359**]{}, 283 (2002).
Y. Huh and S. Sachdev, Phys. Rev. B [**78**]{}, 064512 (2008).
C. Xu, Y. Qi, and S. Sachdev, Phys. Rev. B [**78**]{}, 134507 (2008).
L. Fritz and S. Sachdev, Phys. Rev. B [**80**]{}, 144503 (2009).
J. Wang, G.-Z. Liu, and H. Kleinert, Phys. Rev. B [**83**]{}, 214503 (2011).
G.-Z. Liu, J.-R. Wang, and J. Wang, Phys. Rev. B [**85**]{}, 174525 (2012).
J.-H. She, M. J. Lawler, and E.-A. Kim, Phys. Rev. B [**92**]{}, 035112 (2015).
J.-R. Wang, G.-Z. Liu, and C.-J. Zhang, New. J. Phys. [**18**]{}, 073023 (2016).
I. F. Herbut, V. Juričic, and B. Roy, Phys. Rev. B [**79**]{}, 085116 (2009).
D. A. Siegel, C.-H. Park, C. Hwang, J. Deslippe, A. V. Fedorov, S. G. Louie, and A. Lanzara, Proc. Natl. Acad. Sci. U.S.A. [**108**]{}, 11365 (2011).
J. Chae, S. Jung, A. F. Young, C. R. Dean, L. Wang, Y. Gao, K. Watanabe, T. Taniguchi, J. Hone, K. L. Shepard, P. Kim, N. B. Zhitenev, and J. A. Stroscio, Phys. Rev. Lett. [**109**]{}, 116802 (2012).
G. L. Yu, R. Jalil, B. Belle, A. S. Mayorov, P. Blake, F. Schedin, S. V. Morozov, L. A. Ponomarenko, F. Chiappini, S. Wiedmann, U. Zeitler, M. I. Katsnelson, A. K. Geim, K. S. Novoselov, and D. C. Elias, Proc. Natl. Acad. Sci. U.S.A. [**110**]{}, 3282 (2013).
J. Ye and S. Sachdev, Phys. Rev. Lett. [**80**]{}, 5409 (1998).
J. Ye, Phys. Rev. B [**60**]{}, 8290 (1999).
T. Stauber, F. Guinea, and M. A. H. Vozmediano, Phys. Rev. B [**71**]{}, 041406(R) (2005).
I. F. Herbut, V. Juričić, and O. Vafek, Phys. Rev. Lett. [**100**]{}, 046403 (2008).
O. Vafek and M. J. Case, Phys. Rev. B [**77**]{}, 033410 (2008).
M. S. Foster and I. L. Aleiner, Phys. Rev. B [**77**]{}, 195413 (2008).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $G=(V,E)$ be a graph and $A$ its adjacency matrix. We say that a vertex $y \in V$ is a function of vertices $x_1, \ldots, x_k \in V$ if there exists a Boolean function $f$ of $k$ variables such that for any vertex $z \in V - \{y, x_1, \ldots, x_k\}$, $A(y,z)=f(A(x_1,z),\ldots,A(x_k,z))$. The functionality $fun(y)$ of vertex $y$ is the minimum $k$ such that $y$ is a function of $k$ vertices. The functionality $fun(G)$ of the graph $G$ is $\max\limits_H\min\limits_{y\in V(H)}fun(y)$, where the maximum is taken over all induced subgraphs $H$ of $G$. In the present paper, we show that functionality generalizes simultaneously several other graph parameters, such as degeneracy or clique-width, by proving that bounded degeneracy or bounded clique-width imply bounded functionality. Moreover, we show that this generalization is proper by revealing classes of graphs of unbounded degeneracy and clique-width, where functionality is bounded by a constant. This includes permutation graphs, unit interval graphs and line graphs. We also observe that bounded functionality implies bounded VC-dimension, i.e. graphs of bounded VC-dimension extend graphs of bounded functionality, and this extension is also proper.'
author:
- 'Bogdan Alecu[^1]'
- 'Aistis Atminas[^2]'
- 'Vadim Lozin[^3]'
title: Graph functionality
---
[*Keywords:*]{} clique-width, graph degeneracy, VC-dimension, permutation graph, line graph, graph representation
Introduction
============
Let $G=(V,E)$ be a simple graph, i.e. undirected graph without loops and multiple edges. We denote by $A=A_G$ the adjacency matrix of $G$ and by $A(x,y)$ the element of this matrix corresponding to vertices $x,y\in V$, i.e. $A(x,y)=1$ if $x$ and $y$ are adjacent, and $A(x,y)=0$ otherwise.
We say that a vertex $y \in V$ is a function of vertices $x_1, \ldots, x_k \in V$ if there exists a Boolean function $f$ of $k$ variables such that for any vertex $z \in V - \{y, x_1, \ldots, x_k\}$, $$A(y,z)=f(A(x_1,z),\ldots,A(x_k,z)).$$ The functionality $fun(y)$ of vertex $y$ is the minimum $k$ such that $y$ is a function of $k$ vertices. In particular, the functionality of an isolated vertex is $0$, and the same is true for a dominating (also known as universal) vertex, i.e. a vertex adjacent to all the other vertices of the graph. More generally, the functionality of a vertex $y$ does not exceed the number of its neighbours (the degree of $y$) and the number of its non-neighbours. One more simple example of functional vertices is given by twins, i.e. vertices $x$ and $y$ that have the same set of neighbours different from $x$ and $y$. Twins are functions of each other and their functionality is (at most) 1.
The functionality $fun(G)$ of $G$ is $$\max\limits_H\min\limits_{y\in V(H)}fun(y),$$ where the maximum is taken over all induced subgraphs $H$ of $G$. Similarly to many other graph parameters, this notion becomes valuable when its value is small, i.e. is bounded by a constant independent of the size of the graph. This is important, in particular, for coding of graphs, i.e. representing them by words in a finite alphabet, which is needed for representing graphs in computer memory. Indeed, if a vertex $y$ is a function of only constantly many vertices, then to describe the neighbourhood of $y$ we need $O(\log_2 n)$ bits, regardless of how large the neighbourhood (or non-neighbourhood) of $y$ is.
In the present paper, we explore the relationship between graph functionality and other graph parameters. From the above discussion, it follows that graphs of bounded functionality extend graphs of bounded vertex degree. More generally, they extend graphs of bounded degeneracy, where the [*degeneracy*]{} of $G$ is the minimum $k$ such that every induced subgraph of $G$ has a vertex of degree at most $k$. A notion related to degeneracy is that of [*arboricity*]{}, which is the minimum number of forests into which the edges of $G$ can be partitioned. The degeneracy of $G$ is always between the arboricity and twice the arboricity of $G$ and hence graphs of bounded functionality extend graphs of bounded arboricity too.
One more important graph parameter is [*clique-width*]{}. Many algorithmic problems that are generally NP-hard become polynomial-time solvable when restricted to graphs of bounded clique-width [@CMR00]. Clique-width is a relatively new notion and it generalizes another important graph parameter, [ tree-width]{}, studied in the literature for decades. Clique-width is stronger than tree-width in the sense that graphs of bounded tree-width have bounded clique-width. In Section \[sec:bounded\], we show that functionality is stronger than clique-width by proving that graphs of bounded clique-width have bounded functionality. Moreover, in the same section we reveal three classes of graphs, where functionality is bounded but clique-width and degeneracy are not. These are permutation graphs, unit interval graphs and line graphs.
In [@implicit], it was shown that any class of graphs of bounded functionality contains $2^{O(n\log_2 n)}$ labelled graphs with $n$ vertices. Therefore, functionality is unbounded in any larger class of graphs. In particular, it is unbounded in the classes of bipartite graphs, co-bipartite graphs and split graphs. In [@Lozin], it was show that these are the only three minimal hereditary classes of graphs of unbounded VC-dimension. Therefore, graphs of bounded VC-dimension extend graphs of bounded functionality. Moreover, this extension is proper, as we show in Section \[sec:unbounded\]. Section \[sec:conclusion\] concludes the paper with a number of other open problems. In the rest of the present section we introduce basic terminology and notation used in the paper.
For a simple graph $G$, we denote by $V(G)$ and $E(G)$ the vertex set and the edge set of $G$, respectively. The neighbourhood $N(v)$ of a vertex $v\in V(G)$ is the set of vertices of $G$ adjacent to $v$ and the degree of $v$ is $|N(v)|$. A vertex of degree 0 is called [*isolated*]{}. The closed neighbourhood of $v$ is $N[v]=\{v\}\cup N(v)$.
A [*clique*]{} in a graph $G$ is a set of pairwise adjacent vertices, and an [*independent set*]{} is a set of pairwise non-adjacent vertices. The [*girth*]{} of $G$ is the length of a shortest cycle in $G$. A chordless cycle of length $n$ is denoted $C_n$.
A [*forest*]{} is a graph without cycles and a [*tree*]{} is a connected graph without cycles. A graph $G$ is [*bipartite*]{} if $V(G)$ can be partitioned into two independent sets, [*co-bipartite*]{} if $V(G)$ can be partitioned into two cliques, and [*split*]{} if $V(G)$ can be partitioned into a clique and an independent set.
A graph $H$ is an [*induced subgraph*]{} of a graph $G$ if $H$ can be obtained from $G$ by vertex deletions. A class $X$ of graphs is [*hereditary*]{} if it is closed under taking induced subgraphs, or equivalently, if it is closed under deletion of vertices from graphs in the class. A class $X$ is [*monotone*]{} if it is closed under vertex deletions and edge deletions, and $X$ is [*minor-closed*]{} if it is closed under vertex deletions, edge deletions and edge contractions. Clearly, every minor-closed class is monotone and every monotone class is hereditary.
It is well-known (and not difficult to see) that a class $X$ of graphs is hereditary if and only if it can be described by means of minimal forbidden induced subgraphs, i.e. vertex-minimal graphs that do not belong to $X$. If $M$ is the set of minimal forbidden induced subgraphs for $X$, then we say that graphs in $X$ are $M$-free, and if $M$ is finite, we say that $X$ is [*finitely defined*]{}.
By $A\otimes B$ we denote the symmetric difference of two sets, i.e. $A\otimes B=(A-B)\cup (B-A)$. When taking the symmetric difference of vertex neighbourhoods, we will always exclude the two vertices themselves; for brevity, we will write $N(u) \otimes N(v)$ to mean the set of vertices different from $u$ and $v$ and adjacent to exactly one of $u$ and $v$.
Graphs of small functionality {#sec:bounded}
=============================
As we mentioned in the introduction, functionality is bounded for graphs of bounded degree or degeneracy, which is easy to see. This includes, in particular, all proper minor-closed classes of graphs. The family of monotone classes is larger and not all classes in this family are of bounded functionality. In this paper, we present a dichotomy with respect to bounded/unbounded functionality for monotone classes defined by finitely many forbidden induced subgraphs. The first part of the dichotomy describes monotone classes of bounded functionality without restriction to finitely defined classes and is presented in Section \[sec:monotone\]. The other part of the dichotomy applies to finitely defined monotone classes only and is presented in Section \[sec:unbounded\].
Monotone classes of bounded functionality {#sec:monotone}
-----------------------------------------
\[thm:monotone-bounded\] If $X$ is a monotone class that does not contain at least one forest, then graphs in $X$ have bounded functionality.
Let $F$ be a forest that does not belong to $X$, and let $k$ be the number of vertices in $F$. Assume that $X$ has a graph $G$ every vertex of which has degree at least $k$. Then $G$ contains every tree $T$ with at most $k+1$ vertices as a (not necessarily induced) subgraph, which can be easily shown by induction on the number of vertices in $T$. But then $G$ contains $F$ as a subgraph, which contradicts the assumption that $X$ is a monotone class that does not contain $F$. This shows that every graph in $X$ contains a vertex of degree at most $k-1$. Since $X$ is hereditary, we conclude that the degeneracy of graphs in $X$ is at most $k-1$. Therefore, graphs in $X$ have bounded functionality.
Graphs of bounded clique-width
------------------------------
The notion of clique-width of a graph was introduced in [@CER93]. The clique-width of a graph $G$ is denoted ${\rm cwd}(G)$ and is defined as the minimum number of labels needed to construct $G$ by means of the following four graph operations:
- creation of a new vertex $v$ with label $i$ (denoted $i(v)$),
- disjoint union of two labelled graphs $G$ and $H$ (denoted $G\oplus H$),
- connecting vertices with specified labels $i$ and $j$ (denoted $\eta_{i,j}$) and
- renaming label $i$ to label $j$ (denoted $\rho_{i\to j}$).
Every graph can be defined by an algebraic expression using the four operations above. This expression is called a $k$-expression if it uses $k$ different labels. For instance, the cycle $C_5$ on vertices $a,b,c,d,e$ (listed along the cycle) can be defined by the following 4-expression: $$\eta_{4,1}(\eta_{4,3}(4(e)\oplus\rho_{4\to 3}(\rho_{3\to 2}(\eta_{4,3}(4(d)\oplus\eta_{3,2}(3(c)\oplus\eta_{2,1}(2(b)\oplus 1(a)))))))).$$
Alternatively, any algebraic expression defining $G$ can be represented as a rooted tree, whose leaves correspond to the operations of vertex creation, the internal nodes correspond to the $\oplus$-operations, and the root is associated with $G$. The operations $\eta$ and $\rho$ are assigned to the respective edges of the tree. Figure \[fig:tree\] shows the tree representing the above expression defining a $C_5$.
(370,50) (15,50) (90,50) (165,50) (240,50) (315,50) (355,50) (90,10) (240,10) (315,10) (25,50)[(1,0)[55]{}]{} (100,50)[(1,0)[55]{}]{} (175,50)[(1,0)[55]{}]{} (250,50)[(1,0)[55]{}]{} (325,50)[(1,0)[20]{}]{} (90,40)[(0,-1)[20]{}]{} (165,40)[(0,-1)[20]{}]{} (240,40)[(0,-1)[20]{}]{} (315,40)[(0,-1)[20]{}]{} (85,47)[+]{} (160,47)[+]{} (235,47)[+]{} (311,47)[+]{} (8,47)[$C_5$]{} (83,10)[$_{4(e)}$]{} (158,10)[$_{4(d)}$]{} (233,10)[$_{3(c)}$]{} (308,10)[$_{2(b)}$]{} (348,50)[$_{1(a)}$]{} (101,55)[$_{\rho_{4\to 3}\rho_{3\to 2}\eta_{4,3}}$]{} (36,55)[$_{\eta_{4,1}\eta_{4,3}}$]{} (196,55)[$_{\eta_{3,2}}$]{} (271,55)[$_{\eta_{2,1}}$]{}
Among various examples of graphs of bounded clique-width we mention distance-hereditary graphs. These are graphs of clique-width at most 3 [@perfect-cw]. Every graph in this class can be constructed from a single vertex by successively adding either a pendant vertex or a twin (true or false) [@DH]. From this characterization we immediately conclude that the functionality of distance-hereditary graphs is at most one. More generally, in the next theorem we show that functionality is bounded for all classes of graphs of bounded clique-width.
For any graph $G$, $fun(G)\le 2{\rm cwd}(G)-1$. \[cwd\]
Let $G$ be a graph of clique-width $k$ and let $T$ be a tree corresponding to a $k$-expression that describes $G$. Consider a node $v$ of the tree such that the tree rooted at $v$ has more than $k$ leaves, and no children of $v$ have this property (if no such $v$ exists, we are done, since $G$ has at most $k$ vertices). Denote the children of $v$ by $u_1,\ldots, u_t$ and let $i$ be the minimum index such that $u_1, \ldots, u_i$ have a combined total of more than $k$ leaves amongst their descendants. Consider the subtree $T'$ of $T$ consisting of $v$ and the union of the trees rooted at $u_1, \ldots, u_i$. This subtree has more than $k$ leaves and therefore at least two of them, say $x$ and $y$, have the same label at node $v$. On the other hand, $T'$ has at most $2k$ leaves by the choice of $i$. Therefore, the symmetric difference $N(x)\otimes N(y)$ contains at most $2k-2$ vertices, since $x$ and $y$ are not distinguished outside of the tree rooted at $v$. As a result, the functionality of both $x$ and $y$ is at most $2k-1$ ($x$ is a function of $\{y\}\cup (N(x)\otimes N(y))$ and similarly $y$ is a function of $\{x\}\cup (N(x)\otimes N(y))$).
It is known (see e.g. [@CO00]) that the clique-width of an induced subgraph of $G$ cannot exceed the clique-width of $G$. Therefore, every induced subgraph of $G$ has a vertex of functionality at most $2k-1$. Thus, the functionality of $G$ is at most $2k - 1$.
This result shows that the family of graph classes of bounded functionality extends the family of graph classes of bounded clique-width. Moreover, this extension is proper, because clique-width is known to be unbounded for square grids. This example, however, is not very interesting in the sense that square grids have bounded vertex degree and hence bounded functionality. In the next three sections, we reveal several classes of graphs of bounded functionality, where neither clique-width nor degeneracy is bounded.
Unit interval graphs
--------------------
A unit interval graph is the intersection graph of intervals of the same length on the real line. The class of unit interval graphs is one of the minimal hereditary classes of unbounded clique-width [@Lozin-minimal]. Also, degeneracy is unbounded in this class, since it contains cliques of arbitrarily large size. Our next result shows that functionality is bounded for unit interval graphs.
\[thm:unit-interval\] The functionality of unit interval graphs is at most 2.
Let $G$ be a unit interval graph with $n$ vertices and assume without loss of generality that $G$ has no isolated vertices (since any such vertex has functionality 0). Take a unit interval representation for $G = (V, E)$ with the interval endpoints all distinct. We label the vertices $v_1,\ldots,v_n$ in the order in which they appear on the real line (from left to right), and denote the endpoints of interval $I_i$ corresponding to vertex $v_i$ by $a_i < b_i$. We will bound $$S=\sum\limits_{i=1}^{n - 1} |N(v_i) \otimes N(v_{i+1})|.$$
Note that any neighbour of $v_i$ which is not a neighbour of $v_{i+1}$ needs to have its right endpoint between $a_i$ and $a_{i+1}$. Similarly, any neighbour of $v_{i+1}$ but not of $v_i$ needs to have its left endpoint between $b_i$ and $b_{i+1}$. In other words, $|N(v_i) \otimes N(v_{i+1})|$ is bounded above by the number of endpoints in $(a_i, a_{i+1}) \cup (b_i, b_{i+1})$ (we say bounded above and not equal, since it might happen that $b_i$ lies between $a_i$ and $a_{i+1}$, without contributing to the symmetric difference).
The key is now to note that any endpoint can be counted at most once in the whole sum $S$, since all $(a_i, a_{i+1})$ are disjoint (and the same applies to the $(b_i, b_{i+1})$), and the $a$’s can only appear between $b$’s (and vice-versa). In fact, $a_1$ and $b_n$ are never counted in $S$, and if $a_2$ is between $b_1$ and $b_2$, then $v_1$ must be isolated, so $a_2$ is not counted either. The sum is thus at most $2n - 3$. Since it has $n - 1$ terms, one of the terms, say $|N(v_t) \otimes N(v_{t+1})|$, must be at most 1. Therefore, the functionality of both $v_t$ and $v_{t+1}$ is at most $2$.
We have proved that each unit interval graph has a vertex of functionality at most $2$. Since this class is hereditary, we conclude that the functionality of any unit interval graph is at most $2$.
Permutation graphs
------------------
Let $\pi$ be a permutation of the elements in $\{1,2,\ldots,n\}$. The permutation graph of $\pi$ is a graph with vertex set $\{1,2,\ldots,n\}$ in which two vertices $i$ and $j$ are adjacent if and only if $(i-j)(\pi(i)-\pi(j))<0$. Clique-width is known to be unbounded in the class of permutation graphs [@perfect-cw], and so is degeneracy. In Section \[sec:perm-bound\], we show that functionality is bounded by a constant in this class. A similar result for unit interval graphs and graphs of bounded clique-width was proved by finding a pair of vertices with bounded symmetric difference of their neighbourhoods. This is not the case for permutation graphs, as we show in Section \[sec:perm-sd\]. This result is of independent interest, because in conjunction with Theorem \[cwd\] it provides an alternative proof of the fact that clique-width is unbounded in the class of permutation graphs.
### Functionality is bounded for permutation graphs {#sec:perm-bound}
For the purpose of this section, we associate a permutation $\pi$ with its plot, i.e. the set of points $(i, \pi(i))$ in the plane. We label those points by $\pi(i)$ and define the [*geometric neighbourhood*]{} of a point $k$ to be the union of two regions in the plane: the one above and to its left, and the one below and to its right. Then it is not difficult to see that the set of points of the permutation lying in the geometric neighbourhood of $k$ is precisely the set of neighbours of vertex $k$ in the permutation graph of $\pi$.
(1,6) circle (2pt) node\[below left\] [6]{} (2,1) circle (2pt) node\[below left\] [1]{} (3,4) circle (2pt) node\[below left\] [4]{} (4,2) circle (2pt) node\[below left\] [2]{} (5,5) circle (2pt) node\[below left\] [5]{} (6,3) circle (2pt) node\[below left\] [3]{};‘
(3, 7) – (3, 0); (0, 4) – (7, 4);
(3,4) – (0,4) – (0,7) – (3,7) – cycle; (3,4) – (7,4) – (7,0) – (3,0) – cycle;
The functionality of permutation graphs is at most 8.
Since the class of permutation graphs is hereditary, it suffices to show that every permutation graph contains a vertex of functionality at most 8. Let $G$ be a permutation graph corresponding to a permutation $\pi$. The proof will be given in two steps: first, we show that if there is a vertex with a specific property in $G$, then this vertex is a function of 4 other vertices. Second, we show how to find vertices that are “close enough” to having that property.
1: Consider the plot of $\pi$. Among any 3 horizontally consecutive points, one is vertically between the two others. We call such a point [*vertical middle*]{} (in the permutation from Figure \[examplepermutation\], the vertical middle points are 4, 2 and 3). Similarly, among any 3 vertically consecutive points, one is horizontally between the two others, and we call this point [*horizontal middle*]{} (in Figure \[examplepermutation\], the horizontally middle points are 2, 5 and 4).
Now let us suppose that $\pi$ has a point $x$ that is simultaneously a horizontal and a vertical middle point. Then $x$ is part of a triple $x$, $b$, $t$ (not necessarily in that order) of horizontally consecutive points, where $b$ is the bottom point (the lowest in the triple) and $t$ is the top point (the highest in the triple). Also, $x$ is part of a triple $x$, $l$, $r$ (not necessarily in that order) of vertically consecutive points, where $l$ is the leftmost and $r$ is the rightmost point in the triple (see Figure \[middlepoint\] for an illustration).
In general, $x$ can be at any of the 9 intersection points of pairs of 3 consecutive vertical and horizontal lines, i.e. $x$ is somewhere in $X$ (see Figure \[middlepointareas\]). We also have $l \in L$, $r \in R$, $t \in T$ and $b \in B$ for the surrounding points (see Figure \[middlepointareas\]). The important thing to note is that, since the points are consecutive, those are the [*only*]{} points of the permutation lying in the shaded area $X \cup L \cup R \cup T \cup B$. Any point different from $x,l,r,t,b$ lies in one of $Q_1$, $Q_2,Q_3$ or $Q_4$.
[0.49]{}
(3.5,0) – (3.5,10); (4,0) – (4,10); (4.5,0) – (4.5,10); (0,5.5) – (10,5.5); (0,6) – (10,6); (0,6.5) – (10,6.5);
(3.5,5.5) circle (2pt) node\[below left\] [$x$]{}; (1.5,6.5) circle (2pt) node\[below left\] [$l$]{}; (7,6) circle (2pt) node\[below left\] [$r$]{}; (4,2) circle (2pt) node\[below left\] [$b$]{}; (4.5,8) circle (2pt) node\[below left\] [$t$]{}; (4,6) – (4,10) – (0,10) – (0,6) – cycle; (4.5,6.5) – (10,6.5) – (10,0) – (4.5,0) – cycle;
[0.49]{}
(0,6.5) – (3.5, 6.5) – (3.5,10) – (4.5,10) – (4.5,6.5) – (10,6.5) – (10, 5.5) – (4.5,5.5) – (4.5,0) – (3.5,0) – (3.5, 5.5) – (0,5.5) – cycle;
(3.5,0) – (3.5,10); (4.5,0) – (4.5,10); (0,5.5) – (10,5.5); (0,6.5) – (10,6.5);
(4,6) node[$X$]{}; (1.75,6) node[$L$]{}; (7.25,6) node[$R$]{}; (4,8.25) node[$T$]{}; (4,2.75) node[$B$]{}; (7.25,8.25) node[$Q_1$]{}; (1.75,8.25) node[$Q_2$]{}; (1.75,2.75) node[$Q_3$]{}; (7.25,2.75) node[$Q_4$]{};
(0,6.5) – (3.5, 6.5) – (3.5,10) – (4.5,10) – (4.5,6.5) – (10,6.5) – (10, 5.5) – (4.5,5.5) – (4.5,0) – (3.5,0) – (3.5, 5.5) – (0,5.5) – cycle;
It is not difficult to see that the geometric neighbourhood corresponding to $(N(r) \cap N(b)) \cup (N(l) \cap N(t))$ (see Figure \[middlepoint\]) will always contain $Q_2$ and $Q_4$, and will never intersect $Q_1$ or $Q_3$. Therefore, the function that describes how $x$ depends on $\{l, r, t, b\}$ can be written as follows: $$f(x_r,x_b,x_l,x_t)=x_rx_b\vee x_lx_t,$$ where $x_r,x_b,x_l,x_t$ are Boolean variables corresponding to points $r,b,l,t$, respectively. In other words, a vertex $y\not \in \{x,l, r, t, b\}$ is adjacent to $x$ if and only if $$f(A(y,r),A(y,b),A(y,l),A(y,t))=1.$$
2: Let us relax the simultaneous middle point condition to the following one: amongst every 5 vertically (respectively horizontally) consecutive points, call the middle three [*weak horizontal*]{} (respectively [*vertical*]{}) [*middle points*]{}. Note that if the number of points is divisible by 5, at least $\frac{3}{5}$ of them are weak vertical and at least $\frac{3}{5}$ of them are weak horizontal middle points. Using this observation it is not hard to deduce that if there are at least 13 points, then more than half of them are weak vertical and more than half of them are weak horizontal middle points. Therefore, there must exist a point $x$ that is simultaneously both. We can deal with this case only, as the functionality of any graph on at most 12 vertices is at most 6. If $x$ is simultaneously a weak vertical and weak horizontal middle point, then there must exist quintuples $l$, $x$, $m_1$, $m_2$, $r$ and $t$, $x$, $m_3$, $m_4$, $b$ (not necessarily in that order), where $x$ is a simultaneous weak middle point in both directions, while $m_1$, $m_2$, $m_3$ and $m_4$ are the other weak middle points in their respective quintuples. By removing $m_1$, $m_2$, $m_3$ and $m_4$ from the graph, we find ourselves in the configuration of Step 1 and conclude that $x$ is a function of $\{l, r, t, b\}$ in the reduced graph. Therefore, in the original graph $x$ is a function of $\{l, r, t, b, m_1, m_2, m_3, m_4\}$, concluding the proof.
### Symmetric difference in permutation graphs {#sec:perm-sd}
Given a graph $G$ and a pair of vertices $x,y$ in $G$, let us denote $sd(x,y)=|N(x)\otimes N(y)|$ and $${\rm sd}(G)=\max\limits_H\min\limits_{x,y\in V(H)}sd(x,y),$$ where the maximum is taken over all induced subgraphs $H$ of $G$. With some abuse of terminology we call ${\rm sd}(G)$ the [*symmetric difference*]{} of $G$.
This parameter was used implicitly in Theorems \[cwd\] and \[thm:unit-interval\] to prove results about bounded functionality, because by bounding the symmetric difference we bound the functionality of a graph, which is easy to see. On the other hand, from the proof of Theorem \[cwd\] it follows that graphs of bounded symmetric difference extend graphs of bounded clique-width. Therefore, by showing that this parameter is unbounded in a class $X$ of graphs we prove that the clique-width is unbounded in $X$. Our next result shows the symmetric difference is unbounded for permutation graphs.
\[sdpermutation\] For any $t \in \mathbb N$, there is a permutation graph $G$ with ${\rm sd}(G) \geq t$.
Similarly to the previous section, we make use of the geometric representation of permutations. Given two vertices $x$ and $y$ of a permutation graph $G$, the symmetric difference of their neighbourhoods can be represented geometrically as an area in the plane (see Figure \[geometricsd\]). More precisely, a vertex different from $x$ and $y$ lies in the symmetric difference of their neighbourhoods if and only if the corresponding point of the permutation lies in the shaded area.
(2,1) circle (2pt) node\[below left\] [$x$]{} (5,5) circle (2pt) node\[below left\] [$y$]{};
(2, 0) – (2, 7); (5, 0) – (5, 7); (0, 1) – (7, 1); (0, 5) – (7, 5);
(5,0) – (5,1) – (2,1) – (2,0) – cycle; (0,1) – (2,1) – (2,5) – (0,5) – cycle; (2,7) – (5,7) – (5,5) – (2,5) – cycle; (7,5) – (5,5) – (5,1) – (7,1) – cycle;
In order to prove the theorem, it suffices, for each $t \in \mathbb N$, to exhibit a set $S_t$ of points in the plane (with no two on the same vertical or horizontal line) such that for any pair $x, y \in S_t$, there are at least $t$ other points of $S_t$ lying in the geometric symmetric difference of $x$ and $y$. Such a construction immediately gives rise to a permutation and thus to a permutation graph with symmetric difference at least $t$.
We construct sets $S_t$ in the following way (see Figure \[exampleset\] for an example):
- start with all the points with integer coordinates between 0 and $t$ inclusive;
- apply to the set the rotation sending $(1, 0)$ to $(1, \frac{1}{t+1})$ and $(0, 1)$ to $(-\frac{1}{t+1}, 1)$.
iin [0,...,6]{} [ in [0,...,6]{} [ (i- /7, i/7 + ) circle (2pt) node; ]{} ]{} (0, 0) – (0, 7); (0, 0) – (7, 0);
To see that these sets have indeed the desired property, let $x, y \in S_t$. For simplicity, we will use the coordinates of the points before the rotation. Suppose $x = (x_1, x_2)$ and $y = (y_1, y_2)$. There are two possible cases (after switching $x$ and $y$ if necessary):
- If $x_1 \leq y_1$ and $x_2 \leq y_2$, the following points all lie in the symmetric difference of $x$ and $y$:
- Points $(x_1, k)$ with $k < x_2$ (in the bottom region).
- Points $(x_1, k)$ with $x_2 < k \leq y_2$ (in the left region).
- Points $(y_1, k)$ with $y_2 < k$ (in the top region).
- Points $(y_1, k)$ with $x_2 \leq k < y_2$ (in the right region).
In particular, (1) and (3) account for at least $x_2 + t - y_2$ points, while (2) and (4) account for $2(y_2 - x_2)$ others. We conclude that in total, at least $t + (y_2 - x_2) \geq t$ points lie in the symmetric difference of $x$ and $y$.
- If $x_1 \leq y_1$ and $x_2 > y_2$, the following points all lie in the symmetric difference of $x$ and $y$:
- Points $(k, y_2)$ with $x_1 \leq k < y_1$ (in the bottom region).
- Points $(k, x_2)$ with $k < x_1$ (in the left region).
- Points $(k, x_2)$ with $x_1 < k \leq y_1$ (in the top region).
- Points $(k, y_2)$ with $y_1 < k$ (in the right region).
Summing up, we find again at least $t$ points in the symmetric difference of $x$ and $y$.
This result together with Theorem \[cwd\] give an alternative proof of the following known fact.
The class of permutation graphs has unbounded clique-width.
Line graphs and generalization
------------------------------
The line graph $L(G)$ of a graph $G$ is the graph with vertex set $E(G)$ in which two vertices are adjacent if and only if the corresponding edges of $G$ share a vertex. In other words, $L(G)$ is the intersection graph of edges of $G$. Both clique-width [@four] and degeneracy is unbounded in the class of line graphs. In this section we show that the functionality of a line graph is at most 6.
The functionality of line graphs is at most 6.
Let $G$ be a graph and $H$ be the line graph of $G$. Since the class of line graphs is hereditary, it suffices to prove that $H$ has a vertex of functionality at most 6. We will prove a stronger result showing that [*every*]{} vertex of $H$ has functionality at most $6$.
Let $x$ be a vertex in $H$, i.e. an edge in $G$. We denote the two endpoints of this edge in $G$ by $a$ and $b$. Assume first that both the degree of $a$ and the degree of $b$ is at least 4. Let $Y=\{y_1,y_2,y_3\}$ be a set of any three edges of $G$ incident to $a$, and let $Z=\{z_1,z_2,z_3\}$ be a set of any three edges of $G$ incident to $b$.
We claim that a vertex $v\not \in \{x\}\cup Y\cup Z$ is adjacent to $x$ in $H$ if and only if it is adjacent to every vertex in $Y$ or to every vertex in $Z$. Indeed, if $v$ is adjacent to $x$ in $H$, then the edge $v$ intersects the edge $x$ in $G$. If the intersection consists of $a$, then $v$ is adjacent to every vertex in $Y$ in the graph $H$, and if the intersection consists of $b$, then $v$ is adjacent to every vertex in $Z$ in the graph $H$. Conversely, let $v$ be adjacent to every vertex in $Y$, then $v$ must intersect the edges $y_1,y_2,y_3$ in $G$ at vertex $a$, in which case $v$ is adjacent to $x$ in $H$. Similarly, if $v$ is adjacent to every vertex in $Z$, then $v$ intersects the edges $z_1,z_2,z_3$ in $G$ at vertex $b$ and hence $v$ is adjacent to $x$ in $H$.
Therefore, in the case when both $a$ and $b$ have degree at least 4 in $G$, the function that describes how $x$ depends on $\{y_1,y_2,y_3,z_1,z_2,z_3\}$ in the graph $H$ can be written as follows: $$f(y_1,y_2,y_3,z_1,z_2,z_3)=y_1y_2y_3\vee z_1z_2z_3.$$
If the degree of $a$ is less than 4, we include in $Y$ all the edges of $G$ distinct from $x$ which are incident to $a$ (if there are any) and remove the term $y_1y_2y_3$ from the function. Similarly, if the degree of $b$ is less than 4, we include in $Z$ all the edges of $G$ distinct from $x$ which are incident to $b$ (if there are any) and remove the term $z_1z_2z_3$ from the function. If both terms have been removed, the function is defined to be identically 0, i.e. no vertices are adjacent to $x$ in $H$, except for those in $Y\cup Z$.
Having proved that the intersection graph of edges, i.e. the intersection graph of a family of $2$-subsets, has bounded functionality, it is natural to ask whether the intersection graph of a family of $k$-subsets has bounded functionality for $k>2$. This question is substantially harder and we present a solution only for $k=3$.
### Line graphs of 3-uniform hypergraphs
In this section we will show that intersection graphs of 3-uniform hypergraphs is a class of bounded functionality. We will denote a 3-uniform hypergraph with the ground set $V$ by $(V, \mathcal{S})$, where $\mathcal{S} \subseteq V \times V \times V$. We will use variables $s, s', s_1, s_2, \ldots$ to denote hyperedges, i.e. the elements of $\mathcal{S}$, and variables $v, v_1, v_2,\ldots $ to denote the elements of $V$. We will say that two hyperedges $s$ and $s'$ intersect if $s \cap s' \neq \emptyset$. We start with a preparatory result.
\[lemma1\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph and $v \in V$. Then one of the following holds:
- There are 3 hyperedges $s_1, s_2, s_3$ such that $s_i \cap s_j = \{v\}$ for all $1\leq i<j \leq 3$.
- There are 4 vertices $v_1, v_2, v_3, v_4$ such that each hyperedge $s \in \mathcal{S}$ that contains $v$ also contains at least one of the $v_1, v_2, v_3$ or $v_4$.
Consider the set $\mathcal{E}=\{s \backslash \{v\} : s \in \mathcal{S}, v \in s\}$. This is the set of pairs of vertices that are obtained by removing vertex $v$ from the hyperedges that contain $v$. Therefore, $(V, \mathcal{E})$ can be viewed as a graph. The lemma now says that either this graph contains a matching with 3 edges (as a subgraph) or it contains 4 vertices that any edge is adjacent to (vertex cover of size 4). The proof of this is now easy. One can take a maximal matching $M$, and if it has at least 3 edges, then we are done. In the other case, when the maximal matching $M$ has at most two edges, take $v_i$’s to be the vertices of the matching. If needed, add arbitrary vertices to obtain a set of 4 vertices. By maximality of the matching, every hyperedge contains at least one of the vertices selected, hence we are done as well.
The following two easy observations will be needed in the course of the proof.
\[obs1\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph. Suppose hyperedges $s_1, s_2, s_3 \in \mathcal{S}$ pairwise intersect at exactly one vertex, say $\{v\} = s_1 \cap s_2 = s_2 \cap s_3 = s_3 \cap s_1$. In other words, $s_1=\{v, v_1, v_2\}$, $s_2=\{v, v_3, v_4\}$, $s_1=\{v, v_5, v_6\}$, for some distinct vertices $v_1, v_2, \ldots, v_6$. Let $F'=\{(v_i, v_j, v_k): 1 \leq i \leq 2, 3 \leq j \leq 4, 5 \leq k \leq 6\}$ be the set of 8 hyperedges that intersect each of $s_1, s_2, s_3$ in exactly one vertex that is different from $v$. Then one can easily determine whether a given edge $s' \in \mathcal{S} \backslash F'$ contains vertex $v$ or not by looking at the intersection of $s'$ with $s_1, s_2, s_3$. Indeed, $s'$ contains $v$ if and only if $s'$ intersects each of $s_1$, $s_2$ and $s_3$.
\[obs2\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph. Suppose hyperedges $s_1, s_2, s_3 \in \mathcal{S}$ pairwise intersect at exactly 2 vertices. In other words, $s_1=\{v_1, v_2, v_3\}$, $s_2=\{v_1, v_2, v_4\}$ and $s_3=\{v_1, v_2, v_5\}$, for some distinct vertices $v_1, v_2, v_3, v_4, v_5 \in V$. Let $F'$ be the set containing the hyperedge $\{v_3, v_4, v_5\}$. Then one can easily determine whether a given edge $s' \in \mathcal{S} \backslash F'$ contains at least one of the vertices $v_1, v_2$ or not by looking at the intersection of $s'$ with $s_1, s_2, s_3$. Indeed, $s'$ contains $v_1$, $v_2$ or both if and only if $s'$ intersects each of $s_1$, $s_2$ and $s_3$.
Let $(V, \mathcal{S})$ be a 3-uniform hypergraph and let $v_1, v_2 \in V$. We will call the pair $v_1v_2$ [*thick*]{} if there are at least 32 hyperedges in $\mathcal{S}$ that contain $\{v_1, v_2\}$.
We will split our analysis into two cases. In the first lemma we will show that the intersection graphs of 3-uniform hypergraphs without thick pairs have bounded functionality. In the second lemma we will provide a structural theorem about hypergraphs containing thick pairs, from which a bounded functionality result follows easily as well. We note that in the case without thick pairs, we provide a bound on functionality for [*any*]{} vertex of the intersection graph. Meanwhile, in the case of hypergraphs with thick pairs, for any given bound $M$ one can find a hypergraph and a hyperedge such that corresponding vertex in the intersection graph has functionality at least $M$. Thus a structural result is needed in this case, to show that we can find a [*particular*]{} hyperedge in any given hypergraph with thick pairs, such that the functionality of the vertex corresponding to the hyperedge is bounded by a constant, that does not depend on the hypergraph.
We start with the case when there are no thick pairs.
\[lemmawithoutthick\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph without thick pairs. Then for any hyperedge $s \in \mathcal{S}$ there is a set of hyperedges $F \subset \mathcal{S} \backslash \{s\}$ of size $|F| \leq 462$ such that for any $s' \in \mathcal{S}\backslash (F \cup \{s\})$, one can determine whether $s'$ intersects $s$ by looking at the intersections of $s'$ with the hyperedges of $F$.
Let $s$ be any hyperedge in the hypergraph. Since we assume that there are no thick pairs, there are at most $30 \times 3$ hyperedges in $(V,\mathcal{S})$ that intersect $s$ in exactly 2 vertices. We denote this set of at most 90 hyperedges by $F_1$. Let $v \in s$, and consider the hyperedges in $(V, \mathcal{S} \backslash (F_1 \cup \{s\}))$ that contain vertex $v$. By Lemma \[lemma1\] we can distinguish between the following two cases.
- Assume there exist 3 hyperedges $s_1, s_2, s_3$ in $(V, \mathcal{S} \backslash (F_1 \cup \{s\}))$ that pairwise intersect at vertex $v$ only. In this case, we denote by $F_2$ the set of at most 11 hyperedges consisting of $s_1, s_2, s_3$ and all the hyperedges in $\mathcal{S}$ that have exactly one vertex in each of $s_1 \backslash \{v\}$, $s_2 \backslash \{v\}$ and $s_3 \backslash \{v\}$. According to Observation \[obs1\] we can determine whether a given hyperedge $s'\in \mathcal{S} \backslash (F_1 \cup F_2 \cup \{s\})$ contains $v$ or not by looking at the intersection of $s'$ with $s_1, s_2, s_3$.
- Suppose now that there exists a set of vertices $\{v_1, v_2, v_3, v_4\}$ such that every hyperedge in $(V, \mathcal{S} \backslash (F_1 \cup \{s\}))$ that contains $v$ also contains at least one of $v_1, v_2, v_3$ or $v_4$. In this case, we denote by $F_2$ the set of all the hyperedges that contain at least one of the pairs $\{v, v_1\}$, $\{v, v_2\}$, $\{v, v_3\}$ or $\{v, v_4\}$. By our assumption on no thick pairs, the set $F_2$ contains at most $31\times 4=124$ edges. Observe that no hyperedge $s' \in \mathcal{S} \backslash (F_1 \cup F_2 \cup \{s\})$ intersects $v$.
By analogy with building the set $F_2$ for the vertex $v$, we build two more sets $F_3$ and $F_4$ for the other two vertices contained in the hyperedge $s$, i.e. for the vertices in $s\backslash \{v\}$. Now it is easy to see that the $F=F_1 \cup F_2 \cup F_3 \cup F_4$ allows us to determine whether a given hyperedge $s' \in S \backslash (F \cup \{s\})$ intersects $s$ or not. Note that $F$ has size at most $90+3 \times 124=462$.
In our next result, we will show that a 3-uniform hypergraph with a thick pair contains one of the structures presented in Figure \[thickedge\], which we call “fly”, “windmill”, and “broken windmill”.
[0.3]{}
in [0]{}
\(1) at (+0, 2) [ ]{}; (2) at (-1, 0) [ ]{}; (3) at (+1, 0) [ ]{};
\(4) at (+1, 1) [ ]{}; (5) at (+2,1) [ ]{}; (6) at (+3,1) [ ]{};
\(7) at (-1,1) [ ]{}; (8) at (-2,1) [ ]{}; (9) at (-3,1) [ ]{};
(+0, 2) node\[anchor=south\] [$v_{1}$]{}; (-1, 0) node\[anchor=north\] [$v_{2}$]{}; (+1, 0) node\[anchor=north\] [$v_{3}$]{};
/in [1/2,2/3,3/1,1/4,1/5,1/6,1/7,1/8, 1/9, 3/4, 3/5, 3/6, 2/7, 2/8, 2/9]{} () – ();
[0.3]{}
in [0]{}
\(1) at (+0, 2) [ ]{}; (2) at (-1, 0) [ ]{}; (3) at (+1, 0) [ ]{};
\(4) at (+0, -0.4) [ ]{}; (5) at (+0, -0.8) [ ]{}; (6) at (+0, -1.2) [ ]{};
\(7) at (-2, 2.2) [ ]{}; (8) at (-2, 3) [ ]{};
\(9) at (+2, 2.2) [ ]{}; (10) at (+2, 3) [ ]{};
\(11) at (-0.5, 4) [ ]{}; (12) at (+0.5, 4) [ ]{};
(+0.4, 2) node\[anchor=south\] [$v_{1}$]{}; (-1.2, 0) node\[anchor=north\] [$v_{2}$]{}; (+1.2, 0) node\[anchor=north\] [$v_{3}$]{};
/in [1/2,2/3,3/1,2/4,2/5,2/6, 3/4, 3/5, 3/6, 1/7, 1/8, 1/9, 1/10, 1/11, 1/12, 7/8, 9/10, 11/12]{} () – ();
[0.3]{}
in [0]{}
\(1) at (+0, 2) [ ]{}; (2) at (-1, 0) [ ]{}; (3) at (+1, 0) [ ]{};
\(4) at (+0, -0.4) [ ]{}; (5) at (+0, -0.8) [ ]{}; (6) at (+0, -1.2) [ ]{};
(+1.3, 2.1) node\[anchor=south\] [$v_{1}$ [bounded degree]{}]{}; (-1.2, 0) node\[anchor=north\] [$v_{2}$]{}; (+1.2, 0) node\[anchor=north\] [$v_{3}$]{};
/in [1/2,2/3,3/1,2/4,2/5,2/6, 3/4, 3/5, 3/6]{} () – ();
\(1) – (-0.4, 2.2); (1) – (-0.3, 2.3); (1) – (-0.1, 2.3); (1) – (+0.1, 2.3); (1) – (+0.3, 2.3); (1) – (+0.4, 2.2);
To prove the result about these three structures, we need the following observation.
\[obs3\] Let $(V,\mathcal{S})$ be a 3-uniform hypergraph and let $v \in V$ be a vertex that does not belong to any thick pair. Then one of the following holds:
- Either there are 3 hyperedges $s_1, s_2, s_3$ that pairwise intersect only at vertex $v$.
- Or vertex $v$ is contained in at most 124 hyperedges of $(V,\mathcal{S})$.
From Lemma \[lemma1\], it follows that either there are 3 hyperedges $s_1, s_2, s_3$ that pairwise intersect only at vertex $v$, or there are 4 vertices $v_1, v_2, v_3$ and $v_4$ such that every hyperedge that contains $v$ also contains at least one of $v_1, v_2, v_3$ or $v_4$. Note that in the second case, since neither of $vv_1, vv_2, vv_3$ and $vv_4$ is thick, there are at most 31 hyperedges containing one of these pairs. Therefore, there are at most $31 \times 4 = 124$ hyperedges that contain $v$. This finishes the proof of the observation.
\[lemmawiththick\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph that contains a thick pair. Then it contains one of the following:
- A “fly”, which is a hyperedge $s=\{v_1, v_2, v_3\}$ together with hyperedges $s_1, s_2, s_3, s_4, s_5, s_6$ such that $s_1, s_2, s_3$ intersect $s$ at $\{v_1,v_2\}$ and $s_4, s_5, s_6$ intersect $s$ at $\{v_1, v_3\}$.
- A “windmill”, which is a hyperedge $s=\{v_1, v_2, v_3\}$ together with hyperedges $s_1, s_2, s_3, s_4, s_5$, $s_6$ such that $s_1, s_2, s_3$ intersect $s$ at $\{v_2, v_3\}$ and $s_4, s_5, s_6$ intersect $s$ at $v_1$ and the pairwise intersection of $s_4, s_5, s_6$ is vertex $v_1$ as well.
- A “broken windmill”, which is a hyperedge $s=\{v_1, v_2, v_3\}$ together with hyperedges $s_1, s_2, s_3$ such that $s_1, s_2, s_3$ intersect $s$ at $\{v_2, v_3\}$ and there are only at most $124$ hyperedges in $(V, \mathcal{S}\backslash \{s\})$ that contain vertex $v_1$.
Let $T=\{(v, v_1v_2): \{v, v_1, v_2\} \in \mathcal{S}, v_1v_2$ is a thick pair$\}$. The main idea of the proof is counting the elements in $T$. Let $E$ be the set of thick pairs and let $W=\{v \in V : (v, v_1v_2) \in T$ for some $ v_1, v_2 \in V\}$.
Note that any thick pair belongs to at least 32 hyperedges, hence $|T| \geq 32 |E|$. Also note that each vertex of $W$ belongs to some thick pair, or else a “windmill” or a “broken windmill” appears. Indeed, assume $s=\{v, v_1, v_2\}$ is a hyperedge with a thick pair $v_1v_2$ and $v$ does not belong to any thick pair. Then, by Observation \[obs3\], either there are at most 124 hyperedges containing vertex $v$ in $(V, \mathcal{S} \backslash \{s\})$ or there are three hyperedges $s_1$, $s_2$, $s_3$ in $(V, \mathcal{S} \backslash \{s\})$ that pairwise intersect only at vertex $v$. Together with any three hyperedges different from $s$ that contain the vertices of the thick pair $\{v_1,v_2\}$ and are different from $s$, this gives us either a “windmill” or a “broken windmill”.
Thus, from now on, we will assume that each vertex of $W$ belongs to a thick pair. As each thick pair contains at most two vertices, we have $|W| \leq 2 |E|$. We conclude that $|T| \geq 32 |E| \geq 16 |W|$. Hence, by the pigeonhole principle, there are 16 elements of $T$ that contain the same element $v \in W$. In other words, we have 16 different thick pairs $e_1, e_2, \ldots, e_{16}$ that make a hyperedge with vertex $v$, i.e. $(v, e_i) \in T$ for all $i$.
Consider first the case when four of these pairs, say $e_1, e_2, e_3, e_4$, have a vertex $w$ in common. Denote these pairs by $e_1=wz$, $e_2=wu_1$, $e_3=wu_2$ and $e_4=wu_3$. Then $s=\{v,w,z\}$, with $s_1=\{v,w,u_1\}$, $s_2=\{v,w,u_2\}$, $s_3=\{v,w,u_3\}$, together with any three hyperedges that contain the thick pair $e_1=wz$ and are different from $s$, gives us a “fly”.
Finally, consider the case when no four pairs of $e_1, e_2, \ldots, e_{16}$ share a vertex in common. Then a graph $G$ with the edges set $\{e_1, e_2, \ldots, e_{16}\}$ has degree at most 3 and hence it must contain a matching of size at least 4. Indeed, by observing that each edge of $G$ is incident to at most 4 other edges, one can pick any edge and remove all incident edges repeatedly at least 4 times to obtain the required matching of size 4. Now, as each of these four edges is a thick pair and forms a hyperedge with $v$, we can easily see that a “windmill” appears. This finishes the proof.
\[cor1\] Let $(V, \mathcal{S})$ be a 3-uniform hypergraph that contains a thick pair. Then there is a hyperedge $s \in \mathcal{S}$ and a set of hyperedges $F \subset \mathcal{S} \backslash \{s\}$ of size $|F| \leq 128$ such that such that for any $s' \in \mathcal{S} \backslash (F \cup \{s\})$, one can determine whether $s'$ intersects $s$ by looking at the intersections of $s'$ with the hyperedges of $F$.
We use the notation of the statement of Lemma \[lemmawiththick\]. Let $s$ be a hyperedge given by Lemma \[lemmawiththick\] belonging either to a “fly” or to a “windmill” or to a “broken windmill”.
If $s$ belongs to a “fly”, then we can take $F$ to consist of 6 hyperedges $s_1, s_2, \ldots, s_6$ together with 2 further possible hyperedges on the wings of the fly (if the hypergraph contains it) on vertices $(s_1 \cup s_2 \cup s_3) \backslash \{v_1,v_2\}$ and $(s_4 \cup s_5 \cup s_6) \backslash \{v_1,v_3\}$. It now follows from Observation \[obs2\] that intersecting any hyperedge $s' \in S \backslash (F \cup s)$, with $s_1, s_2$ and $s_3$ one can determine whether $s'$ contains either $v_1$ or $v_2$. Similarly, intersecting with $s_4, s_5, s_6$, determines whether $s'$ contains either $v_1$ or $v_3$. Hence, by looking at the intersection of the edges of $F$ with $s'$ we can determine whether $s'$ intersects $s$ or not.
If $s$ belongs to a “windmill”, we can take $F$ to consist of 6 hyperedges $s_1, s_2, \ldots, s_6$ together with one possible hyperedge on $(s_1 \cup s_2 \cup s_3) \backslash \{v_2, v_3\}$ and 8 possible hyperedges on $(s_4 \cup s_5 \cup s_6) \backslash \{v_1\}$ that have one vertex in each wing of the “windmill”. By Observation \[obs2\] intersection of $s'$ with $s_1,s_2, s_3$ determines whether $s'$ contains either $v_2$ or $v_3$, while Observation \[obs1\] allows us to determine whether $s'$ contains $v_1$ by looking at the intersection of $s'$ with $s_4, s_5$ and $s_6$. Thus, again we can determine whether $s'$ intersects $s$.
If $s$ belongs to a “broken windmill”, we can take $F$ to consist of all the hyperedges that contain vertex $v_1$, of which there are at most 124, also with $s_1, s_2, s_3$ and one further possible hyperedge on the set $(s_1 \cup s_2 \cup s_3) \backslash \{v_2, v_3\}$. Now $F$ contains at most 128 edges and it is clear by Observation \[obs2\] that this set determines whether $s'$ intersects $s$ or not.
From Lemma \[lemmawithoutthick\] and Corollary \[cor1\] we deduce our main result of this section.
Intersection graphs of 3-uniform hypergraphs have functionality bounded by 462.
Graphs of large functionality {#sec:unbounded}
=============================
Knowing what is good without knowing what is bad is just half-knowledge. Therefore, in this section we turn to graphs of large functionality.
When we talk about graphs of large functionality we assume that we deal with an infinite family $X$ of graphs, because in any finite collection of graphs functionality is bounded by a constant. Moreover, we can further assume that $X$ is hereditary. Indeed, if $X$ is not hereditary, we can extend it to a hereditary class by adding all induced subgraphs of graphs in $X$, and this extension has (un)bounded functionality if and only if $X$ has, because by definition the functionality of an induced subgraph of a graph $G$ is never larger than the functionality of $G$.
In [@implicit], it was shown that any hereditary class of graphs of bounded functionality has $2^{O(n\log_2 n)}$ labelled graphs with $n$ vertices. In the terminology of [@SpHerProp] these are classes with (at most) factorial speed of growth, or simply (at most) factorial classes. Therefore, in every superfactorial class functionality is unbounded. This is the case, for instance, for bipartite, co-bipartite and split graphs, since each of these classes contains at least $2^{n^2/4}$ labelled graphs with $n$ vertices. We state this formally as a lemma.
\[lem:three-classes\] Functionality is unbounded in the classes of bipartite, co-bipartite and split graphs.
This conclusion allows us to establish a relationship between functionality and one more important graph parameter known as VC-dimension.
A set system $(X,S)$ consists of a set $X$ and a family $S$ of subsets of $X$. A subset $A\subseteq X$ is [*shattered*]{} if for every subset $B\subseteq A$ there is a set $C\in S$ such that $B=A\cap C$. The VC-dimension of $(X,S)$ is the cardinality of a largest shattered subset of $X$.
The VC-dimension of a graph $G=(V,E)$ was defined in [@VC] as the VC-dimension of the set system $(V,S)$, where $S$ the family of closed neighbourhoods of vertices of $G$, i.e. $S=\{N[v]\ :\ v\in V(G)\}$. We denote the VC-dimension of $G$ by $vc(G)$.
There exists a function $f$ such that for any graph $G$, $vc(G)\le f(fun(G))$.
Fix a $k$ and consider the class $X_k$ of all graphs of functionality at most $k$. Clearly, $X_k$ is hereditary. Assume $X_k$ contains graphs of arbitrarily large VC-dimension and let $G_1,G_2,\ldots$ be an infinite sequence of graphs from $X_k$ with strictly increasing values of the VC-dimension. Let $Y$ be the hereditary class containing all these graphs and all their induced subgraphs. Then $Y$ is a hereditary subclass of $X_k$ with unbounded VC-dimension. It is was shown in [@Lozin] that the only minimal hereditary classes of graph of unbounded VC-dimension are bipartite, co-bipartite and split graphs. But then $Y$ and hence $X_k$ contains one of these three classes, which is a contradiction to Lemma \[lem:three-classes\]. Therefore, there is a constant $f(k)$ bounding the VC-dimension of graphs in $X_k$, which defines the function $f$.
This theorem shows that the family of classes of bounded VC-dimension extends the family of classes of bounded functionality. The next result shows that this extension is proper and reveals several classes of unbounded functionality and bounded VC-dimension.
\[thm:five-classes\] The following classes have bounded VC-dimension and unbounded functionality:
- chordal bipartite graphs,
- complements of chordal bipartite graphs,
- strongly chordal graphs,
- bipartite graphs of girth at least $k$, for a fixed value of $k$,
- finitely defined monotone classes containing all forests.
To see that VC-dimension is bounded in all these classes, observe that none of them contains any of the three minimal classes of unbounded VC-dimension (bipartite, co-bipartite and split graphs).
To prove the unboundedness of functionality in these classes, we will show that all of them are superfactorial. For the first three classes a superfactorial bound on the number of $n$-vertex labelled graphs was shown in [@Spinrad] and it equals $2^{\Theta(n\log^2_2 n)}$.
Now we turn to the last two families of graph classes and observe that all of them are monotone. It is known (see e.g. [@girth]) that for each $k$ there exist $(C_3,C_4,\ldots,C_k)$-free bipartite graphs with $n$ vertices and $\Omega (n^{1+1/k})$ edges. Since the class of $(C_3,C_4,\ldots,C_k)$-free bipartite graphs is monotone, we conclude that the number of $n$-vertex labelled graphs in this class is at least $2^{\Omega (n^{1+1/k})}$, i.e. the class is superfactorial.
Finally, let $X$ be a finitely defined monotone class containing all forests. Since $X$ contains all forests, every forbidden graph for $X$ contains a cycle, and since the number of forbidden graphs is finite, there is a largest $k$ such that every forbidden graph contains an induced cycle of length at most $k$. Therefore, $X$ contains all $(C_3,C_4,\ldots,C_k)$-free graphs, and hence, as before, $X$ contains graphs with $n$ vertices and $\Omega (n^{1+1/k})$ edges. Since $X$ is monotone, it contains at least $2^{\Omega (n^{1+1/k})}$ labelled graphs with $n$ vertices, i.e. $X$ is superfactorial.
We observe that this theorem cannot be extended to the family of all monotone classes, as the example of forests shows. Obviously, this class is monotone (and contains all forests), but functionality is bounded by 1 in this class, since every forest has a vertex of degree at most 1. Nevertheless, in conjunction with Theorem \[thm:monotone-bounded\] the above result provides the following dichotomy for finitely defined monotone classes.
A finitely defined monotone class has bounded functionality if and only if it does not contain all forests.
So far, we have identified some classes containing graphs of large functionality. However, presenting specific constructions of graphs of large functionality is a task, which is not so straightforward. We solve it in the following section.
Constructing graphs of large functionality
------------------------------------------
Since large VC-dimension implies large functionality, it would be natural to construct graphs of large functionality through constructing graphs of large VC-dimension. The latter is an easy task. Indeed, consider the bipartite graph $D_n=(A,B,E)$ with two parts $|A|=n$ and $|B|=2^n$. For each subset $C\subseteq A$ we create a vertex in $B$ whose neighbourhood coincide with $C$. Clearly, the VC-dimension of $D_n$ is $n$ and hence with $n$ growing the functionality of $D_n$ grows as well.
However, this example is not very interesting in the sense that $D_n$ contains vertices of low functionality (of low degree) and hence graphs of large functionality are hidden in $D_n$ as proper induced subgraphs. A much more interesting task is constructing graphs where [*all*]{} vertices have large functionality. In what follows, we show that this is the case for hypercubes.
Let $V_n=\{0,1\}^n$ be the set of binary sequences of length $n$ and let $v,w\in V_n$. The Hamming distance $d(v,w)$ between $v$ and $w$ is the number of positions in which the two sequences differ. A [*hypercube*]{} $Q_n$ is the graph with vertex set $V_n=\{0,1\}^n$, in which two vertices are adjacent if and only if the Hamming distance between them equals 1.
Functionality of the hypercube $Q_n$ is at least $(n-1)/3$.
By symmetry, it suffices to show that the vertex $v=00 \ldots 0 \in V_n$ has functionality at least $(n-1)/3$. Let $v$ be a function of vertices in a set $S \subseteq V_n \backslash \{v\}$. To provide a lower bound on the size of $S$, and hence a lower bound on the functionality of $v$, for each $i=1, 2, \ldots, n$ consider the set $S_i=\{w \in S: d(w, v)=i\}$, i.e. the set of all binary sequences in $S$ that contain exactly $i$ 1s. Also, consider the following set: $$I=\{i \in \{1,2, \dots, n\} : \exists z = z_1z_2 \ldots z_n \in S_1 \cup S_2 \cup S_3 \ \mbox{with} \ z_i=1\}.$$ Suppose $|I|\le n-2$. Then there exist two positions $i$ and $j$ such that for any sequence $z = z_1z_2 \ldots z_n \in S_1 \cup S_2 \cup S_3$, we have $z_i=0$ and $z_j=0$. Consider the following two vertices:
- $u=u_1u_2\ldots u_n$ with $u_k=1$ if and only if $k=i$,
- $w=w_1w_2 \ldots w_n$ with $w_k=1$ if and only if $k=i$ or $k=j$.
We claim that $u$ and $w$ are not adjacent to any vertex $z \in S$. First, it is not hard to see that for any $z \in S_1 \cup S_2 \cup S_3$ we have $d(z, u) \geq 2$ and $d(z, w) \geq 2$. Indeed, any $z \in S_1 \cup S_2 \cup S_3$ differs from $u$ and $w$ in position $i$, i.e. $z_i=0$ and $u_i=w_i=1$, and there must exist a $k\ne i,j$ with $z_k=1$ and $u_k=w_k=0$. Also, it is not difficult to see that $d(z, u) \geq 2$ and $d(z, w) \geq 2$ for any vertex $z \in S \backslash (S_1 \cup S_2 \cup S_3)$, because any such $z$ has at least four 1s, while $u$ and $w$ have at most two 1s. Therefore, by definition, $u$ and $w$ are not adjacent to any vertex in $S$.
We see that the assumption that $|I|\le n-2$ leads to the conclusion that there are two vertices $u, w \in Q_n \backslash (S \cup \{v\})$ which are non-adjacent to any vertex in $S$, but have different adjacencies to $v$. This contradicts the fact that $v$ is a function of the vertices in $S$. So, we must conclude that $I$ has size at least $n-1$. As each vertex in $S_1 \cup S_2 \cup S_3$ has at most three 1s, we conclude that $S_1 \cup S_2 \cup S_3$ must contain at least $|I|/3=(n-1)/3$ vertices. This completes the proof of the theorem.
We conclude this section by observing that the hereditary closure of the set of hypercubes, i.e. the hereditary class containing all the hypercubes and all their induced subgraphs, is one more example of a hereditary class of unbounded functionality and bounded VC-dimension. The difference between this example and the classes in Theorem \[thm:five-classes\] is that the speed of the hereditary closure of hypercubes is an open question. We discuss some other open questions related to the topic of the paper in the concluding section.
Concluding remarks and open problems {#sec:conclusion}
====================================
In this paper, we proved a number of results about graph functionality. However, many questions on this topic remain unanswered.
Bounded functionality, implicit representation and factorial properties of graphs
---------------------------------------------------------------------------------
Let us repeat that any hereditary class of graphs of bounded functionality is at most factorial [@implicit]. It is natural to ask whether all factorial classes are of bounded functionality.
Is it true that for any hereditary class with at most factorial speed of growth there exists a constant bounding the functionality of graphs in the class?
To emphasize the importance of the family of factorial classes let us mention that it contains many classes of theoretical or practical importance, such as line graphs, interval graphs, permutation graphs, threshold graphs, forests, planar graphs and, even more generally, all proper minor-closed graph classes, all classes of graphs of bounded vertex degree, of bounded clique-width, etc.
There is one more important notion associated with factorial classes of graphs, namely, the notion of implicit representation of graphs, which was introduced in [@implicit-0] and then further developed in [@implicit-1]. Similarly to bounded functionality, any hereditary class that admits an implicit representation is at most factorial. However, the question whether all factorial classes admit implicit representations, also known as [*implicit graph representation conjecture*]{}, is widely open. We ask whether there is any relationship between the two notions.
Does implicit representation implies bounded functionality and/or vice versa?
Other open questions
--------------------
We conclude the paper with a number of other open questions related to the notion of graph functionality. Some of them are motivated by the results presented in the paper. We list them in no particular order. The first of them is inspired by a result in [@implicit] showing that if the family of prime (with respect to modular decomposition) graphs in a hereditary class $X$ is factorial, then the entire class $X$ is factorial.
Is it true that if prime (with respect to modular decomposition) graphs in a hereditary class $X$ have bounded functionality, then all graphs in $X$ have bounded functionality?
Describe explicitly any function satisfying $vc(G)\le f(fun(G))$.
Are there any minimal hereditary classes of unbounded functionality?
Is functionality bounded for interval graphs, or more generally, for graphs of bounded boxicity?
What is the time complexity of computing the functionality of a graph?
Are there NP-hard problems that admit polynomial-time or fixed-parameter tractable algorithms for graphs of bounded functionality?
[99]{} A. Atminas, A. Collins, V. Lozin, and V. Zamaraev, Implicit representations and factorial properties of graphs, [*Discrete Mathematics*]{}, 338 (2015) 164–179.
N. Alon, G. Brightwell, H. Kierstead, A. Kostochka, P. Winkler, Dominating sets in $k$-majority tournaments. [*Journal of Combinatorial Theory, Series B*]{}. 96 (2006), 374–387.
J. Balogh, B. Bollobás, D. Weinreich, The speed of hereditary properties of graphs, *J. Combin. Theory Ser. B* 79 (2000) 131–156.
H.-J. Bandelt, H.M. Mulder, Distance-hereditary graphs, [*Journal of Combinatorial Theory, Series B*]{}, 41 (2) (1986) 182–208.
A. Brandstädt, J. Engelfriet, Hoang-Oanh Le and V. Lozin, Clique-Width for Four-Vertex Forbidden Subgraphs, [*Theory of Computing Systems*]{}, 34 (2006) 561–590.
B. Courcelle, J. Engelfriet and G. Rozenberg, Handle-rewriting hypergraph grammars, *Journal of Computer and System Sciences* 46 (1993) 218–270.
B. Courcelle, J.A. Makowsky and U. Rotics, Linear time solvable optimization problems on graphs of bounded clique-width, [*Theory Comput. Syst.*]{} 33 (2000) 125–150.
B. Courcelle and S. Olariu, Upper bounds to the clique-width of a graph, [*Discrete Applied Math.*]{} 101 (2000) 77–114.
M.C. Golumbic, U. Rotics, On the clique-width of some perfect graph classes. [*Internat. J. Found. Comput. Sci.*]{} 11 (2000), no. 3, 423–443.
S. Kannan, M. Naor, S. Rudich, Implicit representation of graphs. In 20th Annual ACM Symposium on Theory of Computing (STOC), pages 334–343. ACM Press, May 1988.
S. Kannan, M. Naor, S. Rudich, Implicit representation of graphs, [*SIAM Journal on Discrete Mathematics*]{}, 5 (1992) 596–603.
F. Lazebnik, V. A. Ustimenko, and A. J.Woldar, A new series of dense graphs of high girth. [*Bull AMS*]{} 32 (1995), 73–79.
V. Lozin, Minimal classes of graphs of unbounded clique-width, [*Annals of Combinatorics*]{}, 15 (2011) 707–722.
V. Lozin, Graph Parameters and Ramsey Theory, [*Lecture Notes in Computer Science*]{}, 10765 (2018) 185–194.
J. P. Spinrad, Nonredundant 1’s in $\Gamma$-Free Matrices, [*SIAM J. Discrete Mathematics*]{}, 8 (1995) 251–257.
[^1]: Mathematics Institute, University of Warwick, Coventry, CV4 7AL, UK. [email protected]
[^2]: Department of Mathematics, London School of Economics, London WC2A 2AE, UK, [email protected]
[^3]: Mathematics Institute, University of Warwick, Coventry, CV4 7AL, UK. [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The XENON100 experiment, located at the Laboratori Nazionali del Gran Sasso (LNGS), aims to directly detect dark matter in the form of Weakly Interacting Massive Particles (WIMPs) via their elastic scattering off xenon nuclei. We present a comprehensive study of the predicted electronic recoil background coming from radioactive decays inside the detector and shield materials, and intrinsic radioactivity in the liquid xenon. Based on GEANT4 Monte Carlo simulations using a detailed geometry together with the measured radioactivity of all detector components, we predict an electronic recoil background in the WIMP-search energy range and 30 kg fiducial mass of less than 10$^{-2}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$, consistent with the experiment’s design goal. The predicted background spectrum is in very good agreement with the data taken during the commissioning of the detector in Fall 2009.'
author:
- 'E. Aprile'
- 'K. Arisaka'
- 'F. Arneodo'
- 'A. Askin'
- 'L. Baudis'
- 'A. Behrens'
- 'K. Bokeloh'
- 'E. Brown'
- 'J.M.R. Cardoso'
- 'B. Choi'
- 'D. Cline'
- 'S. Fattori'
- 'A.D. Ferella'
- 'K.-L. Giboni'
- 'A. Kish'
- 'C.W. Lam'
- 'J. Lamblin'
- 'R.F. Lang'
- 'K.E. Lim'
- 'Q. Lin'
- 'S. Lindemann'
- 'M. Lindner'
- 'J.A.M. Lopes'
- 'K. Lung'
- 'T. Marrodán Undagoitia'
- 'Y. Mei'
- 'A.J. Melgarejo Fernandez'
- 'K. Ni'
- 'U. Oberlack'
- 'S.E.A. Orrigo'
- 'E. Pantic'
- 'G. Plante'
- 'A.C.C. Ribeiro'
- 'R. Santorelli'
- 'J.M.F. dos Santos'
- 'M. Schumann'
- 'P. Shagin'
- 'H. Simgen'
- 'A. Teymourian'
- 'D. Thers'
- 'E. Tziaferi'
- 'H. Wang'
- 'M. Weber'
- 'C. Weinheimer'
title: Study of the electromagnetic background in the XENON100 experiment
---
Introduction {#intro}
============
For all experiments dealing with very low signal rates, such as dark matter or double beta decay searches, the reduction and discrimination of the background is one of the most important and difficult tasks. As the sensitivity of these experiments keeps increasing, the fight against the background remains crucial.
The XENON100 detector, which is installed in the Laboratori Nazionali del Gran Sasso (LNGS), Italy, is the second generation detector within the XENON program, dedicated to the direct detection of particle dark matter in the form of Weakly Interacting Massive Particles (WIMPs) [@wimps]. It is the successor of XENON10 [@xe10-instrument], which has set some of the best limits on WIMP-nucleon scattering cross-sections [@xe10-independent; @xe10-dependent]. XENON100 aims to improve this sensitivity due to an increase of the target mass and a significant reduction of the background in the target volume.
In the standard scenario, WIMPs are expected to elastically scatter off xenon nuclei resulting in low energy nuclear recoils. Neutrons passing through the detector also produce nuclear recoils of similar energy, whereas gamma rays and electrons produce electronic recoils. This opens the possibility to efficiently reject the electromagnetic background using various discrimination techniques. Experiments like XENON100 distinguish electronic interactions from nuclear recoils based on a different ratio in the yield of scintillation light (primary signal, S1) and ionization charge (secondary signal, S2). Using this discrimination technique, XENON10 and XENON100 reached an electronic recoil rejection efficiency better than 99% at 50% nuclear recoil acceptance [@xe100-independent; @xe10-independent].
The main sources of electronic recoil background in XENON100 are radioactive contamination of the materials used to construct the detector and the shield, intrinsic radioactivity in the LXe target, and the decays of $^{222}$Rn and its progeny inside the detector shield. Even if the electronic recoil rejection efficiency, based on the ratio of the scintillation and ionization signals, is high, a potential statistical leakage of electronic recoil events into the nuclear recoil region can mimic a dark matter signal. One way to handle this background is a well-planned detector design avoiding the presence of radioactive materials close to the active volume.
During the design phase of XENON100, all detector materials and components have been carefully selected based on measurements of their radioactive contamination in order to achieve a low background level. The background is further suppressed by improvements on the passive shield and by surrounding the target volume with an active LXe veto layer.
In this paper we summarize the effort to use extensive Monte Carlo simulations to predict the electronic recoil background of XENON100 from natural radioactivity in the detector and shield components, and to study the background reduction by applying fiducial volume and veto coincidence cuts. Section \[secDetectorModel\] describes the detector model which has been used in the simulations. The predicted electronic recoil background from the detector and shield materials is discussed in Section \[secDetectorMaterials\], the background from the decays of $^{222}$Rn and its progenies in the shield cavity in Section \[secRadonCavity\], and the background from radon and krypton in LXe in Section \[secIntrinsicBG\]. The comparison of the background model with the measured background spectrum is presented in Section \[secComparison\], and conclusions are drawn in Section \[secConclusions\].
XENON100 detector design and model simulated with the GEANT4 toolkit {#secDetectorModel}
====================================================================
The XENON100 detector is a dual phase time-projection chamber (TPC). The total amount of 161 kg of LXe is enclosed in the vacuum insulated cryostat, made from the low activity stainless steel of type 1.4571/316Ti (316Ti SS). The target consists of 62 kg of LXe, defined by a structure made from polytetrafluoroethylene (PTFE, [*Teflon*]{}) and copper. The target volume is viewed by two arrays of photomultiplier tubes (PMT), one on the bottom immersed in LXe, and one in the gas phase above the target volume. The electric fields in the TPC are generated by applying potential differences across the electrodes, which are made of stainless steel meshes welded onto 316Ti SS rings. They include two electrodes on the bottom of the TPC above the bottom PMT array, and a stack of three electrodes at the liquid-gas interface.
In order to simulate the response of the detector to various types of particles and to predict the intrinsic and ambient electronic recoil background, a detailed model (Figure \[DetectorModel\]) has been created with the GEANT4 toolkit [@g4]. Table \[tab:ScreeningResults\] shows the amount of materials used for the detector construction, computed from the model and in agreement with the actual detector.
--------------------------------------- ------------------------- ------------------------ --------------------- --------------------- --------------------- ----------------------------------
Component Amount
$^{238}$U / $^{226}$Ra $^{232}$Th $^{60}$Co $^{40}$K other nuclides
Cryostat and ’diving bell’ (316Ti SS) 73.61 kg 121.46 147.23 404.87 662.52
Support bars (316Ti SS) 49.68 kg 64.58 144.07 69.55 352.73
Detector PTFE 11.86 kg 0.71 1.19 0.36 8.89
Detector copper 3.88 kg 0.85 0.62 5.21 0.78
PMTs 242 pieces 60.50 111.32 181.50 1972.30 $^{137}$Cs: 41.14
PMT bases 242 pieces 38.72 16.94 2.42 38.72
TPC resistor chain 1.5$\times$10$^{-3}$ kg 1.11 0.57 0.12 7.79
Bottom electrodes (316Ti SS) 0.23 kg 0.43 0.45 2.14 2.36
Top electrodes (316Ti SS) 0.24 kg 0.85 0.43 1.73 1.16
PMT cables 1.80 kg 0.85 1.97 0.37 18.65 $^{108m}$Ag: 2.67
Copper shield 2.1$\times$10$^{3}$ kg 170.80 24.69 6.59 80.26
Polyethylene shield 1.6$\times$10$^{3}$ kg 368.0 150.4 - 1120.0
Lead shield (inner layer) 6.6$\times$10$^{3}$ kg 4.3$\times$10$^{3}$ 3.6$\times$10$^{3}$ 7.2$\times$10$^{2}$ 9.6$\times$10$^{3}$ $^{210}$Pb: 1.7$\times$10$^{8}$
Lead shield (outer layer) 27.2$\times$10$^{3}$ kg 1.1$\times$10$^{5}$ 1.4$\times$10$^{4}$ 2.9$\times$10$^{3}$ 3.8$\times$10$^{5}$ $^{210}$Pb: 1.4$\times$10$^{10}$
--------------------------------------- ------------------------- ------------------------ --------------------- --------------------- --------------------- ----------------------------------
PMT part Material Weight \[g\]
----------------------------- -------------------- --------------
Metal package and stem pins Kovar metal 13.0
Electrodes stainless steel 7.0
Glass for window synthetic silica 2.0
Glass in stem borosilicate glass 1.0
Aluminum ring Al 0.1
Insulator ceramic 0.04
Getter ZrAl 0.02
: Mass model of the R8520-06-AL PMT [@PMTmassModel]. The last two materials are not included in the GEANT4 model due to their low mass. []{data-label="tab:PMTmassModel"}
The passive shield, with 4$\pi$ coverage of the detector, is installed on a 25 cm polyethylene slab (not shown). From outside to inside, it consists of tanks filled with water (thickness 20 cm, not shown) to shield against ambient neutrons, placed on 4 sides of the shield box. After the water shield, there are two layers of lead: a 15 cm outer layer and a 5 cm inner layer with low level of the radioactive isotope ${^{210}}$Pb (Table \[tab:ScreeningResults\]). Inside the lead box, there are 20 cm of polyethylene against neutron background. The innermost shield layer consists of 5 cm thick (0.5 cm on the bottom) copper plates. It reduces the gamma background from the outer shield layers. Highly radioactive detector components are mounted outside of the shield, for example signal and high voltage feedthroughs, vacuum pumps, pressure sensors and associated electronics. An innovative detector design feature, which has contributed to the low background rate of XENON100, is the mounting of the cryogenics system, based on a pulse tube refrigerator, outside the passive shield, far from the LXe target.
The cylindrical TPC is formed by 24 interlocking PTFE panels. PTFE reflects scintillation light with high efficiency for vacuum ultraviolet [@yamashita], and optically separates the 62 kg target volume from the surrounding LXe, which is in average 4 cm thick and has a total mass of 99 kg. This allows to exploit the self-shielding capability of LXe due to its high density and high atomic number. In addition, this LXe volume around the target is instrumented with PMTs, becoming an active veto for background reduction by rejecting events in which a particle deposits part of its energy in the veto volume.
The cryostat is supported inside the shield by the 316Ti SS bars, which are mounted onto the movable shield door. The thickness of the inner and outer cryostat walls is 1.5 mm and the total weight of the vessel is 70.0 kg, which is only 30% of that of the XENON10 detector’s cryostat [@xe10-instrument]. The inner vessel containing the LXe is lined on the walls and the bottom with a 1.5 mm thick PTFE layer in order to increase the light collection efficiency in the active veto volume.
Electrons created by ionization in the LXe target are drifted upwards by a strong electric field created by applying voltage on the cathode, installed on the bottom of the TPC. In order to shield the bottom PMTs from this electric field, an additional grounded electrode is installed below the cathode.
The gas phase for charge amplification is maintained using a ’diving bell’ system, made from 316Ti SS with a total weight 3.6 kg. It allows to keep the liquid level constant at a precise height while having an additional layer of LXe above the TPC. A slight overpressure in the bell is provided by the gas returning from the continuous recirculation system, and the height of the gas outlet from the bell can be changed by a motion feedthrough in order to adjust the liquid level.
An extraction field is created across the liquid-gas interface by applying high voltage on the anode, which is placed inside the ’diving bell’. Two additional electrodes are installed below and above the anode and are kept at ground potential in order to close the field cage and shield the top PMT array from the high electric field. Ionization electrons are extracted into the gas phase and accelerated, producing the proportional scintillation (S2) signal [@S2]. The gaps between the top electrodes are 5 mm, and the liquid level is adjusted between the lower two of them. In the background model, only the 316Ti SS support rings for the electrodes are considered, given that the meshes are $\sim$100 $\mu$m thick and have a very low mass, leading to a negligible background from their radioactivity.
The prompt (S1) and the proportional scintillation light (S2) is detected by 242 1"-square R8520-06-AL Hamamatsu PMTs. They are among the lowest radioactivity PMTs, and are optimized to operate in LXe. The top PMT array consists of 98 PMTs, mounted in a concentric pattern in a PTFE support structure inside the ’diving bell’. The bottom PMT array consists of 80 PMTs, mounted below the cathode in the LXe, and arranged in a rectangular pattern in order to maximize the photocathode coverage. Additionally, 64 PMTs view LXe of the veto volume: 16 PMTs above and below the TPC, and 32 observing the sides. The components of the PMT are shown in Table \[tab:PMTmassModel\]. In the GEANT4 model, a PMT is simplified with a stainless steel case and a synthetic silica window inside a thin aluminum ring. A PMT is supplied with high voltage through the voltage divider circuit mounted on a base made from $Cirlex$. The base is approximated in the model as a homogenous unit.
The XENON100 data acquisition system (DAQ) digitizes the full waveform of the 242 PMTs at 100 MHz, where the time window for an event is 400 $\mu$s, more than twice the maximum electron drift time. If a particle has deposited energy at multiple places in the target, then two or more S2 pulses are recorded in the trace. Such an event is a multiple scatter event and is rejected in the analysis since the predicted behavior of the WIMP, due to its very low scattering cross-section, is a single scatter event.
For the calculation of the final background rate in the Monte Carlo simulations, multiple scatter events are rejected taking into account the finite position resolution of the detector. A multiple scatter event is considered as a single scatter event if the interactions happen less than 3 mm apart in $Z$. This position resolution is given by the width of the S2 signals and the peak separation efficiency of the S2 peak finder algorithm.
Background due to radioactive contamination in the detector and shield materials {#secDetectorMaterials}
================================================================================
Special care has been taken to select detector and shield materials according to their radioactive contamination. Before detector construction, the majority of materials planned to be used were screened with low background Ge detectors in order to determine their intrinsic radioactivity, mostly due to residual $^{238}$U, $^{232}$Th, $^{40}$K, and $^{60}$Co contamination. XENON has access to a dedicated screening facility underground at LNGS, the Gator detector [@gator]. Moreover, the LNGS screening facility, with some of the most sensitive Ge detectors in the world [@LNGSfacility], has also been used.
The radioactive contamination of the materials used for detector and shield construction is shown in Table \[tab:ScreeningResults\]. The results of the screening [@screening] do not exclude the possibility of the radioactive disequilibrium in $^{238}$U chain in some materials. However, isotopes before $^{226}$Ra in the decay chain have a very low intensity of $\gamma$-rays, and their contribution to the electronic recoil background is not significant. Thus, secular equilibrium has been assumed for the present study.
Decays of the radioactive isotopes in the materials listed in Table \[tab:ScreeningResults\] have been simulated with GEANT4, and the corresponding background rates have been calculated. The measured activities have been used as an input information for the Monte Carlo simulations and background predictions. For the analysis presented here, the upper limits are treated as detection values.
Figure \[figSpectraFS\] shows the predicted spectra in the entire energy range, and Figure \[figSpectraWS\] in the region of interest. The energy range for the background rate calculation is chosen to be sufficiently wide, up to 100 keV, to include the signal region for inelastic dark matter which is predicted to be in a higher energy range than the one from standard elastic WIMP scattering [@inelastic]. The effect of the discrimination between multiple and single scatter events on the background rate can be seen in Figure \[figSpectraFS\]: the multiple scatter behavior of incident gamma rays is typical for higher energies, whereas single scatter events dominate in the low energy region, and the multiple scatter cut does not yield a significant reduction of the background rate. Further background reduction can be achieved with fiducial volume cuts.
![(Color online) Predicted background from the detector and shield materials: energy spectra of all events (thin dashed line) and single scatters (solid line) in the entire 62 kg LXe target, and single scatters in the 40 kg (thick dashed line) and 30 kg fiducial volumes (dotted line), with infinite energy resolution.[]{data-label="figSpectraFS"}](Figure2.pdf){height="0.73\linewidth"}
![Zoom into the low energy region of the Monte Carlo spectra shown in Figure \[figSpectraFS\]. The spectra of all scatters and single scatter events in the entire target volume overlap.[]{data-label="figSpectraWS"}](Figure3.pdf){height="0.73\linewidth"}
In Figure \[figSpectraWS\] several characteristic X-rays can be seen. The xenon K-shell fluorescence peaks appear at 30 keV and 34 keV. The X-ray peaks at 15, 75, 85 and 90 keV are from Pb and Bi close to the target volume, for example in PTFE walls. In addition, there is a 46 keV gamma line from $^{210}$Pb decay, and 63 keV gamma line from the decay of $^{234}$Th. Due to their short mean free path, these low energy lines can be observed only at the edge of the LXe volume. After applying a cut on the position of the interactions, the peaks disappear and the spectrum becomes relatively flat in the low energy region. The background rate is thus presented as the average below 100 keV.
The spatial distribution of the single scatter electronic recoil events in the region of interest is presented in Figure \[figPositionDistribution\]. The radial cut rejects events at the edge of the target volume, originating mostly from radioactive decays in the PTFE of the TPC and the 316Ti SS of the cryostat vessels. The background from the PMTs, PMT bases, ’diving bell’, and the electrodes can be efficiently reduced by rejecting events within the top and bottom layers of LXe.
![(Color online) Predicted electronic recoil background from detector and shield materials, excluding intrinsic radioactivity in the LXe. Shown are single scatter events with energy below 100 keV in the TPC, without veto cut. $Z$ = 0 cm corresponds to the liquid-gas interface. The cathode mesh is located at $Z$ = $-$304.5 mm. The dashed line shows the 40 kg fiducial volume, and the solid line illustrates the 30 kg fiducial volume optimized to minimize the background.[]{data-label="figPositionDistribution"}](Figure4.pdf){height="0.73\linewidth"}
--------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- ---------------------- ----------------------
Volume
Veto cut
Cryostat and ’diving bell’ (316Ti SS) 21.00 6.77 2.63 0.65 1.78 0.47
Support bars (316Ti SS) 1.05 0.24 0.19 0.05 0.12 0.04
Detector PTFE 3.47 2.89 0.05 1.5$\times$10$^{-2}$ 3.4$\times$10$^{-2}$ 0.01
Detector copper 0.31 0.13 0.02 4.7$\times$10$^{-3}$ 1.2$\times$10$^{-2}$ 2.6$\times$10$^{-3}$
PMTs 89.13 51.97 7.86 2.16 3.98 1.13
PMT bases 15.95 10.26 0.86 0.22 0.40 0.12
TPC resistor chain 1.7$\times$10$^{-4}$ 1.2$\times$10$^{-4}$ 2.7$\times$10$^{-6}$ 7.1$\times$10$^{-7}$ 2.1$\times$10$^{-6}$ 5.7$\times$10$^{-7}$
Bottom electrodes (316Ti SS) 0.93 0.46 0.04 6.4$\times$10$^{-3}$ 0.02 4.1$\times$10$^{-3}$
Top electrodes (316Ti SS) 1.02 0.55 0.03 7.0$\times$10$^{-3}$ 0.01 4.6$\times$10$^{-3}$
PMT cables 0.56 0.08 0.10 0.02 0.07 0.02
Copper shield 0.64 0.22 0.10 0.04 0.08 0.02
Polyethylene shield 0.33 0.09 0.05 0.01 0.03 0.01
Total 134.39 73.66 11.93 3.18 6.54 1.83
--------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- ---------------------- ----------------------
The effect of the active LXe veto is presented in Figure \[figVetoCuts\], showing the total rate as a function of the energy threshold in the veto volume. The measured efficiency of the veto coincidence cut has been implemented in the Monte Carlo simulations. The average energy threshold measured with a collimated $^{137}$Cs source is about 100 keV. This allows to reduce the background rate in the entire target volume by $\sim$50%. Background reduction is even more efficient if the veto cut is combined with a fiducial volume cut, which results in a $>$90% reduction of the background rate. The reduction of the background rate remains almost constant when the energy threshold in the veto is below 100 keV. This is explained by an anti-correlation of the energy deposition in the active veto and target volume: events that deposit a small amount of energy in the target volume are likely to have deposited a larger amount of energy in the veto volume.
![(Color online) Predicted background rate from the detector and shield materials in the energy range below 100 keV, as a function of the energy threshold in the active veto. The average energy threshold in the veto measured with a collimated $^{137}$Cs source is about 100 keV.[]{data-label="figVetoCuts"}](Figure5.pdf){height="0.73\linewidth"}
![(Color online) The predicted background from the detector and shield materials (thick black line) in the 30 kg fiducial mass without veto cut, together with the individual contributions from the PMTs (solid line), the cryostat with pipes and ’diving bell’ (dash dotted line), PMT bases (long dashed line). The short dashed line shows the summed background from all other components: detector PTFE and copper, cryostat support bars, TPC resistor chain, top and bottom electrodes, PMT cables, and copper and polyethylene shield.[]{data-label="figSpectraDetectorMaterials"}](Figure6.pdf){height="0.73\linewidth"}
Table \[tab:gamma-rates\] presents the energy averaged rates of single scatter electronic recoil from detector materials in the region of interest, below 100 keV. The background rate has been predicted for the entire 62 kg LXe target, and for two fiducial volumes: a simple 40 kg cylindrical fiducial volume used in the analysis of the first XENON100 data [@xe100-independent], and a 30 kg fiducial volume cut optimized to minimize the background.
The low energy Monte Carlo spectrum of the background from the detector and shield materials is shown together with the individual contributions in Figure \[figSpectraDetectorMaterials\], for the 30 kg fiducial mass without veto cut. The background rate is dominated by the PMTs ($\sim$65% of the total background from all detector and shield materials), and the 316Ti SS cryostat, pipes and ’diving bell’ (other $\sim$25%). The dominant contribution to the background from the PMTs is originating from the $^{60}$Co and $^{40}$K contamination (50% and 34%, respectively). The main contaminant in the 316Ti SS is $^{60}$Co, which is responsible for almost 70% of the total background from this material. Components as detector PTFE and copper, cryostat support bars, TPC resistor chain, top and bottom electrodes, PMT cables, copper and polyethylene shield contribute $<$10% to the total background rate from the detector and shield materials.
Background from radon in the shield cavity {#secRadonCavity}
==========================================
A potentially dangerous background for XENON100 is the gamma background from the decay of ${^{222}}$Rn daughters inside the shield cavity. The average measured radon activity in the LNGS tunnel at the location of the experiment is $\sim$350 Bq/m$^3$ (Figure \[figRadonA\], top). Therefore, the shield cavity with a total volume 0.58 m$^3$ is constantly flushed with nitrogen gas when the shield door is closed. Nevertheless, a certain amount of radon can still be present. During the science runs, a low and constant $^{222}$Rn concentration is kept inside the shield. It is continuously monitored as shown in Figure \[figRadonA\], bottom. The measured values are at the limit of the sensitivity of the radon monitor. No correlation can be seen between the radon concentration inside and outside the shield.
![(Color online) Predicted rate of single electronic recoils with energy below 100 keV as a function of $^{222}$Rn concentration in the shield cavity for different fiducial masses. As a reference value, the horizontal dashed line corresponds to a background rate of 10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$.[]{data-label="figRadonB"}](Figure8.pdf){height="0.73\linewidth"}
Figure \[figRadonB\] shows the predicted background rate due to $^{222}$Rn as a function of its concentration inside the shield. Without veto cut, the background rate from 1 Bq/m$^{3}$ of $^{222}$Rn in the shield is 6$\times$10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ for the entire target mass of 62 kg, 9$\times$10$^{-4}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ in the 40 kg fiducial volume, and 2$\times$10$^{-4}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ in the 30 kg fiducial volume. For the 30 kg fiducial mass, this is less than 2% of the background from the detector and shield materials. Moreover, the measured radon concentration is well below 1 Bq/m$^{3}$.
Background due to intrinsic krypton and radon radioactivity {#secIntrinsicBG}
===========================================================
There is no long-lived radioactive xenon isotope, with the exception of the potential double beta emitter $^{136}$Xe, with the half-life limits of $>$7$\times$10$^{23}$ years and $>$1.1$\times$10$^{22}$ years for the neutrinoless and 2$\nu$ double beta decay, respectively [@DoubleBetaLimit].
Commercially available xenon gas, where purification is performed by distillation and adsorption-based chromatography, has a concentration of natural krypton at the ppm level. Natural krypton contains about 10$^{-11}$ g/g of radioactive $^{85}$Kr. The background from the beta decay of $^{85}$Kr, with T$_{1/2}$ = 10.76 years and endpoint energy 687 keV, is a potential limitation in the sensitivity of rare-event searches using xenon targets.
The gas used in the XENON100 experiment has been processed at a commercial distillation plant to reduce the concentration of krypton to $<$10 ppb. The high-temperature getter used in the experiment to purify xenon from water and electronegative contaminants does not remove the noble gas krypton. Therefore, an additional gas purification has been performed by cryogenic distillation. The reduction of krypton concentration down to a few ppt has been reported in Ref. [@distillation], with a distillation column very similar to the one procured by XENON100.
![Rate of single electronic recoils from $^{85}$Kr decay in the energy region below 100 keV as a function of the concentration of natural Kr in the LXe. Fiducial and veto cuts are inefficient for this intrinsic background source. As a reference value, the horizontal dashed line corresponds to a background rate of 10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$. []{data-label="figLXeKr"}](Figure9.pdf){height="0.73\linewidth"}
Levels of radioactive trace contaminations in xenon might vary at different stages of the experiment, as they strongly depend on purification processes. The background rate from intrinsic radioactivity in LXe has thus been predicted for different concentrations of $^{nat}$Kr in LXe, and is shown in Figure \[figLXeKr\].
Another intrinsic source of background is the decay of $^{222}$Rn daughters in the LXe. Radon is present in the LXe due to emanation from detector materials and the getter, and diffusion of the gas through the seals.
![(Color online) Predicted background rate below 100 keV as a function of $^{222}$Rn concentration in the LXe. As a reference value, the horizontal dashed line corresponds to a background rate of 10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$. []{data-label="figLXeRn"}](Figure10.pdf){height="0.73\linewidth"}
{width="0.95\linewidth"}
In the Monte Carlo simulation, $^{222}$Rn decays are generated uniformly in the LXe. Only the part of the chain before $^{210}$Pb is considered, since the relatively long half-life time of 22.3 years for $^{210}$Pb results in radioactive disequilibrium in the decay chain. The predicted background rate in the energy region below 100 keV is shown in Figure \[figLXeRn\] as a function of the $^{222}$Rn concentration in the LXe.
A background contribution from each intrinsic radioactive source of less than 10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$, which is used as a reference value, translates into a concentration of $^{nat}$Kr below 50 ppt, and an activity of $^{222}$Rn in LXe of $<$20 $\mu$Bq/kg in the entire target mass of 62 kg. The background from $^{222}$Rn daughters in the LXe can be reduced by a fiducial volume cut, removing decays at the edge of the target volume which are likely to produce high energy gamma rays with a longer mean free path which escape the target volume. For the 40 kg and 30 kg fiducial volumes, a background level of 10$^{-3}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ corresponds to 35 $\mu$Bq/kg.
Comparison of the predictions with the measured data {#secComparison}
====================================================
During the commissioning run in Fall 2009 [@xe100-independent], the level of krypton in the LXe has been measured with a delayed coincidence analysis using a decay channel where $^{85}$Kr undergoes a beta-decay with E$_{max}$ = 173.4 keV to $^{85m}$Rb ($\tau$ = 1.46 $\mu$s), which in turn decays to the ground state emitting a gamma-ray with an energy of 514 keV. The concentration of $^{nat}$Kr in the LXe determined with this technique is 143$_{-90}^{+130}$ ppt \[mol/mol\], assuming a $^{85}$Kr abundance of 10$^{-11}$.
The $^{222}$Rn level in the LXe has been determined using a beta-alpha time coincidence analysis, where events corresponding to the decays of $^{214}$Bi (T$_{1/2}$ = 19.7 min, E$_{max}$ = 3.27 MeV) and $^{214}$Po (T$_{1/2}$ = 164 $\mu$s, E$_{\alpha}$ = 7.69 MeV) are tagged. Based on this analysis, the upper limit on the $^{222}$Rn activity in LXe is $<$21 $\mu$Bq/kg.
The volumetric activity of $^{222}$Rn inside the shield cavity has been continuously monitored during this commissioning run and was always below 1 Bq/m$^{3}$.
------------------------------------------ -------- ------- ------- ------ ------ ------
Volume
Veto cut
Detector and shield materials 134.39 73.66 11.93 3.18 6.54 1.83
$^{222}$Rn in the shield (1 Bq/m$^{3}$) 5.95 1.72 0.92 0.16 0.16 0.02
$^{85}$Kr in LXe (120 ppt of $^{nat}$Kr) 2.35 2.35 2.35 2.35 2.35 2.35
$^{222}$Rn in LXe (21 $\mu$Bq/kg) 1.04 0.51 0.56 0.38 0.53 0.37
All sources 143.73 78.24 15.76 6.07 9.58 4.57
------------------------------------------ -------- ------- ------- ------ ------ ------
A comparison of the measured background spectrum and the Monte Carlo simulation for the 30 kg fiducial volume without veto cut is shown in Figure \[figDataMC\]. The energy region below 100 keV is shown separately in Figure \[figDataMCzoom\]. For optimal energy resolution and improved linearity, the energy scale of the measured spectrum exploits the anti-correlation between the light and the charge [@CombinedEnergy]: S1 and S2 are combined according to E = S1/4.4 + S2/132.6 \[keV\]. The simulated spectrum is smeared with a Gaussian function using the energy resolution measured with calibration sources: $\frac{\sigma(E)}{E}$ = 0.009 + 0.485/$\sqrt{E~\textnormal{[keV]}}$. The contribution from the detector and shield materials is scaled based on the screening values shown in Table \[tab:ScreeningResults\]. The upper limits from materials screening are used as fixed values for the scaling. For the level of $^{222}$Rn in the shield cavity we used the value measured with a dedicated radon monitor, and for the $^{222}$Rn level in the LXe - the value determined with the delayed coincidence analysis. The level of krypton has been inferred from the best fit of the simulated to the measured spectrum, and is in agreement with the value obtained from the delayed coincidence method.
![(Color online) Zoom into the low energy region of Figure \[figDataMC\]: energy spectra of the measured background and Monte Carlo simulations in the 30 kg fiducial volume without veto cut. The 2$\nu~\beta\beta$ decay of $^{136}$Xe has negligible contribution to the background below 100 keV.[]{data-label="figDataMCzoom"}](Figure12.pdf){height="0.73\linewidth"}
Very good agreement of the background model with the data is achieved for the low energy region, below 700 keV, and for the main peaks: $^{214}$Pb (352 keV), $^{208}$Tl (583 keV), $^{137}$Cs (662 keV), $^{60}$Co (1173 and 1332 keV), and $^{40}$K (1460 keV). In particular, simulated and measured background spectra agree well in the energy region of interest, below 100 keV (Figure \[figDataMCzoom\]). The predicted rates of single scatter electronic recoil events in the energy region of interest are presented in Table \[tab:summaryElectronRecoils\]. In the 30 kg fiducial volume, $^{85}$Kr contributes $\sim$30% to the total background without veto cut, and 55% when a veto coincidence cut with an average energy threshold of 100 keV is applied. The contribution from $^{222}$Rn in the LXe is $<$7%, from $^{222}$Rn in the shield cavity $<$2% of the total background rate in the energy region of interest.
The disagreement between simulated and measured spectra above $\sim$1.5 MeV is caused by non-linear effects in the PMT response, which results in a worse performance of the position reconstruction algorithms, changing the rate in the fiducial volumes and leading to a worsening of the position dependent signal corrections.
Using only the sources described in the previous sections, the background model shows a deficit in the 700-1100 keV range. Most of this deficit can be explained by cosmogenic activation of the stainless steel parts during materials storage and the detector construction at the ground level, in particular $^{54}$Mn isotope with the half-life time of 312 days. The activity assumed in the present study is 1.25 mBq/kg, and the decays have been generated uniformly in all parts made of 316Ti SS. This value is well below the conservative limit (2.14 mBq/kg), assuming the production rate of $^{54}$Mn in 316Ti SS at LNGS altitude of 6.5$\pm$0.7 mBq/kg [@heusser], and taking into account that in Fall 2009 the detector had been underground almost 500 days. Radioactive isotopes $^{58}$Co, $^{56}$Co, and $^{46}$Sc can be also produced by cosmic rays in the stainless steel, and emit high energy $\gamma$-rays. Their decays have been included in the background model assuming saturation activities at LNGS altitude [@heusser]. However, due to short half-life times of $\sim$80 days their contribution is negligible. The predicted background from the cosmogenic activation in the stainless steel is at the level of 10$^{-4}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$, thus about 5% of that from natural radioactivity in the same components.
The cosmogenic activation of natural xenon during storage at Earth’s surface might be responsible for the remaining discrepancy. This has been studied assuming 1 year of activation and 2 years of cooldown time with the ACTIVIA [@ACTIVIA] and COSMO [@COSMO] simulation packages. Both use semi-empirical formulae [@SilberbergTsao] to estimate the cross-sections of nuclear processes. The production rates predicted by the packages differ by one order of magnitude and more, and the remaining discrepancy around 1100 keV seen in Figure \[figDataMC\] cannot be explained without destroying the remarkable agreement in other energy ranges. A similar calculation has been performed for a natural xenon target in Ref. [@CosmogenicProduction] using the TALYS code [@TALYS], and the published results do not agree with either ACTIVIA or COSMO in terms of isotopes produced by cosmogenic activation and their production yields.
The theoretical spectrum of the 2$\nu$ double beta decay of $^{136}$Xe is also shown in Figure \[figDataMC\], assuming the half-life limit of 1.1$\times$10$^{22}$ years [@DoubleBetaLimit]. Its contribution does not change the total background spectrum significantly, thus it can be concluded that the small remaining discrepancy between measured and simulated spectra cannot be explained by this potential background source. The predicted energy averaged background rate from the 2$\nu~\beta\beta$ decay of $^{136}$Xe is at the level of 10$^{-6}$ events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ below 100 keV, three orders of magnitude lower than the background from other components.\
\
\
Conclusions {#secConclusions}
===========
An extensive study to predict the electronic recoil background of the XENON100 experiment has been performed . The study is based on Monte Carlo simulations with GEANT4 using a detailed mass model of the detector and its shield, and the measured radioactivity values of all relevant detector components.
The design goal of XENON100, to gain a factor of 100 reduction in background rate compared to XENON10 (0.6 events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ [@xe10-independent]), has been achieved. This has been possible thanks to a selection of all detector materials, an innovative design of the cryogenic system, the use of an active LXe veto, and an improved passive shield.
The predicted rate of single scatter electronic recoil events in the energy region below 100 keV, without veto coincidence cut is 15.8$\times$10$^{-3}$ (9.6$\times$10$^{-3}$) events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ for 40 kg (30 kg) fiducial mass (Table \[tab:summaryElectronRecoils\]). By applying a veto cut with an average energy threshold of 100 keV, these rates are reduced to 6.1$\times$10$^{-3}$ (4.6$\times$10$^{-3}$) events$\cdot$kg$^{-1}\cdot$day$^{-1}\cdot$keV$^{-1}$ for 40 kg (30 kg) fiducial mass. The discrimination between electron and nuclear recoils based on the ratio of proportional to primary scintillation light (S2/S1) is not considered in this paper, and provides a further background reduction of $>$99%.
From the good agreement between Monte Carlo simulation and measured data, as shown in Figures \[figDataMC\] and \[figDataMCzoom\], and the predicted background rates from Table \[tab:summaryElectronRecoils\], it can be concluded that the electronic recoil background in the XENON100 experiment during the commissioning run in Fall 2009 [@xe100-independent] is dominated by the natural radioactivity in the detector materials. With an optimized fiducial volume cut and an active veto cut, the background rate in the energy region of interest can be reduced down to a level where radioactive $^{85}$Kr in LXe starts to dominate.
The results of the present work are not only important for understanding the electromagnetic background in the XENON100 experiment and the validation of the background model, but can be also useful for the design of next-generation detectors for dark matter searches, such as XENON1t or DARWIN [@darwin].
Acknowledgements
================
This work has been supported by the National Science Foundation Grants No. PHY-03-02646 and PHY-04-00596, the Department of Energy under Contract No. DE-FG02-91ER40688, the CAREER Grant No. PHY-0542066, the Swiss National Foundation SNF Grant No. 20-118119, the Volkswagen Foundation, and the FCT Grant No. PTDC/FIS/100474/2008.
[21]{}
J.D. Lewin and P.F. Smith, Astropart. Phys. [**6**]{} 87 (1996);\
G. Jungman, M. Kamionkowski, and K. Griest, Phys. Rept. [**267**]{} 195 (2006). E. Aprile [*et al.*]{} (XENON10 Collaboration), arXiv:1001.2834v1 (2010). J. Angle [*et al.*]{} (XENON10 Collaboration), Phys. Rev. Lett. [**100**]{} 021303 (2008). J. Angle [*et al.*]{} (XENON10 Collaboration), Phys. Rev. Lett. [**101**]{} 091301 (2008). E. Aprile [*et al.*]{} (XENON100 Collaboration), Phys. Rev. Lett. [**105**]{} 131302 (2010). J. Sulkimo [*et al.*]{} GEANT4 Collaboration, Nucl. Instrum. Methods Phys. Res. Sect. A [**506**]{} 250 (2003). M. Yamashita [*et al.*]{}, Nucl. Instrum. Methods Phys. Res. Sect. A [**535**]{} 692 (2004). Hamamatsu Photonics, private communication. R.A. Muller [*et al.*]{}, Phys. Rev. Lett. [**27**]{} 532 (1971). L. Baudis [*et al.*]{}, Submitted to JINST, arXiv:1103.2125 (2011). C. Arpesella, Appl. Radiat. Isot. [**47**]{} 991 (1996). E. Aprile [*et al.*]{}, ’Material screening and selection for the XENON100 experiment’. Manuscript in preparation. D. Smith and N. Weiner, Phys. Rev. D [**64**]{} 043502 (2001);\
S. Chang, ,G.D. Kribs, D. Tucker-Smith, N. Weiner, Phys. Rev. D [**79**]{} 043513 (2009). R. Bernabei [*et al.*]{}, Nucl. Phys. B [**110**]{} (Proc. Suppl.) 88 (2002). K. Abe [*et al.*]{} (XMASS Collaboration) Astropart. Phys. [**31**]{} 290 (2009). E. Conti [*et al.*]{}, Phys. Rev. B [**68**]{} 054201 (2003);\
T. Doke, NIM [**B234**]{} 203 (2005);\
E. Aprile, K.L. Giboni, P. Majewski, K. Ni, M. Yamashita, Phys. Rev. B [**76**]{} 014115 (2007). M. Laubenstein and G. Heusser, Appl. Radiat. Isotopes [**67**]{} 5 (2009). J.J. Back and Y.A. Ramachers, Nucl. Instrum. Methods Phys. Res. Sect. A [**586**]{} 286 (2008). J. Martoff and P.D. Lewin, Comput. Phys. Commun. [**72**]{} 96 (1992). R. Silberberg and C.H. Tsao, Astrophys. J. Suppl. [**220**]{} 315 (1973);\
R. Silberberg and C.H. Tsao, Astrophys. J. [**501**]{} 911 (1998). D.M. Mei, Z.B. Yin, and S.R. Elliott, Astropart. Phys. [**31**]{} 417 (2006). A.J. Koning, S. Hilaire, and M. C. Duijvestijn, AIP Conf. Proc. [**769**]{} 1154 (2005). L. Baudis (DARWIN Consortium), PoS (IDM2010) 122 (2010).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Convergence and stability results for the inverse Born series \[Moskow and Schotland, Inverse Problems, 24:065005, 2008\] are generalized to mappings between Banach spaces. We show that by restarting the inverse Born series one obtains a class of iterative methods containing the Gauss-Newton and Chebyshev-Halley methods. We use the generalized inverse Born series results to show convergence of the inverse Born series for the Schrödinger problem with discrete internal measurements. In this problem, the Schrödinger potential is to be recovered from a few measurements of solutions to the Schrödinger equation resulting from a few different source terms. An application of this method to a problem related to transient hydraulic tomography is given, where the source terms model injection and measurement wells.'
address: 'Mathematics Department, University of Utah, 155 S 1400 E RM 233, Salt Lake City, UT 84112-0090, USA'
author:
- Patrick Bardsley and Fernando Guevara Vasquez
bibliography:
- 'itborn.bib'
title: |
Restarted inverse Born series for the\
Schrödinger problem with\
discrete internal measurements
---
Discussion {#sec:discussion}
==========
We show here that with little modification, the inverse Born series convergence results of @Moskow:2008:CSI can be generalized to mappings between Banach spaces. With this abstraction, we only need to show that the forward Born operators are bounded as in to obtain convergence, stability and error estimates for the inverse Born series. Such results are then proven for the problem of finding the Schrödinger potential from discrete internal measurements. A nice byproduct of our approach is that we can relate forward and inverse Born series coefficients (up to a symmetrization) to the Taylor series coefficients of an analytic map and its inverse (provided it exists).
Since the cost of computing the $n-$th term of the inverse Born series increases exponentially in $n$, we also consider the iterative method obtained by restarting the inverse Born series after summing the first $k$ terms. We obtain a class of methods that we call RIBS($k$) and that includes the well-known Gauss-Newton and Chebyshev-Halley iterative methods. Our numerical results show these methods give reconstructions comparable to those obtained with the inverse Born series.
Among the future directions of this work would be to show the RIBS($k$) method is convergent. We conjecture that the convergence rate of RIBS($k$) is of order $k$. The RIBS($k$) method is only locally convergent, meaning that we need to be already close to the solution for the method to converge. Globalization strategies that keep, when possible, this higher order convergence rate are needed.
The application we use to illustrate our method is a problem related to transient hydraulic tomography. Since we convert this problem to the problem of finding a Schrödinger potential and all the methods we use here are locally convergent, the contrasts that we can deal with are far from realistic ones. We believe that a proper globalization strategy will allow us to deal with higher contrasts. Another important question that we have not dealt with here is that of regularization. The only regularization that we consider here is the choice of the linear operator that primes the inverse Born series. By analogy with what can be done with the Gauss-Newton method, we believe it is possible to include specific a priori information about the true parameters by formulating the problem as minimizing the misfit plus a penalty term that takes into account the a priori information.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank Liliana Borcea, Alexander V. Mamonov, Shari Moskow and John Schotland for insightful conversations on this subject. FGV is grateful to Otmar Scherzer for pointing out reference [@Hettlich:2000:SMIP]. The work of the authors was partially supported by the National Science Foundation grant DMS-0934664.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Majorization-minimization algorithms consist of successively minimizing a sequence of upper bounds of the objective function. These upper bounds are tight at the current estimate, and each iteration monotonically drives the objective function downhill. Such a simple principle is widely applicable and has been very popular in various scientific fields, especially in signal processing and statistics. We propose an incremental majorization-minimization scheme for minimizing a large sum of continuous functions, a problem of utmost importance in machine learning. We present convergence guarantees for non-convex and convex optimization when the upper bounds approximate the objective up to a smooth error; we call such upper bounds “first-order surrogate functions”. More precisely, we study asymptotic stationary point guarantees for non-convex problems, and for convex ones, we provide convergence rates for the expected objective function value. We apply our scheme to composite optimization and obtain a new incremental proximal gradient algorithm with linear convergence rate for strongly convex functions. Our experiments show that our method is competitive with the state of the art for solving machine learning problems such as logistic regression when the number of training samples is large enough, and we demonstrate its usefulness for sparse estimation with non-convex penalties.'
author:
- 'Julien Mairal[^1]'
title: 'Incremental Majorization-Minimization Optimization with Application to Large-Scale Machine Learning[^2]'
---
non-convex optimization, convex optimization, majorization-minimization.
90C06, 90C26, 90C25
Introduction {#sec:intro}
============
The principle of successively minimizing upper bounds of the objective function is often called *majorization-minimization* [@lange2] or *successive upper-bound minimization* [@razaviyayn]. Each upper bound is locally tight at the current estimate, and each minimization step decreases the value of the objective function. Even though this principle does not provide any theoretical guarantee about the quality of the returned solution, it has been very popular and widely used because of its simplicity. Various existing approaches can indeed be interpreted from the majorization-minimization point of view. This is the case of many gradient-based or proximal methods [@beck; @combette; @hale; @nesterov; @wright], expectation-maximization (EM) algorithms in statistics [@dempster; @neal], difference-of-convex (DC) programming [@horst], boosting [@collins; @pietra], some variational Bayes techniques used in machine learning [@wainwright2], and the mean-shift algorithm for finding modes of a distribution [@tomasi]. Majorizing surrogates have also been used successfully in the signal processing literature about sparse estimation [@candes4; @daubechies; @gasso], linear inverse problems in image processing [@ahn; @erdogan2], and matrix factorization [@lee2; @mairal7].
In this paper, we are interested in making the majorization-minimization principle scalable for minimizing a large sum of functions: $$\min_{\theta \in \Theta} \left[ f(\theta) \defin \frac{1}{T}\sum_{t=1}^T
f^t(\theta)\right], \label{eq:prob}$$ where the functions $f^t: \Real^p \to \Real$ are continuous, and $\Theta$ is a convex subset of $\Real^p$. When $f$ is non-convex, exactly solving (\[eq:prob\]) is intractable in general, and when $f$ is also non-smooth, finding a stationary point of (\[eq:prob\]) can be difficult. The problem above when $T$ is large can be motivated by machine learning applications, where $\theta$ represents some model parameters and each function $f^t$ measures the adequacy of the parameters $\theta$ to an observed data point indexed by $t$. In this context, minimizing $f$ amounts to finding parameters $\theta$ that explain well some observed data. In the last few years, stochastic optimization techniques have become very popular in machine learning for their empirical ability to deal with a large number $T$ of training points [@bottou2; @duchi2; @shalev2; @xiao]. Even though these methods have inherent sublinear convergence rates for convex and strongly convex problems [@lan; @nemirovski], they typically have a cheap computational cost per iteration, enabling them to efficiently find an approximate solution. Recently, incremental algorithms have also been proposed for minimizing finite sums of functions [@blatt; @defazio2; @defazio; @schmidt2; @shalev2]. At the price of a higher memory cost than stochastic algorithms, these incremental methods enjoy faster convergence rates, while also having a cheap per-iteration computational cost.
Our paper follows this approach: in order to exploit the particular structure of problem (\[eq:prob\]), we propose an incremental scheme whose cost per iteration is independent of $T$, as soon as the upper bounds of the objective are appropriately chosen. We call the resulting scheme “MISO” (*Minimization by Incremental Surrogate Optimization*). We present convergence results when the upper bounds are chosen among the class of “first-order surrogate functions”, which approximate the objective function up to a smooth error—that is, differentiable with a Lipschitz continuous gradient. For non-convex problems, we obtain almost sure convergence and asymptotic stationary point guarantees. In addition, when assuming the surrogates to be strongly convex, we provide convergence rates for the expected value of the objective function. Remarkably, the convergence rate of MISO is linear for minimizing strongly convex composite objective functions, a property shared with two other incremental algorithms for smooth and composite convex optimization: the *stochastic average gradient* method (SAG) of Schmidt, Le Roux and Bach @schmidt2, and the *stochastic dual coordinate ascent* method (SDCA) of Shalev-Schwartz and Zhang @shalev2. Our scheme MISO is inspired in part by these two works, but yields different update rules than SAG or SDCA, and is also appropriate for non-convex optimization problems.
In the experimental section of this paper, we show that MISO can be useful for solving large-scale machine learning problems, and that it matches cutting-edge solvers for large-scale logistic regression [@beck; @schmidt2]. Then, we show that our approach provides an effective incremental DC programming algorithm, which we apply to sparse estimation problems with nonconvex penalties [@candes4].
The paper is organized as follows: Section \[sec:generic\] introduces the majorization-minimization principle with first-order surrogate functions. Section \[sec:incremental\] is devoted to our incremental scheme MISO. Section \[sec:exp\] presents some numerical experiments, and Section \[sec:ccl\] concludes the paper. Some basic definitions are given in Appendix \[appendix:background\].
Majorization-minimization with first-order surrogate functions {#sec:generic}
==============================================================
In this section, we present the generic majorization-minimization scheme for minimizing a function $f$ without exploiting its structure—that is, without using the fact that $f$ is a sum of functions. We describe the procedure in Algorithm \[alg:generic\_batch\] and illustrate its principle in Figure \[fig:mm\]. At iteration $n$, the estimate $\theta_n$ is obtained by minimizing a surrogate function $g_n$ of $f$. When $g_n$ uniformly majorizes $f$ and when $g_n(\theta_{n-1})=f(\theta_{n-1})$, it is clear that the objective function value monotonically decreases.
$\theta_0 \in \Theta$ (initial estimate); $N$ (number of iterations). Compute a surrogate function $g_n$ of $f$ near $\theta_{n-1}$; Minimize the surrogate and update the solution: $\theta_n \in \argmin_{\theta \in \Theta} g_n(\theta).$
$\theta_{N}$ (final estimate);
(-5,-0.5)(5,4) (-5,0)(5,0) (0,-0.5)(0,3) (-4,0.9)(-4,2.5) (3.84,2.4)[ ${ f(\theta)}$]{} (-3.33,1.9)[ ${ g_n(\theta)}$]{} (-5.3,1.6)[ ${ h_n(\theta)}$]{} (0.62,1.40)(-1.3,0.3) (0.9,1.2) [$\theta_{n-1}$]{} (-1.4,0.6) [$\theta_{n}$]{} (2.1,1.0) [ $f(\theta_n) \leq f(\theta_{n-1})$]{}
For this approach to be effective, intuition tells us that we need functions $g_n$ that are easy to minimize and that approximate well the objective $f$. Therefore, we measure the quality of the approximation through the smoothness of the error $h_n \defin g_n-f$, which is a key quantity arising in the convergence analysis. Specifically, we require $h_n$ to be $L$-smooth for some constant $L>0$ in the following sense:
\[def:lsmooth\] A function $f: \Real^p \to \Real$ is called $L$-smooth when it is differentiable and when its gradient $\nabla f$ is $L$-Lipschitz continuous.
With this definition in hand, we now introduce the class of “first-order surrogate functions”, which will be shown to have good enough properties for analyzing the convergence of Algorithm \[alg:generic\_batch\] and other variants.
\[def:surrogate\_batch\] A function $g: \Real^p \to \Real$ is a first-order surrogate function of $f$ near $\kappa$ in $\Theta$ when
$g(\theta') \geq f(\theta')$ for all minimizers $\theta'$ of $g$ over $\Theta$. When the more general condition $g \geq f$ holds, we say that $g$ is a *majorizing* surrogate;
the approximation error $h\defin g-f$ is $L$-smooth, $h(\kappa)=0$, and $\nabla h(\kappa)=0$.
We denote by $\S_{L}(f,\kappa)$ the set of first-order surrogate functions and by $\S_{L,\rho}(f,\kappa) \subset \S_{L}(f,\kappa)$ the subset of $\rho$-strongly convex surrogates.
First-order surrogates are interesting because their approximation error—the difference between the surrogate and the objective—can be easily controlled. This is formally stated in the next lemma, which is a building block of our analysis:
\[lemma:basic\] Let $g$ be a surrogate function in $\S_{L}(f,\kappa)$ for some $\kappa$ in $\Theta$. Define the approximation error $h\defin g-f$, and let $\theta'$ be a minimizer of $g$ over $\Theta$. Then, for all $\theta$ in $\Theta$,
- $|h(\theta)| \leq \frac{L}{2}\|\theta-\kappa\|_2^2$;
- $f(\theta') \leq f(\theta) + \frac{L}{2}\|\theta-\kappa\|_2^2$.
Assume that $g$ is $\rho$-strongly convex, i,e., $g$ is in $\S_{L,\rho}(f,\kappa)$. Then, for all $\theta$ in $\Theta$,
- $f(\theta') + \frac{\rho}{2}\|\theta'-\theta\|_2^2 \leq f(\theta) + \frac{L}{2}\|\theta-\kappa\|_2^2$.
The first inequality is a direct application of a classical result (Lemma 1.2.3 of [@nesterov4]) on quadratic upper bounds for $L$-smooth functions, when noticing that $h(\kappa)=0$ and $\nabla h(\kappa)=0$. Then, for all $\theta$ in $\Theta$, we have $f(\theta') \leq g(\theta') \leq g(\theta) =
f(\theta)+h(\theta),$ and we obtain the second inequality from the first one.
When $g$ is $\rho$-strongly convex, we use the following classical lower bound (see [@nesterov]): $$g(\theta') +
\frac{\rho}{2}\|\theta-\theta'\|_2^2 \leq g(\theta).$$ Since $f(\theta') \leq g(\theta')$ by Definition \[def:surrogate\_batch\] and $g(\theta)=f(\theta)+h(\theta)$, the third inequality follows from the first one.
We now proceed with a convergence analysis including four main results regarding Algorithm \[alg:generic\_batch\] with first-order surrogate functions $g_n$. More precisely, we show in Section \[subsec:nonconvex\] that, under simple assumptions, the sequence of iterates asymptotically satisfies a stationary point condition. Then, we present a similar result with relaxed assumptions on the surrogates $g_n$ when $f$ is a composition of two functions, which occur in practical situations as shown in Section \[subsec:surrogates\]. Finally, we present non-asymptotic convergence rates when $f$ is convex in Section \[subsec:convex\]. By adapting convergence proofs of proximal gradient methods [@nesterov] to our more general setting, we recover classical sublinear rates $O(1/n)$ and linear convergence rates for strongly convex problems.
Non-convex convergence analysis {#subsec:nonconvex}
-------------------------------
For general non-convex problems, proving convergence to a global (or local) minimum is impossible in general, and classical analysis studies instead asymptotic stationary point conditions (see, [*e.g.*]{}, [@bertsekas]). To do so, we make the following mild assumption when $f$ is non-convex:
- $f$ is bounded below and for all $\theta,\theta'$ in $\Theta$, the directional derivative $\nabla f(\theta,\theta'-\theta)$ of $f$ at $\theta$ in the direction $\theta'-\theta$ exists.
The definitions of directional derivatives and stationary points are provided in Appendix \[appendix:background\]. A necessary first-order condition for $\theta$ to be a local minimum of $f$ is to have $\nabla
f(\theta,\theta'\!-\!\theta) \geq 0$ for all $\theta'$ in $\Theta$ (see, [*e.g.*]{}, [@borwein]). In other words, there is no feasible descent direction $\theta'\!-\!\theta$ and $\theta$ is a stationary point. Thus, we consider the following condition for assessing the quality of a sequence $(\theta_n)_{n \geq 0}$ for non-convex problems:
Under assumption , a sequence $(\theta_n)_{n \geq 0}$ satisfies the asymptotic stationary point condition if $$\liminf_{n \to +\infty} \inf_{\theta \in \Theta} \frac{\nabla
f(\theta_{n},\theta-\theta_n)}{ \|\theta-\theta_n\|_2} \geq 0. \label{eq:stationary}$$ Note that if $f$ is differentiable on $\Real^p$ and $\Theta = \Real^p$, $\nabla f(\theta_{n},\theta-\theta_n) =\nabla
f(\theta_{n})^\top(\theta-\theta_n)$, and the condition (\[eq:stationary\]) implies that the sequence $(\nabla f(\theta_n))_{n \geq 0}$ converges to $0$.
As noted, we recover the classical definition of critical points for the smooth unconstrained case. We now give a first convergence result about Algorithm \[alg:generic\_batch\].
\[prop:conv1\] Assume that holds and that the surrogates $g_n$ from Algorithm \[alg:generic\_batch\] are in $\S_L(f,\theta_{n-1})$ and are either majorizing $f$ or strongly convex. Then, $\!(f(\theta_n))_{n \geq 0}$ monotonically decreases, and $(\theta_n)_{n\geq 0}$ satisfies the asymptotic stationary point condition.
The fact that $(f(\theta_n))_{n \geq 0}$ is non-increasing and convergent because bounded below is clear: for all $n\geq 1$, $f(\theta_n) \leq g_n(\theta_n) \leq g_n(\theta_{n-1}) = f(\theta_{n-1})$, where the first inequality and the last equality are obtained from Definition \[def:surrogate\_batch\]. The second inequality comes from the definition of $\theta_n$.
Let us now denote by $f^\star$ the limit of the sequence $(f(\theta_n))_{n \geq
1}$ and by $h_n\defin g_n-f$ the approximation error function at iteration $n$, which is $L$-smooth by Definition \[def:surrogate\_batch\] and such that $h_n(\theta_n) \geq 0$. Then, $h_n(\theta_n) = g_n(\theta_n) - f(\theta_n) \leq f(\theta_{n-1}) - f(\theta_n)$, and $$\sum_{n=1}^{\infty} h_n(\theta_n) \leq f(\theta_0) - f^\star.$$ Thus, the non-negative sequence $(h_n(\theta_n))_{n \geq 0}$ necessarily converges to zero. Then, we have two possibilities (according to the assumptions made in the proposition).
- If the functions $g_n$ are majorizing $f$, we define $\theta' = \theta_{n}-\frac{1}{L}\nabla h_n(\theta_{n})$, and we use the following classical inequality for $L$-smooth functions [@nesterov4]: $$h_n(\theta') \leq h_n(\theta_n) - \frac{1}{2L}\|\nabla h_n(\theta_n)\|_2^2.$$ Therefore, we may use the fact that $h_n(\theta') \geq 0$ because $g_n \geq f$, and $$\|\nabla h_n(\theta_n)\|_2^2 \leq 2L (h_n(\theta_n)-h_n(\theta')) \leq 2L h_n(\theta_n) \underset{n \to +\infty}{\longrightarrow} 0.$$
- If instead the functions $g_n$ are $\rho$-strongly convex, the last inequality of Lemma \[lemma:basic\] with $\kappa = \theta = \theta_{n-1}$ and $\theta' = \theta_n$ gives us $$\frac{\rho}{2}\|\theta_n-\theta_{n-1}\|_2^2 \leq f(\theta_{n-1})-f(\theta_n).$$ By summing over $n$, we obtain that $\|\theta_n-\theta_{n-1}\|_2^2$ converges to zero, and $$\|\nabla h_n(\theta_n)\|_2 = \|\nabla h_n(\theta_n)-\nabla h_n(\theta_{n-1})\|_2 \leq L\|\theta_n-\theta_{n-1}\|_2 \underset{n \to +\infty}{\longrightarrow} 0,$$ since $\nabla h_n(\theta_{n-1})=0$ according to Definition \[def:surrogate\_batch\].
We now consider the directional derivative of $f$ at $\theta_n$ and a direction $\theta-\theta_n$, where $n \geq 1$ and $\theta$ is in $\Theta$, $$\nabla f(\theta_{n},\theta-\theta_n) = \nabla g_n(\theta_{n},\theta-\theta_n) - \nabla h_n(\theta_n)^\top (\theta-\theta_n).$$ Note that $\theta_n$ minimizes $g_n$ on $\Theta$ and therefore $\nabla g_n(\theta_{n},\theta-\theta_n) \geq 0$. Therefore, $$\nabla f(\theta_{n},\theta-\theta_n) \geq - \|\nabla h_n(\theta_n)\|_2 \|\theta-\theta_n\|_2,$$ by Cauchy-Schwarz’s inequality. By minimizing over $\theta$ and taking the infimum limit, we finally obtain $$\liminf_{n \to +\infty} \inf_{\theta \in \Theta} \frac{\nabla f(\theta_{n},\theta-\theta_n)}{ \|\theta-\theta_n\|_2} \geq - \lim_{n\to+\infty} \|\nabla h_n(\theta_n)\|_2 = 0.$$
This proposition provides convergence guarantees for a large class of existing algorithms, including cases where $f$ is non-smooth. In the next proposition, we relax some of the assumptions for objective functions that are compositions $f=f' \circ e$, where $\circ$ is the composition operator. In other words, $f(\theta) =
f'(e(\theta))$ for all $\theta$ in $\Real^p$.
\[prop:conv1\_separable\] Assume that holds and that the function $f$ is a composition $f = f' \circ e$, where $e: \Real^p \to \Real^d$ is $C$-Lipschitz continuous for some constant $C>0$, and $f': \Real^d \to \Real$. Assume that the function $g_n$ in Algorithm \[alg:generic\_batch\] is defined as $g_{n}\defin
g_{n}' \circ e$, where $g_{n}'$ is a majorizing surrogate in $\S_L(f',e(\theta_{n-1}))$. Then, the conclusions of Proposition \[prop:conv1\] hold.
We follow the same steps as the proof of Proposition \[prop:conv1\]. First, it is easy to show that $(f(\theta_n))_{n \geq 0}$ monotonically decreases and that $h_n(\theta_n) \defin g_n(\theta_n)-f(\theta_n)$ converges to zero when $n$ grows to infinity. Note that since we have made the assumptions that $g_n = g_n' \circ e$ and that $f = f' \circ e$, the function $h_n\defin g_n - f$ can be written as $h_n = h_n' \circ e$, where $h_n'\defin g_n'- f'$ is $L$-smooth. Proceeding as in the proof of Proposition \[prop:conv1\], we can show that $\|\nabla
h_n'(e(\theta_n))\|_2$ converges to zero.
Let us now fix $n \geq 1$ and consider $\delta$ such that $\theta_n+\delta$ is in $\Theta$. We have $$h_n(\theta_n+\delta) = h_n'( e(\theta_n+\delta)) = h_n'( e(\theta_n) + \|\delta\|_2 \z),$$ where $\z$ is a vector whose $\ell_2$-norm is bounded by a universal constant $C > 0$ because the function $e$ is Lipschitz continuous. Since $h_n'$ is $L$-smooth, we also have $$h_n(\theta_n+\delta) = h_n'( e(\theta_n) + \|\delta\|_2 \z) = h_n( \theta_n) + \|\delta\|_2 \nabla h_n^{\prime}(e(\theta_n))^\top \z + O(\|\delta\|_2^2).$$ Plugging this simple relation with $\delta=t(\theta-\theta_n)$, for some $0 < t < 1$ and $\theta$ in $\Theta$, into the definition of the directional derivative $\nabla h_n(\theta_n, \theta-\theta_n)$, we obtain the relation $$|\nabla h_n(\theta_n, \theta-\theta_n)| \leq C\| \nabla h_n^{\prime}(e(\theta_n))\|_2{\|\theta-\theta_n\|_2},$$ and since $\nabla f(\theta_{n},\theta-\theta_n) = \nabla g_n(\theta_{n},\theta-\theta_n) - \nabla h_n(\theta_n,\theta-\theta_n)$, and $\nabla g_n(\theta_{n},\theta-\theta_n) \geq 0$, $$\liminf_{n \to +\infty} \inf_{\theta \in \Theta} \frac{\nabla f(\theta_{n},\theta-\theta_n)}{ \|\theta-\theta_n\|_2} \geq -C \lim_{n \to +\infty} \| \nabla h_n^{\prime}(e(\theta_n))\|_2 = 0.$$
In this proposition, $g_n$ is an upper bound of $f=f' \circ e$, where the part $e$ is Lipschitz continuous but $g_n-f$ is not $L$-smooth. This extension of Proposition \[prop:conv1\] is useful since it provides convergence results for classical approaches that will be described later in Section \[subsec:surrogates\]. Note that convergence results for non-convex problems are by nature weak, and our non-convex analysis does not provide any convergence rate. This is not the case when $f$ is convex, as shown in the next section.
Convex analysis {#subsec:convex}
---------------
The next proposition is based on a proof technique from Nesterov @nesterov, which was originally designed for the proximal gradient method. By adapting it, we obtain the same convergence rates as in [@nesterov].
\[prop:conv2\] Assume that $f$ is convex, bounded below, and that there exists a constant $R >0$ such that $$\|\theta-\theta^\star\|_2 \leq R ~~~\text{for all}~ \theta \in \Theta \st f(\theta) \leq f(\theta_0),\label{eq:bounded}$$ where $\theta^\star$ is a minimizer of $f$ on $\Theta$. When the functions $g_n$ in Algorithm \[alg:generic\_batch\] are in $\S_{L}(f,\theta_{n-1})$, we have for all $n \geq 1$, $$f(\theta_n)-f^\star \leq \frac{2LR^2}{n+2},$$ where $f^\star \defin f(\theta^\star)$. Assume now that $f$ is $\mu$-strongly convex. Regardless of condition (\[eq:bounded\]), we have for all $n \geq 1$, $$f(\theta_n) - f^\star \leq \beta^n( f(\theta_0)-f^\star),$$ where $\beta \defin \frac{L}{\mu}$ if $\mu >2L$ or $\beta \defin \left(1-\frac{\mu}{4L}\right)$ otherwise.
We successively prove the two parts of the proposition.\
[[*Non-strongly convex case:*]{} ]{} Let us consider the function $h_n \defin g_n - f$ at iteration $n \geq 1$. By Lemma \[lemma:basic\], $$f(\theta_n) \leq \min_{\theta \in \Theta} \left[ f(\theta) + \frac{L}{2} \|\theta-\theta_{n-1}\|_2^2\right].$$ Then, following a similar proof technique as Nesterov in [@nesterov], $$\begin{split}
f(\theta_n) & \leq \min_{\alpha \in [0,1]} \left[ f(\alpha\theta^\star+(1-\alpha)\theta_{n-1}) + \frac{L\alpha^2}{2} \|\theta^\star-\theta_{n-1}\|_2^2\right] \\
& \leq \min_{\alpha \in [0,1]} \left[ \alpha f(\theta^\star) +(1-\alpha)f(\theta_{n-1}) + \frac{L\alpha^2}{2} \|\theta^\star-\theta_{n-1}\|_2^2\right],
\end{split} \label{eq:tmp_rate2}$$ where the minimization over $\Theta$ is replaced by a minimization over the line segment $\alpha\theta^\star+(1-\alpha)\theta_{n-1} :
\alpha \in [0,1]$. Since the sequence $(f(\theta_n))_{n \geq 0}$ is monotonically decreasing we may use the bounded level set assumption and we obtain $$f(\theta_n) - f^\star \leq \min_{\alpha \in [0,1]} \left[ (1-\alpha)(f(\theta_{n-1}) - f^\star) + \frac{LR^2\alpha^2}{2}\right].\label{eq:tmp_rate3}$$ To simplify, we introduce the notation $r_n \defin f(\theta_n)- f^\star$, and we consider two cases:
- [*first case:*]{} if $r_{n-1} \geq LR^2$, then the optimal value $\alpha^\star$ in (\[eq:tmp\_rate3\]) is $1$ and we consequently have $r_n \leq \frac{LR^2}{2}$;
- [*second case:*]{} otherwise $\alpha^\star =\frac{r_{n-1}}{LR^2}$ and $r_n \leq r_{n-1}\left(1-\frac{r_{n-1}}{2LR^2}\right)$. Thus, $r_n^{-1} \geq r_{n-1}^{-1} \left(1-\frac{r_{n-1}}{2LR^2}\right)^{-1} \geq r_{n-1}^{-1} + \frac{1}{2LR^2}$, where the second inequality comes from the convexity inequality $(1-x)^{-1} \geq 1+x$ for $x \in (0,1)$.
We now apply recursively the previous inequalities, starting with $n=1$. If $r_0 \geq LR^2$, we are in the first case and then $r_1 \leq \frac{LR^2}{2}$; Then, we will subsequently be in the second case for all $n \geq 2$ and thus $r_n^{-1} \geq r_1^{-1} + \frac{n-1}{2LR^2} \geq \frac{n+3}{2LR^2}$. Otherwise, if $r_0 < LR^2$, we are always in the second case and $r_n^{-1} \geq r_0^{-1} + \frac{n}{2LR^2} \geq \frac{n+2}{2LR^2}$, which is sufficient to obtain the first part of the proposition.
[[*$\mu$-strongly convex case:*]{} ]{} Let us now assume that $f$ is $\mu$-strongly convex, and let us drop the bounded level sets assumption. The proof again follows @nesterov for computing the convergence rate of proximal gradient methods. We start from (\[eq:tmp\_rate2\]). We use the strong convexity of $f$ which implies that $f(\theta_{n-1}) \geq f^\star + \frac{\mu}{2}\|\theta_{n-1}-\theta^\star\|_2^2$, and we obtain $$f(\theta_n) - f^\star \leq \left(\min_{\alpha \in [0,1]} 1-\alpha + \frac{L\alpha^2}{\mu}\right)(f(\theta_{n-1})-f^\star).$$ At this point, it is easy to show that if $\mu \geq 2L$, the previous binomial is minimized for $\alpha^\star = 1$, and if $\mu \leq 2L$, then we have $\alpha^\star = \frac{\mu}{2L}$. This yields the desired result.
The result of Proposition \[prop:conv2\] is interesting because it does not make any strong assumption about the surrogate functions, except the ones from Definition \[def:surrogate\_batch\]. The next proposition shows that slightly better rates can be obtained with additional strong convexity assumptions.
\[prop:conva\] Assume that $f$ is convex, bounded below, and let $\theta^\star$ be a minimizer of $f$ on $\Theta$. When the surrogates $g_n$ of Algorithm \[alg:generic\_batch\] are in $\S_{L,\rho}(f,\theta_{n-1})$ with $\rho \geq L$, we have for all $n \geq 1$, $$f(\theta_n) - f^\star \leq \frac{L\|\theta_{0}-\theta^\star\|_2^2}{2n},\label{eq:rate1}$$ where $f^\star \defin f(\theta^\star)$. When $f$ is $\mu$-strongly convex, we have for all $n\geq 1$, $$f(\theta_n) - f^\star \leq \left(\frac{L}{\rho+\mu}\right)^{n-1}\frac{L\|\theta_0-\theta^\star\|_2^2}{2}.$$
As before, we successively prove the two parts of the proposition.
[[*Non-strongly convex case:*]{} ]{} From Lemma \[lemma:basic\] (with $g=g_n$, $\kappa=\theta_{n-1}$, $\theta'=\theta_n$, $\theta=\theta^\star$), we have for all $n\geq 1$, $$f(\theta_n) - f(\theta^\star) \leq \frac{L}{2}\|\theta_{n-1}-\theta^\star\|_2^2 - \frac{\rho}{2}\|\theta_n-\theta^\star\|_2^2\leq \frac{L}{2}\|\theta_{n-1}-\theta^\star\|_2^2 - \frac{L}{2}\|\theta_n-\theta^\star\|_2^2.\label{eq:rate1_proof}$$ After summation, $$n(f(\theta_n) - f(\theta^\star)) \leq \sum_{k=1}^n (f(\theta_k) - f(\theta^\star))\leq \frac{L}{2}(\|\theta_{0}-\theta^\star\|_2^2-\|\theta_{n}-\theta^\star\|_2^2) \leq \frac{L\|\theta_{0}-\theta^\star\|_2^2}{2},$$ where the first inequality comes from the inequalities $f(\theta_k) \geq f(\theta_n)$ for all $k \leq n$. This is sufficient to prove the first part. Note that proving convergence rates for first-order methods by finding telescopic sums is a classical technique (see, [*e.g.*]{},[@beck]).
[[*$\mu$-strongly convex case:*]{} ]{} Let us now assume that $f$ is $\mu$-strongly convex. The strong convexity implies that $f(\theta_n)-f^\star \geq \frac{\mu}{2}\|\theta_n-\theta^\star\|_2^2$ for all $n$. Combined with (\[eq:rate1\_proof\]), this yields $$\frac{\mu+\rho}{2}\|\theta_n-\theta^\star\|_2^2 \leq \frac{L}{2}\|\theta_{n-1}-\theta^\star\|_2^2,$$ and thus $$f(\theta_n) - f(\theta^\star) \leq \frac{L}{2}\|\theta_{n-1}-\theta^\star\|_2^2 \leq \left(\frac{L}{\rho+\mu}\right)^{n-1}\frac{L\|\theta_0-\theta^\star\|_2^2}{2}.$$
Even though the constants obtained in the rates of Proposition \[prop:conva\] are slightly better than the ones of Proposition \[prop:conv2\], the condition $g_n$ in $\S_{L,\rho}(f,\kappa)$ with $\rho \geq L$ is much stronger than the simple assumption that $g_n$ is in $\S_L(f,\kappa)$. It can indeed be shown that $f$ is necessarily $(\rho\!-\!L)$-strongly convex if $\rho \!>\! L$, and convex if $\rho\!=\!L$. In the next section, we give some examples where such a condition holds.
Examples of first-order surrogate functions {#subsec:surrogates}
-------------------------------------------
We now present practical first-order surrogate functions and links between Algorithm \[alg:generic\_batch\] and existing approaches. Even though our generic analysis does not always bring new results for each specific case, its main asset is to provide a unique theoretical treatment to all of them.
### Lipschitz gradient surrogates {#subsubsec:gradient}
When $f$ is $L$-smooth, it is natural to consider the following surrogate: $$g: \theta \mapsto f(\kappa) + \nabla f(\kappa)^\top (\theta-\kappa) + \frac{L}{2}\|\theta-\kappa\|_2^2.$$ The function $g$ is an upper bound of $f$, which is a classical result [@nesterov4]. It is then easy to see that $g$ is $L$-strongly convex and $L$-smooth. As a consequence, the difference $g-f$ is $2L$-smooth (as a sum of two $L$-smooth functions), and thus $g$ is in $\S_{2L,L}(f,\kappa)$.
When $f$ is convex, it is also possible to show by using Lemma \[lemma:convexerror\] that $g$ is in fact in $\S_{L,L}(f,\kappa)$, and when $f$ is $\mu$-strongly convex, $g$ is in $\S_{L-\mu,L}(f,\kappa)$. We remark that minimizing $g$ amounts to performing a gradient descent step: $\theta' \leftarrow \kappa - \frac{1}{L}\nabla f(\kappa)$.
### Proximal gradient surrogates {#subsubsec:proximal}
Let us now consider a composite optimization problem, meaning that $f$ splits into two parts $f = f_1 + f_2$, where $f_1$ is $L$-smooth. Then, a natural surrogate of $f$ is the following function: $$g: \theta \mapsto f_1(\kappa) + \nabla f_1(\kappa)^\top (\theta-\kappa) + \frac{L}{2}\|\theta-\kappa\|_2^2 + f_2(\theta).$$ The function $g$ majorizes $f$ and the approximation error $g-f$ is the same as in Section \[subsubsec:gradient\]. Thus, $g$ in in $\S_{2L}(f,\kappa)$ or in $\S_{2L,L}(f,\kappa)$ when $f_2$ is convex. Moreover,
- when $f_1$ is convex, $g$ is in $\S_{L}(f,\kappa)$. If $f_2$ is also convex, $g$ is in $\S_{L,L}(f,\kappa)$;
- when $f_1$ is $\mu$-strongly convex, $g$ is in $\S_{L-\mu}(f,\kappa)$. If $f_2$ is also convex, $g$ is in $\S_{L-\mu,L}(f,\kappa)$.
Minimizing $g$ amounts to performing one step of the proximal gradient algorithm [@beck; @nesterov; @wright]. It is indeed easy to show that the minimum $\theta'$ of $g$—assuming it is unique—can be equivalently obtained as follows: $$\theta' = \argmin_{\theta \in \Theta} \left[\frac{1}{2}\left\| \theta - \left(\kappa -\frac{1}{L}\nabla f_1(\kappa)\right)\right\|_2^2 + \frac{1}{L}f_2(\theta)\right],$$ which is often written under the form $\theta' =
\text{Prox}_{f_2/L}[\kappa-(1/L)\nabla f_1(\kappa)]$, where “$\text{Prox}$” is called the “proximal operator” [@moreau]. In some cases, the proximal operator can be computed efficiently in closed form, for example when $f_2$ is the $\ell_1$-norm; it yields the iterative soft-thresholding algorithm for sparse estimation [@daubechies]. For a review of proximal operators and their computations, we refer the reader to [@bach8; @combettes2005signal].
### Linearizing concave functions and DC programming {#subsec:dc}
Assume that $f = f_1 + f_2$, where $f_2$ is concave and $L$-smooth. Then, the following function $g$ is a majorizing surrogate in $\S_{L}(f,\kappa)$: $$g: \theta \mapsto f_1(\theta) + f_2(\kappa) + \nabla f_2(\kappa)^\top (\theta-\kappa).$$ Such a surrogate appears in DC (difference of convex) programming [@horst]. When $f_1$ is convex, $f$ is indeed the difference of two convex functions. It is also used in sparse estimation for dealing with some non-convex penalties [@bach8]. For example, consider a cost function of the form $\theta \mapsto f_1(\theta) + \lambda \sum_{j=1}^p \log(|\theta[j]| +
\varepsilon)$, where $\theta[j]$ is the $j$-th entry in $\theta$. Even though the functions $\theta \mapsto \log(|\theta[j]|
+\varepsilon)$ are not differentiable, they can be written as the composition of a concave smooth function $u \mapsto \log(u + \varepsilon)$ on $\Real^+$, and a Lipschitz function $\theta \mapsto |\theta[j]|$. By upper-bounding the logarithm function by its linear approximation, Proposition \[prop:conv1\_separable\] justifies the following surrogate: $$g: \theta \mapsto f_1(\theta) + \lambda \sum_{j=1}^p \log(|\kappa[j]| + \varepsilon) + \lambda\sum_{j=1}^p \frac{|\theta[j]|-|\kappa[j]|}{|\kappa[j]|+\varepsilon}, \label{eq:upperbounddc}$$ and minimizing $g$ amounts to performing one step of a reweighted-$\ell_1$ algorithm (see @candes4 and references therein). Similarly, other penalty functions are adapted to this framework. For instance, the logarithm can be replaced by any smooth concave non-decreasing function, or group-sparsity penalties [@turlach; @yuan] can be used, such as $\theta \mapsto \sum_{g \in {\mathcal{G}}}
\log(\|\theta_g\|_2 + \varepsilon)$, where ${\mathcal{G}}$ is a partition of $\{1,\ldots,p\}$ and $\theta_g$ records the entries of $\theta$ corresponding to the set $g$. Proposition \[prop:conv1\_separable\] indeed applies to this setting.
### Variational surrogates {#subsubsec:variational}
Let us now consider a real-valued function $f$ defined on $\Real^{p_1} \times \Real^{p_2}$. Let $\Theta_1 \subseteq \Real^{p_1}$ and $\Theta_2 \subseteq \Real^{p_2}$ be two convex sets. Minimizing $f$ over $\Theta_1 \times \Theta_2$ is equivalent to minimizing the function $\tilde{f}$ over $\Theta_1$ defined as $\tilde{f}(\theta_1) \defin \mapsto \min_{\theta_2
\in \Theta_2} f(\theta_1,\theta_2)$. Assume now that
- $\theta_2 \mapsto f(\theta_1,\theta_2)$ is $\mu$-strongly convex for all $\theta_1$ in $\Real^{p_1}$;
- $\theta_1 \mapsto f(\theta_1,\theta_2)$ is differentiable for all $\theta_2$;
- $(\theta_1,\theta_2) \mapsto \nabla_{1} f(\theta_1,\theta_2)$ is $L'$-Lipschitz with respect to $\theta_1$ and $L$-Lipschitz with respect to $\theta_2$.[^3]
Let us fix $\kappa_1$ in $\Theta_1$. Then, the following function is a majorizing surrogate in $\S_{L''}(\tilde{f},\kappa)$: $$g: \theta_1 \mapsto f(\theta_1,\kappa_2^\star) ~\text{with}~~ \kappa_2^\star \defin \argmin_{\theta_2 \in \Theta_2} f(\kappa_1,\theta_2),$$ with $L'' = 2L'+ L^{2}/\mu$. We can indeed apply Lemma \[lemma:danskin\], which ensures that $\tilde{f}$ is differentiable with $\nabla
\tilde{f}(\theta_1) = \nabla_1 f(\theta_1,\theta_2^\star)$ and $\theta_2^\star
\defin \argmin f(\theta_1,\theta_2)$ for all $\theta_1$. Moreover, $g$ is $L'$-smooth and $\tilde{f}$ is $L'+L^2/\mu$-smooth according to Lemma \[lemma:danskin\], and thus $h \defin g-\tilde{f}$ is $L''$-smooth. Note that a better constant $L''=L'$ can be obtained when $f$ is convex, as noted in the appendix of [@mairal17].
The surrogate $g$ leads to an alternate minimization algorithm; it is then interesting to note that Proposition \[prop:conv2\] provides similar convergence rates as another recent analysis [@beck2013convergence], which makes slightly different assumptions on the function $f$. Variational surrogates might also be useful for problems of a single variable $\theta_1$. For instance, consider a regression problem with a Huber loss function $H$ defined for all $u$ in $\Real$ as $$H(u) \defin \left\{
\begin{array}{lr} \frac{u^2}{2\delta} +\frac{\delta}{2} & \text{if}~~ |u| \leq \delta, \\
|u| & \text{otherwise}, \end{array} \right. \label{eq:huber}$$ where $\delta$ is a positive constant.[^4] The Huber loss can be seen as a smoothed version of the $\ell_1$-norm when $\delta$ is small, or simply a robust variant of the squared loss $u \mapsto \frac{1}{2}u^2$ that asymptotically grows linearly. Then, it is easy to show that $$H(u) = \frac{1}{2} \min_{w \geq \delta} \left[\frac{u^2}{w} + w\right]. \label{eq:huber_var}$$ Consider now a regression problem with $m$ training data points represented by vectors $\x_i$ in $\Real^p$, associated to real numbers $y_i$, for $i=1,\ldots,m$. The robust regression problem with the Huber loss can be formulated as the minimization over $\Real^p$ of $$\tilde{f}: \theta_1 \mapsto \sum_{i=1}^m H(y_i- \x_i^\top \theta_1) = \min_{\theta_2 \in \Real^m : \theta_2 \geq \delta} \left[f(\theta_1,\theta_2) \defin \frac{1}{2} \sum_{i=1}^m\frac{(y_i- \x_i^\top \theta_1)^2}{\theta_2[i]} + \theta_2[i]\right],$$ where $\theta_1$ is the parameter vector of a linear model. The conditions described at the beginning of this section can be shown to be satisfied with a Lipschitz constant proportional to $(1/\delta)$; the resulting algorithm is the iterative reweighted least-square method, which appears both in the literature about robust statistics [@lange2], and about sparse estimation where the Huber loss is used to approximate the $\ell_1$-norm [@bach8].
### Jensen surrogates
Jensen’s inequality also provides a natural mechanism to obtain surrogates for convex functions. Following the presentation of Lange, Hunger and Yang @lange2, we consider a convex function $f: \Real \mapsto \Real$, a vector $\x$ in $\Real^p$, and define $\tilde{f}:
\Real^p \to \Real$ as $\tilde{f}(\theta) \defin f(\x^\top \theta)$ for all $\theta$. Let $\w$ be a weight vector in $\Real_+^p$ such that $\|\w\|_1=1$ and $\w[i] \neq 0$ whenever $\x[i] \!\neq\! 0$. Then, we define for any $\kappa$ in $\Real^p$: $$g: \theta \mapsto \sum_{i=1}^p \w[i] f \left(\frac{\x[i]}{\w[i]}( \theta[i]-\kappa[i]) + \x^\top \kappa\right),$$ When $f$ is $L$-smooth, and when $\w[i] \defin |\x[i]|^\nu / \|\x\|_\nu^\nu$, $g$ is in $\S_{L'}(\tilde{f},\kappa)$ with
- $L' = L\|\x\|_\infty^2\|\x\|_0$ for $\nu=0$;
- $L' = L\|\x\|_\infty\|\x\|_1$ for $\nu=1$;
- $L' = L\|\x\|_2^2$ for $\nu=2$.
To the best of our knowledge, non-asymptotic convergence rates have not been studied before for such surrogates, and thus we believe that our analysis may provide new results in the present case. Jensen surrogates are indeed quite uncommon; they appear nevertheless in a few occasions. In addition to the few examples given in [@lange2], they are used for instance in machine learning by Della Pietra @pietra for interpreting boosting procedures through the concept of *auxiliary functions*.
Jensen’s inequality is also used in a different fashion in EM algorithms [@dempster; @neal]. Consider $T$ non-negative functions $f^t:
\Real^p \mapsto \Real_+$, and, for some $\kappa$ in $\Real^p$, define some weights $\w[t]= f^t(\kappa) / \sum_{t'=1}^T f^{t'}(\kappa)$. By exploiting the concavity of the logarithm, and assuming hat $\w[t] > 0$ for all $t$ to simplify, Jensen’s inequality yields $$- \log \left( \sum_{t=1}^T f^t(\theta) \right) \leq - \sum_{t=1}^T \w[t] \log \left(\frac{f^t(\theta)}{\w[t]}\right), \label{eq:em}$$ The relation (\[eq:em\]) is key to EM algorithms minimizing a negative log-likelihood. The right side of this equation can be interpreted as a majorizing surrogate of the left side since it is easy to show that both terms are equal for $\theta=\kappa$. Unfortunately the resulting approximation error functions are not $L$-smooth in general and these surrogates do not follow the assumptions of Definition \[def:surrogate\_batch\]. As a consequence, our analysis may apply to some EM algorithms, but not to all of them.
### Quadratic surrogates
When $f$ is twice differentiable and admits a matrix $\HH$ such that $\HH- \nabla^2 f$ is always positive definite, the following function is a first-order majorizing surrogate: $$g: \theta \mapsto f(\kappa) + \nabla f(\kappa)^\top (\theta-\kappa) + \frac{1}{2}(\theta-\kappa)^\top\HH(\theta-\kappa).$$ The Lipschitz constant of $\nabla(g-f)$ is the largest eigenvalue of $\HH- \nabla^2 f(\theta)$ over $\Theta$. Such surrogates appear frequently in the statistics and machine learning literature [@bohning; @jebara; @khan]. The goal is to to model the global curvature of the objective function during each iteration, without resorting to the Newton method. Even though quadratic surrogates do not necessarily lead to better theoretical convergence rates than simpler Lipschitz gradient surrogates, they can be quite effective in practice [@jebara].
An incremental majorization-minimization algorithm: MISO {#sec:incremental}
========================================================
In this section, we introduce an incremental scheme that exploits the structure (\[eq:prob\]) of $f$ as a large sum of $T$ components. The most popular method for dealing with such a problem when $f$ is smooth and $\Theta=\Real^p$ is probably the *stochastic gradient descent* algorithm (SGD) and its variants (see [@nemirovski]). It consists of drawing at iteration $n$ an index ${\hat
t}_n$ and updating the solution as $\theta_n \leftarrow \theta_{n-1} - \eta_n
\nabla f^{{\hat t}_n}(\theta_{n-1})$, where the scalar $\eta_n$ is a step size. Another popular algorithm is the *stochastic mirror descent* algorithm (see [@juditsky]) for general non-smooth convex problems, a setting we do not consider in this paper since non-smooth functions do not always admit practical first-order surrogates.
Recently, linear convergence rates for strongly convex functions $f^t$ have been obtained in @schmidt2 and @shalev2 by using randomized incremental algorithms whose cost per iteration is independent of $T$. The method SAG @schmidt2 for smooth unconstrained convex optimization is a randomized variant of the incremental gradient descent algorithm of Blatt, Hero and Gauchman [@blatt], where an estimate of the gradient $\nabla f$ is incrementally updated at each iteration. The method SDCA @shalev2 for strongly convex composite optimization is a dual coordinate ascent algorithm that performs incremental updates in the primal (\[eq:prob\]). Unlike SGD, both SAG and SDCA require storing information about past iterates, which is a key for obtaining fast convergence rates.
In a different context, incremental EM algorithms have been proposed by Neal and Hinton @neal, where upper bounds of a non-convex negative log-likelihood function are incrementally updated. By using similar ideas, we introduce the scheme MISO in Algorithm \[alg:generic\_incremental\]. At every iteration, a single function is observed, and an approximate surrogate of $f$ is updated. Note that in the same line of work, Ahn et al. [@ahn] have proposed a block-coordinate descent majorization-minimization algorithm, which corresponds to MISO when the variational surrogates of Section \[subsubsec:variational\] are used.
$\theta_0 \in \Theta$ (initial estimate); $N$ (number of iterations). Initialization: choose some surrogates $g_0^t$ of $f^t$ near $\theta_0$ for all $t$; Randomly pick up one index $\hat{t}_n$ and choose a surrogate $g_n^{\hat{t}_n}$ of $f^{\hat{t}_n}$ near $\theta_{n-1}$; set $g^t_n \defin g^t_{n-1}$ for all $t \neq \hat{t}_n$. Update the solution: $ \theta_n \in {\displaystyle \argmin_{\theta \in \Theta}} \frac{1}{T} \sum_{t=1}^T g_n^t(\theta)$.
$\theta_{N}$ (final estimate);
In the next two sections, we study the convergence properties of the scheme MISO. We proceed as in Section \[sec:generic\]. Specifically, we start with the non-convex case, focusing on stationary point conditions, and we show that similar guarantees as for the batch majorization-minimization algorithm hold. Then, for convex problems, we present convergence rates that essentially apply to the proximal gradient surrogates. We obtain sublinear rates $O(T/n)$ for the general convex case, and linear ones for strongly convex objective functions. Even though these rates do not show any theoretical advantage over the batch algorithm, we also present a more surprising result in Section \[subsec:strong\]; in a large sample regime $T \geq 2L/\mu$, for $\mu$-strongly convex functions $f^t$, minorizing surrogates may be used and faster rates can be achieved.
Convergence analysis {#subsec:convmiso}
--------------------
We start our analysis with the non-convex case, and make the following assumption:
- $f$ is bounded below and for all $\theta,\theta'$ in $\Theta$ and all $t$, the directional derivative $\nabla f^t(\theta,\theta'-\theta)$ of $f^t$ at $\theta$ in the direction $\theta'-\theta$ exists.
Then, we obtain a first convergence result.
\[prop:conv13\] Assume that holds and that the surrogates $g_n^{{\hat t}_n}$ from Algorithm \[alg:generic\_incremental\] are majorizing $f^{{\hat t}_n}$ and are in $\S_{L}(f^{{\hat t}_n},\theta_{n-1})$. Then, the conclusions of Proposition \[prop:conv1\] hold with probability one.
We proceed in several steps.
[[*Almost sure convergence of $(f(\theta_n))_{n \geq 0}$:*]{} ]{} Let us define $\barg_n \defin \frac{1}{T}\sum_{t=1}^T g_n^t$. We have the following relation for all $n \geq 1$, $$\barg_n = \barg_{n-1} + \frac{g_n^{{\hat t}_n} - g_{n-1}^{{\hat t}_n}}{T}, \label{eq:recur}$$ where the surrogates and the index ${\hat t}_n$ are chosen in the algorithm. Then, we obtain the following inequalities, which hold with probability one for all $n \geq 1$, $$\begin{split}
\barg_{n}(\theta_n) & \leq \barg_n(\theta_{n-1}) = \barg_{n-1}(\theta_{n-1}) + \frac{g_n^{{\hat t}_n}(\theta_{n-1}) - g_{n-1}^{{\hat t}_n}(\theta_{n-1})}{T} \\
& = \barg_{n-1}(\theta_{n-1}) + \frac{f^{{\hat t}_n}(\theta_{n-1}) - g_{n-1}^{{\hat t}_n}(\theta_{n-1})}{T} \leq \barg_{n-1}(\theta_{n-1}).
\end{split}$$ The first inequality is true by definition of $\theta_n$ and the second one because $\barg_{n-1}^{{\hat t}_n}$ is a majorizing surrogate of $f^{{\hat
t}_n}$. The sequence $(\barg_n(\theta_n))_{ n \geq 0}$ is thus monotonically decreasing, bounded below with probability one, and thus converges almost surely. By taking the expectation of these previous inequalities, we also obtain that the sequence $(\E[\barg_{n}(\theta_n)])_{n \geq 0}$ monotonically converges. Thus, the non-positive quantity $\E[f^{{\hat
t}_n}(\theta_{n-1}) - g_{n-1}^{{\hat t}_n}(\theta_{n-1})]$ is the summand of a converging sum and we have $$\begin{split}
\E\left[ \sum_{n=0}^{+\infty} g_{n}^{{\hat t}_{n+1}}(\theta_{n}) - f^{{\hat t}_{n+1}}(\theta_{n})\right] & = \sum_{n=0}^{+\infty} \E[g_{n}^{{\hat t}_{n+1}}(\theta_{n}) - f^{{\hat t}_{n+1}}(\theta_{n})] \\
& = \sum_{n=0}^{+\infty} \E[\E[g_{n}^{{\hat t}_{n+1}}(\theta_{n}) - f^{{\hat t}_{n+1}}(\theta_{n})| \FF_{n}]] \\
& = \sum_{n=0}^{+\infty} \E[\barg_{n}(\theta_{n}) - f(\theta_{n})] \\
& =\E\left[\sum_{n=0}^{+\infty} \barg_{n}(\theta_{n}) - f(\theta_{n})\right] < +\infty, \\
\end{split}$$ where we use Beppo-Lévy theorem to interchange the expectation and the sum in front of non-negative quantities, and $\FF_n$ is the filtration representing all information up to iteration $n$ (including $\theta_n$). As a result, the sequence $(\barg_{n}(\theta_{n}) - f(\theta_{n}))_{n \geq 0}$ converges almost surely to $0$, implying the almost sure convergence of $(f(\theta_n))_{n \geq 0}$.
[[*Asymptotic stationary point conditions:*]{} ]{} Let us define ${\bar h}_n \defin \barg_n-f$, which is $L$-smooth. Then, for all $\theta$ in $\Theta$ and $n \geq 1$, $$\nabla f(\theta_n,\theta-\theta_n) = \nabla \barg_n(\theta_n,\theta-\theta_n) - \nabla {\bar h}_n(\theta_n)^\top(\theta-\theta_n).$$ We have $\nabla \barg_n(\theta_n,\theta-\theta_n) \geq 0$ by definition of $\theta_n$, and $\|\nabla {\bar h}_n(\theta_n)\|_2^2 \leq 2L {\bar h}_n(\theta_n)$, following similar steps as in the proof of Proposition \[prop:conv1\]. Since we have previously shown that $({\bar h}_n(\theta_n))_{n \geq 0}$ almost surely converges to zero, we conclude as in the proof of Proposition \[prop:conv1\], replacing $h_n$ by $\barh_n$ and $g_n$ by $\barg_n$.
We also give the counterpart of Proposition \[prop:conv1\_separable\] for Algorithm \[alg:generic\_incremental\].
\[prop:conv13b\] Assume that is satisfied and that the functions $f^t$ are compositions $f^t = f^{\prime t} \circ e^t$, where the functions $e^t$ are $C$-Lipschitz continuous for some $C>0$. Assume also that the functions $g_n^{{\hat t}_n}$ in Algorithm \[alg:generic\_incremental\] are also compositions $g_n^{{\hat t}_n} = g_n^{\prime {\hat t}_n} \circ e^{{\hat t}_n}$, where $g_n^{\prime {\hat t}_n}$ is majorizing $f^{\prime {\hat t}_n}$ and is in $\S_{L}(f^{\prime {\hat t}_n},e^{{\hat t}_n}(\theta_{n-1}))$. Then, the conclusions of Proposition \[prop:conv13\] hold.
We first remark that the first part of the proof of Proposition \[prop:conv13\] does not exploit the fact that the approximation errors $g_n^t - f^t$ are $L$-smooth, but only the fact that $g_n^t$ is majorizing $f^t$ for all $n$ and $t$. Thus, the first part of the proof of Proposition \[prop:conv13\] holds in the present case, such that $(f(\theta_n))_{n \geq 0}$ almost surely converges, and the sequence $(\barg_n(\theta_n)- f(\theta_n))_{n \geq 0}$ almost surely converges to zero, where $\barg_n$ is defined in the proof of Proposition \[prop:conv13\].
It remains to show that the asymptotic stationary point conditions are satisfied. To that effect, we follow the proof of Proposition \[prop:conv1\_separable\]. We first have, for all $n \geq 1$, $$\nabla f(\theta_n,\theta-\theta_n) = \nabla \barg_n(\theta_n,\theta-\theta_n) - \frac{1}{T}\sum_{t=1}^T \nabla {\bar h}^t_n(\theta_n,\theta-\theta_n),$$ with $\nabla \barg_n(\theta_n,\theta-\theta_n) \geq 0$ and $\barh_n^t \defin \barg_n^t-f^t$. Then, following the proof of Proposition \[prop:conv1\_separable\], it is easy to show that $$|\nabla \barh_n^t(\theta_n, \theta-\theta_n)| \leq C\| \nabla \barh_n^{\prime t}(e^t(\theta_n))\|_2{\|\theta-\theta_n\|_2},$$ where $\barh_n^{\prime t} = \barg_n^{\prime t}-f^{\prime t}$, and we conclude as in Proposition \[prop:conv1\_separable\].
The next lemma provides convergence rates for the convex case, under the assumption that the surrogate functions are $\rho$-strongly convex with $\rho
\geq L$. The result notably applies to the proximal gradient surrogates of Section \[subsubsec:proximal\].
\[prop:conv16\] Assume that $f$ is convex and bounded below, let $\theta^\star$ be a minimizer of $f$ on $\Theta$, and let us define $f^\star \defin \min_{\theta \in \Theta} f(\theta)$. When the surrogates $g_n^t$ in Algorithm \[alg:generic\_incremental\] are majorizing $f^t$ and are in $\S_{L,\rho}(f^t,\theta_{n-1})$ with $\rho \geq L$, we have for all $n \geq 1$, $$\E[ f(\bartheta_n)-f^\star] \leq \frac{LT\|\theta^\star-\theta_0\|_2^2}{2n}, \label{eq:incr:rate}$$ where $\bartheta_n \defin \frac{1}{n} \sum_{i=1}^n \theta_i$ is the average of the iterates. Assume now that $f$ is $\mu$-strongly convex. For all $n \geq 1$, $$\E[ f(\theta_n) - f^\star] \leq \left(1- \frac{2\mu}{T(\rho+\mu)}\right)^{n-1}\frac{L\|\theta^\star-\theta_0\|_2^2}{2}.\label{eq:incr:ratemu}$$
We proceed in several steps.
[[*Preliminaries:*]{} ]{} For all $n \geq 1$, we introduce the point $\kappa_{n-1}^t$ in $\Theta$ such that $g_n^t$ is in $\S_{L,\rho}(f^t,\kappa_{n-1}^t)$. We remark that such points are drawn recursively according to the following conditional probability distribution: $$\PPP(\kappa_{n-1}^t = \theta_{n-1} | \FF_{n-1}) = \delta ~~\text{and}~~ \PPP(\kappa_{n-1}^t = \kappa_{n-2}^t | \FF_{n-1})=1-\delta,$$ where $\delta \defin 1/T$, $\FF_n$ is the filtration representing all information up to iteration $n$ (including $\theta_n$), and $\kappa_0^t \defin \theta_0$ for all $t$. Thus we have for all $t$ and all $n \geq 1$, $$\E[\|\theta^\star-\kappa_{n-1}^t\|_2^2]= \E[\E[\|\theta^\star-\kappa_{n-1}^t \|_2^2| \FF_{n-1}]] = \delta\E[\|\theta^\star-\theta_{n-1}\|_2^2] + (1-\delta) \E[\|\theta^\star-\kappa_{n-2}^t\|_2^2].\label{eq:incr:tmp2}$$ We also need the following extension of Lemma \[lemma:basic\] to the incremental setting: for all $\theta$ in $\Theta$ and $n \geq 1$, $$f(\theta_n) \leq f(\theta) + \frac{1}{T}\sum_{t=1}^T \left( \frac{L}{2}\|\theta-\kappa_{n-1}^t\|_2^2 -\frac{\rho}{2}\|\theta-\theta_n\|_2^2 \right). \label{eq:incr:tmp4}$$ The proof of this relation is similar to that of Lemma \[lemma:basic\], exploiting the $\rho$-strong convexity of ${\bar g}_n \defin (1/T)\sum_{t=1}^T g_n^t$. We can now study the first part of the proposition.
[[*Non-strongly convex case ($\rho=L$):*]{} ]{} Let us define the quantities $A_n \defin \E[\frac{1}{2T}\sum_{t=1}^T \|\theta^\star-\kappa_n^t\|_2^2]$ and $\xi_n \defin \frac{1}{2}\E[\|\theta^\star-\theta_n\|_2^2]$. Then, we have from (\[eq:incr:tmp4\]) with $\theta=\theta^\star$, and by taking the expectation $$\E[f(\theta_n)-f^\star] \leq LA_{n-1} - L \xi_n.$$ It follows from (\[eq:incr:tmp2\]) that $A_n = \delta\xi_{n} + (1-\delta) A_{n-1}$ and thus, for all $n \geq 1$, $$\E[f(\theta_n)-f^\star] \leq \frac{L}{\delta}(A_{n-1} - A_n).$$ By summing the above inequalities, and using Jensen’s inequality, we obtain that $$\E[f(\bartheta_n)-f^\star] \leq \frac{1}{n}\sum_{i=1}^n\E[f(\theta_i)-f^\star] \leq \frac{L A_0}{\delta},$$ leading to the convergence rate of Eq. (\[eq:incr:rate\]), since $A_0 = \frac{1}{2}\|\theta^\star-\theta_0\|_2^2$.
[[*$\mu$-strongly convex case:*]{} ]{} Assume now that the functions $f^t$ are $\mu$-strongly convex. For all $n \geq 1$, the strong convexity of $f$ and (\[eq:incr:tmp4\]) give us the following inequalities $$\mu \xi_n \leq \E[f(\theta_n) - f^\star] \leq L A_{n-1} - \rho\xi_n,$$ Combining this last inequality with (\[eq:incr:tmp2\]), we obtain that for all $n \geq 1$, $$A_n = \delta \xi_n + (1-\delta) A_{n-1} \leq \left( \frac{\delta L}{\mu+\rho} + (1-\delta) \right) A_{n-1}.$$ Thus, $A_n \leq \beta^n A_0$ with $\beta\defin \frac{ (1-\delta)(\rho+\mu)+\delta L}{\rho+\mu}$. Since $A_0=\xi_0$, $\E[f(\theta_n) - f^\star] \leq L A_{n-1}$, and $\beta \leq 1- 2\delta \mu/(\rho+\mu)$, we finally have shown the desired convergence rate (\[eq:incr:ratemu\]).
The convergence rate of the previous proposition in the convex case suggests that the incremental scheme and the batch one of Section \[sec:generic\] have the same overall complexity, assuming that each iteration of the batch algorithm is $T$ times the one of MISO. For strongly convex functions $f^t$, we obtain linear convergence rates, a property shared by SAG or SDCA; it is thus natural to make a more precise comparison with these other incremental approaches, which we present in the next two sections.
MISO for smooth unconstrained optimization {#subsec:strong}
------------------------------------------
In this section, we assume that the optimization domain is unbounded—that is, $\Theta=\Real^p$, and that the functions $f^t$ are $L$-smooth. When using the Lipschitz gradient surrogates of Section \[subsubsec:gradient\], MISO amounts to iteratively using the following update rule: $$\theta_n \leftarrow \frac{1}{T} \sum_{t=1}^T \kappa_{n-1}^t -
\frac{1}{LT}\sum_{t=1}^T \nabla f^t(\kappa_{n-1}^t), \label{eq:MISO1}$$ where the vectors $\kappa_{n-1}$ are recursively defined for $n \geq 2$ as $\kappa_{n-1}^{\hat{t}_{n}}=\theta_{n-1}$ and $\kappa_{n-1}^{{t}}=\kappa_{n-2}^{{t}}$ for $t \neq \hat{t}_{n}$, with $\kappa_{0}^t=\theta_0$ for all $t$. It is then easy to see that the complexity of updating $\theta_n$ is independent of $T$, by storing the vectors $\z_n^t=\kappa_{n-1}^t-(1/L)\nabla f^t(\kappa_{n-1}^t)$ and performing the update $\theta_n = \theta_{n-1} + (1/T)(\z_n^t-\z_{n-1}^t)$. In comparison, the approach SAG yields a different, but related, update rule: $$\theta_n \leftarrow \theta_{n-1} -
\frac{\alpha}{T}\sum_{t=1}^T \nabla f^t(\kappa_{n-1}^t), \label{eq:SAG1}$$ where the value $\alpha=1/(16L)$ is suggested in [@schmidt2]. Even though the rules (\[eq:MISO1\]) and (\[eq:SAG1\]) seem to be similar to each other at first sight, they behave differently in practice and do not have the same theoretical properties. For non-convex problems, MISO is guaranteed to converge, whereas it is not known whether it is the case for SAG or not. For convex problems, both methods have a convergence rate of the same nature—that is, $O(T/n)$. For $\mu$-strongly-convex problems, however, the convergence rate of SAG reported in [@schmidt2] is substantially better than ours. Whereas the expected objective of SAG decreases with the rate $O(\rho^n)$ with $\rho_\text{SAG}=1-\min( \mu/(16L), 1/(8T))$, ours decreases with $\rho_\text{MISO}=1-2\mu/(T(L+\mu))$, which is larger than $\rho_\text{SAG}$ unless the problem is very well conditioned.
By maximizing the convex dual of (\[eq:prob\]) when the functions $f^t$ are $\mu$-strongly convex, the approach SDCA yields another update rule that resembles (\[eq:MISO1\]) and (\[eq:SAG1\]), and offers similar convergence rates as SAG. As part of the procedure, SDCA involves large primal gradient steps $\theta_{n-1}-(1/\mu)\nabla f^{\hat{t}_n}(\theta_{n-1})$ for updating the dual variables. It is thus appealing to study whether such large gradient steps can be used in (\[eq:MISO1\]) in the strongly convex case, regardless of the majorization-minimization principle. In other words, we want to study the use of the following surrogates within MISO: $$g_n^t : \theta \mapsto f^t(\kappa_{n-1}^t) + \nabla f^t(\kappa_{n-1}^t)^\top (\theta-\kappa_{n-1}^t) + \frac{\mu}{2}\|\theta-\kappa_{n-1}^t\|_2^2, \label{eq:lower_surrogates}$$ which are lower bounds of the functions $f^t$ instead of upper bounds. Then, minimizing $(1/T)\sum_{t=1}^T g_n^t$ amounts to performing the update (\[eq:MISO1\]) when replacing $L$ by $\mu$. The resulting algorithm is slightly different than SDCA, but resembles it. As shown in the next proposition, the method achieves a fast convergence rate when $T \geq 2L/\mu$, but may diverge if $T$ is small. Note that at the same time as us, a similar result was independently obtained by Defazio et al. [@defazio], where a refined analysis provides a slightly better rate, namely the constant $1/3$ in (\[eq:incr:ratemu2\]) may be replaced by $1/2$.
\[prop:conv17\] Assume that the functions $f^t$ are $\mu$-strongly convex, $L$-smooth, and bounded below. Let $\theta^\star$ be a minimizer of $f$ on $\Theta$. Assume that $T \geq 2L/\mu$. When the functions $g_n^t$ of Eq. (\[eq:lower\_surrogates\]) are used in Algorithm \[alg:generic\_incremental\], we have for all $n \geq 1$, $$\E[ f(\theta_n) - f^\star] \leq \left(1 - \frac{1}{3T}\right)^{n} \frac{2T}{\mu}\|\nabla f(\theta_0)\|_2^2.\label{eq:incr:ratemu2}$$ When the functions $f^t$ are lower-bounded by the function $\theta \mapsto
(\mu/2)\|\theta\|_2^2$, we can use the initialization $\theta_0=0$ and $g_0^t: \theta \mapsto (\mu/2)\|\theta\|_2^2$ for all $t$. Then, the quantity $({2T}/{\mu})\|\nabla f(\theta_0)\|_2^2$ in (\[eq:incr:ratemu2\]) can be replaced by $T f^\star$.
As in the proof of Proposition \[prop:conv13\], we introduce the function $\barg_n \defin
\frac{1}{T}\sum_{t=1}^T g_n^t$, which is minimized by $\theta_n$ for $n \geq 1$. Since $\barg_n$ is a lower bound on $f$, we have the relation $\barg_n(\theta_n) \leq \barg_n(\theta^\star) \leq f^\star$. Inspired by the convergence proof of SDCA [@shalev2], which computes an convergence rate of an expected duality gap, we proceed by studying the convergence of the sequence $(f^\star-\E[\barg_n(\theta_n)])_{n \geq 1}$.
On the one hand, we have for all $n \geq 1$, $$\begin{split}
\barg_n(\theta_n) & = \barg_n(\theta_{n-1}) - \frac{\mu}{2}\|\theta_n-\theta_{n-1}\|_2^2 \\
& = \barg_{n-1}(\theta_{n-1}) + \delta( g_n^{{\hat t}_n}(\theta_{n-1}) - g_{n-1}^{{\hat t}_n}(\theta_{n-1})) - \frac{\mu}{2}\|\theta_n-\theta_{n-1}\|_2^2,
\end{split} \label{eq:tmpconv13aa}$$ where $\delta = 1/T$. The first equality is true because $\barg_n$ is quadratic and is minimized by $\theta_n$, and the second one uses the relation (\[eq:recur\]). By definition of $g_n^{{\hat t}}$, we have that $g_n^{{\hat t}_n}(\theta_{n-1}) = f^{{\hat t}_n}(\theta_{n-1})$, and by taking the expectation, $\E[g_n^{{\hat t}_n}(\theta_{n-1})]=
\E[f^{{\hat t}_n}(\theta_{n-1})] = \E[\E[f^{{\hat t}_n}(\theta_{n-1})|
\FF_{n-1}]] = \E[f(\theta_{n-1})]$, where $\FF_n$ is the the filtration representing all information up to iteration $n$. We also have that $\E[g_{n-1}^{{\hat t}_n}(\theta_{n-1})] = \E[\E[g_{n-1}^{{\hat t}_n}(\theta_{n-1}) | \FF_{n-1}]] = \E[\barg_{n-1}(\theta_{n-1})]$. Thus, we obtain a first useful relation: $$\E[\barg_n(\theta_n)] = (1-\delta)\E[\barg_{n-1}(\theta_{n-1})] + \delta \E[f(\theta_{n-1})] - \frac{\mu}{2}\E\left[\|\theta_n-\theta_{n-1}\|_2^2\right].\label{eq:tmpconv13a}$$ On the other hand, for all $n\geq 2$, $$\begin{split}
\barg_n(\theta_n) & = \barg_{n-1}(\theta_n) + \delta (g_n^{{\hat t}_n}(\theta_{n}) - g_{n-1}^{{\hat t}_n}(\theta_{n})) \\
& = \barg_{n-1}(\theta_{n-1}) \!+\! \frac{\mu-\delta L}{2}\|\theta_{n}\!-\!\theta_{n-1}\|_2^2 \!+\! \delta \left(g_n^{{\hat t}_n}(\theta_{n}) \!+\! \frac{L}{2}\|\theta_n\!-\!\theta_{n-1}\|_2^2 - g_{n-1}^{{\hat t}_n}(\theta_{n})\right) \\
& \geq \barg_{n-1}(\theta_{n-1}) + \frac{\mu-\delta L}{2}\|\theta_{n} - \theta_{n-1}\|_2^2.
\end{split} \label{eq:tmpconv13c}$$ We have used the fact that $\theta \mapsto g_n^{{\hat t}_n}(\theta) +
(L/2)\|\theta-\theta_{n-1}\|_2^2$ is a majorizing surrogate of $f^{{\hat
t}_n}$, whereas $g_{n-1}^{{\hat t}_n}$ is minorizing $f^{{\hat t}_n}$. By adding twice (\[eq:tmpconv13c\]) after taking the expectation and once (\[eq:tmpconv13a\]), we obtain that for all $n \geq 2$, $$\begin{split}
3\E[\barg_n(\theta_n)] & \geq (3 - \delta)\E[\barg_{n-1}(\theta_{n-1})] + \delta\E[ f(\theta_{n-1})] + \left(\frac{\mu}{2}-\delta L\right)\E[\|\theta_n-\theta_{n-1}\|_2^2] \\
& \geq (3 - \delta)\E[\barg_{n-1}(\theta_{n-1})] + \delta\E[ f(\theta_{n-1})],
\end{split}
\label{eq:tmpconv13b}$$ where the second inequality comes from the large sample size condition $\delta L \leq \mu/2$. Since $\E[f(\theta_{n-1})] \geq f^\star$, this immediately gives for $n\geq 2$, $$f^\star - \E\left[\barg_n(\theta_n)\right] \leq \left( 1- \frac{1}{3T} \right) \left(f^\star - \E\left[\barg_{n-1}(\theta_{n-1})\right]\right).$$ To obtain a convergence rate for $\E[f(\theta_n)]-f^\star$, we use again Eq. (\[eq:tmpconv13b\]). For $n \geq 2$, $$\begin{split}
\delta(\E[ f(\theta_{n-1})] - f^\star) & \leq \delta(\E[ f(\theta_{n-1})] - \E[\barg_{n-1}(\theta_{n-1})]) \\
& \leq 3 (\E[\barg_n(\theta_n)]- \E[\barg_{n-1}(\theta_{n-1})]) \\
& \leq 3 (f^\star- \E[\barg_{n-1}(\theta_{n-1})]) \\
& \leq 3 \left( 1- \frac{1}{3T} \right)^{n-2}\left(f^\star - \barg_{1}(\theta_{1})\right),
\end{split} \label{eq:tmpincr1}$$ and we obtain the convergence rate (\[eq:incr:ratemu2\]) by first noticing that $$\begin{split}
f^\star - \barg_1(\theta_1) & = f^\star - \barg_1(\theta_0) + \frac{\mu}{2}\|\theta_0-\theta_1\|_2^2 \\
& = f^\star - f(\theta_0) + \frac{\mu}{2}\left\|\frac{1}{\mu} \nabla f(\theta_0)\right\|_2^2 \\
& \leq \frac{1}{2\mu}\|\nabla f(\theta_0)\|_2^2,
\end{split}$$ where we use the relation $\barg_1 = \barg_0$ and $\barg_0(\theta_0)=f(\theta_0)$. Then, we use the fact that $(1-1/3T) \geq
5/6$ since $T \geq 2L/\mu \geq 2$, such that $3(1-1/3T)^{-1}/(2\mu) \leq
9/(5\mu) \leq 2/\mu$.
To prove the last part of the proposition, we remark that all inequalities we have proved so far for $n\geq 2$, become true for $n=1$. Thus, the last inequality in (\[eq:tmpincr1\]) is also true when replacing $n-2$ by $n-1$ and $\barg_1(\theta_1)$ by $\barg_0(\theta_0)=0$.
The proof technique is inspired in part by the one of SDCA [@shalev2]; the quantity $\sum_{t=1}^T g_n^t(\theta_n)$ is indeed a lower bound of $f^\star$, and plays a similar role as the dual value in SDCA. We remark that the convergence rate (\[eq:incr:ratemu2\]) improves significantly upon the original one (\[eq:incr:ratemu\]), and is similar to the one of SAG when $T$ is larger than $2L/\mu$.[^5] However, Proposition \[prop:conv17\] only applies to strongly convex problems. In other cases, the more conservative rule (\[eq:MISO1\]) should be preferred in theory, even though we present heuristics in Section \[subsec:heuristics\] that suggest using larger step sizes than $1/L$ in practice.
MISO for composite optimization
-------------------------------
When $f$ can be written as $f=(1/T)\sum_{t=1}^T f_1^t + f_2$, where the functions $f_1^t$ are $L$-smooth, we can use the proximal gradient surrogate presented in Section \[subsubsec:proximal\]; it yields the following rule: $$\theta_n \in \argmin_{\theta \in \Theta} \frac{1}{2}\left\| \theta -
\left(\frac{1}{T}\sum_{t=1}^T\kappa_{n-1}^t - \frac{1}{LT} \sum_{t=1}^T
\nabla f_1^t ( \kappa_{n-1}^t)\right) \right\|_2^2 +
\frac{\lambda}{L}f_2(\theta), \label{eq:MISO_composite}$$ where the vectors $\kappa_{n-1}^t$ are defined as in Section \[subsec:strong\]. This update is related to SDCA, as well as to stochastic methods for composite convex optimization such as the regularized dual averaging algorithm of Xiao [@xiao]. As in the previous section, we obtain guarantees for non-convex optimization, but our linear convergence rate for strongly convex problems is not as fast as the one of SDCA. Even though we do not have a similar result as Proposition \[prop:conv17\] for the composite setting, we have observed that using a smaller value for $L$ than the theoretical one could work well in practice. We detail such an empirical strategy in the next section.
Practical implementation and heuristics {#subsec:heuristics}
---------------------------------------
We have found the following strategies to improve the practical performance of MISO.
#### Initialization
A first question is how to initialize the surrogates $g_0^t$ in practice. Even though we have suggested the functions $g_0^t$ to be in $\S_{L}(f^t,\theta_0)$ in Algorithm \[alg:generic\_incremental\], our analysis weakly relies on this assumption. In fact, most of our results hold when choosing surrogates computed at points $\kappa_0^t$ that are not necessarily equal to $\theta_0$; at most only constants from the convergence rates would be affected by such a change. An effective empirical strategy is inspired by the second part of Proposition \[prop:conv17\]: we first define functions $g_{0}^t: \theta
\mapsto (L/2)\|\theta-\theta_0\|_2^2$, and perform $T$ iterations of MISO without randomization, selecting the function $f^t$ at iteration $t$, such that each surrogate is updated exactly once. Then, we use these updated surrogates for initializing the regular randomized scheme.
#### Warm restart and continuation
When available, warm restart can be used for initializing the surrogates. Assume that we are interested in minimizing a composite function $(1/T)\sum_{t=1}^T f_1^t(\theta) + \lambda f_2(\theta)$, which is parameterized by a scalar $\lambda$, and that we want to obtain a minimizer for several parameter values $\lambda_1 < \lambda_2 < ... < \lambda_M$. We first solve the problem for $\lambda=\lambda_M$, and then use the surrogates obtained at the end of the optimization for initializing the algorithm when addressing the problem with $\lambda=\lambda_{M-1}$. We proceed similarly going from larger to smaller values of $\lambda$. We have empirically observed that the warm restart strategy could be extremely efficient in practice, and would deserve further study in a future work.
#### Heuristics for selecting step sizes
Choosing proximal gradient surrogates $g^t$ requires choosing some Lipschitz constant $L$ (or a strong convexity parameter $\mu$ for Proposition \[prop:conv17\]), which leads to a specific step size in (\[eq:MISO\_composite\]). However, finding an appropriate step size can be difficult in practice for several reasons: (i) in some cases, these parameters are unknown; (ii) even though a global Lipschitz constant might be available, a local Lipschitz constant could be more effective; (iii) the convergence rates of Proposition \[prop:conv16\] can be obtained by choosing a smaller value for $L$ than the “true” Lipschitz constant, as long as the inequality $\E[f(\theta_n)] \leq \E[\barg_n(\theta_n)]$ is always satisfied, where $\barg_n \defin (1/T)\sum_{t=1}^T g_n^t$. This motivates the following heuristics:
- first perform one pass over $\eta\!=\!5\%$ of the data to select a constant $L_1 = 2^{-k}L_0$ with $k$ chosen among positive integers, yielding the smallest objective on the data subset, where $L_0$ is an upper bound of the true Lipschitz constant.
- proceed as in MISO1, but choose a more aggressive strategy $L_2=L_1 \eta$; during the optimization, compute the quantities $a_n^t$ and $b_n^t$ defined as $a_n^t = a_{n-1}^t$, $b_n^t = b_{n-1}^t$ if $t \neq \hat{t}_n$, and otherwise $a_n^{{\hat t}_n} = f^{{\hat t}_n}(\theta_{n-1})$, $b_n^{{\hat
t}_n} = g_{L_2}^{{\hat t}_n}(\theta_{n-1})$, where we have parameterized the surrogates $g^t$ by $L_2$. Every $T$ iterations, compare the sums $A_n = \sum_{t=1}^T a_n^t$ and $B_n = \sum_{t=1}^T b_n^t$. If $A_n \leq B_n$, do nothing; otherwise, increase the value of $L_2$ until this inequality is satisfied.
The heuristic MISO2 is more aggressive than MISO1 since it starts with a smaller value for $L$. After every iteration, this value is possibly increased such that on average, the surrogates “behave” as majorizing functions. Even though this heuristic does not come with any theoretical guarantee, it was found to perform slightly better than MISO1 for strongly-convex problems.
#### Using a different parameter $L_t$ for every function $f_t$
Even though our analysis was conducted with a global parameter $L$ for simplicity, it is easy to extend the analysis when the parameter $L$ is adjusted individually for every surrogate. This is useful when the functions $f_t$ are heterogeneous.
#### Parallelization with mini-batches
The complexity of MISO is often dominated by the cost of updating the surrogates $g_n^{{\hat t}_n}$, which typically requires computing the gradient of a function. A simple extension is to update several surrogates at the same time, when parallel computing facilities are available.
Experimental validation {#sec:exp}
=======================
In this section, we evaluate MISO on large-scale machine learning problems. Our implementation is coded in C++ interfaced with Matlab and is freely available in the open-source software package SPAMS [@mairal7].[^6] All experiments were conducted on a single core of a 2GHz Intel CPU with $64$GB of RAM.
#### Datasets
We use six publicly available datasets, which consist of pairs $(y_t,\x_t)_{t=1}^T$, where the $y_t$’s are labels in $\{-1,+1\}$ and the $\x_t$’s are vectors in $\Real^p$ representing data points. The datasets are described in Table \[table:datasets\]. , , , and are obtained from the 2008 Pascal large-scale learning challenge.[^7] and are obtained from the LIBSVM website.[^8] The datasets are pre-processed as follows: all dense datasets are standardized to have zero-mean and unit variance for every feature. The sparse datasets are normalized such that each $\x_t$ has unit $\ell_2$-norm.
name $T$ $p$ storage density size (GB)
------ ---------------- ---------------- --------- --------- -----------
$581\,012$ $54$ dense 1 $0.23$
$500\,000$ $500$ dense 1 $1.86$
$ 2\,500\,000$ $1\,155$ dense 1 $21.5$
$ 72\,309$ $20\,958$ sparse 0.0024 $0.056$
$781\,265$ $47\,152$ sparse 0.0016 $0.89$
$250\,000$ $16\,091\,143$ sparse 0.0002 $13.90$
: Description of datasets used in our experiments.[]{data-label="table:datasets"}
$\ell_2$-logistic regression {#subsec:explog}
----------------------------
We consider the $\ell_2$-regularized logistic regression problem, which can be formulated as follows: $$\min_{\theta \in \Real^p} \frac{1}{T} \sum_{t=1}^T \ell(y_t ,\x_t^{\top}\theta) + \frac{\lambda}{2}\|\theta\|_2^2, \label{eq:logistic}$$ where $\ell(u,\hat{u}) = \log(1+e^{-u \hat{u}})$ for all $(u,\hat{u})$. Following [@schmidt2], we report some results obtained with different methods with the parameter $\lambda=1/T$, which is argued to be of the same order of magnitude as the smallest value that would be used in practice for machine learning problems. We also performed experiments with the values $\lambda=0.1/T$ and $\lambda=10/T$ to study the impact of the strong convexity parameter; the output of these two additional experiments is not reported in the present paper for space limitation reasons, but it will be discussed and taken into account in our conclusions. The algorithms included in the comparison are:
- the stochastic gradient descent algorithm with a heuristic for choosing the step-size similar to MISO1, and inspired by Leon Bottou’s sgd toolbox for machine learning.[^9] A step-size of the form $\rho/\sqrt{n+n_0}$ is automatically adjusted when performing one pass on $\eta=5\%$ of the training data. We obtain consistent results with the performance of SGD reported by Schmidt et al. [@schmidt2] when the step-size is chosen from hindsight. Based on their findings, we do not include in our figures other variants of SGD, [*e.g.*]{}, [@duchi4; @ghadimi2; @hazan2; @xiao].
- the accelerated gradient method proposed by Beck and Teboulle [@beck] with a line-search for automatically adjusting the Lipschitz constant.
- the algorithm of Shalev-Schwartz and Zhang [@shalev2], efficiently implemented in the language C by Mark Schmidt.[^10]
- a fast implementation in C also provided by Mark Schmidt [@schmidt2]. We use the step-size $1/L$ since it performed similar to their heuristic line search.
- the majorization-minimization algorithm MISO, using the trivial upper bound $L^t=0.25\|\x_t\|_2^2$ on the Lipschitz constant for example $t$.
- the majorization-minimization heuristic MISO1 described in Section \[subsec:heuristics\].
- the heuristic MISO2, also described in Section \[subsec:heuristics\].
- the update rule corresponding to Proposition \[prop:conv17\].
For sparse datasets, MISO0, MISO1, and MISO2 are not practical since they suffer from a $O(Tp)$ memory cost. Their update rules can indeed be rewritten $$\theta_n \leftarrow \theta_{n-1} - \frac{1}{T}\left( \left(\theta_{n-1} - \frac{1}{L}\nabla f^{{\hat t}_n}(\theta_{n-1})\right) - \left( \kappa_{n-1}^{{\hat t}_n} - \frac{1}{L}\nabla f^{{\hat t}_n}(\kappa_{n-1}^{{\hat t}_n}) \right)\right),$$ where $f^t: \theta \mapsto \ell(y_t ,\x_t^{\top}\theta) +
\frac{\lambda}{2}\|\theta\|_2^2$. Thus, for every example $t$, the algorithm requires storing the dense vector $\kappa_{n-1}^{t} - ({1}/{L})\nabla
f^{t}(\kappa_{n-1}^{{t}})$. Therefore, we use mini-batches of size $\lfloor1/d\rfloor$, where $d$ is the density of the dataset; the resulting algorithms, which we denote by MISO0-mb, MISO1-mb, and MISO2-mb, have a storage cost equal to $O(d pT)$, which is the same as the dataset.
On the other hand, the update rule MISO$\mu$ applied to the $\lambda$-strongly convex functions $f^t$ admits a simpler and computationally cheaper form. Since $\kappa_{n-1}^{t} - ({1}/{\lambda})\nabla f^{t}(\kappa_{n-1}^{{t}}) = -(1/\lambda)\ell'(y_{{t}} ,\x_{{t}}^{\top}\kappa_{n-1}^{{t}}) \x_{{t}}$, the update becomes $$\theta_n \leftarrow \theta_{n-1} - \frac{1}{T\lambda}\left( \ell'(y_{\hat{t}_n} ,\x_{\hat{t}_n}^{\top}\theta_{n-1}) - \ell'(y_{\hat{t}_n} ,\x_{\hat{t}_n}^{\top}\kappa_{n-1}^{{\hat t}_n}) \right) \x_{\hat{t}_n}, \label{eq:MISO4b}$$ where $\ell'$ denotes the derivative of $\ell$ with respect to its second argument. Assuming that the dataset fits into memory, the only extra quantities to store are the scalars $\ell'(y_{\hat{t}_n}
,\x_{\hat{t}_n}^{\top}\kappa_{n-1}^{{\hat t}_n})$, and the resulting memory cost is simply $O(T)$.
We present our comparison of the above methods with $\lambda=1/T$ on Figures \[fig:l2epochs\] and \[fig:l2time\], where we plot the relative duality gap defined as $(f(\theta_n)-g^\star)/g^\star$, where $g^\star$ is the best value of the Fenchel dual that we have obtained during our experiments. The conclusions of our empirical study are the following:
- [*SAG, SDCA and MISO$\mu$:*]{} these methods perform similarly and were consistently the fastest, except in the regime $T < 2L/\mu$ where MISO$\mu$ can diverge;
- [*the four variants of MISO:*]{} as predicted by its theoretical convergence rate, MISO0 does not perform better than ISTA [@beck] without line-search (not reported in the figures). MISO1 and MISO2 perform significantly better. MISO$\mu$ is always better or as good as MISO1 and MISO2, except for sparse datasets with $\lambda=0.1/T$ where the condition $T \geq
2L/\mu$ is not satisfied;
- [*influence of mini-batch:*]{} whereas MISO2 performs equally well as SAG/SDCA for dense datasets, mini-batches for sparse datasets makes it slower;
- [*stochastic gradient descent:*]{} SGD-h performs always well at the beginning of the procedure, but is not competitive compared to incremental approaches after a few passes over the data.
Note that an evaluation of a preliminary version of MISO2 is presented in [@mairal17] for the $\ell_1$-regularized logistic regression problem, where the objective function is not strongly convex. Our experimental findings showed that MISO2 was competitive with state-of-the-art solvers based on active-set and coordinate descent algorithms [@fan2].
![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f1.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f2.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f3.eps "fig:"){width="0.33\linewidth"}\
![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f4.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f5.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the number of passes over the data.[]{data-label="fig:l2epochs"}](095763f6.eps "fig:"){width="0.33\linewidth"}
![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f7.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f8.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f9.eps "fig:"){width="0.33\linewidth"}\
![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f10.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f11.eps "fig:"){width="0.33\linewidth"} ![Relative duality gap obtained for logistic regression with respect to the CPU time.[]{data-label="fig:l2time"}](095763f12.eps "fig:"){width="0.33\linewidth"}
Non-convex sparse estimation
----------------------------
The majorization-minimization principle is appealing for non-convex and non-smooth optimization, where only few algorithms apply. Here, we address a sparse estimation problem presented in Section \[subsec:dc\]: $$\min_{\theta \in \Real^p} \frac{1}{T} \sum_{t=1}^T \frac{1}{2}(y_t -
\x_t^{\top}\theta)^2 + \lambda\sum_{j=1}^p \log(|\theta[j]|+\varepsilon),
\label{eq:expdc}$$ where the scalars $y_t$ and the vectors $\x_t$ are the same as in the previous section, and $\varepsilon$ is set to $0.01$. The model parameter $\lambda$ controls the sparsity of the solution. Even though (\[eq:expdc\]) is non-convex and non-smooth, stationary points can be obtained in various ways. In this section, we consider majorization-minimization approaches where the penalty function $\theta \mapsto \sum_{j=1}^p \log(|\theta[j]|+\varepsilon)$ is upper-bounded as in Eq. (\[eq:upperbounddc\]), whereas the functions $\theta
\mapsto (1/2)(y_t - \x_t^{\top}\theta)^2$ are upper-bounded by the Lipschitz gradient surrogates of Section \[subsubsec:gradient\]. We compare five approaches:
- Algorithm \[alg:generic\_batch\] with the trivial Lipschitz constant $L=(1/T)\sum_{t=1}^T0.25\|\x_t\|_2^2$.
- Algorithm \[alg:generic\_batch\] with the line-search scheme of ISTA [@beck] for adjusting $L$.
- we compare MISO0, MISO1, and MISO2, as in the previous section.
We choose a parameter $\lambda$ for each dataset, such that the solution with the lowest objective function obtained by any of the tested method has approximately a sparsity of $10$ for datasets and , 100 for and , and $1\,000$ for and . The methods are initialized with $\theta_0= (\|\y\|_2/\|\X\X^\top \y\|_2)\X^\top \y$; indeed, the initialization $\theta_0=0$ that was a natural choice in Section \[subsec:explog\] appears to be often a bad stationary point of problem (\[eq:expdc\]) and thus an inappropriate initial point. We report the objective function values for different passes over the data in Figure \[fig:dcepochs\], and the sparsity of the solution in Figure \[fig:dcspars\]. Our conclusions are the following:
- methods with line searches do significantly better than those without, showing that adjusting the constant $L$ is important for these datasets;
- MISO1 does asymptotically better than MM-LS for five of the datasets after $50$ epochs and slightly worse for ; in general, MISO1 seems to converge substantially faster than other approaches, both in terms of objective function and in terms of the support of the solution.
- the performance of MISO2 is mitigated. In one case, it does better than MISO1, but in some others, it converges to the stationary point $\theta\!=\!0$.
![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f13.eps "fig:"){width="0.33\linewidth"} ![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f14.eps "fig:"){width="0.33\linewidth"} ![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f15.eps "fig:"){width="0.33\linewidth"}\
![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f16.eps "fig:"){width="0.33\linewidth"} ![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f17.eps "fig:"){width="0.33\linewidth"} ![Objective function during the sparse estimation experiment.[]{data-label="fig:dcepochs"}](095763f18.eps "fig:"){width="0.33\linewidth"}
![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f19.eps "fig:"){width="0.33\linewidth"} ![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f20.eps "fig:"){width="0.33\linewidth"} ![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f21.eps "fig:"){width="0.33\linewidth"}\
![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f22.eps "fig:"){width="0.33\linewidth"} ![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f23.eps "fig:"){width="0.33\linewidth"} ![Sparsity of the solution during the sparse estimation experiment.[]{data-label="fig:dcspars"}](095763f24.eps "fig:"){width="0.33\linewidth"}
Conclusion {#sec:ccl}
==========
In this paper, we have presented new algorithms based on the majorization-minimization principle for minimizing a large sum of functions. The main asset of our approach is probably its applicability to a large class of non-convex problems, including non-smooth ones, where we obtain convergence and asymptotic stationary point guarantees. For convex problems, we also propose new incremental rules for composite optimization, which are competitive with state-of-the-art solvers in the context of large-scale machine learning problems such as logistic regression.
We note that other majorization-minimization algorithms have recently been analyzed, such as block coordinate variants in [@mairal17; @razaviyayn2] and stochastic ones in [@choromanska; @mairal18; @razaviyayn]. In particular, we have proposed in [@mairal18] a stochastic majorization-minimization algorithm that does not require to store information about past iterates, when the objective function is an expectation. Since the first version of our work was published in [@mairal18], MISO has also been extended by other authors in [@zhong] using the alternating direction method of multipliers framework.
For future work, we are currently investigating extensions of the scheme MISO$\mu$ for strongly convex objective functions. We believe that the algorithm can be modified to remove the large sample condition $T \geq 2L/\mu$, that the convergence proof can be extended to the proximal setting, and that it is possible to use acceleration techniques in the sense of Nesterov [@nesterov4]. Another interesting direction of research would be to study the stability of our result to inexact minimization of surrogate functions following for instance the analysis of [@schmidt] for proximal gradient methods.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author would like to thank Zaid Harchaoui, Francis Bach, Simon Lacoste-Julien, Mark Schmidt, Martin Jaggi, the associate editor, and the anonymous reviewers for their useful comments.
Basic definitions and useful results {#appendix:background}
====================================
The following definitions can be found in classical textbooks, e.g, [@bertsekas; @borwein; @boyd; @nocedal]. For the sake of completeness, we briefly introduce them here.
\[def:derivative\] Let us consider a function $f: \Real^p \to \Real$ and $\theta, \theta'$ be in $\Real^p$. When it exists, the following limit is called the directional derivative of $f$ at $\theta$ in the direction $\theta'-\theta$: $ \nabla f(\theta,\theta'-\theta) \defin \lim_{t \to 0^+}
{(f(\theta+t(\theta'-\theta)) - f(\theta)})/{t}.$ When $f$ is differentiable at $\theta$, directional derivatives exist in every direction, and $\nabla
f(\theta,\theta'-\theta)=\nabla f(\theta)^\top (\theta'-\theta)$.
\[def:stationary\] Let us consider a function $f: \Theta \subseteq \Real^p \to \Real$, where $\Theta$ is a convex set, such that $f$ admits a directional derivative $\nabla f(\theta,\theta'-\theta)$ for all $\theta,\theta'$ in $\Theta$. We say that $\theta$ in $\Theta$ is a stationary point if for all $\theta'$ in $\Theta$, $ \nabla f(\theta, \theta'-\theta) \geq 0$.
\[def:lipschitz\] A function $f: \Real^p \to \Real$ is $L$-Lipschitz continuous for some $L > 0$ when for all $\theta,\theta'$ in $\Real^p$, $ |f(\theta')-f(\theta)| \leq L \|\theta-\theta'\|_2.$
\[def:strong\_convexity\] Let $\Theta$ be a convex set. A function $f: \Theta \subseteq \Real^p \to \Real$ is called $\mu$-strongly convex when there exists a constant $\mu > 0$ such that for all $\theta'$ in $\Theta$, the function $\theta \mapsto f(\theta)-\frac{\mu}{2}\|\theta-\theta'\|_2^2$ is convex.
We now present two lemmas that are useful for characterizing first-order surrogate functions. Their proofs can be found in the appendix of [@mairal17].
\[lemma:convexerror\] Let $f,g: \Real^p \to \Real$ be two functions. Define $h\defin g-f$. Then, if $g$ is $\rho$-strongly convex and $f$ is $L$-smooth, with $\rho \!\geq\! L$, $h$ is $(\rho\!-\!L)$-strongly convex; \[step:convexerror\] if $g$ and $f$ are convex and $L$-smooth, $h$ is also $L$-smooth; if $g$ and $f$ are $\mu$-strongly convex and $L$-smooth, $h$ is $(L-\mu)$-smooth.
\[lemma:danskin\] Let $f: \Real^{p_1} \times \Theta_2 \to \Real$ be a function of two variables where $\Theta_2 \subseteq \Real^{p_2}$ is a convex set. Assume that
- $\theta_1 \mapsto f(\theta_1,\theta_2)$ is differentiable for all $\theta_2$ in $\Theta_2$;
- $\theta_2 \mapsto \nabla_{1} f(\theta_1,\theta_2)$ is $L$-Lipschitz continuous for all $\theta_1$ in $\Real^{p_1}$;
- $\theta_2 \mapsto f(\theta_1,\theta_2)$ is $\mu$-strongly convex for all $\theta_1$ in $\Real^{p_1}$.
Also define $\tilde{f}(\theta_1) \defin
\min_{\theta_2 \in \Theta_2} f(\theta_1,\theta_2)$. Then, $\tilde{f}$ is differentiable and $\nabla \tilde{f} (\theta_1) = \nabla_{1} f(\theta_1,\theta_2^\star)$, where $\theta_2^\star \defin \argmin_{\theta_2 \in \Theta_2} f(\theta_1,\theta_2)$. Moreover, if $\theta_1 \mapsto \nabla_{1} f(\theta_1,\theta_2)$ is $L'$-Lipschitz continuous for all $\theta_1$ in $\Real^{p_1}$, the gradient $\nabla \tilde{f}$ is $(L'+L^2/\mu)$-Lipschitz.
[10]{}
, [*Convergent incremental optimization transfer algorithms: Application to tomography*]{}, IEEE T. Med. Imaging, 25 (2006), pp. 283–296.
, [*Optimization with sparsity-inducing penalties*]{}, Found. Trends Mach. Learn., 4 (2012), pp. 1–106.
, [*[A fast iterative shrinkage-thresholding algorithm for linear inverse problems]{}*]{}, SIAM J. Imaging Sci., 2 (2009), pp. 183–202.
, [*On the convergence of block coordinate descent type methods*]{}, SIAM J. Optimiz., 23 (2013), pp. 2037–2060.
, [*Nonlinear programming*]{}, Athena Scientific Belmont, 1999. 2nd edition.
, [*A convergent incremental gradient method with a constant step size*]{}, SIAM J. Optimiz., 18 (2007), pp. 29–51.
, [*Monotonicity of quadratic-approximation algorithms*]{}, Ann. I. Stat. Math., 40 (1988), pp. 641–663.
, [*Convex analysis and nonlinear optimization: [T]{}heory and examples*]{}, Springer, 2006.
, [*Online algorithms and stochastic approximations*]{}, in Online Learning and Neural Networks, David Saad, ed., 1998.
, [*Convex Optimization*]{}, Cambridge University Press, 2004.
, [*Enhancing sparsity by reweighted $\ell_1$ minimization*]{}, J. Fourier Anal. Appl., 14 (2008), pp. 877–905.
, [*Stochastic bound majorization*]{}, arXiv:1309.5605, (2013).
, [*Logistic regression, [AdaBoost and Bregman]{} distances*]{}, Mach. Learn., 48 (2002), pp. 253–285.
, [*[Proximal splitting methods in signal processing]{}*]{}, in Fixed-Point Algorithms for Inverse Problems in Science and Engineering, Springer, 2010.
, [*Signal recovery by proximal forward-backward splitting*]{}, Multiscale Model. Simul., 4 (2005), pp. 1168–1200.
, [*An iterative thresholding algorithm for linear inverse problems with a sparsity constraint*]{}, Commun. Pur. Appl. Math., 57 (2004), pp. 1413–1457.
, [*[SAGA]{}: A fast incremental gradient method with support for non-strongly convex composite objectives*]{}, in Adv. NIPS, 2014.
, [*Finito: A faster, permutable incremental gradient method for big data problems*]{}, in Proc. ICML, 2014.
, [*Duality and auxiliary functions for [B]{}regman distances*]{}, tech. report, CMU-CS-01-109, 2001.
, [*Maximum likelihood from incomplete data via the [EM]{} algorithm*]{}, J. Roy. Stat. Soc. B, 39 (1977), pp. 1–38.
, [*Adaptive subgradient methods for online learning and stochastic optimization*]{}, J. Mach. Learn. Res., 12 (2011), pp. 2121–2159.
, [*Efficient online and batch learning using forward backward splitting*]{}, J. Mach. Learn. Res., 10 (2009), pp. 2899–2934.
, [*Ordered subsets algorithms for transmission tomography*]{}, Phys. Med. Biol., 44 (1999), pp. 2835–2851.
, [ *[LIBLINEAR]{}: A library for large linear classification*]{}, J. Mach. Learn. Res., 9 (2008), pp. 1871–1874.
, [*Mean shift is a bound optimization*]{}, IEEE T. Pattern Anal., 27 (2005), pp. 471–474.
, [*Recovering sparse signals with non-convex penalties and [DC]{} programming*]{}, IEEE T. Signal Process., 57 (2009), pp. 4686–4698.
, [*Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization i: A generic algorithmic framework*]{}, SIAM J. Optimiz., 22 (2012), pp. 1469–1492.
, [*Fixed-point continuation for $\ell_1$-minimization: Methodology and convergence*]{}, SIAM J. Optimiz., 19 (2008), pp. 1107–1130.
, [*Beyond the regret minimization barrier: an optimal algorithm for stochastic strongly-convex optimization*]{}, in Proc. COLT, 2011.
, [*[DC]{} programming: overview*]{}, J. Optim. Theory App., 103 (1999), pp. 1–43.
, [*Majorization for [CRF]{}s and latent likelihoods*]{}, in Adv. NIPS, 2012.
, [*First order methods for nonsmooth convex large-scale optimization*]{}, in Optimization for Machine Learning, MIT Press, 2011.
, [*Variational bounds for mixed-data factor analysis*]{}, in Adv. NIPS, 2010.
, [*An optimal method for stochastic composite optimization*]{}, Math. Program., 133 (2012), pp. 365–397.
, [*Optimization transfer using surrogate objective functions*]{}, J. Comput. Graph. Stat., 9 (2000), pp. 1–20.
, [*A stochastic gradient method with an exponential convergence rate for finite training sets*]{}, in Adv. NIPS, 2012.
, [*Algorithms for non-negative matrix factorization*]{}, in Adv. NIPS, 2001.
, [*Optimization with first-order surrogate functions*]{}, in Proc. ICML, 2013. The appendix is available as the technical report arXiv:1305.3120.
, [*Stochastic majorization-minimization algorithms for large-scale optimization*]{}, in Adv. NIPS, 2013.
, [*Online learning for matrix factorization and sparse coding*]{}, J. Mach. Learn. Res., 11 (2010), pp. 19–60.
, [*Fonctions convexes duales et points proximaux dans un espace hilbertien*]{}, C. R. Acad. Sci. Paris Sér. A Math., 255 (1962), pp. 2897–2899.
, [*[A view of the EM algorithm that justifies incremental, sparse, and other variants]{}*]{}, Learning in graphical models, 89 (1998).
, [*Robust stochastic approximation approach to stochastic programming*]{}, SIAM J. Optimiz., 19 (2009), pp. 1574–1609.
, [*Introductory lectures on convex optimization*]{}, Kluwer Academic Publishers, 2004.
height 2pt depth -1.6pt width 23pt, [*Gradient methods for minimizing composite objective functions*]{}, Math. Program., 140 (2012), pp. 125–161.
, [*[Numerical optimization]{}*]{}, Springer Verlag, 2006. 2nd edition.
, [*A unified convergence analysis of block successive minimization methods for nonsmooth optimization*]{}, SIAM J. Optimiz., 23 (2013), pp. 1126–1153.
, [*A stochastic successive minimization method for nonsmooth nonconvex optimization*]{}, arXiv:1307.4457v2, (2013).
, [*Convergence rates of inexact proximal-gradient methods for convex optimization*]{}, in Adv. NIPS, 2011.
, [*Minimizing finite sums with the stochastic average gradient*]{}, arXiv:1309.2388, (2013).
, [*Proximal stochastic dual coordinate ascent*]{}, arXiv:1211.2717, (2012).
, [*Simultaneous variable selection*]{}, Technometrics, 47 (2005), pp. 349–363.
, [*Graphical models, exponential families, and variational inference*]{}, Found. Trends Mach. Learn., 1 (2008), pp. 1–305.
, [*Sparse reconstruction by separable approximation*]{}, IEEE T. Signal Process., 57 (2009), pp. 2479–2493.
, [*Dual averaging methods for regularized stochastic learning and online optimization*]{}, J. Mach. Learn. Res., 11 (2010), pp. 2543–2596.
, [*Model selection and estimation in regression with grouped variables.*]{}, J. Roy. Stat. Soc. B, 68 (2006), pp. 49–67.
, [*Fast stochastic alternating direction method of multipliers*]{}, in Proc. ICML, 2014.
[^1]: Inria, LEAR Team, Laboratoire Jean Kuntzmann, CNRS, Univ. Grenoble Alpes. 655, avenue de l’Europe, 38330 Montbonnot, France. ().
[^2]: This work was partially supported by the Gargantua project (program Mastodons - CNRS), the Microsoft Research-Inria joint centre, and Agence Nationale de la Recherche (MACARON project ANR-14-CE23-0003-01 and the LabEx PERSYVAL-Lab ANR-11-LABX-0025). A short version of this work was presented at the International Conference of Machine Learning (ICML) in 2013 [@mairal17].
[^3]: The notation $\nabla_1$ denotes the gradient with respect to $\theta_1$.
[^4]: To simplify the notation, we present a shifted version of the traditional Huber loss, which usually satisfies $H(0)=0$.
[^5]: Note that a similar assumption appears in the first analysis of SAG published in [@leroux] before its refinement in [@schmidt2].
[^6]: <http://spams-devel.gforge.inria.fr/>.
[^7]: <http://largescale.ml.tu-berlin.de>.
[^8]: <http://www.csie.ntu.edu.tw/~cjlin/libsvm/>.
[^9]: available here: <http://leon.bottou.org/projects/sgd>.
[^10]: available here: <http://www.di.ens.fr/~mschmidt/Software/SAG.html>.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Nano resonators in which mechanical vibrations and spin waves can be coupled are an intriguing concept that can be used in quantum information processing to transfer information between different states of excitation. Until now, the fabrication of free standing magnetic nanostructures which host long lived spin wave excitatons and may be suitable as mechanical resonators seemed elusive. We demonstrate the fabrication of free standing monocrystalline yttrium iron garnet (YIG) 3D nanoresonators with nearly ideal magnetic properties. The freestanding 3D structures are obtained using a complex lithography process including room temperature deposition and lift-off of amorphous YIG and subsequent crystallization by annealing. The crystallization nucleates from the substrate and propagates across the structure even around bends over distances of several micrometers to form e.g. monocrystalline resonators as shown by transmission electron microscopy. Spin wave excitations in individual nanostructures are imaged by time resolved scanning Kerr microscopy. The narrow linewidth of the magnetic excitations indicates a Gilbert damping constant of only $\alpha = 2.6 \times 10^{-4}$ rivalling the best values obtained for epitaxial YIG thin film material. The new fabrication process represents a leap forward in magnonics and magnon mechanics as it provides 3D YIG structures of unprecedented quality. At the same time it demonstrates a completely new route towards the fabrication of free standing crystalline nano structures which may be applicable also to other material systems.
Keyword: Magnonics, 3D nano-fabrication, Magnon resonators, Magnon mechanics, Spin cavitronics, YIG nanostructures
author:
- 'F. Heyroth'
- 'C. Hauser'
- 'P. Trempler'
- 'P. Geyer'
- 'F. Syrowatka'
- 'R. Dreyer'
- 'S.G. Ebbinghaus'
- 'G. Woltersdorf'
- 'G. Schmidt'
bibliography:
- 'sampleAPS.bib'
title: Monocrystalline free standing 3D yttrium iron garnet magnon nano resonators
---
Introduction
============
Nanomechanical oscillators are useful tools for quantum information processing. Over the past decade numerous groups have for example demonstrated the conversion of quantum information from the microwave to the optical regime by means of a micromechanical resonator[@Reed2017; @Andrews2015; @Lecocq2016; @Ockeloen2016]. By coupling of electrical excitations in superconducting qubits to mechanical oscillators[@OConnell2010] even readout of quantum information has been demonstrated[@Lahaye2009; @Chu2017]. The necessary interaction was often obtained by electric fields as in capacitive drum resonators. Another suitable mechanism for information transfer, however, can make use of the coupling of magnetic fields to spin wave modes in a magnon resonator. Indeed the coupling of a magnon mode in a macroscopic yttrium iron garnet (YIG) sphere to a single qubit has already been demonstrated in 2015[@Tabuchi2015]. For downscaling and integration, however, smaller YIG structures are needed. Taking these results into account it is a promising perspective to realize a new transfer mechanism by coupling magnons to mechanical oscillations in a nanomechanical resonator via magnetoelastic coupling. Obviously, YIG would be an ideal candidate for these resonators since YIG is the material with the lowest known Gilbert damping[@Kasuya1961] and it exhibits extremely long lifetimes for spin waves (magnons) in the $\mu$s regime. As a single crystalline garnet material with a Young’s modulus of the same order of magnitude as that of silicon carbide it is expected to also provide low losses for mechanical waves (phonons) and may yield nanoresonators with high quality factors. Again in macroscopic YIG spheres in the sub-mm range the coupling of magnons to phonons has already been demonstrated[@Zhang2016]. However, up to now no method was known to shape three-dimensional nanostructures from monocrystalline YIG. Nanopatterning of thin films with reasonable quality has been demonstrated[@Jungfleisch2015; @Li2016; @Collet2016; @Zhu2017; @Collet2017], but no patterning of nano-sized free standing resonators has been put forward. Nevertheless, it would be extremely attractive if micron- or sub-micron sized YIG bridges or cantilevers were available. The mechanical resonance frequencies in such structures may be easily engineered to fall in the range of typical magnon frequencies[@Huang2003]. As a first step in this direction we have realized the fabrication of freely suspended YIG microbridges with very low damping for spin waves. Although the mechanical properties could not yet be investigated in detail, mechanical resonance frequencies calculated for their dimensions using the elastic properties of YIG fall into the range of several hundred MHz and may even reach the GHz regime.
3D nano fabrication
===================
Fabrication techniques for suspended single crystal nanostructures mostly use subtractive processing by removing material from a single crystal (bulk or layer). The most straight forward method uses focused ion beam (FIB) lithography to directly shape the desired structure from bulk or thin film[@Babinec2011]. Although very flexible in terms of possible geometries this technique suffers from the possible damage to the crystal structure by extended beam tails which might be detrimental for the magnetic properties of YIG. Also it requires lateral access for the beam in order to remove the material underneath the suspended structure preventing the creation of multiple structures in close vicinity. Alternatively a crystalline film (resonator material) may be deposited on top of a sacrificial layer. The resonator itself is shaped by lithography and dry etching and only becomes free-standing when the underlying sacrificial layer is removed by highly selective wet chemical etching[@Schwarz2000; @Carr1997]. The resulting geometry, however, has several limitations. It is not truly three dimensional but only a partly suspended two dimensional structure. Also the suspended resonator must be more narrow than the un-suspended pads to which it is attached. Otherwise the pads are under-etched during the removal of the sacrificial layer. Unfortunately no sacrificial layers are known for high quality crystalline YIG films which can only be deposited on garnet surfaces (especially gallium gadolinium garnet, GGG) and no selective wet etchants are available for these materials.
On the other hand nanoscale additive fabrication of polycristalline materials is achieved by electron beam lithography, evaporation, and lift-off. A typical example is the fabrication of metallic air bridges, well known since more than a decade [@Yacoby1995; @Sherwin1994; @Borzenko2004]. The process allows for densely packed structures with high flexibility in terms of geometry. However, it requires low temperature deposition of the material because of the limited thermal stability of electron beam resists. This prevents its use for the patterning of monocrystalline materials such as YIG, which in most cases need to be deposited at elevated temperatures.
A new kind of deposition method for thin film YIG has recently been demonstrated. Amorphous YIG films are deposited at room temperature on GGG using either pulsed laser deposition[@Hauser2016; @Hauser2017] or sputtering[@Chang2014]. In a subsequent annealing step the material adapts to the lattice structure of the substrate resulting in thin single-crystalline YIG films. Surprisingly, the quality of these films in terms of damping surpasses the quality of thin films deposited at high temperature[@Hauser2016; @Hauser2017; @Chang2014]. Because deposition is done at room temperature this deposition method is compatible with electron beam lithography. In this way the fabrication of laterally nanopatterned YIG with reasonably small Gilbert damping constants has been demonstrated recently[@Jungfleisch2015; @Li2016; @Zhu2017]. Theoretically, this process also allows the fabrication of beams and bridges when it is adapted to the patterning process used for metal bridges described above. Nevertheless, the higher kinetic energies of the deposited particles in pulsed laser deposition compared to evaporation may necessitate a specially adapted resist profile to guarantee a successful lift-off. Further on the recrystallization is more challenging. In a thin film, crystallization needs to progress only vertically from the substrate to the film surface (with a typical distance of 100 nm or less). In a bridge structure, however, the crystallization starts at the base of the supporting pillars which are in contact with the substrate and then needs to progress around bends across the entire span of the bridge in order to achieve a monocrystalline structure. Any additional nucleation site for crystallization may disturb the process and introduce an additional grain boundary. As we show in the following, it is possible to realize such a 3D lift-off process for YIG with the crystallization (which indeed starts at the substrate) extending throughout the complete bridge structure even over distances of several micrometers.
Processing
==========
Figs.1a-d schematically show the applied process flow. A thick PMMA layer on a <111>oriented GGG substrate is patterned using electron beam lithography at different electron acceleration voltages for the span (low voltage/LV) and pillars (high voltage/HV) of the bridges, respectively (Fig.\[process\]a). Further details are provided in the methods section. The resulting structure after development of the e-beam resist is shown in Fig.\[process\]b. It exhibits holes down to the substrate for the pillars and a groove for the span of the bridge. At the sides the groove has a slight undercut which later facilitates the lift-off process. Onto the developed structure the amorphous YIG material is deposited by PLD at room temperature (Fig.\[process\]c). Subsequent lift-off and resist removal results in a bridge structure (Fig.\[process\]d) which is finally annealed. Fig.\[SEM\]a shows a scanning electron microscopy (SEM) image of a YIG bridge prior to (a) and after annealing (b). The bridge has a nominal span length of $\rm2\,\mu m$ and a YIG layer thickness of approximately $\rm 110\,nm$. The length of the span does not change during the annealing step within the measurement accuracy of the SEM. For the experiment shown here the pillars are not placed at the end of the bridges. This design yields an overhang at the end to combine the investigation of short cantilevers fixed on one end only with that of bridge structures which are clamped at both ends. The resulting bridges and cantilevers are flat and strain free after the lift-off. Subsequent to annealing the bridge itself remains mostly unchanged, however, the overhang is bent upward (Fig.\[SEM\]b) indicating the presence of strain.
During the crystallization at more than than 800$^\circ$C the lattice can reorder and a structure with very little or no strain is created. During cool-down, however, the difference in thermal expansion coefficient of YIG and GGG can lead to a small deformation. The YIG now exhibits tensile strain. While in a continuous layer on a substrate this strain would lead to a change in lattice constant the bridge can now follow the strain by deformation. By tilting the feet inward, the length of the span can be decreased while the tilting can of the feet can lead to the small upward bend of the overhang. The thermal expansion coefficients for YIG is smaller than that of GGG by $\rm\sim 2\times10^{-6}\,K^{-1}$. By cooling from 800 $^{\circ}$C to room temperature the contraction of the YIG lattice would be approximately 0.1% larger than for GGG. It should be noted that any resulting shortening of the bridge is too small to be measured with the accuracy of our electron microscope.
Fig.\[SEM\]c shows a close up view of an annealed YIG bridge with a span of 750 nm also after annealing. The deposited YIG has a nominal thickness of 110 nm. The edges of this bridge are quite rough and show a lot of residue from the lift-off process. Obviously, these can be detrimental for the quality of mechanical resonances. As we show later, these residues can mostly be avoided or removed.
Structural characterization
===========================
While the SEM images show that the molding of the material is successful, the local crystalline quality can only be assessed by transmission electron microscopy (TEM). Atomic resolution TEM has been performed on different bridges after annealing (details described in the methods section). Fig.\[TEM\]a shows a cross-sectional view of a small bridge with a span of approximately 850nm and a height between span and substrate of 75nm. The sample was prepared using a focused ion beam and cut along a {011} plane perpendicular to the surface. The viewing direction of the TEM is along <011>with a small tilt angle.
The pillars which are in direct contact with the substrate show an epitaxial monocrystalline lattice as also observed for large area deposition by Hauser *et al.* [@Hauser2016]. The transition to the span where the material is thinner shows a number of defects likely due to partially relieved shear strain that can be expected in this location.
The span of the bridge, however, appears monocrystalline and of perfect crystallinity except for a single defect in the center (Fig.\[TEM\]b). This defect is a consequence of the crystallization process as described below. To investigate possible differences in lattice orientation of substrate and bridge FFts of TEM images were taken at different spots of the sample. A comparison of FFTs from the substrate and the bridge shows that except for a minute lattice rotation the lattice parameter and orientation are identical for substrate and bridge. This is expected due to the excellent lattice match between YIG and GGG. (mismatch $\sim 0.06\%$). In addition FFTs from different points of the bridge are superimposed to see whether the lattice orientation varies along the bridge (Fig.\[FFT\]). A color coded overlay of the FFTs on left, right, and center of the span shows that the crystal orientations on both sides are tilted with respect to each other with a tilt angle of about $1\,^\circ$. A similarly small rotation is observed when comparing FFTs from bridge and substrate. From these results we can deduce that crystallization starts simultaneously at both pillars, where the material is strained. Thus the two crystallization fronts may be slightly tilted with respect to each other. When they meet at the center of the span the resulting mismatch can only be compensated for by the formation of the crystal defect such as a small angle grain boundary observed at the center of the bridge. In addition, this mechanism explains the small rotation of left and right hand part of the bridge with respect to each other and with respect to the substrate.
To investigate the influence of the bridge size on crystallinity also cross-sectional TEM images of longer bridges are studied (Fig.\[TEM\]c). Even for a length of $2.8\,\rm\mu$m a similar quality of the span (which is the functional part of the resonator) is obtained.
Spin dynamics
=============
Because of the reduced amount of material it is not possible to measure the saturation magnetization M$_{\rm S}$ of the bridges directly with magnetometry methods. From previous experiments we know that YIG layers fabricated by room temperature deposition and annealing under similar conditions exhibit M$_{\rm S}$ up to 27% below the bulk value of $\mu_0 M_{\rm S}\approx 180\,mT$[@Hansen1974]. We would like to note that the M$_{\rm S}$-value used for the micromagnetic simulations (132mT) is in excellent agreement with these results.
In order to obtain a detailed and accurate measurement of the local dynamic properties we perform time-resolved scanning Kerr microscopy (TR-MOKE) experiments on a 110 nm thick YIG-bridge. Using this method it is possible to image directly the different resonant magnon modes in individual bridge structures. To achieve the necessary high frequency excitation of the YIG structures an impedance matched coplanar wavegude (CPW) is deposited by electron beam lithoghraphy and lift-off processes onto the sample. The CPW is positioned such that an array of bridges is located in the gap between signal line and ground plane (inset of Fig.\[LINEDAMPING\]a). The investigated bridge has a width of 600nm and a span length of $ 3\,\rm\mu m$. The thickness of the deposited YIG film is 110nm and the gap under the span is 100nm. The sample was deposited using the parameters described in the methods section.
The spatially resolved measurements are performed with the external magnetic field oriented along the bridge allowing for the excitation of the backward volume modes (BVM) with k-vectors along the bridge and the Damon Eshbach modes (DEM) with k-vectors at an angle of 90$^\circ$. Fig.\[MOKE\] (top row) shows a number of different modes for increasing magnetic field. The fundamental mode with only one antinode is shown in Fig.\[MOKE\]b. Three standing BVM with nodes distributed along the bridge are shown in Fig.\[MOKE\]c-e, while a DEM mode shows a node extending along the bridge (Fig.\[MOKE\]a). It is clearly visible that the magnons are localized in the span of the bridge and no direct coupling to the pillars or beyond is observed.
We have also modelled the different magnon modes using MuMax3[@Vansteenkiste2014]. Fig.\[MOKE\]$\\$(bottom row) shows the respective simulations, which are in good agreement with our experiments. Like the bridge investigated by TRMOKE the simulated bridge has a width of 600nm and a span length of $ 3\,\rm\mu m$. The thickness of the deposited YIG film is 110nm and the gap under the span is 100nm. The gyromagnetic ratio $\gamma$ obtained in the simulations is 178GHz/T which is close to the value obtained from the MOKE data (171GHz/T, Fig.\[Kittel\]). The saturation magnetization was fitted to match the spin wave patterns resulting in a value of $\rm\mu_0 M_{S}\approx 132\,mT$ which is in good agreement with that of large area films deposited by the same method[@Hauser2016].
In order to obtain a better understanding in terms of the magnetization dynamics in the YIG nano bridges FMR spectra are measured by TRMOKE on a single spot in the center of the bridge for several frequencies. Such a resonance spectrum is shown in Fig.\[LINEDAMPING\]a where the main resonance peak has a line width of approximately 140 $\rm\mu$T at 8 GHz. This value is among the smallest values reported for PLD grown thin film material so far. Only material grown by liquid phase epitaxy exhibits smaller linewidths. From our data we find a Gilbert damping value of the main resonance of $\alpha\approx$(2.6$\pm$0.7)$\times$10$^{-4}$ (Fig.\[LINEDAMPING\]b). Also this value is lower than all values reported for YIG grown by PLD at elevated temperatures. The inhomogeneous line width at zero field is $\mu_0\Delta H_0=75\pm10\rm\mu$T which is lower than anything reported for PLD grown thin film YIG so far. For the given configuration these numbers can also be translated into spin wave life times resulting in 220 ns (3.2 GHz), 160 ns (5.2 GHz), and 120 ns (8.4 GHz).
We also determine the effective saturation magnetization M$\rm_{eff}$ which also contains any anisotropy and the gyromagnetic ratio $\gamma$ the resonance fields of the main FMR line are determined as a function of frequency (Fig. \[Kittel\]) and the data is fitted by the Kittel formula:
$$\label{KE1}
\omega =\mu_{0}\gamma\sqrt{H\rm_{{FMR}}(H\rm_{{FMR}}+M\rm_{eff})}$$
The fit yields $\gamma$=(180.3$\pm$0.6)GHz/T and $\rm\mu_0 M_{eff}$=(0.125$\pm$0.003)T.
In addition to the dynamic properties TR-MIKE also allows to investigate the static switching behavior of individual nanobridges. For this we use the method in an off-resonant fashion around zero field (Fig.\[Switching\]). Here the phase and the magnitude of the rf-susceptibility are used to detect the switching as first demonstrated in [@Woltersdorf2007b]. For the measurement the microwave frequency is set to 1 GHz. The probing light spot is placed at the center of the same bridge. The magnitude of the response depends on the internal magnetic field and is therefore sensitive to the relative alignment of magnetization and applied magnetic field. Hysteretic behavior is found when magnetization and applied magnetic field are antiparallel. From this we determine a coercive field of $\mu_0 H_{\rm C}\approx2\,\rm mT$ for the bridge (lateral dimensions of the span: 600nm$\times$ 3$\mu$m) when the magnetic field is aligned with the long axis of the bridge structure (easy axis). For the magnetic field aligned along the short axis (hard axis) of the bridge we find no hysteretic behavior as expected. This coercive field is considerably larger than for comparable continuous YIG films of the same thickness where we find coercive fields of less than 0.1 mT [@Hauser2016]. The enhanced coercive fields in the YIG nano bridges are expected and a consequence of the shape anisotropy and the increased contribution of the domain wall nucleation energy to the magnetization reversal in nanostructures.
In the FMR spectrum also a second line is visible which partly overlaps with the main peak. Spatially resolved measurements indicate that the two halves of the span which are separated by the central crystalline defect differ in resonance field by approx. 100$\rm\mu$T at 8 GHz. This can be explained by the rotation of the two sides observed in transmission electron microscopy. When the field is applied exactly along one half of the bridge, the small tilt of the other half can shift the resonance field in the order of 100$\rm\mu$T at 8 GHz simply because a very small demagnetizing field is added to the external field. For one degree of tilt this modification can be as large as 0.05% of the resonance field which is sufficient to explain the observed resonance line shift.
In addition the spatial resolution of the TR-MOKE also allows to investigate the variation of the resonance field between different bridges and between different parts of a single bridge (namely span and overhang), respectively. Fig.\[Variation\] shows TR-MOKE images of five different bridges obtained simultaneously and repeated for two different magnetic fields but at the same excitation frequency. For individual bridges the main resonance (only one antinode) appears at fields that vary by almost 0.8mT, respectively.
The resonance in the overhang of a bridge can only be imaged by sweeping the field over a wider range. Fig.\[End\] shows that the overhang also exhibits a localized resonance. The resonance field, however, is offset by approx. 7 mT from the main resonance field of the corresponding span. This shift can be caused by the different strain in the span which is pinned on both sides and the overhang which is pinned only on one side as well as by the different size of the two regions. As the resonance does not extend into the foot of the bridge the k-vector is determined by the length of the area on resonance as we no longer observe a true uniform mode but a standing spin wave with zero nodes.
It should be noted that the fabrication process is not limited to simple bridge geometries but highly flexible and can be extended to more complex structures as shown in the examples of Fig.\[Outlook\] paving the way to a number of applications and experiments. Again, also the magnetic excitations are well defined and can be directly imaged. SEM image and MOKE data in Fig.\[Outlook\]c are obtained from the very same structure. We have also tried to reduce edge and surface roughness which may deteriorate the mechanical resonance properties by using an optimized multi-layer resist and a post-annealing wet-etch step. As a result an improved bridge with smoother edges is shown in Fig.\[Outlook\]d.
Discussion
==========
It is possible to fabricate 3D YIG nanobridges using electron beam lithography, room temperature PLD and lift-off. The structural characterization shows that crystallization during the annealing process progresses throughout the bridge on a length scale of more than one $\mu$m leading to an undisturbed lattice with only very few defects. The span of the bridge typically contains a single crystal defect. To the best of our knowledge, until now this kind of long range crystallization process throughout a 3D nanostructure has not been reported. The damping does not reach the record values of low temperature grown YIG layers but is still in the range of high quality PLD grown YIG films. The minimum line width of 140$\mu$T at 8 GHz for a single bridge is well in the range of high quality thin film material and various resonant magnon modes can be identified in scanning TR-MOKE. Both line width and damping thus rival those obtained for large area thin films deposited at higher temperatures. The mechanical resonances of the YIG bridges have yet to be characterized, nevertheless, an estimate of possible resonance frequencies can be given. According to Yang *et al.*[@Yang2001] the resonance frequency of a so called doubly clamped beam which corresponds to the span of our bridge is approximately
$$\label{KE1}
f_{res}\approx1.03 \frac{t}{L^2}\sqrt{\frac{E}{\rho}}$$
with E the Young’s modulus of YIG ($2\times10^{11}$Pa)[@Clark1961], $\rho$ the density (5.17 g/cm$^2$)[@Clark1961], $t$ the thickness and $L$ the length of the beam. Using these parameters with a thickness of 150 nm and a length of 1 $\mu$m a resonance frequency of 964 MHz is expected while the same beam with a length of 500 nm resonates at 3.86 GHz, which is well in the range of typical magnons as measured in our experiments. For further development of the method the next steps will be to fabricate more complex resonators that can not only host magnons with high quality factors but are also suitable for the characterization of mechanical vibrational modes. In addition statistics need to be obtained by TR-MOKE on the variation and reproducibility of resonance frequencies in nominally identical resonators, which are crucial for applications where the exact behavior needs to be predictable.
Possible applications for the nanoresonators can be found in various areas. Spin cavitronics for example investigates strong coupling of cavity resonator modes to magnon modes in macroscopic magnetic samples (typically YIG). In these experiments current technology uses large volume YIG samples coupled to macroscopic planar superconducting microwave resonators [@Huebl2013] or large cavities [@Tabuchi2014]. Our YIG nano resonators might be deposited over micron sized superconducting coplanar waveguides allowing for more complex experiments. In spin caloritronics YIG bridges may be used to create large and extremely and well defined temperature gradients because the span of the bridge is thermally decoupled from the substrate. It should be noted that a temperature difference of only 1 K over a bridge with a length of 1$\mu$m corresponds to a temperature gradient of 10$^6$ K/m. If coupling between phonons and magnons in the nanoresonators can be established even an application for readout of qubits or conversion of quantum information between the microwave and the optical regime may be possible. Therefore the new technology platform presented here may pave the way for downscaling allowing these schemes to be realized on the micron scale or below and facilitating future integration of qubits.
Methods
=======
Electron beam lithography
-------------------------
The pillars and the span of the bridge are exposed using PMMA as a resist and two different respective acceleration voltages. The span is exposed at 2.8kV while the acceleration voltage for the span is 4.5kV. For both exposures the area dose is 100$\mu$C/cm$^{2}$. The structures are developed for 60 s in isopropanol.
Pulsed laser deposition of YIG
------------------------------
The YIG is deposited in 0.025mbar of oxygen from a home-made target. Laser parameters are 248nm wavelength, fluence of 2.5Jcm$^{-2}$, and a repetition rate of 5Hz. Annealing is performed in an oxygen atmosphere (99.997%) at ambient pressure and 800$^\circ$C for 3hours.
TEM preparation
---------------
TEM samples from bridges are prepared using a focused gallium ion beam ‘FEI VERSA 3D’ dual beam microscope by the classical FIB in-situ lift-out technique as described for instance by Bals *et al.*[@Bals2007]. Due to the electrically isolating substrate this procedure is extended for the preparation of the sample after thermal treatment by depositing a thin conductive carbon layer via ion sputtering before transferring the sample to the FIB. As the first step in the preparation procedure inside the FIB a 200nm thick carbon layer is deposited locally using the electron beam at 5kV from the top through the bridge to fill the space under the bridge with carbon. The hole under the bridge is filled by locally cracking the organometallic complex gas from the platinum Gas Injection System of the FIB with a 5kV electron beam. After lift-out the TEM lamellae are mounted to a grid, thinned down to a thickness below 150nm, and stepwise cleaned on both sides from amorphous material by operating the ion beam of the FIB at 5kV, 2kV and 1kV. HRTEM images from these samples are obtained using a JEOL JEM-4010 TEM operated at 400kV.
TR-MOKE
-------
For the time resolved magneto optic Kerr (TR-MOKE) measurements we use a frequency doubled fs-laser operating at 520 nm to illuminate the sample in a scanning optical microscope with polarization analysis. A detailed description of this method is presented in the work of Farle [et al.]{}[@Farle2013]. In our TR-MOKE measurements the magnetization is excited by continuous wave microwave magnetic field which is phase synchronized to the optical probe pulses, i.e. the sampling is stroboscopic. In order to allow for lock-in amplification of the magneto-optical signal the rf-excitation is modulated[@Woltersdorf2007; @Stigloher2016]. The spatial resolution of the measurements presented in this manuscript is diffraction-limited to about 300 nm.
Micromagnetic simulation
------------------------
The simulations were carried out using MuMax3. The simulated structure is a bridge with a rectangular span of 2700nm x 600nm x 110nm (l x w x h). The pillars are 300nm x 600nm x 110nm. After relaxing magnetization of the structure in the external magnetic field H$_0$ in x-direction a small field step perpendicular to the surface (z) is applied. The following precession in the z/y-plane is recorded and a FFT is performed. In the FFT the main oscillations are identified as peaks in the amplitude. The images are obtained by locally evaluating the amplitude and phase of the precession and transforming them into a color (intensity = amplitude, positive phase: red, negative phase: blue).
Acknowledgment
--------------
This work was supported by the German research foundation (DFG) via collaborative research centers SFB762 (TP B9) and TRR227 (TP B02). We thank the Max-Planck-Institut für Mikrostrukturphysik for making the JEOL JEM-4010 TEM available for our experiments.
![Schematic drawing of the patterning process. (a) A resist (green) is exposed with two different acceleration voltages. A low voltage exposure is used for the span of the bridge (yellow) and a high acceleration voltage (red) exposes the pillars down to the substrate. (b) After development the void in the resist has the shape of the bridge and a slight undercut which later facilitates the lift-off. (c) The YIG is deposited and the shape of the bridge becomes visible. It is important that the YIG on the resist surface is well separated from the bridge itself. (d) After lift-off a free-standing bridge is obtained. []{data-label="process"}](PROCESS.jpg){width="80.00000%"}
![SEM images of two different bridges. (a) and (b) show a larger bridge before and after annealing , respectively. (c) shows a smaller bridge after annealing. The deposited YIG has a nominal thickness of 110 nm.[]{data-label="SEM"}](SEM.jpg){width="60.00000%"}
![Transmission electron micrographs for bridges with a nominal thickness of the span of 110 nm. (a) Shows a bridge with a span of approximately 850 nm length and a height of 75 nm underneath the span. (b) Higher magnification shows single crystalline material with a single defect in the center of the bridge. (c) TEM cross section of a bridge with increased length. Even for a length of 2.8 $\mu$m the bridge is free of defects except for the central defect. Above and below the bridge a carbon film is visible which has been deposited using the electron beam during TEM preparation to protect the surface of the bridge.[]{data-label="TEM"}](TEM.jpg){width="80.00000%"}
![Fast Fourier transforms of different parts of the lattice of a single bridge. (a) shows a TEM image of a bridge with three different square areas color coded in red, green, and blue. (b) shows an FFT of the substrate. (c) For the color coded areas the FFT of the lattice is superimposed using the same color code. A zoom into the superposition (frame) shows that a very small rotation of the lattice has taken place which is in the range of $\sim 1^\circ$. All FFTs are obtained from images with the same orientation and magnification.[]{data-label="FFT"}](FFT.jpg){width="90.00000%"}
![Time-resolved scanning Kerr microscopy (TR-MOKE) images of standing spin-wave modes and simulations. The top row shows TR-MOKE results for the main mode (b), one Damon Eshbach mode (a), and three different backward volume modes (c-e). Measurement parameters are (magnetic field/excitation frequency) 11.96mT/2GHz (a), 21.95mT/2GHz (b), 25.61mT/2GHz (c), 89.72mT/4GHz (d), and 92.52mT/4GHz (e). The modes were imaged at the peak amplitude of the respective resonance. The bottom row shows the corresponding simulation results from simulations at fixed respective magnetic fields (see also methods section). Simulation parameters are 19.4mT/2.32GHz (a), 19.4mT/2.00GHz (b), 19.4mT/1.85GHz (c), 83.8mT/3.73GHz (d), and 83.8mT/3.66GHz (e). The coordinate system on the left hand side shows the orientation of the external magnetic field H$_0$.[]{data-label="MOKE"}](MOKEneu.jpg){width="80.00000%"}
![Resonance frequency plotted as a function of applied magnetic field. The results nicely agree except for small deviations at low magnetic fields. The red circles show the measured data while the blue line is the respective fit using the Kittel formula.[]{data-label="Kittel"}](Kittel.jpg){width="60.00000%"}
![TRMOKE measurement of the static switching behavior. While sweeping the field through the static hysteresis the magnitude of the rf-suceptibility is determined as a function of the applied magnetic field. The hysteretic part of the measurement represents the hysteresis of the static switching of the magnetization.[]{data-label="Switching"}](Switching.jpg){width="60.00000%"}
![(a) FMR spectrum obtained by TR-MOKE at the center of a single bridge, excited at 8 GHz. The red circles show the measured data while the blue line is a fit using three Lorentzian line shapes. The arrows are a guide to the eye showing an upper limit for the full width at half maximum which is 2$\mu_0\Delta$H. The half width at half maximum $\mu_0\Delta$H is mostly referred to in literature as the line width. The measurement which is performed on a single spot with a diameter of approx. 300nm shows two very sharp lines with a small overlap. The line width $\mu_0\Delta$H is smaller than 140$\mu$T. The insert shows a sketch of an array of bridges located between signal line and ground of a CPW. (b) Line width plotted versus frequency. A least mean square fit yields a slope corresponding to a Gilbert damping of (2.6$\pm$0.7)$^{-4}$.[]{data-label="LINEDAMPING"}](Linedamp.jpg){width="80.00000%"}
![Two TR-MOKE images obtained at a frequency of 2 GHz showing five adjacent bridges at two different magnetic fields, respectively. In both images at least one of the bridges shows an intense resonance of the mode with one antinode only. Apparently the resonance field between bridges can vary at least by 0.8mT.[]{data-label="Variation"}](S2.jpg){width="80.00000%"}
![Optical topography image of several bridges (a) and two TRMOKE images of the same area acquired at a frequency of 6 GHz at different respective magnetic fields (b-c). In the topography image we can clearly discern the span of the bridge, the base which is slightly darker and the overhang at the end. (b) Shows the main resonance of the span with one antinode while in (c) a similar mode for the overhang of the same bridge is observed. The resonance fields differ by 7mT.[]{data-label="End"}](S3.jpg){width="40.00000%"}
![SEM images of more complex resonators. The process allows to fabricate various shapes such as open squares (a) or disks and triangles (b). (c) TR-MOKE image of a standing backward volume mode measured on a disk resonator overlayed to an SEM image of the same structure. For these structures the nominal YIG thickness is 210 nm. (d) shows a bridge on which a post-annealing wet-etch was applied. The artifacts at the seam of the structure are strongly reduced.[]{data-label="Outlook"}](Outlook.jpg){width="80.00000%"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study phase diagrams of the Hubbard model on anisotropic triangular lattices, which also represents a model for $\kappa$-type BEDT-TTF compounds. In contrast with mean-field predictions, path-integral renormalization group calculations show a universal presence of nonmagnetic insulator sandwitched by antiferromagnetic insulator and paramagnetic metals. The nonmagnetic phase does not show a simple translational symmetry breakings such as flux phases, implying a genuine Mott insulator. We discuss possible relevance on the nonmagnetic insulating phase found in $\kappa$-(ET)$_2$Cu$_2$(CN)$_3$.'
address: ' Institute for Solid State Physics, University of Tokyo, Kashiwanoha, Kashiwa, Chiba, 277-8581, Japan '
author:
- 'Hidekazu Morita, Shinji Watanabe and Masatoshi Imada'
title: 'Nonmagnetic Insulating States near the Mott Transitions on Lattices with Geometrical Frustration and Implications for $\kappa$-(ET)$_2$Cu$_2$(CN)$_3$'
---
[999]{} P. Fazekas and P.W. Anderson, Phil. Mag. [**30**]{} 423 (1974). M.Imada and T.Kashima, J. Phys. Soc. Jpn. [**69**]{} 2723 (2000); ibid. [**70**]{} 2287 (2001). K. Oshima et al., Phys. Rev. B [**38**]{} 938 (1988). R. H. Mackenzie, Comm. Cond. Mat. Phys.[**18**]{} 309 (1998). H. Kino and H. Fukuyama, J. Phys. Soc. Jpn. [**65**]{} 2158 (1996). H. Kondo and T.Moriya, [**68**]{} 3170 (1999). K. Kuroki and H. Aoki, Phys. Rev. B [**60**]{} 3060 (1999). T. Kashima and M.Imada, J. Phys. Soc. Jpn. [**70**]{} 3052 (2001). D.A. Huse and V. Elser, Phys. Rev. Lett. [**60**]{} 2531 (1988); B. Bernu et al. Phys. Rev. B [**50**]{} 10048 (1994). Z. Weihong R.H. McKenzie and R.P. Singh, Phys. Rev. B [**59**]{} 14367 (1999). N. Read and S. Sachdev, Phys. Rev. Lett.[**62**]{} (1989) 1694; O.P. Sushkov et al., Phys. Rev. B[**63**]{} 104420 (2001). L. Capriotti and S. Sorella, Phys. Rev. Lett.[**84**]{} 3173 (2000). I. Affleck and Marston, Phys. Rev. B [**34**]{} 3774 (1988). For details, S. Watanabe and M. Imada, unpublished. A. Kobayashi et al., Chem. Lett. 459 (1987). T. Komatsu, N. Matsukawa, T. Inoue and G. Saito, J. Phys. Soc. Jpn. [**65**]{} 1340 (1996). A. Fortunelli, and A. Painelli, J. Chem. Phys. [**106**]{} 8051 (1997). Y. Shimizu, M. Maeda, G. Saito, K. Miyagawa and K. Kanoda, private commun. J. Schmalian, Phys. Rev. Lett. [**81**]{} 4232 (1998). K. Kuroki et al., cond-mat/0108506.
|
{
"pile_set_name": "ArXiv"
}
|
\
\
[**in two and higher dimensions**]{} .5in [** **]{}\
\
.2in [*Physics and Astronomy Department, Lehman College, CUNY\
Bronx, NY 10468*]{}\
.1in
--------- --
E-mail:
--------- --
1.5in
**Abstract**
We analyze the entanglement entropy, in real space, for the higher dimensional integer quantum Hall effect on ${\mathbb {CP}}^k$ (any even dimension) for abelian and nonabelian magnetic background fields. In the case of $\nu=1$ we perform a semiclassical calculation which gives the entropy as proportional to the phase-space area. This exhibits a certain universality in the sense that the proportionality constant is the same for any dimension and for any background, abelian or nonabelian. We also point out some distinct features in the profiles of the eigenfunctions of the two-point correlator that underline the difference in the value of entropies between $\nu=1$ and higher Landau levels.
Introduction
============
Entanglement has been used to explore properties of quantum states in a variety of condensed matter systems. Typically a system is divided into two subsystems and the entanglement is calculated in terms of the von Neumann entropy of the reduced density matrix of one of the subsystems. For gapped two-dimensional systems, the leading order contribution to the entanglement entropy is proportional to the perimeter of the boundary separating the two subsystems, in particular $S= c L + \gamma + {\cal{O}}(1/L)$, where $L$ is the length of the boundary, $c$ is a non-universal coefficient and $\gamma$ is a universal quantity called topological entanglement entropy [@kitaev].
Of particular interest among two-dimensional gapped systems are the quantum Hall systems whose entanglement entropy has been widely studied under different partitions. For a real-space partition $\gamma=0$ for fully filled integer Quantum Hall states and nonzero for fractional quantum Hall states [@Sierra]-[@krempa]. The entanglement entropy in the case of integer quantum Hall states is amenable to analytical calculations due to the fact that the many-body ground state is in terms of free fermions. The area-law entropy behavior for the two-dimensional integer QHE was studied in different geometries analytically for $\nu=1$ and numerically up to $\nu=5$ in [@Sierra] and the coefficient $c$ was identified in these cases.
In this paper we extend the calculation of the entanglement entropy in the case of higher dimensional integer quantum Hall effect (any even dimension), in particular quantum Hall effect on ${\mathbb {CP}}^k$ [@KN]-[@Kar1]. For $k=1$ this reduces to the well known case of QHE on $S^2$ where the magnetic field is created by a monopole at the center [@Haldane]. The formulation of QHE on ${\mathbb {CP}}^k$ for $k >1$ displays two interesting features: higher dimensionality and the possibility of introducing both abelian and nonabelian magnetic fields. In the latter case one deals with a many-body system of free fermions with internal degrees of freedom which is amenable to analytical calculations.
The paper is organized as follows. In section 2 we give a brief description of the integer quantum Hall effect on ${\mathbb {CP}}^k$, the corresponding spectrum and the Landau level wavefunctions. In section 3 we focus on the lowest Landau level and derive analytical expressions for the entropy for a) arbitrary $k$ and abelian $U(1)$ magnetic field and b) $k=2$ for $U(1) \times SU(2)$ nonabelian magnetic field with fermions in the triplet representation. The entropy is expressed in terms of the eigenvalues of the two-point correlation function. We perform a semiclassical calculation for the entropy and find that the area law as expressed in terms of a “phase-space" area has the same coefficient $c$ for $\nu=1$ for any dimension and any abelian or nonabelian background. In section 4 we derive analytical expressions for the eigenvalues of the two-point correlator for the first Landau level and the $\nu=2$ quantum Hall system on $S^2$ and comment on how the different profiles account for the different values of $c$ in each case. We end with a short discussion.
Quantum Hall effect on ${\mathbb {CP}}^k$
==========================================
In this section we will give a brief description of the Landau level states and wavefunctions for the quantum Hall effect on ${\mathbb {CP}}^k$, following a group theoretic analysis developed in [@KN]-[@Kar1]. ${\mathbb {CP}}^k$ is a $2k$-dimensional manifold which can be thought of as a coset space, = [SU(k+1) U(k)]{} \[coset\] The Landau wavefunctions can be obtained as functions of $SU(k+1)$ with specific transformation properties under the $U(k)$ subgroup. A basis for such functions is given by the so-called Wigner $\cal{D}$-functions, which are the matrices corresponding to the group elements in the unitary irreducible representations, namely \^[J]{}\_[L,R]{}(g) = J , l\_A gJ, r\_A \[wigner\] where $J$ denotes the representation and $l_A, ~r_A$ stand for two sets of quantum numbers specifying the states within the representation. On an element $g\in SU(k+1)$, we can define left and right $SU(k+1)$ actions by \_A g = T\_A g, \_A g = g T\_A where $T_A$ are the $SU(k+1)$ generators in the representation to which $g$ belongs. The left transformations correspond to magnetic translations. There are $2k$ right generators of $SU(k+1)$ which are not in $U(k)$; these can be separated into $T_{+i}$, $i=1,2 \cdots ,k$, which are of the raising type and $T_{-i}$ which are of the lowering type. These generate translations while $U(k)$ generates rotations at a point. The covariant derivatives on ${\mathbb {CP}}^k$ are given by \_[i]{} = i[[R]{}\_[i]{} r]{} \[covariant\] where $r$ can be thought of as the radius of ${\mathbb {CP}}^k$. This is consistent with the fact that the commutator of covariant derivatives is the magnetic field. The commutators of ${\hat R}_{+i}$ and ${\hat R}_{-i}$ are in the Lie algebra of $U(k)$; in the case of ${\mathbb {CP}}^k$ these correspond to constant magnetic fields. In particular we can specify the background field by specifying the right action of $U(k)$ on the wavefunctions. \_a \^J\_[m; ]{} (g) &=& (T\_a\^)\_ \^J\_[m; ]{} (g) \[right1\]\
[R]{}\_[k\^2 +2k]{} \^J\_[m; ]{} (g) &=& - [n k]{} \^J\_[m; ]{} (g) \[right2\] where the index $m=1,\cdots, {\rm dim}J$ represents the state within the $SU(k+1)$ representation $J$ and therefore counts the degeneracy of the Landau level. The first of these equations shows that the wavefunctions $\Psi^J_{m; \alpha}$ transform, under right rotations, as a representation $\tilde{J}$ of $SU(k)$. $(T^{{\tilde J}}_a)_{\alpha \beta}$ are the representation matrices for the generators of $SU(k)$ in the representation ${\tilde J}$ and $n$ is an integer characterizing the abelian part of the background field. $\alpha ,\beta$ label states within the $SU(k)$ representation ${\tilde J}$ (which is itself contained in the representation $J$ of $SU(k+1)$). The index $\alpha$ in the wavefunctions $\Psi^J_{m; \alpha} (g)$ characterizes the nonabelian charge of the underlying fermion fields. In terms of $\D$-functions, the correctly normalized wavefunctions are given by \^J\_[m; ]{} (g) = g = [D]{}\^J\_[m; ]{}(g) \[normalization\] where $N= {\rm dim} J$ and the following orthogonality theorem has been used d(g) \^[\*J]{}\_[m;]{} (g) \^[J]{}\_[m’;’]{} (g) = [\_[mm’]{}\_[’]{}N]{} \[orthogonality\] $d\mu (g)$ is the Haar measure on $SU(k+1)$ normalized to unity.
In the absence of a confining potential, the Hamiltonian $H$ for the Landau problem is proportional to the covariant Laplacian on ${\mathbb{CP}}^k$, namely H = - [14 m]{} (\_[+i]{} \_[-i]{} + \_[-i]{} \_[+i]{} ) \[hamiltonian\] which apart from additive constants can be reduced to the form $\sum_{i} {\hat R}_{+i} {\hat R}_{-i}$. Thus the lowest Landau level wavefunctions satisfy the holomorphicity condition \_[-i]{} = 0 \[holo\] The conditions (\[right1\]), (\[right2\]) and (\[holo\]) completely fix the representation $J$ and therefore the degeneracy of the lowest Landau level.
First we consider the lowest Landau wavefunctions for the case of an abelian background magnetic field. In that case the state $\ket{J, n}$ corresponds to the singlet representation of $SU(k) \in SU(k+1)$ with a $U(1)$ charge proportional to $n$ as specified in (\[right2\]), namely $R_3=-n/2$. These can be thought of as the coherent states for ${\mathbb {CP}}^k$, written explicitly in terms of complex coordinates, \_[i\_1 i\_2 i\_k]{}&=& \^ [z\_1\^[i\_1]{} z\_2\^[i\_2]{}z\_k\^[i\_k]{}(1+z )\^[n 2]{}]{} , \
s &=& i\_1 +i\_2 + +i\_k , 0i\_i n , 0 s n \[wav\] These wavefunctions form a symmetric, rank $n$ representation $J$ of $SU(k+1)$. The dimension of this representation, which is also the LLL degeneracy, is N=[dim]{} J = [[(n+k)!]{} ]{} \[dim\] The volume element for ${\mathbb{CP}^k}$ is d= [k! \^k]{} [d\^2z\_1 d\^2z\_k (1+ z)\^[k+1]{}]{} \[vN40\] We have chosen the normalization such that the total volume, $\int d\mu$, is $1$.
In the case of a $U(1) \times SU(k)$ nonabelian background, it is convenient to label the irreducible representation of $SU(k+1)_R$ by $(p+l,
q+l')$ corresponding to the tensor [@KN1] \^[a\_1...a\_q \_[1]{}...\_[l’]{}]{}\_[b\_1...b\_p \_[1]{}...\_[l]{}]{} \^[q,l’]{}\_[p,l]{} \[7a\] where $p,q$ indicate $U(1)$ indices and $l,l'$ indicate $SU(k)$ indices, namely $a$’s and $b$’s take the value $(k+1)$ and $\gamma$’s and $\delta$’s take values $1,\cdots,k$.
The right hypercharge corresponding to (\[right2\]) is R\_[k\^2+2k]{} = -k(p-q)+l-l’ = -nk \[7b\] The fact that $n$ has to be integer implies that $(l-l')/k$ is an integer, thus constraining the possible $SU(k)_R$ representations $\tilde{J}$.
Further, as explained in detail in [@KN1], the lowest Landau level states correspond to $q=0,~l=0$. So the LLL states we consider correspond to the tensor ${\cal T}^{l'}_p$, where $p=n-{l' \over k}$ and $l'=jk,~j=1,2,\cdots$.
Entanglement Entropy for $\nu=1$
================================
The entanglement entropy $S$ for the $\nu =1$ lowest Landau level quantum Hall states is given by S = - \_[m=1]{}\^[N]{} \[entropy\] where the index $m$ counts the degeneracy and $\lambda$’s are the eigenvalues of the two-point correlator $C(z, z')$ [@Sierra], C(z, z’) = \_[m=1]{}\^[N]{} \_m\^[\*]{} (z) \_m (z’) where $z,z'$ are restricted to be inside the domain $D$. We choose $D$ to be the spherically symmetric region of ${\mathbb {CP}}^k$ satisfying $z'\cdot \bar{z} \le R^2$. For ${\mathbb {CP}}^1 \sim S^2$, this region is a polar cap bounded by a latitude at $\theta$, with $R = \tan \theta/2$ via stereographic projection.
The diagonalization of $C(z, z')$ gives the result C(z, z’) \^\*\_l (z’) d(z’) & = & \_[m=0]{}\^[N]{} \_m\^[\*]{} (z) \_m (z’) \^\*\_l (z’) d(z’)\
&=& \_[l]{} \_l\^[\*]{} (z) \[diagonal\] where \_l = \_[D]{} |\_[l]{}|\^2 d \[lambda\] The second line in (\[diagonal\]) is due to the fact that the angular integration over the spherically symmetric region $D$ will give $\int_{D} \Psi_m (z') ~\Psi_l^*(z') d \mu (z') = \delta_{lm}~\lambda_l$.
We now proceed to calculate the eigenvalues $\lambda$ and subsequently the entanglement entropy for the case of an abelian and nonabelian magnetic field backgrounds.
${\mathbb {CP}}^k$ and abelian magnetic field background
---------------------------------------------------------
The lowest Landau level wavefunctions for ${\mathbb {CP}}^k$ in the case of an abelian background magnetic field are given in (\[wav\]). The corresponding eigenvalues of the two-point correlator are \_[i\_1 i\_2 i\_k]{} &=& \_D d\^\*\_[i\_1 i\_2 i\_k]{} (r) \_[i\_1 i\_2 i\_k]{} (r)\
&=& [[(n+k)! ]{}]{} \_D [[(|[z]{}\_1 z\_1)\^[i\_1]{} (|[z]{}\_2 z\_2)\^[i\_2]{} (|[z]{}\_k z\_k)\^[i\_k]{}]{} ]{} [d\^2z\_1 d\^2z\_k ]{} \[lambda1\] where $s= i_1 +i_2 + \cdots +i_k$. We perform the angular integration using the parametrization $z_i = x_{2i-1} + i x_{2i}$, where x\_1 &=& (\_1)\
x\_2 &=& (\_1) (\_2)\
\
x\_[2k-1]{} &=& (\_1) (\_2) (\_[2k-2]{}) (\_[2k-1]{})\
x\_[2k]{} &=& (\_1) (\_2) (\_[2k-1]{}) \[para\] and $0 \le \phi_1, \phi_2, \cdots , \phi_{2k-2} \le \pi~~,~~0 \le \phi_{2k-1} \le 2 \pi$. Using the fact that in terms of this parametrization d\^2z\_1 d\^2z\_k = \^[2k-1]{} d ( \_1)\^[2k-2]{} d\_1 (\_2)\^[2k-3]{} d\_2 (\_[2k-2]{}) d\_[2k-2]{} d\_[2k-1]{} \[22\] and \_[0]{}\^ ()\^[2i]{} d = [ (i + ) (i + 1)]{} \[23\] we find, after doing the angular integrations, that \_[i\_1 i\_2 i\_k]{} \_s = [(n+k)! ]{} \_0\^[R\^2]{} [x\^[s+k-1]{} ]{} dx \[24\] For each value of $s= i_1 +i_2 + \cdots +i_k$, the eigenvalue $\lambda_s$ has a degeneracy $d_s = {(s+k-1)! / {s! (k-1)!}}$.
The expression for the entanglement entropy is S= \_[s=0]{}\^[n]{} [(s+k-1)! ]{} \[25\]
We will now evaluate the entanglement entropy using a semiclassical approximation and relate this to the area of the region $D$. This is possible when the $U(1)$ charge $n$, which controls the dimensionality of the lowest Landau Hilbert space, becomes very large.
Making a change of variables to $t = x/(1+ x)$, the expression for the eigenvalues $\lambda$’s in (\[24\]) can be written as, \_s &=& [[ (n+k)!]{} ]{} \_0\^[t\_0]{} dt t\^[s+k-1]{} (1-t)\^[n-s]{}\
&=& [[ (n+k)!]{} ]{} B(t\_0; s+k, n-s+1) \[beta\] where $t_0 = R^2/(1+R^2)$ and $B(z; m_1,m_2)$ is the incomplete beta function. For large $n$ this is amenable to a semiclassical calculation as shown in [@VPN]. We will follow that derivation here. Eq. (\[beta\]) can be written as \_s &=& [ (n+k)! ]{} \_0\^[t\_0]{} dt e\^[F(t)]{}\
F(t)&=&(s+k-1) t + (n-s) (1-t) \[49\] The maximum of $F(t)$ occurs at $t^* = s+k-1/(n+k-1)$. Expanding $F(t)$ around $t^*$ we find that $e^{F}$ becomes a Gaussian function centered around $t^*$. In fact, \_[t\^[\*]{} ]{}= - [(n+k-1)\^3 ]{} which implies that the width of the Gaussian is very narrow for all $s$. For small $s$ the center of the Gaussian, $t^* \sim 0$, falls within the range of integration and we find that $\lambda_s \sim 1$. For large $s \sim n$ the center of the Gaussian, $t^* \sim 1 > t_0$, falls outside the range of integration and therefore $\lambda_s \sim 0$. The middle of the transition occurs at $s^*$ such that $t^* = t_0$, namely t\^\*= [[s\^\*+k-1]{} ]{} = t\_0 && s\^\* = t\_0(n+k-1) - (k-1)\
&&n-s\^\* = (n+k-1)(1-t\_0) \[t\*\] Expanding $F(t)$ around $t_0$ in (\[t\*\]) we find F(t) = (n+k -1) - [(n+k-1)2 t\_0 (1-t\_0) ]{} (t-t\_0)\^2 + \[52\] Using this expression we find that for large $n$ \_0\^[t\_0]{} e\^[F(t)]{} \~e\^[F(t\_0)]{} \_0\^[t\_0]{} = e\^[F(t\_0)]{} \[53\] Substituting this in (\[49\]) and using Stirling’s formula $n! = \sqrt{2 \pi n} ~(n/e)^n$, we find that \_[s\_\*]{} \[54\] . The value of $t_0$ is controlled by $R$, which characterizes the size of the spherical domain $D$ and the above calculation shows that $\lambda_s$ is significantly different from 0 or 1 only for $s$ such that the corresponding wavefunctions are localized very near the boundary of the entangling surface.
For large $n$ we can define a variable $y = s/(n+k-1)$, $0 \le y \le 1$, and consider $\lambda$ as a continuous function of $y$. From what we have seen before $\lambda \rightarrow 1$ as $y \rightarrow 0$, $\lambda \rightarrow 0$ as $y \rightarrow 1$ and $\lambda \rightarrow 1/2$ as $y \rightarrow s^* /(n+k-1)=t_0$. In deriving a semiclassical expression for the entanglement entropy we will also need to calculate the derivative of $\lambda$ at the transition region, namely ${d \lambda \over dy}|_{y =t_0}$. For that we have to calculate the difference $\lambda_{s^* +i} - \lambda_{s^*}$. For $s=s^* +i$, the maximum of $F(t)$ occurs at t\_1 = [s\_\* +k -1 n+k-1]{} + [i n+k-1]{} = t\_0 + , \[55\] where $\epsilon = i / (n+k-1) \ll 1$ for small $i$ and large $n$. We now expand $F$ in (\[49\]) around $t_1$, but because the peak has been shifted beyond the upper limit of integration, (\[53\]) will give an extra contribution proportional to $\epsilon$ for small $\epsilon$, \_0\^[t\_0]{} \~ - \[56\] Using Stirling’s formula and taking $\epsilon \rightarrow 0$ we find \_[y=t\_0]{} = lim\_[0]{} [[\_[s\^\* + i]{} - \_[s\^\*]{}]{} ]{} \~- + ( [1 ]{}) \[57\] Figures 1 and 2 show plots of $\lambda_s$ for different values of $k$ and $t_0=R^2/(1+R^2)$.
We found in (\[25\]) that the expression for the entanglement entropy is S &= &\_[s=0]{}\^[n]{} [(s+k-1)! ]{} H\_s\
H\_s & = & -\_s \_s - (1 - \_s) (1-\_s) \[58\] It is clear from the Figures 1 and 2 that $H_s$ is nonzero only for values of $s$ very near the transition region where $\lambda_{s^*} = 1/2$. We can then expand $H(\lambda(y))$ around the value $\lambda=1/2$, H((y)) = H(t\_0) + [1 2]{} [d\^2 H ]{}(y-t\_0)\^2 + \[59\] where |\_[y=t\_0]{} & =& [dH d]{} [ddy]{}|\_[=1/2]{} =0\
[ d\^2 H ]{}|\_[y=t\_0]{} & =& [dH d]{} [d\^2 dy\^2]{} + [d\^2 H d\^2]{}([ddy]{})\^2 |\_[=1/2]{} = - [4(n+k-1) 2t\_0 (1-t\_0)]{} \[60\] Since $H$ has a narrow support around $\lambda = 1/2$ it can be approximated by the Gaussian H(y) & = & H\_0 \
&=& 2 e\^[-[(n+k-1) ]{} (y-t\_0)\^2]{} \[61\] We can rescale to $s=y(n+k-1)$ and $s^* = t_0(n+k-1)-(k-1)$ to obtain the semiclassical Gaussian approximation to $H_s$ as H\_[s,k]{} = 2 \[62\] Figures 3 and 4 show the validity of the Gaussian approximation by comparing (\[62\]) to the exact expression (\[58\]), (\[beta\]).
We can now use (\[62\]) to analytically calculate the entropy in (\[58\]) for large $n$ by converting the sum into an integral over the variable $y$ S & \~& n [s\^[\*k-1]{} (k-1)!]{} 2 \_0\^1 dy\
& \~& n\^[k - ]{} [t\_0\^[k-1]{} (k-1)!]{} ( 2)\^[3/2]{}\
& \~& n\^[k - ]{} [ ( 2)\^[3/2]{} (k-1)!]{} [R\^[2k-1]{} ]{} \[60a\]
The fact that the entropy is proportional to the entangling area $R^{2k-1} \over {(1+R^2)^k}$ has to do with the fact that only wavefunctions localized around the entanglement boundary with corresponding eigenvalues $\lambda \sim 1/2$ contribute to the entropy.
For $k=1$ this agrees with the result found in [@Sierra]. In the case of the QHE on the sphere the entangling surface is a circle of perimeter $L= 2 \pi \sin\theta$, where, based on the stereographic projection $2R /(1+R^2) = \sin\theta$. Scaling the radius of the entangling surface by $\sqrt{n/2}$ (for QHE on $S^2$ the monopole charge, magnetic field and radius of the sphere are related by $n=2B r^2$) we find the area law quoted in [@Sierra] S(k=1) = [[ (2) (2)\^[3/2]{}]{} 4]{} L = 0.204 L \[64\]
The normalized volume element (\[vN40\]), upon angular integration can be written in terms of the radial variable $\rho$ defined in (\[para\]) d= [k! \^k]{} [d\^2z\_1 d\^2z\_k (1+ z)\^[k+1]{}]{} = 2k [ \^[2k-1]{} ]{} [ d ]{} \[61\] where $e_{\rho} = d \rho / (1 + \rho^2) $ is the vierbein along the radial direction $\rho$. This defines the geometric area of the entangling surface (with volume normalized to 1) to be $A_{\rm geom} = 2k { R^{2k-1} \over {(1 + R^2)^k}} $. On the other hand the phase-space volume which is proportional to the degrees of freedom is $V_{\rm phase~space} = { n^k \over k!} \int d \mu$. This then defines a phase-space surface area A\_[phase space]{} = [n\^[k-]{} k!]{} A\_[geom]{} = 2 [n\^[k - ]{} (k-1)!]{} [R\^[2k-1]{} ]{} \[62a\] Scaling the entanglement entropy in (\[60a\]) in terms of this phase-space area we derive a universal expression valid in all dimensions, with a proportionality constant independent of $k$, namely S \~[2]{} ( 2)\^[3/2]{} A\_[phase space]{} \[63\]
${\mathbb {CP}}^2$ and nonabelian magnetic field background
------------------------------------------------------------
The derivation of the entanglement entropy in the case of a nonabelian background magnetic field is more involved. As mentioned in section 2 the LLL states form irreducible representations of $SU(k+1)$ of the form ${\cal T}^{l'}_p$, where $p=n-{l' \over k}$ and $l'=jk,~j=1,2,\cdots$. We will elucidate the calculation of the entanglement entropy for the special case of ${\mathbb {CP}}^2$ with a nonabelian magnetic field $U(1) \times SU(2)$ for the lowest value of $l'$ , namely $l'=k=2$ and $p=n-1$. The derivation for other values of $k$ and $l'$ follows similar ideas. The dimension of this representation and therefore the degeneracy of the corresponding LLL is [@KN1] N = [3 n (n+3) 2]{} \[26\] In identifying the corresponding wavefunctions we consider the states $\bra{ m} \hat{g } \ket{w} $, where the states on the right are of the form ${\cal T}^2_p$ with two up indices and transforming as the $\tilde{J}=1$ triplet representation of $SU(2) \in SU(3)$. (Since the lowest allowed value for $l'$ is 2, based on (\[7b\]) and following comments, the doublet representation is not allowed for ${\mathbb {CP}}^2$.) The corresponding group elements in the appropriate representation can be constructed in terms of products of elements of the $3 \times 3$ matrix $g$ which forms the fundamental representation of $SU(3)$ and its conjugate $g^*$. We need $p$ copies of $g$ and two copies of $g^*$ to match the structure of the ${\cal T}^2_p$ representation. In terms of these matrices, choosing the state $\ket{w}$ as explained above, we get \~g\^[\*i\_1]{}g\^[\*i\_2]{}g\_[j\_1 3]{} g\_[j\_p 3]{} \[general\] where $i,j =1,2,3$ and $\alpha, \beta = 1,2$. Within (\[general\]) there are three distinct series, each one forming an $SU(2)$ multiplet under the left transformations, and all of them together comprising the full $\nu=1$ lowest Landau level $SU(3)$ representation. The three such series are of the form:
\^[(1)]{}\_[()]{} \~g\^[\*3]{} g\^[\*3]{} (g\_[13]{})\^[l]{} (g\_[23]{})\^[m-l]{} (g\_[33]{})\^[n-1-m]{} \[27\] where $l = 0,\cdots,m$ and $m=0,\cdots,n-1$. For each $m$, (\[27\]) form an $SU(2)$ left representation with $j= m/2$. There are $\sum_0^{n-1} (m+1) = n(n+1) /2$ such states.
\^[(2)]{}\_[()]{} \~( g\^[\*]{} g\^[\*3]{} +g\^[\*3]{} g\^[\*]{}) (g\_[13]{})\^[l]{} (g\_[23]{})\^[m-l]{} (g\_[33]{})\^[n-1-m]{} \[28\] For each $m$, (\[28\]) form an $SU(2)$ left representation with $j= (m+1)/2$. There are $\sum_0^{n-1} (m+2)= n(n+3) /2$ such states.
These are of the form \^[(3)]{}\_[()]{} \~( g\^[\*]{} g\^[\*]{} + g\^[\*]{} g\^[\*]{}) (g\_[13]{})\^[l]{} (g\_[23]{})\^[m-l]{} (g\_[33]{})\^[n-1-m]{} \[29\] For each $m$, (\[29\]) form an $SU(2)$ left representation with $j= (m/2)+1$. There are $\sum_0^{n-1} (m+3)= n(n+5) /2$ such states.
Considering all three series together, the total number of states are $N = 3n(n+3) /2$ confirming the result in (\[26\]). We now proceed to normalize the above wavefunctions. In doing so we will use the fact that the elements $g_{i 3}$ can be written in terms of the complex coordinates parametrizing ${\mathbb {CP}}^2$, namely g\_[3]{}& = & [z\_ ]{} , =1,2\
g\_[3 3]{} & = & [1 ]{} \[31a\] where $\bz \cdot z = \bz_1 z_1 + \bz_1 z_1$.
States, with the correct normalization within each series, can be explicitly constructed by starting with the highest weight state and applying the lowering operator $J_{-}$ as follows J\_[-]{} g\_[13]{} = g\_[23]{} &,& J\_[-]{} g\_[23]{} =0\
J\_ [-]{} g\^[\*2i]{} = - g\^[\*1i]{} &,& J\_ [-]{} g\^[\*1i]{} = 0 \[30\]
The highest weight state within this $SU(2)$ multiplet is the state of the form = C\_1 g\^[\*3]{} g\^[\*3]{} (g\_[13]{})\^[m]{} (g\_[33]{})\^[n-1-m]{} \[31\] where $C_1$ is the normalization factor to be determined. The rest of the states are obtained by applying the lowering operator $J_{-}$ whose action is indicated in (\[30\]), namely = C\_1 g\^[\*3]{} g\^[\*3]{} (g\_[13]{})\^[m-l]{} (g\_[23]{})\^[l]{} (g\_[33]{})\^[n-1-m]{} \[32\] where $l=0,1,\cdots, m$. Using (\[31\]) and the fact that $g^{\dagger} g = 1$ we find that \_ \_\^[\*(1)]{}\_\^[(1)]{} = |C\_1|\^2 \^2 [[ (\_1 z\_1)\^[m-l]{} (\_2 z\_2)\^[l]{}]{} ]{} \[32a\]
The ${\mathbb {CP}}^2$ volume element is $d\mu = {2 \over \pi^2} {d^2z_1 d^2z_2
\over (1+ \bz \cdot z)^{3}}$. Using the relation d = 2 [l! m! (n+1-l-m)! (n+3)!]{} \[33\] we find that the correctly normalized wavefunction is of the form \^[(1)]{}\_[(l,m;)]{} = g\^[\*3]{} g\^[\*3]{} (g\_[13]{})\^[m-l]{} (g\_[23]{})\^[l]{} (g\_[33]{})\^[n-1-m]{} \[34\] where $l=0,1,\cdots, m$.
The highest weight state within this $SU(2)$ multiplet is given by = C\_2 ( g\^[\*2]{} g\^[\*3]{} +g\^[\*3]{} g\^[\*2]{}) (g\_[13]{})\^[m]{} (g\_[33]{})\^[n-1-m]{} \[35\] Acting with the lowering operator $J_{-}$ as before we obtain the rest of the states which are of the form & = & C\_2 (g\_[33]{})\^[n-1-m]{} \[36\] where $l=0,1,\cdots, m+1$. Using again the relation (\[33\]) we find the normalized wavefunctions to be of the form \^[(2)]{}\_[(l,m;)]{} &=&\
&& (g\_[33]{})\^[n-1-m]{} \[37\] $l=0,1,\cdots, m+1$.
The highest weight state in series 3 is given by = C\_3 g\^[\*2]{} g\^[\*2]{} (g\_[13]{})\^[m]{} (g\_[33]{})\^[n-1-m]{} \[38\] Every other state in this multiplet is constructed as before by applying $l$ times the lower operator $J_{-}$, producing & = & C\_3 (g\_[33]{})\^[n-1-m]{} \[39\] where $l=0,1,\cdots, m+2$. Using (\[33\]) we find the normalized wavefunctions to be of the form \^[(3)]{}\_[(l,m;)]{} &=&\
&& (g\_[33]{})\^[n-1-m]{} \[40\] $l=0,1,\cdots, m+2$.
The two-point correlator carries nonabelian indices and is defined as C\_[ab]{}(r,r’) = \_[A]{} \^\*\_[A;a]{}(r) \_[A;b]{}(r’) \[cab\] where we have denoted collectively the left indices by $A=(l,m)$ and the right (nonabelian) indices by $a=(\alpha\beta)$. The diagonalization of $C_{ab}(r,r')$ gives \_[b]{} C\_[ab]{}(r,r’) \^\*\_[A;b]{}(r’) d’ = \^\*\_[A;a]{}(r) \[eigen\] where the eigenvalues $\lambda$ are defined = \_a \_D \_[A;a]{}\^[\*]{}(r) \_[A,a]{}(r) d(r) \[41\] In deriving this we used the fact the the wavefunctions (\[34\]), (\[37\]), (\[40\]) are orthogonal to each other.
We find that there are three distinct expressions for $\lambda$’s; one for each of the $SU(2)$ multiplets described above. After performing the angular integration in (\[41\]) using (\[22\])-(\[23\]) we find \_s\^[(1)]{} & = & [(n+3)! ]{} \^[R\^2]{} [ dx (1+x)\^[n+4]{}]{} x\^[s+3]{}\
\_s\^[(2)]{} & = & [(n+3)! ]{} \^[R\^2]{} [ dx (1+x)\^[n+4]{}]{} \[42\]\
\_s\^[(3)]{} & = & [(n+3)! ]{} \^[R\^2]{} [ dx (1+x)\^[n+4]{}]{} with the corresponding degeneracy $s+1$, $s+2$ and $s+3$. As $R^2 \rightarrow \infty$, $\lambda_s^{(I)} \rightarrow 1$ confirming the correct normalization for the wavefunctions.
The expression for the entanglement entropy for the nonabelian lowest Landau level states for ${\mathbb {CP}}^2$ can now be written as S &=& \_[s=0]{}\^[n-1]{}\
&& H\_s\^[(I)]{} = -\_s\^[(I)]{} \_s\^[(I)]{} - (1 - \_s\^[(I)]{}) (1-\_s\^[(I)]{}) \[43\] with the $\lambda^{(I)}$’s given in (\[42\]).
We will now show that $\lambda^{(I)}$’s in (\[42\]) can be related to the abelian ones in (\[beta\]) making a semiclassical calculation of (\[43\]) similar to the abelian case. Making a change of variables to $t = x/(1+x)$ as before we find that (\[42\]) can be written as \_s\^[(1)]{} & = & [(n+3)! ]{} \^[t\_0]{}\_0 dt t\^[s+3]{} (1-t)\^[n-s-1]{}\
\_s\^[(2)]{} & = & [(n+3)! ]{} \^[t\_0]{}\_0 dt\
\_s\^[(3)]{} & = & [(n+3)! ]{} \^[t\_0]{}\_0 dt \[61a\] Comparing these to the abelian $\mathbb{CP}^k$ values (\[beta\]) which we denote by $\lambda^{(\rm Ab)}$ we find the following relations, \^[(1)]{}\_[s, k=2]{} &=& \^[(Ab)]{}\_[s+1, k=3]{}\
\^[(2)]{}\_[s, k=2]{} &=& [ n+3 n+1]{} \^[(Ab)]{}\_[s+1, k=2]{} -[2 n+1]{} \^[(Ab)]{}\_[s+1, k=3]{}\
\^[(3)]{}\_[s, k=2]{} &=& [ n+3 n+1]{} \^[(Ab)]{}\_[s+1, k=1]{} -[2 (n+3) ]{} \^[(Ab)]{}\_[s+1, k=2]{} +[2 ]{} \^[(Ab)]{}\_[s+1, k=3]{} \[62b\] At the large $n$ limit the nonabelian eigenvalues $\lambda^{(I)}$ for $k=2$ coincide with the abelian ones for $k=1,2,3$ correspondingly, \^[(1)]{}\_[s, k=2]{} &=& \^[(Ab)]{}\_[s+1, k=3]{}\
\^[(2)]{}\_[s, k=2]{} & & \^[(Ab)]{}\_[s+1, k=2]{}\
\^[(3)]{}\_[s, k=2]{} & & \^[(Ab)]{}\_[s+1, k=1]{} \[63b\] Similarly, at the large $n$ limit, the nonabelian entropy (\[43\]) becomes a multiple of the abelian one in (\[60a\]) S &=& \_[s=0]{}\^[p]{}\
&& \_[s=0]{}\^[p]{}\
& & 3 n\^[3/2]{} ( 2)\^[3/2]{} [R\^[3]{} ]{} = 3 S\^[(Ab)]{} \[64b\] The overall factor of 3 relating the nonabelian entanglement entropy to the abelian one above has to do with the fact that each lowest Landau state is an $SU(2)$ triplet, ${\rm dim} \tilde{J} = 3$. Although the calculation of the entropy in the case of a nonabelian background was explicitly done for $\mathbb{CP}^2$ and the triplet representation, one expects a more general statement to hold. In the large $n$ limit the degeneracy of the LLL in a case of a nonabelian background is [@KN; @KN1] N \~[dim]{} [n\^k k!]{} \[65\] The corresponding phase-space volume in this case is $V_{\rm phase~space} = {\rm dim}\tilde{J}~{ n^k \over k!} \int d \mu$ and the corresponding phase-space surface area is A\_[phase space]{} = n\^[2k-1 2]{} [ [dim]{} k!]{} A\_[geom]{} = n\^[k - ]{} [ [2 [dim]{}]{} (k-1)!]{} [R\^[2k-1]{} ]{} \[66\] Expressed in terms of the phase-space surface area the overall coefficient in the expression for the entanglement entropy is the same for [*any abelian or nonabelian background*]{} at large $n$ S \~[2]{} ( 2)\^[3/2]{} A\_[phase space]{} \[67\]
Higher Landau levels
====================
In this section we will consider the entropy for higher Landau levels focusing in particular on some of the differences in the behavior of the eigenvalues $\lambda$ between the lowest Landau level $q=0$, the first excited Landau level $q=1$ and the case of $\nu=2$ where both levels are filled. We will only consider the $k=1$ case, QHE on the sphere. Similar features apply for higher $k$.
The wavefunctions for the $q$-th Landau level are of the form \^J\_[m]{} (g) = g \[68\] where $J= {n/2} + q$ and ${\rm dim} J = n+2q+1$. The state $\ket{J, n} $ is not the lowest weight state of the $J$ representation. The lowest weight state is the LLL state with $n \rightarrow n+2q$. The $q=1$ states can therefore be generated by the action of $\hat{R}_{+}$ on the LLL states with $n \rightarrow n+2$. In the case of the sphere the representation of the $\hat{R}_i$ operators is of the form \_[+]{} &= & -\_ u\^\*\_ , \_[-]{} = \_ u\_\
\_[3]{}& =& \_ [1 2]{} \[-u\_ +u\^\*\_ \]\[69\] where u\_ = [1 ]{} (
z\
1\
) \[70\] The $\hat{R}$-operators satisfy the $SU(2)$ algebra = 2 \_3 \[71\] Based on the argument above the correctly normalized wavefunctions of the $q=1$ Landau level are \_s\^[q=1]{} &=& \_[+]{} ( u\_1\^s u\_2\^[n+2-s]{} )\
&=& \[72\] The corresponding eigenvalues of the two-point correlator are now of the form \_s\^[(q=1)]{} = [(n+3)! ]{} \^[R\^2]{} dx [ x\^[s-1]{} (1+x)\^[n+4]{}]{} \[(n+2-s)x-s\]\^2 \[73\] Changing variables to $x={t \over (1-t)}$ as before we can rewrite the eigenvalues as \_s\^[(q=1)]{} = [[ (n+3)! (n+2)]{} ]{} \_0\^[t\_0]{} dt t\^[s-1]{} (1-t)\^[n-s+1]{} \[t-[s ]{}\]\^2 \[74\] The eigenvalue $\lambda_s^{(q=1)}$ as a function of $s$ is similar to $\lambda_s^{(q=0)}$ away from the transition region, but it displays a distinct step-like pattern around the transition $s =t_0~(n+2)$, as shown in Figures 5 and 6. The reason for this has to do with the fact that the wavefunctions (\[72\]) have a node. Since they are generated by the action of $R_{+}$ on the LLL wavefunctions of monopole charge $n+2$ they are necessarily orthogonal to them. Since the LLL wavefunctions are nonzero and have no node, othogonality requires that the first level Landau wavefunctions must have a node. Higher Landau level wavefunctions acquire more nodes and one expects more steps around the transition region for the corresponding eigenvalues $\lambda$. In fact based on the observation that the $q$-th level states can be written, up to normalization, as $\hat{R}^{q}_{+} | LLL, n \rightarrow n+2q > $ one can argue that the wavefunctions will have $q$ nodes and the profile of the corresponding $\lambda$ will display $q$ distinct steps. A similar step-like pattern was observed in [@Dunne] for the higher Landau edge density functions for circular samples.
One can try to repeat the semiclassical analysis we did before for the first Landau level. Eq. (\[74\]) can be written as \_s &=& [[ (n+3)! (n+2)]{} ]{} \_0\^[t\_0]{} dt e\^[F(t)]{} \[t-[s ]{}\]\^2 \[90\]\
F(t)&=&(s-1) t + (n-s+1) (1-t) The maximum of $F(t)$ occurs at $t^* = s-1/n$. Expanding $F(t)$ around $t^*$ we find that $e ^{F}$ becomes a Gaussian function of narrow width centered around $t^*$. In fact, \_[t\^[\*]{} ]{}= - [n\^3 ]{} \[91\] Around the transition region the main contribution of the integral comes from the range of $s$ around $s^*$ such that $t^*=t_0$, namely t\^\*= [[s\^\*-1]{} ]{} = t\_0 s\^\* = t\_0 n +1 , n-s\^\*+1 = n(1-t\_0) \[92\] We now evaluate the integral in (\[90\]) by expanding the integrand around $t_0$. In expanding $(t-s/(n+2))^2$ around $t_0$ we find that the large-$n$ contribution comes from the $(t-t_0)^2$ term. The constant and linear term in $t$ are suppressed by powers of $n$. \_0\^[t\_0]{} e\^[F(t)]{} (t-[s ]{})\^2 & \~&e\^[F(t\_0)]{} \_0\^[t\_0]{} (t-t\_0)\^2\
&=&e\^[F(t\_0)]{} [ ]{} (2 t\_0 (1-t\_0))\^[3/2]{} \[93\] Substituting this in (\[90\]) and using Stirling’s formula $n! = \sqrt{2 \pi n} ~(n/e)^n$, we find that \_[s\_\*]{}\^[(q=1)]{} = [ 1 2]{} \[94\] independent of $t_0$ which is of course what is expected.
Although the semiclassical treatment above is sufficient to capture the value of $\lambda^{(q=1)}$ at the transition point, the evaluation of $H_s^{(q=1)}$ is more involved since it cannot be approximated by a simple Gaussian due to the step like pattern for $\lambda^{(q=1)}$. $H_s^{(q=1)}$ will remain approximately flat in the step-like region, so higher derivatives around $s^*$ will be important to capture the correct behavior around the transition region. Figures 7 and 8 display the plots of $H_s^{(q=1)}$ around $\lambda=1/2$ based on the numerical evaluation of the exact expressions in (\[58\]) and (\[74\]). This clearly shows a deviation from the Gaussian distribution (see also Figure 13).
As a result the entropy for the first Landau level is larger than the entropy of the LLL even though the number of states are approximately the same at large $n$ ($n+1$ states for $q=0$ and $n+3$ states for $q=1$). A numerical evaluation of the entropy shows that it obeys an area law and it gives S\^[(q=1)]{} =1.65 S\^[(q=0)]{} \[82\]
When both $q=0$ and $q=1$ levels are filled, namely $\nu=2$, the situation is more involved as there are overlaps between the wavefunctions of different Landau levels. In particular, \_[s,s’]{} & = & \_0\^[R\^2]{} \^[\*(q=0)]{}\_s (r) \^[(q=1)]{}\_[s’]{} (r) d\
& = & \_[s+1,s’]{} [(n+1)! ]{} \_0\^[R\^2]{} \[83\] The two-point correlator now is C(r,r’) = \_[s=0]{}\^[n]{} \^[\*0]{}\_s (r) \^[0]{}\_s (r’) + \_[s=0]{}\^[n+2]{} \^[\*1]{}\_s (r) \^[1]{}\_s (r’) \[84\] and C(r,r’) (
\_s\^[\*0]{}(r’)\
\_[s+1]{}\^[\*1]{} (r’)\
) d’ = (
\_s\^[0]{}& \_[s,s+1]{}\
\_[s,s+1]{} & \_[s+1]{}\^[1]{}
) (
\_s\^[\*0]{}(r)\
\_[s+1]{}\^[\*1]{} (r)\
) \[85\] where $\lambda^{0}~,~\lambda^{1}$ are the eigenvalues we derived earlier for the lowest and first Landau level and $\delta\lambda$ is the overlap in (\[83\]). There are $2n+4$ eigenvalues for the two-point correlator given by: $\lambda_0^1~, \tilde{\lambda}_{s}^{\pm} ~,~ \lambda^1_{n+2}$, where $s=0,\cdots,n$ and \_s\^ = [[\_s\^0 + \_[s+1]{}\^1 ]{}2]{} \[86\] The interesting feature here is that once both Landau levels are included the step like pattern in the profile of $\lambda^1$ disappears. The profile of the new $\tilde{\lambda}^{\pm}$ resembles that of $\lambda^0$ but shifted with respect to $\lambda^0$, see Figures 9 and 10.
As a result the corresponding entropy per mode $\tilde{H}^{\pm}_s$, where \^\_s = -\^\_s \^\_s - (1 - \^\_s) (1-\^\_s) \[87\] are Gaussian distributions each centered around the value of s for which $\tilde{\lambda}^{\pm} =1/2$ as shown in Figure 11 and 12.
Figure 13 shows a comparison between $H^{(q=0)}~, H^{(q=1)}$ and $H^{(\nu=2)}$ which explains the differences in the values of the corresponding entropies, namely S\^[(=2)]{} > S\^[(q=1)]{} > S\^[(=1)]{} \[88\]
A numerical evaluation of the entropy for the $\nu=2$ case gives S\^[(=2)]{} = 1.76 S\^[(=1)]{} \[89\] This agrees with the result in [@Sierra].
Discussion
==========
In this paper we have analyzed the entanglement entropy for fully filled $\nu=1$ higher dimensional quantum Hall effect on ${\mathbb {CP}}^k$ for abelian and nonabelian magnetic fields. The analytical calculation is based on a semiclassical analysis and we showed that the entropy satisfies the area law. In fact the entropy as expressed in terms of a phase-space entangling surface area has the same proportionality constant for all higher dimensions irrespective of the abelian or nonabelian nature of the background magnetic field. It will be interesting to see if a similar universal formula can be obtained for higher Landau levels.
In the presence of edge degrees of freedom the entanglement entropy for the two-dimensional integer quantum Hall effect develops subleading logarithmic contributions [@other1]. It has been shown in the two-dimensional $\nu=1$ quantum Hall effect that when the edge boundary intersects the boundary of the entangling surface there is an additional logarithmic contribution whose coefficient is determined by the central charge of the gapless edge modes [@estienne]- [@krempa]. In the context of higher dimensional quantum Hall effect we have previously analyzed the analogs of higher dimensional chiral abelian and nonabelian droplets, the edge spectrum and corresponding effective actions [@KN1]. It would be interesting therefore to extend the analysis of the entanglement entropy to these cases where the entangling surface and edge boundary overlap in higher dimensions and calculate the corresponding subleading corrections to the area law for the entanglement entropy.
Similar considerations for higher Landau levels in both two and higher dimensions are also worth pursuing.
[**Acknowledgements**]{}
I thank V.P. Nair for helpful discussions. This research was supported in part by the U.S. National Science Foundation grant PHY-1915053 and by PSC-CUNY awards.
[99]{} A. Kitaev and J. Preskill, [**96**]{}, 110404 (2006); M. Levin and X.G. Wen, [**96**]{}, 110405 (2006).
I.D. Rodriguez and G. Sierra, [**B80**]{}, 15303 (2009); [*J. Stat. Mech.* ]{} [**12**]{}, 12033 (2010).
A. Sterdyniak, A. Chandran, N. Regnault, B. A. Bernevig and P. Bonderson, [**B85**]{}, 125308 (2012).
J. Dubail, N. Read and E.H. Rezayi, [**B85**]{}, 115321 (2012); [**B86**]{}, 245310 (2012).
A. Petrescu, H. F. Song, S. Rachel, Z. Ristivojevic, C. Flindt, N. Laflorencie, I. Klich, N. Regnault and K. Le Hur, [*J. Stat. Mech.*]{} [**10**]{}, 10005 (2014).
H.Li and F. D. Haldane, [**101**]{}, 010504 (2008). B. Estienne and J-M. Stephan, [**B101**]{}, 115136 (2020).
P-G. Rozon, P-A. Bolteau and W. Witczak-Krempa, arXiv:1911.1128.
D. Karabali and V.P. Nair, [ **B641**]{}, 533 (2002).
D. Karabali and V.P. Nair [ **B679**]{}, 427 (2004); [ **B697**]{}, 513 (2004).
D. Karabali, [ **B726**]{}, 407 (2005); [ **B750**]{}, 265 (2006); D. Karabali and V.P. Nair, [**D94**]{}, 024022 (2016).
D.D.M. Haldane, , 605 (1983).
G.V. Dunne, [**B8**]{}, 1625 (1994).
V.P. Nair, arXiv:2001.04957.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Motivated by recent experimental activities on surface critical phenomena, we present a detailed theoretical study of the near-surface behavior of the local order parameter $m(z)$ in Ising-like spin systems. Special attention is paid to the [*crossover*]{} regime between “ordinary” and “normal” transition in the three-dimensional semi-infinite Ising model, where a finite magnetic field $H_1$ is imposed on the surface which itself exhibits a reduced tendency to order spontaneously. As the theoretical foundation, the spatial behavior of $m(z)$ is discussed by means of phenomenological scaling arguments, and a finite-size scaling analysis is performed. Then we present Monte Carlo results for $m(z)$ obtained with the Swendsen-Wang algorithm. In particular the sharp power-law increase of the magnetization, $m(z)\sim H_1\,z^{1-\eta^{ord}_{\perp}}$, predicted for a [*small*]{} $H_1$ by previous work of the authors, is corroborated by the numerical results. The relevance of these findings for experiments on critical adsorption in systems where a small effective surface field occurs is pointed out.'
address: 'Fachbereich Physik, Universität GH Essen, 45117 Essen (F R Germany)'
author:
- Peter Czerner and Uwe Ritschel
title: 'Near-Surface Long-Range Order at the Ordinary Transition: Scaling Analysis and Monte Carlo Results'
---
epsf 23.45cm
Surface critical phenomena, critical adsorption, Ising model, Monte Carlo simulation
Introduction {#intro}
============
A great deal of current experimental activity concentrates on the investigation of critical phenomena near surfaces. After the impressive confirmation of the theoretical predictions [@binder; @diehl; @didi; @diwa] in experiments with binary alloys [@mail], the more recent experimental efforts focus on binary mixtures near their consolute point [@law; @beysens; @franck] and near-critical fluids [@findenegg]. In most of these experiments the order parameter, the concentration difference in fluid mixtures or the density difference between liquid and gaseous phase in fluids, plays a central role. For instance the reflectivity and ellipticity measured in light-scattering experiments are directly related to the order-parameter profile [@reflec; @ellipt]. Hence, quantitative information about the local order parameter near the boundary is necessary for the interpretation of the experimental data.
While a very well-developed theory exists for the individual [*surface universality classes*]{}, corresponding to fixed point of the renormalization-group flow, the picture in the crossover regions between the fixed points is less complete. The experiments are generically not carried at the fixed points, however, and so a detailed understanding of the crossover region is particularly important.
Consider for example the semi-infinite three-dimensional (3-$d$) Ising system with spin-spin interaction $J$. In this model the influence of the surface is usually taken into account by means of a modified exchange interaction $J_1$ in the surface and a magnetic field $H_1$ imposed on the surface spins [@binder; @diehl]. While the former models modifications due to the surface within the critical medium, the latter represents the influence of the adjacent (noncritical) medium, as the container wall for example, on the system.
For a brief (and necessarily incomplete) summary on surface critical phenomena, let us first set $H_1=0$. Then at the bulk critical point $T_c$ the tendency to order near the surface can be reduced, increased, or unchanged compared with the bulk. Which case is realized depends on the ratio $J_1/J$. At a particular value, $J_1^{sp}\simeq 1.5\,J$ [@binder; @labi; @ruge] the third case is realized. This corresponds to the [*surface universality class*]{} of the “special transition”. For $J<J_1^{sp}$ the surface has a reduced tendency to order but nevertheless becomes (passively) ordered at the bulk phase transition. In the opposite case, $J>J_1^{sp}$, the surface orders at a temperature [*above*]{} $T_c$, and [*at*]{} $T_c$ the bulk undergoes a phase transition in the presence of an already ordered surface. From the viewpoint of the renormalization group the special transition is an unstable fixed point [@diehl]. For a start value $J_1 < J_1^{sp}$ the (running) surface coupling is driven to the stable fixed point $J_1=0$ corresponding to the universality class of the “ordinary transition". For $J_1 > J_1^{sp}$ it is driven to $J_1=\infty$, again a stable fixed point, corresponding to the universality class of the “extraordinary transition” [@diehl; @smock].
Next we consider the effects of $H_1$ in a system with $J_1<J_1^{sp}$. This is, for example, the situation generically met in experiments with binary fluids. In particular we are interested in the behavior of the order parameter in this situation. The universality classes are determined by the fixed-point values $H_1=0$ and $H_1=\infty$ of the renormalization-group transformations. For $H_1=0$, at the ordinary transition, the order parameter $m(z)$ simply vanishes since, in terms of Ising spins, the symmetry under the reversal $s_i\to -s_i$ is not broken, neither in the bulk nor in the surface. For $H_1=\infty$ the universality class is called the “normal transition”. The normal transition is known to be equivalent to the extraordinary transition [@bray; @dibu]. In both cases $m(z)$ starts from a large $m_1$ at the surface and then decays to the bulk equilibrium value (being zero for $T\ge T_c$ and nonzero for $T<T_c$). At $T_c$, i.e. for infinite correlation length $\xi$, the decay is described by a universal power law $m\sim z^{-\beta/\nu}$ for macroscopic distances $z$. For instance for the 3-$d$ Ising model $\beta/\nu \simeq 0.52$ [@ferlan]. For $T\neq T_c$ a crossover to the exponential decay $\sim \exp(-z/\xi)$ takes place in a distance $z\simeq \xi$ from the surface.
What happens in the crossover region between $H_1=0$ and $H_1=\infty$? Mean-field theory predicts a profile that starts from some finite $m_1$ and then monotonously decays to the equilibrium value. In Ref.[@czeri] the present authors have shown that, contrary to the naive (mean-field) expectation, fluctuations may cause the order parameter to steeply [*increase*]{} to values $m(z)\gg m_1$ in a surface-near regime. This growth is described by a universal power law $$\label{power}
m(z) \sim H_1\,z^{\kappa}\quad \mbox{with} \quad \kappa=1-n_{\perp}^{ord}\>,$$ where $\eta^{ord}_{\perp}$ is the anomalous dimension pertaining to the ordinary transition (governing the decay of correlations in the direction perpendicular to the surface [@diehl]). For instance for the 3-$d$ Ising model $\kappa\simeq 0.21$.
The scenario for the crossover between ordinary and normal transition developed in [@czeri] is the following: At bulk criticality and $J_1<J_1^{sp}$ for any finite $H_1$ the order-parameter profile increases up to a certain length scale $l^{ord}$ and then crosses over to the power law $\sim z^{-\beta/\nu}$. The scale $l^{ord}$ is given by an [*inverse*]{} power of $H_1$ and, thus, becomes smaller for increasing $H_1$ such that in the limit $H_1\to\infty$ the maximum has moved to the surface. In this limit only the previously mentioned monotonous power-law decay characteristic for the normal transition is left over. A qualitative sketch of typical crossover profiles is shown in Fig.1. In this plot the axes are logarithmic and both $m(z)$ and $z$ are measured in arbitrary units. The individual curves have the correct asymptotics, $m(z)\sim z^{0.21}$ for $z\to 0$ and $m(z)\sim z^{-0.52}$ for $z\to \infty$. However, the (yet unknown) real crossover function is replaced by a simple substitute.\
\#1\#2[0.6\#1]{}\
In Ref.[@twod] it was demonstrated by means of MC simulations and the comparison with exact results that also in the 2-$d$ Ising model the crossover between ordinary and normal transition is qualitatively of the same form as in $d=3$. However, the simple power law (\[power\]) is modified by a logarithm in $d=2$. The main purpose of the present work is to verify the results for the order parameter obtained in [@czeri], where scaling and heuristic arguments were used and a quantitative calculation in the framework of renormalization-group improved perturbation theory was performed, by means of Monte Carlo (MC) simulation explicitly for the three-dimensional system.
The MC studies devoted to (static) critical phenomena near surfaces that are contained in the literature concentrated mainly on the dependence of thermodynamic variables on $J_1/J$ and, in particular, on critical adsorption for large $J_1$ [@extraord] as well as on the precise location of $J_1^{sp}$ [@ruge; @bila1]. $H_1$ was set to zero in these studies. Numerical studies of the influence of $H_1$ concentrated on the surface layer magnetization [@bila2] and, in the context of wetting, on systems [*below*]{} the critical temperature [@wett]. To our knowledge, there is no work in the literature where order-parameter profiles at or above $T_c$ with non-vanishing $H_1$ were studied by MC methods and which could have been directly compared with the analytic results reported in Ref.[@czeri].
The rest of this paper is organized as follows: In Sec.\[two\] we summarize and supplement the main results of [@czeri], especially the phenomenological scaling analysis which allows to make quite precise predictions for $m(z)$ in the crossover regime. In Sec.\[three\] our MC procedure, essentially the Swendsen-Wang algorithm slightly modified to allow for the inclusion of a surface field, is described. The MC results are presented in Sec.\[results\]. Eventually, the last section contains besides a short summary remarks on the relevance of our results for experiments.
Theory {#two}
======
Model
-----
We consider the semi-infinite Ising system with a free boundary on a plane square lattice. The exchange coupling between direct neighbors in the bulk is $J$. In the surface the nearest-neighbor coupling is $J_1$. A surface magnetic field $H_1$ is imposed on the boundary spins and bulk magnetic fields are set to zero such that the Hamiltonian of the model reads $$\label{ising}
{\cal H}_{\rm Ising}= -J\!\sum_{<ij>\in V}\,s_is_j-J_1\!\sum_{<ij>\in\partial V}\,s_is_j-H_1\sum_{i\in\partial V}\,s_i\>,$$ where $\partial V$ and $V$ stand for the boundary and the rest of the system (without the boundary), respectively. Below we mainly work with the dimensionless variables $$\label{Kandh1}
K=J/k_BT\>,\quad K_1=J_1/k_BT\>,\quad\mbox{and}\quad h_1=H_1/k_BT\>.$$ For the (reduced) critical bulk coupling we took $K_c\equiv J/k_BT_c=0.22165$ from the literature [@ferlan]. The value of $K_1$ that corresponds to the special transition is $K_1^{sp}=1.5\,K_c$ [@bila1; @ruge].
Scaling Analysis {#scalan}
----------------
In the critical regime thermodynamic quantities are described by homogeneous functions of the scaling fields. As a consequence, the behavior of the local magnetization under rescaling of distances should be described by $$\label{scal}
m(z,\tau,{h}_1)\sim b^{-x_{\phi}}\,m(zb^{-1},\,\tau b^{1/\nu},\,
{h}_1\,b^{y^{ord}_1})\>,$$ where $x_{\phi}=\beta/\nu$ and $y^{ord}_1=\Delta^{ord}_1/\nu$ are the scaling dimensions of the equilibrium magnetization $m(z\to \infty)$ and the surface field $h_1$, respectively [@diehl]. In general the surface exponents have different values for different surface universality scales [@diehl], and so these quantities are additionally marked by ‘[*ord*]{} ’ for belonging to the ordinary transition. The (MC) literature values for the 3-$d$ Ising model are $x_{\phi}=0.518(7)$ [@ferlan] and $y_1^{ord}=0.73$. The value $y_1^{ord}$ was obtained by employing the scaling relation $y_1+x_1=d-1$ together with the recent Monte Carlo result $x_1^{ord}=\beta_1^{ord}/\nu=1.27$ [@ruge2].
Removing the arbitrary rescaling parameter $b$ in Eq.(\[scal\]) by setting it $\sim z$, one obtains the scaling form of the magnetization $$\label{scalm}
m(z,\tau,h_1)\sim z^{-x_{\phi}}\,{\cal M}(z/\xi, z/l^{ord})\>,$$ where $$\label{length}
l^{ord}\sim h_1^{-1/y_1^{ord}}$$ is the length scale determined by the surface field. The second length scale pertinent to the semi-infinite system and occurring in (\[scalm\]) is the bulk correlation length $\xi=\tau^{-\nu}$. Regarding the interpretation of MC data, which are normally obtained from finite lattices, one has to take into account a third length scale, the characteristic dimension $L$ of the system, and a finite-size scaling analysis has to be performed. The latter will be described in Sec.\[fisi\].
Going back to the semi-infinite case and setting $\tau=0$, the only remaining length scale is $l^{ord}$, and the order-parameter profile can be written in the critical-point scaling form $$\label{h1}
m(z,{h}_1)\sim z^{-x_{\phi}}\,{\cal M}_c(z/l^{ord})\>.$$ As said above, for $z\to \infty$ the magnetization decays as $\sim
z^{-x_{\phi}}$ and, thus, ${\cal M}_c(\zeta)$ should approach a constant for $\zeta\to \infty$. In order to work out the [*short-distance*]{} behavior of the scaling function ${\cal M}_c(\zeta)$, we demand that $m(z)\sim m_1$ as $z\to 0$. This means that in general, in terms of macroscopic quantities, the boundary value of $m(z)$ is [*not*]{} $m_1$. If the $z$-dependence of $m(z)$ is described by a power law, it cannot approach any value different from zero or infinity as $z$ goes to zero. However, the somewhat weaker relation symbolized by “$\sim$” should hold, stating that the respective quantity asymptotically (up to constants) “behaves as" or “is proportional to". This is in accord with and actually motivated by the field-theoretic short-distance expansion [@syma; @diehl], where operators near a boundary are represented in terms of boundary operators multiplied by $c$-number functions.
In the case of the 3-$d$ Ising model the foregoing discussion leads to the conclusion that $m(z)\sim h_1$ because the “ordinary” surface with $K_1<K_1^{sp}$ is paramagnetic and responds linearly to a small magnetic field [@bray]. This is different in $d=2$ where an additional logarithmic factor occurs [@fishau], $m_1\sim h_1\,\mbox{ln}\,h_1$, and this logarithm also leaves its fingerprint on the near-surface behavior of the magnetization [@twod].
The immediate consequence of the simple [*linear*]{} response in $d=3$ for the scaling function ${\cal M}_c(\zeta)$ occuring in (\[h1\]) is that it has to behave as $\sim \zeta^{y^{ord}_1}$ in the small-$\zeta$ limit. After inserting this in (\[h1\]), we obtain that the exponent governing the short-distance behavior of $m(z)$ is given by the difference between $y_1^{ord}$ and $x_{\phi}$, such that for $z\ll l^{ord}$ the magnetization is described by $$\label{power2}
m(z)\sim h_1 \,z^{y^{ord}_1-x_{\phi}}\>.$$ Using the scaling relations $\eta_{\perp}=(\eta+\eta_{\parallel})/2$ and $y_1=(d-\eta_{\parallel})/2$ [@diehl], we eventually obtain $\kappa$ as expressed in (\[power\]).
In the mean-field approximation the result for $\kappa$ is zero. Thus, in this case one really has $m(z\to 0)=m_1$ and the monotonously decaying order-parameter profile mentioned earlier. However, a positive value is obtained when fluctuations are taken into account below the upper critical dimensionality $d^*=4$. Taking $y_1^{ord}\simeq 0.73$ from above, the value for $\kappa$ is 0.21.
Phenomena to some extent analogous to the ones discussed above were reported for the crossover between [*special*]{} and normal transition [@brezin]. Also near the special transition the surface field $h_1$ gives rise to a length scale. However, the respective exponent, the analogy to $\kappa$, is negative, and, thus, one encounters a profile that [*monotonously*]{} decays for all (macroscopic) $z$, with different power laws in the short-distance and the long-distance regime and a crossover at distances comparable to the length scale set by $h_1$. However, [*non-monotonous*]{} behavior in the crossover region as described above for $m(z)$ is a common feature in the case of the energy density in $d=2$ [@mifi] as well as in higher dimensionality [@eisenriegler].
Relation to Critical Dynamics {#critdyn}
-----------------------------
The [*spatial*]{} variation of the magnetization discussed so far strongly resembles the [*time*]{} dependence of the order parameter in relaxational processes at the critical point. If a system with nonconserved order parameter (model A) is quenched from a high-temperature initial state to the critical point, with a small initial magnetization $m^{(i)}$, the order parameter behaves as $m \sim
m^{(i)}\,t^{\theta}$ [@jans], where the short-time exponent $\theta$ is governed by the difference between the scaling dimensions of initial and equilibrium magnetization divided by the dynamic (equilibrium) exponent [@own1]. Like the exponent $\kappa$ in (\[power\]), the exponent $\theta$ vanishes in MF theory, but becomes positive below $d^*$. For example, its value in the 3-$d$ Ising model with Glauber dynamics is $\theta=0.108$ [@theta].
The high-temperature initial state of the relaxational process is to some extent analogous to the surface that strongly disfavors the order and that (for $h_1=0$) belongs to the universality class of the ordinary transition. Further expanding this analogy, heating a system from a low-temperature (ordered) initial state to the critical point would be similar to the situation at the extraordinary transition. Eventually, analogous to the special transition would be a “relaxation process” that starts from an equilibrium state at $T_c$.
Heuristic Argument {#heuristic}
------------------
There is also a heuristic argument for the growth of the magnetization in the near-surface regime expressed in (\[power\]). As said above, a small $h_1$ generates a surface magnetization $m_1\sim h_1$. Regions that are close to the surface will respond to this surface magnetization by ordering as well. How strong this influence is depends on two factors.
First, it is proportional to the correlated area in a plane parallel to the surface at a distance $z$. While correlations in the surface are asymptotically (for $J_1\to 0$) suppressed, for $z> 0$ the range of correlations between spins located in a plane parallel to the surface in a distance $\rho$ from each other can be regarded as finite, because for $\rho> z$ the parallel correlation function is governed by surface exponents and decays much faster than in the bulk. Hence the corresponding effective correlation length, $\xi_{\parallel}(z)$, should behave as $z$. Referring once more to critical dynamics as discussed in the previous section, $\xi_{\parallel}(z)$ is analogous to the time dependent (growing) correlation length $\xi(t)\sim t^{1/\zeta}$ (where $\zeta$ here stands for the dynamic equilibrium exponent).
Second, it depends on the probability that a given spin orientation has “survived” in a distance $z$. For small $h_1$ and $z<l^{ord}$, the latter is governed by the [*perpendicular*]{} correlation function $C(z)\sim z^{-(d-2+\eta_{\parallel}^{ord})}$. Taking into account both factors, we obtain $$\label{heuris}
m(z)\sim h_1\,C(z)\,\xi_{\parallel}^{d-1}=h_1 z^{1-\eta_{\perp}^{ord}},$$ the short-distance power law reported in (\[power\]). Qualitatively speaking, the surface when carrying a small $m_1$ induces a much larger magnetization in the adjacent layers, which are much more susceptible and capable of responding with a magnetization $m\gg m_1$.
This simple picture for the anomalous short-distance behavior holds for dimensions $2<d<4$. At and above the the upper critical dimension $d^*=4$, where the mean-field theory starts to provide the correct description, the power-law growth of magnetization is not observed, since there the increase of the correlated surface area is compensated by the decay of the perpendicular correlation function. In the case of the two-dimensional Ising model the assumption that $m_1\sim h_1$ is no longer valid, and logarithmic terms occur [@twod].
Modifications at $T\neq T_c$ {#temp}
----------------------------
The phenomenological scaling analysis presented above can be straightforwardly extended to the case $\tau>0$. In $d>2$, we may assume that the behavior near the surface for $z<<\xi$ is unchanged compared to (\[power\]), and, thus, the increasing profiles are also expected slightly above the critical temperature. The behavior farther away from the surface depends on the ratio $l^{ord}/\xi$. In the case of $l^{ord}>\xi$ a crossover to an exponential decay will take place for $z\simeq \xi$ and the regime of nonlinear decay does not occur. For $l^{ord}< \xi$ a crossover to the power-law decay $\sim z^{-\beta/\nu}$ takes place and finally at $z\simeq \xi$ the exponential behavior sets in.
An interesting phenomenon can be observed in the case $\xi <l^{ord}$. As discussed above, $m(z)$ then never reaches the regime with power-law decay, but crosses over from the near-surface increase directly to the exponential decay. Since the region where $m(z)$ grows extends up to the distance $\xi$, the magnetization in the maximum has roughly the value $m_{max} \simeq \xi^{\kappa}$. Now, the amplitude of the exponential decay should behave as $\sim m_{max}$ such that for $z\gg \xi$ we have $$\label{expo}
m(z)\sim h_1\,\xi^{\kappa}\,\exp(-z/\xi)\>.$$ In other words, in the case $\xi < l^{ord}$ the short-distance exponent $\kappa$ not only governs the behavior of $m(z)$ near the surface, but also leaves its fingerprint much farther away from the surface in form of an universal dependence of the [*amplitude*]{} of the exponential decay on the correlation length $\sim \xi^{\kappa}$. Nothing comparable occurs when $\xi =\infty$ (compare Sec.\[scalan\] above). When $l^{ord}$ is the only scale, all profiles approach the same curve $m(z)\approx {\cal A}\, z^{-\beta/\nu}$ for $z{\,\vcenter{\hbox{$\buildrel\textstyle>\over\sim$}}}l^{ord}$, with an amplitude ${\cal A}$ [*independent*]{} of $h_1$. An analogous phenomenon, termed “long-time memory” of the initial condition, does also occur in critical dynamics for $T\ge T_c$ [@own2].
Below the critical temperature (and near the ordinary transition), the short-distance behavior of the order parameter is also described by a power law, this time governed by a different exponent, however [@gompper]. The essential point is that below $T_c$ the surface orders spontaneously even for $h_1=0$. Hence, in the scaling analysis the scaling dimension of $h_1$ has to be replaced by the scaling dimension of $m_1$, the conjugate density to $h_1$, given by $x_1^{ord}=\beta^{ord}_1/\nu$ [@diehl]. The exponent that describes the increase of the profile is thus $x_1^{ord}-x_{\phi}$ [@gompper], a number that even in mean-field theory is different from zero ($=1$) and for the $3-d$ Ising model its value is 0.75.
Finite Size Scaling {#fisi}
-------------------
In order to assess the finite size effects to be expected in the MC simulations, we have to take into account the finite-size length scale $L$. The latter is proportional to the linear extension of the lattice (called $N$ below). The generalization of (\[scal\]) reads [@fisi] $$\label{scalfs}
m(z,\tau,{h}_1,L)\sim b^{-x_{\phi}}\,m(zb^{-1},\,\tau b^{1/\nu},\,
{h}_1\,b^{y_1^{ord}}, Lb^{-1})\>,$$ and proceeding as before, we obtain as the generalization of (\[scalm\]) to a system of finite size: $$\label{scalmfs}
m(z,\tau,h_1,L)\sim z^{-x_{\phi}}\,{\cal M}(z/\xi, z/l^{ord},z/L)\>.$$ Thus even at $T_c$ there are two macroscopic length scales, on the one hand $L$ (imposed by the geometry) and on the other hand $l^{ord}$ (the scale set by $h_1$).
It is well known that for large $z{\,\vcenter{\hbox{$\buildrel\textstyle>\over\sim$}}}L$ we have to expect an exponential decay of $m(z)$ on the scale $L$. In the opposite limit, when $z$ is smaller than both $L$ and $l^{ord}$, we expect the short-distance behavior (\[power\]) to occur. That this expectation turns out to be correct is the necessary condition for observing (\[power\]) in MC simulations. Stated in terms of correlation lengths it means that as long as $\xi_{\parallel}$ (cf. discussion in Sec.\[heuristic\]) is smaller than $L$ (and the bulk correlation length $\xi$), the form of the profile is unchanged compared to the one of the semi-infinite system (at bulk criticality). In particular it implies that the surface magnetization $m_1$, whose linear response to $h_1$ was an important ingredient to our scaling analysis of Sec.\[scalan\], should not depend on $L$, as long as $L$ can be regarded as macroscopic.
Farther away from the surface, the form of the profile depends on the ratio between $l^{ord}$ and $L$. In the case of $l^{ord}>L$ a crossover to an exponential decay will take place for $z\simeq L$, and, analogous to the situation with a finite correlation length, also in the finite-size system the amplitude of this exponential decay is governed by the exponent $\kappa$ (compare to (\[expo\]) and the discussion in Sec.\[critdyn\]), such that we have for $z\gg L$ $$\label{expol}
m(z) \sim h_1\,L^{\kappa}\,\exp(-z/L)$$ Again, analogous finite-size effects were reported also in relaxation processes near criticality [@own2]. In the opposite case, $l^{ord}<L$, a crossover to the power-law decay $\sim z^{-x_{\phi}}$ takes place, followed by the crossover to the exponential behavior at $z\simeq L$. Thus, qualitatively, the discussion for systems of finite size is largely analogous to the one in Sec.\[critdyn\] for finite $\xi$.
Monte Carlo Simulation {#three}
======================
System {#geom}
------
The results of the scaling analysis, especially the short-distance law (\[power\]), were checked by MC simulations. To this end we calculated order-parameter profiles for the 3-$d$ Ising model with uniform bulk exchange coupling $K$ and set $K_1=0$, corresponding to the fixed-point value of the ordinary transition.
The geometry of the systems studied was that of a rectangular (cuboidal) lattice with two free surfaces opposite to each other and the other boundaries periodically coupled. The surface field $h_1$ was imposed on both free surfaces. The linear dimension perpendicular to the surfaces was taken to be two times larger than the lateral extension in order to keep corrections due to the second surface, the so-called Fisher-de Gennes effect [@fidege], small [@fidege]. Hence, when we talk about a lattice of size $N$ in this section, we refer to a rectangular system with $N^2\times 2\,N$ spins.
The distance from the surface is still called $z$ in the following, although it is clearly an integer quantity, with $z=0$ corresponding to the location of one of the surfaces. Order parameter profiles were calculated by averaging in individual configurations over planes parallel to the surface and, in turn, we averaged over a large number of configurations generated by the algorithm described in Sec.\[swewa\]. Eventually the symmetry of the system was used and also the average between the left and right half of the lattice was taken.
Procedure {#swewa}
---------
We consider the Ising model, defined by (\[ising\]). The aim of the (equilibrium) Monte Carlo procedure is to generate a representative sample of configurations ${\bf s}$ distributed according to the Boltzmann factor $P({\bf s}) \sim \exp \left[{\cal H}({\bf s})/k_BT\right]$ [@bihe]. Further, it must be guaranteed that, starting from any initial configurations, after a reasonable amount of time such a sample can be extracted. The latter is in principle provided if the algorithm that generates a new configuration ${\bf s}'$ from the old one satisfies [*detailed balance*]{}. In terms of transition probabilities $W({\bf s}\!\to\! {\bf s}')$ this condition can be expressed as $$\label{detbal}
W({\bf s}\!\to \!{\bf s}') \exp\left[-{\cal H}({\bf s})/k_BT\right] = W({\bf s}'\!\to {\bf s}) \exp\left[-{\cal H}({\bf s}')/k_BT\right]\>.$$
For practical purposes, however, not any algorithm satisfying (\[detbal\]) is suitable for MC simulations of critical or near-critical systems. The reason is that physically meaningful algorithms, like Glauber and Kawasaki dynamics [@bihe], are greatly hampered by the [*critical slowing down*]{} upon approaching the equilibrium. One way out of this dilemma is the Swendsen-Wang (SW) algorithm [@swewa], which satisfies (\[detbal\]) but does (probably) not correspond to a physically meaningful dynamics.
The SW algorithm generates a transition (or update) ${\bf s}\to {\bf s}'$ between spin configurations via connected bond clusters. A cluster configuration ${\bf n}$ is constructed from ${\bf s}$ by creating bonds between neighboring spins of [*equal*]{} sign. Then these bonds are “activated” [@janke] with probability $$\label{bondprob}
p=1-e^{-2K}.$$ No bonds are generated between spins of [*opposite*]{} sign. As the next step, bond clusters are defined as connected sets of active bonds. Also isolated spins are identified as a cluster, such that eventually each spin belongs to a cluster.
In order to obtain a new spin configuration ${\bf s}'$ from ${\bf n}$, one assigns to all sites of a given cluster a new spin value with equal probability for each spin direction (independent of the old spin value). The probability for the transition ${\bf s}\to {\bf s}'$ $$\label{p1}
W({\bf s} | {\bf n} | {\bf s}')=p^b(1-p)^mq^{-N_c}$$ where ${\bf n}$ is an intermediate cluster configuration with $N_c$ clusters, and $b$ and $m$ are the numbers of “active” and “inactive” bonds, respectively. This transition corresponds to one Monte Carlo sweep.
In order to verify that the algorithm satisfies detailed balance, we have to consider a transition in the opposite direction. For the transition ${\bf s}'$ from ${\bf s}$ via the [*same*]{} cluster configuration ${\bf n}$, there is a probability $$W({\bf s} |{\bf n} | {\bf s}')=p^b(1-p)^{m'}q^{-N_c},$$ with the same $b$ and $N_c$ as before. However, the number of non-active bonds $m'$ can in general be different, because neighboring clusters can originate from domains with spins of the same or of different sign, in both cases leading to the same cluster configuration.
The total transition probability from ${\bf s}$ to ${\bf s}'$ is given by $$W({\bf s} ,{\bf s}')=\sum_{{\bf n}}W({\bf s}|{\bf n}|{\bf s}'),$$ where the sum runs over all possible intermediate cluster configurations ${\bf n}$. Since the sum $b+m$ is constant for a given spin configuration, it is straightforward to show that $$\frac{W({\bf s} \rightarrow {\bf s}')}{W({\bf s}' \rightarrow {\bf s})}=(1-p)^{m-m'}.$$ Eventually, taking into account that the energy difference between ${\bf s}$ and ${\bf s}'$ is given by $$-2J(m-m')=-\Delta {\cal H}\>,$$ with (\[bondprob\]) one obtains the detailed balance relation (\[detbal\]).
The algorithm presented so far works as long as no magnetic fields are imposed on the spins. To take into account the third term in (\[ising\]) that describes the influence of the surface magnetic field $H_1$, we follow Wang [@wang] and introduce a layer of “ghost” spins next to the surface that couple to the surface spins only. The “ghost” spins all point in the direction of $H_1$ and couple to the “real” spins with coupling strength equal to $H_1$. If one or more “active” bond between a surface and a ghost spin exist, the cluster has to keep its old spin when the system is updated. This prescription preserves detailed balance. In the practical calculation this rule was realized by a modified (reduced) spin-flip probability $$p(n_s)=1-\frac{1}{2} \,\exp(-2\,h_1\, n_s)$$ for clusters pointing in the direction of $h_1$, where $n_s$ is the number of [*surface*]{} spins contained in the cluster. For clusters pointing in opposite direction the probability has to remain unchanged (equal to 1/2).
In order to obtain an equilibrium sample, we discarded several hundred – the precise number depended on the size of the system – configurations after the start of the run. To keep memory consumption low, we used multispin-coding techniques, i.e. groups of 64 spins were coded in one long integer. All calculations were run on a Silicon Graphics computer (Power Challenge) with four Risk 8000 processors. To obtain a profile with reasonable statistics for our largest system (N=256) took about one week of (single-processor) CPU time.
Results
-------
From the magnetization profiles especially the surface magnetization $m_1$ as a function of $h_1$ can be extracted. It is instructive to compare the results for the 3-$d$ Ising model with those for the two-dimensional case obtained in Ref.[@twod]. This is done in Fig.2. The crosses represent the data obtained from a three-dimensional system with $N=256$, the circles stem from the two-dimensional Ising model with lattice size 512$\times$2048. In both cases the statistical errors are smaller than the symbol size.\
\#1\#2[0.55\#1]{}
The situation in three dimensions is obviously quite simple. Up to $h_1\simeq 0.1$ the response of $m_1$ on $h_1$ is just linear. This is the regime where the scaling analysis of Sec.\[scalan\] applies, in particular the basic assumption that $m(z)\sim h_1$ as $z$ goes to zero. The dashed line is a linear fit to the data for $h_1\le 0.1$. For larger values of $h_1$ the surface magnetization saturates, such that for $h_1\to \infty$ the curve has to approach unity
The dependence of $m_1$ on the surface field in two dimensions is more complicated. As discussed in detail in Ref.[@twod] and investigated in many exact calculations [@fishau], there occurs a logarithmic factor in the functional dependence of $m_1$ on $h_1$. The solid line shows the exact result of the semi-infinite system, which, due to the logarithm, never goes through a regime of linear behavior. The MC data (see Ref.[@twod]) deviate from the exact curve for $h_1{\,\vcenter{\hbox{$\buildrel\textstyle<\over\sim$}}}0.02$ and for small $h_1$ indeed show a linear dependence. This observation, linear response for small $h_1$ and an approach to the true semi-infinite behavior for larger $h_1$, is a finite-size effect consistent with exact results [@fishau].
\#1\#2[0.5\#1]{}
The next point are finite-size effects in the vicinity of the surface, especially concerning the dependence of $m_1$ on $N$. As discussed in Sec.\[fisi\], we expect that $m_1$ and the profile up to a certain distance $\sim N$ should not depend on $N$. In Fig. 3 we plotted the data for the local magnetization for four different system sizes ranging from $N=64$ to 256 up to $z=20$. In all cases the surface field was $h_1=0.01$. Quite obviously, $m_1$ itself does not vary with $N$ in the given range of sizes, confirming the finite-size scaling analysis in Sec.\[fisi\] as well as the assumptions underlying the scaling analysis in Sec.\[scalan\]. From the given value of $m_1$ all profiles increase for $z$ increasing away from the surface. For the first few layers the curves lie on top of each other, but in the smaller systems the slopes become smaller compared to larger $N$ at relatively small distances already. For the system with $64^2\times 128$ spins the regime with growing magnetization extends to $z\simeq 7$ only.
\#1\#2[0.5\#1]{}
The variation of the same curves on larger scales is displayed in Fig.4 in double-logarithmic form. Going to larger system size the distance of the maximum $z_{max}$ grows roughly as $\sim N$. For $N=256$ we have $z_{max}\simeq 30$. Recalling the results of the finite-size scaling analysis of Sec.\[fisi\], we conclude that with these parameters the model is in the regime where $L<l^{ord}$, and going to a smaller $h_1$ that would increase $l^{ord}$ would not help to extend the region of growing magnetization. For $z\to 0$ the form of the profiles is consistent with a power law. However, in the small systems the finite size effects cause the profiles to crossover to the exponential decay (compare Sec.\[fisi\]) at a rather small distance. Even in the largest system $N=256$ that could be studied by our present means with reasonable effort, the near-surface power law does not extend beyond $z\simeq 20$. The problem is that we indeed have $L\sim N$, but apparently with a rather small constant of proportionality.
Nevertheless a rough value for the short-distance exponent can be extracted from the profile for $N=256$. The result is $\kappa=0.16(2)$. This is somewhat lower than our expectation. The deviation from the expected value $0.21$ is very likely due to the finite-size effects. We can not claim to see the power law (\[power\]) over a really macroscopic range before the crossover to the finite-size (exponential) behavior sets in. So the determination of a more reliable value of $\kappa$ from the short-distance behavior remains as a task for larger-scale simulations.
\#1\#2[0.5\#1]{}
\#1\#2[0.5\#1]{}
Magnetization profiles for $N=256$ and $h_1$ varying in a wide range between 0.005 and 5 are plotted in Fig.5. The dashed line represents the pure power law $\sim z^{-0.52}$ characteristic for the extraordinary or normal transition. For small $h_1$, up to $h_1\simeq 0.1$, the curves show the near-surface growth consistent with (\[power\]). For $h_1$ up to about 0.02, the location of the maximum (here $z_{max}\simeq 30$) is determined by the finite-size scale $L$. Setting $h_1$ to larger values, the maximum moves closer to the surface. This is the regime where $l^{ord}$ is smaller than $L$ and the location of the maximum is governed by $l^{ord}$. In the case of $h_1{\,\vcenter{\hbox{$\buildrel\textstyle>\over\sim$}}}0.2$, the profiles decay monotonously. Setting $h_1=5.0$, the magnetization at the surface is very close to one, and the decay for $10{\,\vcenter{\hbox{$\buildrel\textstyle<\over\sim$}}}z{\,\vcenter{\hbox{$\buildrel\textstyle<\over\sim$}}}100$ is consistent with the power law $\sim z^{-x_{\phi}}$ The value of the exponent $x_{\phi}=\beta/\nu$ obtained from this curve is $0.51(1)$, which is in excellent agreement with the literature value 0.517 [@ferlan]. The up-bending of the profiles for $z{\,\vcenter{\hbox{$\buildrel\textstyle>\over\sim$}}}100$ is due to the second surface.
Eventually, Fig.6 shows the values of $z_{max}$ as a function of $h_1$ determined from the profiles of Fig.5. From these data, the qualitative picture from above can be made more quantitative. Especially with the result Eq.(\[length\]) we can in principle determine directly the scaling dimension $y_1^{ord}$. A power law fitted to the data for $0.02\le h_1\le 0.1$ yields $1/y_1^{ord}=1.4 (1)$ (where the error was estimated), which, in turn, yields 0.71(4) for the scaling dimension $y_1^{ord}$. As mentioned above the literature value is 0.73. However, as in the case of the short-distance exponent, we also here have to admit that we are not really in a regime where we can call $l^{ord}$ large compared with the lattice spacing, and the good agreement with what we expected from the scaling analysis is actually surprising.
Summary and Concluding Remarks {#summary}
==============================
We studied the near-surface behavior of the order parameter in the three-dimensional Ising system under the influence of a surface magnetic field $H_1$. The anomalous behavior found in [@czeri] by employing scaling arguments and perturbative methods was confirmed in the present work by Monte Carlo simulations. Especially, the short-distance power law (\[power\]) was corroborated.
However, in our Monte Carlo study especially the region with small $H_1$ is severely affected by finite size effects. Even in our largest system ($256^2\times 512$) the increase of $m(z)$ does not extend beyond $z\simeq 30$. In order to obtain reliable results for the exponent $\kappa$ determined with the help of the short-distance power law (\[power\]) and profiles that can be used for the quantitative comparison with experimental data, one has to go to systems beyond the size that we are able to treat by our present means.
Concerning experiments on surface critical phenomena, our results should be of interest especially in those cases where a small $H_1$ occurs, at a surface that disfavors the order. An example where this was obviously realized is the system studied by Desai et al. [@franck]. In their experiment a binary fluid was studied in a container whose walls as a function of time change their preference from one component to the other, the time scale of this change being of the order of days. In other words, the surface field $H_1$ changes sign as time goes by, and during a certain period $H_1$ is small. First substantial steps towards a theoretical explanation of this and other similar experiments on binary mixtures were made already by Ciach et al. [@alina]. A complete theoretical analysis would require a careful derivation of experimentally observably quantities like reflectivity and ellipticity for light scattering experiments on the basis of our results for the order-parameter profile.
Another experiment, discussed already in some detail in [@czeri], is the one by Mailänder et al. [@mail] on Fe$_3$Al. This system undergoes (among other transitions) a second-order phase transition between a phase with $DO_3$ structure and one with $B2$ structure. The near-surface regime was studied by scattering of evanescent x-rays. The exponents observed were consistent with the expectation for the ordinary transition, but Bragg peaks revealed the existence of long-range order near the surface reminiscent to the normal transition. In order to explain the experimental results of [@mail] on the basis of our findings we have to assume that there exits an effective $H_1$ in this system. Then, if the associated length scale $l^{ord}$ is larger than both the scattering depth and the bulk correlation length, the structure function is governed by the anomalous dimension of the ordinary transition. On the other hand, the steep increase of the order parameter should provide the explanation for the observed long-range order near the surface.\
[: We thank A. Ciach for helpful comments and discussions. We are especially indebted to our system administrator R. Oberhage for his professional support concerning computer questions. This work was supported in part by the Deutsche Forschungsgemeinschaft through Sonderforschungsbereich 237.]{}
[99]{} H. W. Diehl and S. Dietrich, Z. Phys. B [**42**]{}, 65 (1981); Erratum B [**43**]{} (1981) 281. S. Dietrich and H. Wagner, Phys. Rev. Lett. [**51**]{} (1983) 1469; Z. Phys. B [**56**]{} (1984) 207. X. Mailänder, H. Dosch, J. Peisl, and R. L. Johnson, Phys. Rev. Lett. [**64**]{} (1990) 2527; see also: H. Dosch, in [*Critical Phenomena at Surfaces and Interfaces, Springer Tracts in Modern Physics*]{}, edited by G. Höhler and E. A. Niekisch, (Springer, Berlin, 1992). B. M. Law, Phys. Rev. Lett. [**67**]{} (1991) 1555; C. L. Caylor and B. M. Law, J. Chem. Phys. [**104**]{} (1996) 2070; S. P. Smith and B. M. Law, Phys. Rev. E [**52**]{} (1996) 580. Hong Zhao, A. Penninckx-Sans, Lay-Theng Lee, D. Beysens, and G. Jannink, Phys. Rev. Lett. [**75**]{} (1995) 1977. N. S. Desai, S. Peach, and C. Franck, Phys. Rev. E [**52**]{} (1995) 4129. S. Blümel and G. H. Findenegg, Phys. Rev. Lett. [**54**]{} (1985) 447; M. Thommes and G. H. Findenegg, Adv. in Space Res. [**16**]{} (1995) 83. M. Schlossmann, X.-L. Wu, and C. Franck, Phys. Rev. B [**31**]{} (1985) 1478; J. A. Dixon, M. Schlossmann, X.-L. Wu, and C. Franck, Phys. Rev. B [**31**]{} (1985) 1509. B. M. Law and D. Beaglehole, J. Phys. D [**14**]{} (1981) 115. D. P. Landau and K. Binder, Phys. Rev. B [**41**]{} (1990) 4633. C. Ruge, S. Dunkelmann, and F. Wagner, Phys. Rev. Lett. [**69**]{} (1992) 2465; C. Ruge, S. Dunkelmann, F. Wagner, and J. Wulf J. Stat. Phys. [**73**]{} (1993) 293. M. Smock and H. W. Diehl, Phys. Rev. B [**47**]{} (1993) 5841. A. J. Bray and M. A. Moore, J. Phys. A [**10**]{} (1977) 1927. H. W. Diehl and T. W. Burkhardt, Phys. Rev. B [**50**]{} (1994) 3894. A. M. Ferrenberg and D. P. Landau, Phys. Rev. B [**44**]{} (1991) 5081. U. Ritschel and P. Czerner, [*Near-surface long-range order at the ordinary transition*]{}, Essen preprint (1996). P. Czerner and U. Ritschel, [*Magnetization Profile in the d=2 Semi-Infinite Ising Model and Crossover between Ordinary and Normal Transition*]{}, Essen preprint (1996). D. P. Landau and K. Binder, Phys. Rev. B [**41**]{} (1990) 4786; M. Smock, H. W. Diehl, and D. P. Landau, Ber. Bunsenges. Phys. Chem. [**98**]{} (1994) 486. K. Binder and D. P. Landau, Phys. Rev. Lett. [**52**]{} (1984) 318. K. Binder and D. P. Landau, Physica A [**163**]{} (1990) 17. K. Binder and D. P. Landau, Phys. Rev. B [**37**]{} (1988) 1745; Phys. Rev. B [**46**]{} (1992) 4844; K. Binder, D. P. Landau and S. Wansleben, Phys. Rev. B [**40**]{} (1989) 6971. C. Ruge, [*Critical Exponents and Universal Amplitude Ratios for the $d=3$ Ising Model with Surfaces*]{}, doctoral thesis (Kiel 1994). K. Symanzyk, Nucl. Phys. B [**190**]{} \[FS3\] (1981) 1; H. W. Diehl and S. Dietrich, Z. Phys. B [**42**]{} (1981) 65. H. Au-Yang, J. Math. Phys. [**14**]{} (1973) 937; H. Au-Yang and M. E. Fisher, Phys. Rev. B [**21**]{} (1980) 3956. E. Brézin and S. Leibler, Phys. Rev. B [**27**]{} (1983) 594; A. Ciach and H. W. Diehl, (unpublished). L. V. Mikheev and M. S. Fisher, Phys. Rev. B [**49**]{} (1994) 378. E. Eisenriegler, J. Chem. Phys. [**79**]{} (1983) 1052; E. Eisenriegler, M. Krech, and S. Dietrich, Phys. Rev. B [**53**]{} (1996) 14377; H. K. Janssen, B. Schaub, and B. Schmittmann, Z. Phys. B [**73**]{} (1989) 539. H. W. Diehl and U. Ritschel, J. Stat. Phys. [**73**]{} (1993) 1; U. Ritschel and H. W. Diehl, Nucl. Phys. B [**464**]{} (1996) 512. Z.-B. Li, U. Ritschel and B. Zheng, J. Phys. A: Math. Gen. [**27**]{} (1994) L837 ; P. Grassberger, Physica A [**214**]{} (1995) 547. U. Ritschel and H. W. Diehl, Phys. Rev. E [**51**]{} (1995) 5392. G. Gompper, Z. Phys. B [**56**]{} (1984) 217. See e.g. J. F. Cardy, [*Finite Size Scaling*]{}, (North-Holland, Amsterdam, 1988) for an overview. M. H. Fisher and P.-G. de Gennes, C. R. Acad. Sci. Ser. B [**287**]{} (1978) 207. K. Binder and D. W. Heermann, [*Monte Carlo Simulation in Statistical Physics*]{}, (Springer, Berlin, 1988). R. H. Swendsen and J.-S. Wang, Phys. Rev. Lett. [**58**]{} (1987) 86. We adopted the terminology of W. Janke and S. Kappler, Phys. Rev. Lett. [**74**]{} (1995) 212. J.-S. Wang, Physica A [**161**]{} (1989) 249. A. Ciach, A. Marciolek, and J. Stecki, [*Critical Adsorption in the Undersaturated Regime - Scaling and Exact Results in Ising Strip*]{}, Warsaw preprint (1996).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the effect of a distorted neutron star dipole magnetic field on pulsar pair cascade multiplicity and pair death lines. Using a simple model for a distorted dipole field that produces an offset polar cap, we derive the accelerating electric field above the polar cap in space charge limited flow. We find that even a modest azimuthally asymmetric distortion can significantly increase the accelerating electric field on one side of the polar cap and, combined with a smaller field line radius of curvature, leads to larger pair multiplicity. The death line for producing pairs by curvature radiation moves downward in the $P$-$\dot P$ diagram, allowing high pair multiplicities in a larger percentage of the radio pulsar population. These results could have important implications for the radio pulsar population, high energy pulsed emission and the pulsar contribution to cosmic ray positrons.'
author:
- 'Alice K. Harding & Alex G. Muslimov'
title: PULSAR PAIR CASCADES IN A DISTORTED MAGNETIC DIPOLE FIELD
---
INTRODUCTION
============
Rotation-powered pulsars are thought to produce electron-positron pairs through electromagnetic cascades in their magnetospheres (Sturrock 1971). Such pair cascades may occur both in the magnetic polar regions, where high energy photons radiated by accelerated particles undergo pair conversion in the strong near-surface magnetic fields (Ruderman & Sutherland 1975, Daugherty & Harding 1982) or in outer gaps, where high energy photons interact with thermal X-rays from the neutron star (NS) surface (Cheng, Ho & Ruderman 1986, Hirotani 2008). For some time though, there has been observational evidence that the number of pairs that can be generated in pair cascades by the existing standard models of particle acceleration in magnetic dipole fields is insufficient to account for the optical to X-ray emission from the synchrotron nebulae powered by the pulsars. Estimates of the pair multiplicity (the number of pairs produced by each primary accelerated particle) of about $10^5 - 10^6$ needed to account for the emission from the Crab pulsar wind nebula (PWN) (DeJager et al. 1996) of about $10^5$ for the Vela PWN (DeJager 2007) are more than one order of magnitude larger than the theoretical pair multiplicities. Synchrotron absorption models for the eclipse in the double pulsar system PSR J0737-3039 (Arons et al. 2005, Lyutikov 2004) require a pair multiplicity of around $10^6$ for the recycled 22 ms pulsar in that system. There has also been a long-standing problem in understanding how long-period pulsars are able to produce coherent radio emission, thought to require electron-positron pairs ( Arons 1983). Again, the standard pulsar models are not able to account for the operation of robust pair cascades in these aging pulsars (Hibschmann & Arons 2001). More recently, the discovery of pulsed gamma-ray emission from a large number of millisecond pulsars by the Large Area Telescope (LAT) on the $Fermi$ Gamma-Ray Space Telescope (Abdo et al. 2009) has revealed light curves that are best modeled by narrow radiation gaps in the outer magnetosphere (Venter et al. 2009, Abdo et al. 2010). Such narrow gaps require screening of the accelerating electric field over most of the magnetosphere by a pair multiplicity that is orders of magnitude larger than standard models are able to produce (Harding & Muslimov 2002).
This motivates us to investigate the effect of a large-scale non-dipolar NS magnetic field geometry on pulsar pair cascades. Off-centered dipole-like magnetic fields seem to be prevalent among planets in the Solar System (Russell & Dougherty 2010), and in some stars. For example, the magnetic fields of isolated magnetic white dwarfs may possess an off-centered dipole or dipole plus quadrupole configurations (see e.g. Putney & Jordan 1995) which is consistent with the idea that the magnetic $Ap$ and $Bp$ stars, the predecessors of highly magnetic white dwarfs, favor an off-centered dipole field. However, it is very likely that the surface magnetic fields of white dwarfs are much more complex (Wickramasinghe 2001). Recent modeling of X-ray pulse profiles from millisecond pulsars (MSPs) shows evidence for offset dipole fields or offset polar caps in PSR J0437-4715 (Bogdanov et al. 2007) and PSR J0030+0451 (Bogdanov & Grindlay 2009). Modeling of the X-ray pulsations of the NSs in Low-Mass X-ray Binaries, the progenitors of millisecond pulsars, show possible evidence of even more extreme magnetic field distortions (Lamb et al. 2009), that could result from distortion of the global magnetic field by e.g. the crustal plate tectonics (Ruderman, 1991). In pulsar theory, non-dipolar magnetic fields have been discussed by many authors and long before the $Fermi$ era. Arons (1983) argued that pair creation in long period ($P > 1$ s) pulsars was difficult and that distortions of polar field lines due to large-scale, non-dipolar surface fields of magnitude comparable to the dipole component, and of much smaller radius of curvature, could enable pair creation in older pulsars. Arons (1997) concluded that “while the frame dragging effect in star centered dipole geometry does improve comparison of the theory with observation, an unacceptably large fraction of the observed stars outside the bounds of pair creation theory still persists". He conjectured that dipole offsetting improves the correspondence between theory and observation, and could account for the “death valley" observed for pulsars approaching the radio death line. In this Letter we report our preliminary assessment of the effect of a distorted dipole magnetic field of a NS on the efficiency of pair creation. A more detailed treatment will appear in a later publication.
Magnetic and Electric Field of a Distorted Dipole
=================================================
We propose the following heuristic model of an asymmetric magnetic field of a pulsar that can be used to approximate plausible distortions of the magnetic field of a NS. We introduce an azimuthal asymmetry to the field lines of an originally symmetric dipole such that the field lines over half of the polar cap (PC) have relatively smaller radius of curvature and over the other half of the PC have larger radius of curvature (see Figure 1 below). Consequently, one side of the PC is larger and the PC is effectively shifted from the \[star\] center of symmetry.
In magnetic spherical polar coordinates ($\theta $, $\phi $) the magnetic field is assumed to have the following form: , \[B\] where $\eta = r/R_{\rm ns} $ is the dimensionless radial coordinate in units of stellar radius, $a$ is the parameter characterizing the distortion of polar field lines. Note that vector $\bf B$ is only approximately solenoidal, since we omitted a small $B_{\phi }$-component ($\sim B_{\theta }\partial a/\partial \phi $) and higher-order corrections in $B_r$ and $B_{\theta }$ that enter only higher-order terms neglected in the formulae below. In subsequent studies we will discuss these terms as well as more general types of field distortion. If the magnetic axis lies in the x-z plane, $a = \epsilon \cos \phi$ produces an effective offset of the PC in the x-z plane and $a = \epsilon \sin \phi $ an effective offset of the PC in the y-z plane, where $0 \leq \epsilon < 1$. By using formula (\[B\]) we get the equation of the magnetic field line, which we give here in the small-angle approximation valid near the NS surface, = x\^[(1+a)/ 2]{}, \[FL\] where $x = r/r_{\rm LC}$ is the radial distance in units of the light cylinder radius, $r_{\rm LC} = c/\Omega$; and $0\leq \xi \leq 1$ is the colatitude of a footpoint of polar field line normalized by the colatitude of the PC boundary. Figure 1 shows the field lines computed from Eqn (2) plotted in the x-z plane. The field line radius of curvature (in units of $r_{\rm LC}$) is x\_c = [43]{}x\^[(1-)/2]{}[1+1+]{} where = [38]{} , = [16]{} , = \^2 x\^[1+a]{}.
The effective shift of the PC on the NS surface is approximately, r\_[PC]{} R\_[ns]{}\_0 , where $\theta_0 = (\Omega R_{\rm ns}/c)^{1/2}$ is the canonical half-angle of the PC, $\Omega$ is the pulsar rotation rate and $R_{\rm ns}$ is the NS radius. Since the effective offset of the PC is a fraction of the PC opening angle, it is a very small fraction of the stellar radius for normal pulsars and somewhat larger fraction for millisecond pulsars. In the rest of the paper we will be referring to $\epsilon $ as the offset parameter.
Neglecting the static general-relativistic corrections, we can write the accelerating electric field, assuming the boundary conditions of space-charge limited flow (SCLF) (cf. Harding & Muslimov 1998), as E\_[||]{} - [12]{}( [[R\_[ns]{}]{}c]{} ) \^2 B\_0 [[x\^a]{}]{} { \[3+ a(\^3 -1)\] + [38]{}[ [x\^[a/2]{}]{}]{}[[\_0]{}]{}} (1-\^2), \[Epar\] where $\chi $ is the pulsar obliquity, $\kappa \approx 0.15~I_{45}/R_6^3$ ($I_{45} = I/10^{45}$ g$\cdot $cm$^2$, $R_6 = R_{\rm ns}/10^6$ cm, $I$ is NS moment of inertia) is the parameter accounting for the general-relativistic frame dragging ( Muslimov & Tsygan 1992). Note that the second term in Eqn (\[Epar\]) is significantly simplified and is not valid for very large values of $\chi$. We also note that the current of primary charges assumed in the SCLF models, which is $j = c\rho_{\rm GJ}(0)$, is not consistent with the polar cap current that has been derived in force-free magnetosphere models (e.g. Contopoulos et al. 1999, Timokhin 2006, Spitkovsky 2006), where $\rho_{\rm GJ}(0) = \Omega B_0/2\pi c$ is the Goldreich-Julian charge density at the surface. The discharge properties under these conditions were discussed by Beloborodov (2008) and may involve non-steady charge outflow and electric fields. We use here the SCLF electric field for comparison with the many previous studies of pulsar pair cascades and pair death lines.
Pair Cascade Simulation and Results
===================================
Using the distorted field structure and accelerating electric field described above, we have simulated a full pair cascade above the pulsar polar cap to compare the pair multiplicity and pair death lines for different offset parameters $\epsilon$. The full pair cascade simulation code, described in Daugherty & Harding (1982) and Daugherty & Harding (1996), has been combined with the acceleration, early cascade development and pair front formation used in Harding & Muslimov (1998). The combined code is thus able to follow a primary particle through its acceleration in the parallel electric field, taking into account inverse Compton and curvature radiation losses, the emission of photons by curvature radiation (CR), the establishment of the pair front (above which the electric field is screened) and the full cascade of pairs and their synchrotron radiation. Since the electric field is screened over a very short length scale by CR pairs (Harding & Muslimov 2001), for the present calculation we simply assume that the electric field is zero above the pair formation front (PFF). Since the results are fairly sensitive to NS equation of state (EoS), we choose for normal pulsars a mass of $M_{\sun} = 1.45$ and radius $R_{\rm ns} = 10$ km, which give a moment of inertia $I = 1.13 \times 10^{45}\,\rm g\,cm^2$ (Lattimer & Prakash 2007), and for MSPs, that may have accreted mass during their spin-up phase, we choose $M_{\sun} = 2.15$, $R_{\rm ns} = 9.9$ km and $I = 1.56 \times 10^{45}\,\rm g\,cm^2$ from a rotating NS model (Friedman et al. 1986).
The resulting pair multiplicities for a range of pulsar periods and surface magnetic field strengths, and for offset parameters $\epsilon = 0, 0.1, 0.2, 0.4$ and $0.6$ are shown in Table 1 for the normal pulsar population and in Table 2 for millisecond pulsars. For these calculations, we have taken $a = \epsilon \sin \phi$ for an effective offset in the y-z plane and magnetic azimuth angle $\phi = 270^\circ$, where the field distortion produces the most favorable conditions for pair creation, i.e. smallest radius curvature, largest PC angle and largest electric field. Pair multiplicities are computed for one representative magnetic colatitude of $\xi = 0.5$. The increase in the parallel electric field, in this case on the trailing side of the PC, by a factor of about $\theta_{0}^{2a}/(1+a)^2$, results in a larger voltage at the PFF and is the most important effect leading to larger pair multiplicities since the maximum curvature radiation energy is proportional to the cube of the particle Lorentz factor. For normal pulsars, even modest values of $\epsilon$ that are equivalent to offsets of a few percent of the stellar radius lead to orders of magnitude increases in the pair multiplicity. Small offsets also enable long period pulsars, that cannot produce any pairs from CR in a pure dipole field, to produce high-multiplicity pair cascades. Most MSPs cannot produce pairs from CR in a dipole field but with some offset of the PC would be able to produce high pair multiplicity. The plateaus and peaks in the multiplicity, that can be seen in Table 1 for $B_0 > 3 \times 10^{12}$ G, are the result of pairs being produced increasingly in the ground Landau state in high fields (Baring & Harding 2001) which inhibits the cascade synchrotron radiation. Photon splitting, which we have not included in these calculations, would further reduce pair multiplicity above $10^{13}$ G. Although the multiplicity grows with increasing $\epsilon$ and increasing field strength, it saturates at a value below $10^5$. This limit, also noted by Medin & Lai (2010), results from pairs being produced with higher and higher average energy at smaller angles to the magnetic field as the primary electron energy, and thus the maximum energy of the CR photons, increases.
The increase in pair multiplicity from a distorted dipole field can move many pulsars above the death line for robust pair creation. Figure 2 shows the death lines for pair creation by CR in the $P$-$\dot P$ diagram for different offset parameters, computed for the two different EoS for normal and MSPs described above. The lines plotted in Figure 2, determined by the method described in Harding & Muslimov (2002), show the lowest $\dot P$ required for a pulsar at a given period to establish a PFF. The death line for a pure dipole cuts through the middle of the normal pulsar population and through the upper edge of the MSP population. As $\epsilon$ increases, the lines move significantly downward through both populations. For $\epsilon = 0.4$, nearly all pulsars including the 8.5 s PSR J2144-3933 (Young, Manchester & Johnston 1999) are able to produce pair cascades which could enable coherent radio emission. The lines decrease in slope as they move from longer to shorter periods and from small to large $\epsilon$ as the particle acceleration becomes more limited by curvature radiation reaction. For the shortest period MSPs, the acceleration is completely radiation-reaction limited (Luo et al. 2000) which causes a slower increase in electron Lorentz factor and impedes the development of pair cascades. The envelope of the observed pulsar population seems to be consistent with the shape predicted from this limit.
Discussion
==========
We have found that distortions of a dipole magnetic field that produce small offsets of the PC can produce large increases in the multiplicity of pulsar pair cascades and allow pair cascades in pulsars not able to initiate cascades in centered dipole fields. Previous estimates of the effect of offset vacuum dipoles on pair death lines included only the decreased field line radius of curvature, concluding that dipole offsets of (0.7 - 0.8)$R_{\rm ns}$ (Arons 1997) or even 0.95 $R_{\rm ns}$ (Medin & Lai 2010), that are large fractions of a stellar radius, are required. Our calculation also takes into account the change in $E_\parallel$ and the particle acceleration energy, which is by far the strongest effect. Since this effect is sensitive to effective offsets that are fractions of a PC radius, we show that many older pulsars and MSPs would be able to sustain robust pair cascades with dipole offsets that are small fractions of a stellar radius. Although distortions of the dipole field could be intrinsic to the NS through asymmetries in the interior currents, present either from birth or as a result of spin-up or spin-down evolution, magnetic dipole fields that are distorted by rotation or currents also produce offset PCs. The sweepback of field lines of a retarded vacuum dipole (Deutsch 1955) causes an offset of the PC towards the trailing side, opposite the direction of rotation (Dyks & Harding 2004), and the force-free magnetosphere, with a similar sweepback of field lines, has even greater offset of the PC (Bai & Spitkovsky 2010).
We will mention here a number of implications of our results. For high enough offsets, the entire radio MSP population could generate pair cascades. For the offset of $\epsilon \sim 0.6$ implied by the apparent shift of $\sim 1$ km of the heated PC of PSR J0437-4715 (Bogdanov et al. 2007), with $P = 5.75$ ms and $\dot P = 1.4 \times 10^{-20}$, a pair cascade multiplicity of about $10^2$ is possible (cf. Table 2). The ability of MSPs with distorted dipoles to generate high-multiplicity pair cascades that can screen the $E_{\parallel}$ over most of the open field could possibly account for the prevalence of sharply-peaked light curves of $\gamma$-ray MSPs that would require narrow acceleration gaps. Higher pair multiplicity in both normal and MSPs could also enhance their contribution to local Galactic cosmic-ray positrons. However, we find that the saturation in pair cascade multiplicity with increasing particle energy limits the extent to which a dipole distortion can produce even higher multiplicity. Therefore, even large values of $\epsilon $ cannot produce the levels of multiplicity in young pulsars required to produce synchrotron emission observed from PWNe. We find for example that the Crab pulsar, with $P = 0.033$ s and $B_0 \sim 3 \times 10^{12}$ G could only produce a pair multiplicity up to $\sim 3 \times 10^4$ for $\epsilon = 0.4$ corresponding to an offset from Eqn (3) of 6% of the stellar radius. A pair multiplicity of $\sim 10^6$ in PSR J0737-3039A would also not be achievable with any degree of offset.
Distorted dipole fields with offset PCs will generally introduce asymmetries in the pulsar emission as well as in the pair multiplicity. Since the particle acceleration and field line curvature vary over the PC, the pair multiplicity will be larger over one half of the PC in the direction of the offset. Pulsars that are below the death line for centered dipole fields, including many MSPs, will have pair cascades only on that side of the PC. The variation in pair multiplicity across the PC could result in asymmetric radio emission depending on how the coherent mechanism is related to the multiplicity. The higher accelerating field and voltage on the offset side of the PC will produce asymmetric heating of the PC and thus inhomogeneous thermal X-ray emission. The structure and energetics of the proposed slot gaps (SG) that form between the boundary of the open magnetic field and the upward curving PFF (Arons & Scharlemann 1979), and can accelerate particles to high altitude (Muslimov & Harding 2004), could be strongly affected by a distorted dipole. The particle Lorentz factor $\gamma$ in the SG , which is expected to reach curvature radiation-reaction limit such that $\gamma \propto E_{\parallel}^{1/4}$, will be larger on one side of the PC, producing CR emission power proportional to $E_\parallel$ that is larger than for a dipole field. An azimuthal asymmetry of both the radiation power and width of the SG would change both the $\gamma$-ray luminosity and light curves. A further application of these results is in pulsar population studies. The decrease in the observed radio pulsar population just above the death line as pulsars spin-down and evolve from left to right across the $\dot P$-$P$ diagram has been difficult to explain. But a range of dipole offsets across the population would produce a spread of pair/radio death lines and thus produce a death valley. It is interesting that the no-offset death line shown in Figure 1 occurs near the densest part of the normal pulsar population. We will explore many of these consequences of increased pair multiplicity in distorted dipole fields in future studies.
AKH thanks the Aspen Center for Physics where fruitful discussions, particularly with J. Arons, A. Timokhin, O. DeJager and A. Spitkovsky, provided stimulation for this work. We also acknowledge support from the NASA Astrophysics Theory and Fundamental Physics Program, the $Fermi$ Guest Investigator Program and the Universities Space Research Association.
Abdo, A. A. et al. 2009, Science, 325, 848. Abdo, A. A. et al. 2010, ApJ, 712, 957 Arons, J. 1983, ApJ, 266, 215. Arons, J. 1997, in Neutron Stars and Pulsars: Thirty Years after the Discovery, eds. N. Shibazaki, N. Kawai, S. Shibata and T. Kifune, University Academy Press, Inc., Tokyo, Japan, p. 339. Arons, J.; Backer, D. C.; Spitkovsky, A.; Kaspi, V. M. 2005, in Binary Radio Pulsars, ASP Conference Series, Vol. 328, Ed. F. A. Rasio and I. H. Stairs (San Francisco: Astronomical Society of the Pacific) p.95 Arons, J. & Scharlemann, E. T. 1979, ApJ, 231, 854 Bai, X-N. & Spitkovsky, A. 2010, ApJ, 715, 1282. Baring, M. G. & Harding, A. K. 2001, ApJ, 547, 929. Beloborodov, A. M. 2008, ApJ, 683, L41. Bogdanov, S. , Rybicki, G. B. & Grindlay, J. E. 2007, ApJ, 670, 668. Bogdanov, S. & Grindlay, J. E. 2009, ApJ, 703, 2259. Cheng, K. S., Ho, C., & Ruderman, M. A. 1986, ApJ, 300, 500. Contopoulos, I., Kazanas, D. & Fendt, C. 1999, ApJ, 511, 351. Daugherty, J. K. & A. K. Harding 1982, ApJ, 252, 337. Daugherty, J. K. & A. K. Harding 1996, ApJ, 458, 278. De Jager, O. C. 2007, ApJ, 658, 1177. De Jager, O. C., A. K. Harding, Michelson, P. F. , Nolan, P. L., Sreekumar, P. & Thompson, D. J. 1996, ApJ, 457, 253. Deutsch, A. J., 1955, Ann. d’Astrophys., 18, 1. Dyks, J. & Harding, A. K. 2004, ApJ, 614, 869. Friedman, J. L., Ipser, J. R., & Parker, L. 1986, ApJ, 304, 115. Harding, A. K., & Muslimov, A. G. 1998, ApJ, 508, 328. Harding, A. K., & Muslimov, A. G. 2001, ApJ, 556, 987. Harding, A. K., & Muslimov, A. G. 2002, ApJ, 568, 862. Hibschman, J. A. & Arons, J. 200, ApJ, 560, 871. Hirotani, K. 2008, Open Astronomy (arXiv:0809.1283) Lamb, F. K. et al, 2009, ApJ, 706, 417. Lattimer, J. M. & Prakash, M. 2007, Physics Reports, 442, 109. Luo, Q.; Shibata, S.; Melrose, D. B. 2000, MNRAS, 318, 943. Lyutikov, M. 2004, MNRAS, 353, 1095. Manchester, R. N., Hobbs, G. B., Teoh, A. & Hobbs, M. 2005, Astron. J., 129, 1993 Medin, Z. & Lai, D. 2010, MNRAS, 406, 1379. Muslimov, A. G. & Tsygan, A. I. 1992, MNRAS, 255, 61 Muslimov, A. G. & Harding, A. K. 2004, ApJ, 606, 1143 Putney, A. & Jordan, S. 1995, ApJ, 449, 863. Ruderman, M. 1991, ApJ, 366, p. 261. Ruderman, M.A. & Sutherland, P. G. 1975, ApJ, 196, 51 Russell, C. T. & Daugherty, M. K. 2010, Space Science Reviews, 152, 251. Spitkovsky, A. ApJ, 648, L51 (2006) Sturrock, P. A. 1971, ApJ, 164, 529. Timokhin, A. MNRAS, 36, 1055 (2006) Venter, C.; Harding, A. K.; Guillemot, L. 2009, ApJ, 707, 800. Wickramasinghe, D. 2001, in Magnetic Fields Across the Hertzsprung-Russell Diagram, AIP Conference Series, Vol. 248, eds. G. Mathys, S. K. Solanki, and D. T. Wickramasinghe. Young, M. D., Manchester, R. N. & Johnston, S. 1999, Nature, 400, 848
[lcccccc]{} $B_0 (10^{12}$ G) & [$\epsilon$]{} & && [Period (s)]{} &&\
&& 0.033 & 0.1 & 0.3 & 1.0 & 3.0\
0.33 & 0.0 & 1.48E3 & 5.66E1& 0.0E0 & 0.0E0 & 0.0E0\
& 0.1& 2.27E3 & 2.47E2 & 0.0E0 & 0.0E0 & 0.0E0\
& 0.2& 3.45E3 & 9.08E2 & 1.3E-5 & 0.0E0 & 0.0E0\
& 0.4 & 7.90E3 & 3.19E3 & 9.74E2 & 1.93E1 & 0.0E0\
1.0 & 0.0 & 4.81E3 & 4.79E2 & 3E-12 & 0.0E0 & 0.0E0\
& 0.1 & 8.13E3 & 1.52E3 & 3.44E1 & 0.0E0 & 0.0E0\
& 0.2 & 1.15E4 & 3.01E3 & 3.41E2 & 0.0E0 & 0.0E0\
& 0.4 & 2.39E4& 9.90E3 & 3.29E3 & 5.72E2 & 5.6E-8\
3.0 & 0.0 & 1.07E4 & 2.08E3 & 4.71E1 & 0.0E0 & 0.0E0\
& 0.1& 1.49E4 & 4.76E3 & 4.42E2 & 0.0E0 & 0.0E0\
& 0.2& 1.90E4 & 8.21E3 & 1.49E3 & 3.32E1 & 0.0E0\
& 0.4 & 2.64E4 & 1.58E4 & 7.86E3 & 2.39E3 & 1.91E2\
10 & 0.0 & 1.26E4 & 7.90E3 & 1.94E3 & 8E-10 & 0.0E0\
& 0.1 & 1.39E4 & 9.97E3 & 4.50E3 & 1.88E2 & 0.0E0\
& 0.2 & 1.46E4 & 1.14E4 & 7.19E3 & 1.90E3 & 0.0E0\
& 0.4 & 1.53E4 & 1.32E4 & 1.11E4 & 7.87E3 & 3.64E3\
For $\chi = 60^0$, $\phi = 270^0$, $\xi = 0.5$ and neutron star parameters $M_\sun = 1.45$, $R_{\rm ns} = 10$ km and $I = 1.13 \times 10^{45}\,\rm g\,cm^2$.
[lcccc]{} $B_0 (10^{9}$ G) & [$\epsilon$]{} & & [Period (ms)]{} &\
&& 2 & 5 & 10\
0.3 & 0.0 & 4.4E-6 & 0.0E0 & 0.0E0\
& 0.1&3.3E-1 & 0.0E0 & 0.0E0\
& 0.2& 2.16E0 & 0.0E0 & 0.0E0\
& 0.4 & 2.30E1 & 9.5E-2 & 1.9E-12\
& 0.6 & 4.96E2 & 1.44E1 & 3.6E-1\
1.0 & 0.0 & 1.43E2 & 7E-4 & 0.0E0\
& 0.1& 2.51E2 & 4.1E-2 & 6E-11\
& 0.2& 3.87E2 & 6.52E0 & 0.0E0\
& 0.4 & 7.33E2 & 2.24E2 & 1.05E1\
& 0.6 & 1.95E3 & 6.82E2 & 3.04E2\
3.0 & 0.0 & 6.46E2 & 1.32E2 & 6.6E-2\
& 0.1& 9.78E2 & 2.03E2 & 3.01E1\
& 0.2& 1.45E3 & 3.04E2 & 1.04E2\
& 0.4 & 2.99E3 & 9.07E2 & 3.22E2\
& 0.6 & 7.12E3 & 2.60E3 & 1.19E3\
10 & 0.0 & 2.48E3 & 6.01E2 & 1.92E2\
& 0.1& 3.28E3 & 9.36E2 & 3.01E2\
& 0.2& 4.64E3 & 1.42E3 & 6.45E2\
& 0.4 & 9.27E3 & 3.05E3 & 1.45E3\
& 0.6 & 2.31E4 & 7.84E3 & 3.82E3\
For $\chi = 45^0$, $\phi = 270^0$, $\xi = 0.5$ and neutron star parameters $M_\sun = 2.15$, $R_{\rm ns} = 9.9$ km and $I = 1.56 \times 10^{45}\,\rm g\,cm^2$.
{width="180mm"}
{width="210mm"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'It is known that more than 50 species use the Earth’s magnetic field for orientation and navigation. Intensive studies particularly behavior experiments with birds, provide support for a chemical compass based on magnetically sensitive free radical reactions as a source of this sense. However, the fundamental question of how quantum coherence plays an essential role in such a chemical compass model of avian magnetoreception yet remains controversial. Here, we show that the essence of the chemical compass model can be understood in analogy to a quantum interferometer exploiting global quantum coherence rather than any subsystem coherence. Within the framework of quantum metrology, we quantify global quantum coherence and correlate it with the function of chemical magnetoreception. Our results allow us to understand and predict how various factors can affect the performance of a chemical compass from the unique perspective of quantum coherence assisted metrology. This represents a crucial step to affirm a direct connection between quantum coherence and the function of a chemical compass.'
author:
- Jianming Cai
- 'Martin B. Plenio'
title: Chemical compass model for avian magnetoreception as a quantum coherent device
---
[*Introduction.—*]{} Despite the growing interest from chemists, biologists and more recently researchers from quantum physics and quantum information, there remain to date only a handful of biological phenomena that are suspected or proven to rely on quantum effects. These include important biological processes such as light harvesting [@Engel07; @Collini10; @LHCT1; @Plenio08; @Scho11; @Chin13], human sense of smell [@SME1; @SME2; @SME3], and avian magnetoreception [@Sch78; @Ritz00; @Sol07; @Rodgers09; @Ritz10; @Ritz09; @Ritz11; @Hore10; @Maeda08; @Kom09; @Cai10; @Gauger11; @Cai12; @Band12; @Hogben12; @Sun12; @Guer2013]. The effects of weak magnetic field in nature have been observed for a long time [@Joh05], ranging from the growth of plants to the remarkable orientation and navigation abilities of animals such as birds and insects. The radical pair mechanism based on anisotropic hyperfine interactions, as a leading theory to explain avian magnetoreception [@Werner77], suggests that the avian compass relies on magnetically sensitive radical pairs formed by photoinduced electron transfer reactions [@Sch78; @Ritz00; @Sol07; @Ritz10; @Ritz09; @Rodgers09]. The cryptochromes in the retina of migratory birds provides a potential physiological implementation of such a mechanism [@Mour04; @Weaver00nature; @Hore03; @Solo07; @Ahmad07; @Ritz00; @Zap09; @Maeda12; @Sol10; @Sol12; @Lied10]. The observations from the behavior experiments [@WiltsRev05; @Phil92; @Ritz04nature] with birds provide corroborating evidence for the idea that the chemical compass mechanism is involved in avian magnetoreception.
In the last few years, the interest in avian magnetoreception has quickly extended from chemists and biologists to quantum physicists [@Ritz11; @Hore10; @Kom09; @Cai10; @Cai12; @Sun12; @Gauger11; @Band12; @Hogben12]. In the chemical compass model, a radical pair is born in spin singlet or triple states [@Sch78; @Steiner89; @Ritz00]. The subsequent dynamics is composed of a quantum coherent interaction with nearby nuclei and the external magnetic field. The former is usually considered as a noise process that suppresses quantum coherence. In previous consideration emphasis was placed on the electronic coherence properties and it remained unclear how this coherence was directly exploited in the function of a chemical compass [@Cai10; @Cai12; @Gauger11; @Band12; @Hogben12]. For example, a chemical compass can still show good sensitivity even though it exhibits negligible electronic coherence [@Cai10; @Hogben12]. It is not clear how to establish a quantitative relation between coherence and compass sensitivity.
![Quantum interferometer model of a chemical compass. (a) A quantum system is prepared ($\mbox{BS}_1$) in a coherent superposition of two quantum states ${\left\vert \alpha \right\rangle}$ and ${\left\vert \beta \right\rangle}$, which gains different dynamical phases dependent on an unknown physical parameter, the information of which is revealed by the measurement after the interference ($\mbox{BS}_2$). (b) A chemical compass is viewed as a quantum interferometer, the hyperfine coupling Hamiltonian $\mathcal{H}_0$ recasts the state $\rho_0$ of the combined system of radical pair and nuclei into a superposition of its eigenstates. The magnetic field induces the change in coherence phases arising from the dynamic phases of individual eigenstate evolution. The spin-dependent chemical reaction channels lead to the interferometric observable which provides the information about the direction of the magnetic field.[]{data-label="fig:INF"}](setup_v2.pdf){width="8.5cm"}
In the present work,we address these fundamental questions and reveal the role of quantum coherence in chemical magnetoreception by appreciating that the quantum dynamics of a chemical compass giving rise to the magnetic sensitivity is akin to a quantum interferometer in quantum metrology [@Wine94; @Budker07; @Caves07]. Such a perspective allows us to establish a quantitative connection between compass sensitivity and the global coherence with respect to the eigenbasis of the hyperfine Hamiltonian of the combined system of radical pair electrons and the surrounding nuclei. We verify that the concept of global coherence is indeed essential for the function of a chemical compass by introducing an appropriate quantification of coherence which has an operational meaning in the context of chemical compass. The results are shown to be valid for general radical pair molecules, and we further explicitly demonstrate the idea for the radical pair inspired by the flavin adenine dinucleotide (FADH$\cdot$) formed photochemically in cryptochromes (which is a very probable candidate for avian magnetoreception [@Ritz00; @Hore03; @Maeda12; @Sol10; @Sol12]) in more detail.
[*Quantum interferometer model of chemical compass.—*]{} In quantum metrology, a system, e.g., a photon or a spin, is initially prepared into the following state \_0=[[\_0 ]{}[\_0 ]{} ]{}, [\_0 ]{}=\_+\_, which is written in the basis of $\{{\left\vert \alpha \right\rangle},{\left\vert \beta \right\rangle}\}$ (determined by the interferometric element BS1, see Fig.\[fig:INF\]). A Hamiltonian $H(\lambda)=\frac{\lambda}{2}{\left( {{\left\vert \alpha \right\rangle}{\left\langle \alpha \right\vert} }-{{\left\vert \beta \right\rangle}{\left\langle \beta \right\vert} } \right)}$ parametrized by a single unknown parameter $\lambda$ results in an evolution that changes the relative phase between the two states ${\left\vert \alpha \right\rangle}$ and ${\left\vert \beta \right\rangle}$ to $\Phi_\lambda={\varphi}_{\alpha}-{\varphi}_{\beta}$, see Fig.\[fig:INF\](a), and leads to the state \_0\_[\_]{}=(
[cc]{} |\_|\^2& \_ \^[\*]{} \_ e\^[i\_]{}\
\_ \^[\*]{}\_ e\^[-i\_]{} & |\_|\^2
). The measurement of the observable $\mathcal{M}={\left\vert \Psi_0 \right\rangle}{\left\langle \Psi_0 \right\vert}$, gives the interference (see BS2 in Fig.\[fig:INF\]) outcome as m\_(\_0,\_) [( \_[\_]{} )]{} = 1-2|\_|\^2|\_|\^2[( 1- )]{}.\[eq:mainf\] The precision of parameter estimation with such a basic quantum interferometer is determined by the contrast of the interference fringe as follows \_(\_0)2|\_|\^2|\_|\^2 . The role of coherence in such a simple interferometric setup can be quantified by |m\_(\_0\^c,\_)|\_[\_=0]{}=2 |\_|\^2|\_|\^2, \[eq:mcqm\] with \_0\^c = (
[cc]{} 0& \_ \^[\*]{} \_\
\_ \^[\*]{}\_ & 0
), which characterizes how coherent the quantum interferometer is. By comparing Eq.(4) and (5), it is clearly seen that how coherence plays its role and determines the precision of parameter estimation in quantum interferometric metrology.
![Effective evolution of a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$). (a) The change of the absolute values of the (coherent) density elements $\langle\epsilon(t) \rangle $ averaging over time, with $\epsilon(t)=[\sum_{m\neq n} (r_{mn}^t-r_{mn})^2]^{1/2}$, see Eq.(\[eq:recast\]-\[eq:totalevolution\]), for different magnetic field directions $(\theta,{\varphi})$. (b) The singlet yield $\mbox{Y}_{\mbox{s}}$ as a function of the magnetic field direction $(\theta,{\varphi})$ including (lower) [*vs.*]{} excluding (upper) the coherence phase changes ${\varphi}_{mn}^{\hat{B}}(t)$ resulting from the Earth’s magnetic field, see Eq.(\[eq:totalevolution\]). The reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:FID"}](FADH_FID_OFF_NEW.pdf "fig:"){width="4.45cm"} ![Effective evolution of a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$). (a) The change of the absolute values of the (coherent) density elements $\langle\epsilon(t) \rangle $ averaging over time, with $\epsilon(t)=[\sum_{m\neq n} (r_{mn}^t-r_{mn})^2]^{1/2}$, see Eq.(\[eq:recast\]-\[eq:totalevolution\]), for different magnetic field directions $(\theta,{\varphi})$. (b) The singlet yield $\mbox{Y}_{\mbox{s}}$ as a function of the magnetic field direction $(\theta,{\varphi})$ including (lower) [*vs.*]{} excluding (upper) the coherence phase changes ${\varphi}_{mn}^{\hat{B}}(t)$ resulting from the Earth’s magnetic field, see Eq.(\[eq:totalevolution\]). The reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:FID"}](FADH_FID_YIELD_NEW.pdf "fig:"){width="4.45cm"}
In the standard model of the chemical compass, the magnetically sensitive radical pairs formed by photoinduced intramolecular electron transfer reactions interact with a few nearby nuclei via hyperfine couplings, as described by the following Hamiltonian \_0=\_[k=]{}\_[j]{}\_[k]{}\_[k\_[j]{}]{}\_[k\_[j]{}]{} where $\vec{\mathbf{I}}_{k_{j}}$ and $\vec{\mathbf{s}}_{k}$ are the nuclear and electron spin operators respectively, $\hat{\mathbf{T}}_{k_{j}}$ denotes the hyperfine coupling tensors [@Steiner89]. For simplicity, we neglect the dipole-dipole and exchange interactions between two radicals, which is valid when the radical-radical distance is sufficiently large or they cancel with each other [@Efi08]. The joint state $\rho_0$ of the radical pair and nuclear spins, written in the eigenbasis $\{{\left\vert \nu_m \right\rangle}\}$ of the hyperfine interaction Hamiltonian $\mathcal{H}_0$, can be expressed as $\rho_0 = \sum_{m,n} r_{mn} {\left\vert \nu_m \right\rangle}{\left\langle \nu_n \right\vert}, \quad \mbox{with} \quad r_{mn}={\left\langle \nu_m \right\vert}\rho_0{\left\vert \nu_n \right\rangle}$. In the absence of a magnetic field, the state evolution solely under the hyperfine interaction Hamiltonian $\mathcal{H}_0$ is \^[\_0]{}(t) = \_[m,n]{} r\_[mn]{} e\^[-i\_[mn]{}\^[0]{}(t)]{}[\_m ]{}[\_n ]{}. \[eq:recast\] For a typical radical pair molecule with a few nuclei, the leading order effect of the Earth’s magnetic field $\mathrm{\hat{B}}=-g\mu_b \vec{b}\cdot (\vec{\mathbf{s}}_{\mathcal{D}}+\vec{\mathbf{s}}_{\mathcal{A}})$ (with $|\vec{b}|=50\mu \mbox{T}$, which is much smaller than the hyperfine interaction strength) is to introduce an additional magnetic field dependent phase on the system coherence according to the perturbation theory, namely \^[\_0+(,)]{}(t) \_[m,n]{} r\_[mn]{}\^t e\^[-i]{}[\_m ]{}[\_n ]{},\[eq:totalevolution\] where $\rho^{\mathcal{H}_0+\mathrm{\hat{B}}(\theta,{\varphi})}(t)$ represents the exact state evolution under the total Hamiltonian including both the hyperfine coupling and the Earth’s magnetic field. The effect of the Earth’s magnetic field might also change the absolute values of the coherent off-diagonal density matrix elements $r_{mn}$ ($m\neq n$), which can be quantified by $\epsilon(t)=[\sum_{m\neq n} (r_{mn}^t-r_{mn})^2]^{1/2}$. In Fig.\[fig:FID\](a), we plot $\epsilon$ averaging over time for the radical pair inspired by the flavin adenine dinucleotide formed photochemically in cryptochromes as an example, which shows that the change of the absolute values of the density matrix elements is very small.
The spin-dependent reactions in a chemical compass, i.e. the singlet and triplet radical pairs will undergo different chemical reaction paths and thereby lead to different chemical consequences [@Ritz00], that witness the magnetic field effect on the radical pair dynamics by the magnetic anisotropy of reaction yield. For simplicity, we consider the scenarios where the reaction rates of the singlet and the triplet radicals are identical, i.e. $k_S=k_T\equiv k$ (see [@SI] for the generalization to more general cases). Treating the system dynamics with the conventional Haberkorn approach [@Steiner89], the singlet yield can be formulated as \_(\_0,\_0,,) = dt \[singletyield\] with the observable $\mathcal{M}=k e^{-kt} ( {\left\vert \mathcal{S} \right\rangle}{\left\langle \mathcal{S} \right\vert}\otimes_{k_j} \mathbb{I}_{d_{k_j}})$, where ${\left\vert \mathcal{S} \right\rangle}=\sqrt{\frac{1}{2}}{\left( {\left\vert \alpha\beta \right\rangle}-{\left\vert \beta\alpha \right\rangle} \right)}$. The singlet yield in Eq.(\[singletyield\]) can be viewed as a continuous generalization of the outcome in a quantum interferometer as $m_{\lambda}(\rho_0,\Phi_{\lambda})$, see Eq.(\[eq:mainf\]). In Fig.\[fig:FID\](b), it can be seen that the magnetic anisotropy of singlet yield mainly comes from the coherent phase changes ${\varphi}_{mn}^{\hat{B}}(t) $ induced by the Earth’s magnetic field, as in the interference-based quantum metrology. The analogy between a quantum interferometer and a chemical compass (a complete list of analogies is included in [@SI]) now allows us to clearly identify the role of coherence in chemical magnetoreception in a quantitative manner.
[*Quantum coherence and compass sensitivity.—*]{} The function of a chemical compass starting from the state $\rho_0$ can be characterized by the magnetic anisotropy of the singlet yield defined as follows \_(\_0,\_0)=\_[,]{} \_(\_0,\_0,,)-\_[,]{} \_(\_0,\_0,,)\[eq:magani\] Henceforth we will call $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as the magnetic sensitivity of a chemical compass. The coherent part of the system state $\rho_0$ is denoted as (\_0,\_0) = \_[mn]{} r\_[mn]{} [\_m ]{}[\_n ]{}. To quantify the role of coherence in chemical magnetoreception, following the inspiration from quantifying the role of coherence in quantum interferometric metrology (see Eq.(\[eq:mcqm\])), we introduce the following measure of global electron-nuclear quantum coherence for a chemical compass with a given hyperfine interaction Hamiltonian $\mathcal{H}_0$ and a system state $\rho_0$ as follows (\_0,\_0)=\_(\_0,(\_0,\_0))\_[b=0]{}\[eq:gcm\] which represents the contribution of coherence to the singlet yield in the absence of the Earth’s magnetic field.
![(a) The magnetic field sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as defined in Eq.(\[eq:magani\]) (upper) and the average value $\langle \mbox{D}_{\mbox{s}} \rangle $ (lower) as a function of coherence $\mathcal{C}(\mathcal{H}_0,\rho_0)$ given by Eq.(\[eq:gcm\]) for $7\times 10^4$ random hyperfine configurations with in total 5-6 nuclei initially in the depolarized state. The violet dots represent these configurations of reference-and-probe type molecules (one radical is free from hyperfine coupling). (b) The magnetic field sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_{\mbox{f}},\rho_0)$ as a function of coherence $\mathcal{C}(\mathcal{H}_{\mbox{f}},\rho_0)$ for a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$) starting from $2\times 10^4$ initial states, each has the singlet born radical pair and random nuclear spin polarizations. In both panels, the reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, and the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:CRDS"}](RAN_A_v1_NEW.pdf "fig:"){width="4.47cm"} ![(a) The magnetic field sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as defined in Eq.(\[eq:magani\]) (upper) and the average value $\langle \mbox{D}_{\mbox{s}} \rangle $ (lower) as a function of coherence $\mathcal{C}(\mathcal{H}_0,\rho_0)$ given by Eq.(\[eq:gcm\]) for $7\times 10^4$ random hyperfine configurations with in total 5-6 nuclei initially in the depolarized state. The violet dots represent these configurations of reference-and-probe type molecules (one radical is free from hyperfine coupling). (b) The magnetic field sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_{\mbox{f}},\rho_0)$ as a function of coherence $\mathcal{C}(\mathcal{H}_{\mbox{f}},\rho_0)$ for a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$) starting from $2\times 10^4$ initial states, each has the singlet born radical pair and random nuclear spin polarizations. In both panels, the reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, and the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:CRDS"}](FADH_A_v1_NEW.pdf "fig:"){width="4.4cm"}
To demonstrate the connection between global coherence and the magnetic sensitivity in a chemical compass for general molecules, we randomly sample a large number of hyperfine configurations [@RS] and plot in Fig.\[fig:CRDS\](a) the compass sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as a function of coherence $\mathcal{C}(\mathcal{H}_0,\rho_0)$ and , where $\rho_0={{\left\vert \mathcal{S} \right\rangle}{\left\langle \mathcal{S} \right\vert} }\bigotimes_{k_j} (\mathbb{I}_{d_{k_j}}/d_{k_j})$. The result shows that typically the global coherence is a [*resource*]{} for the function of a chemical compass, namely the larger coherence the better magnetic sensitivity. We find that global coherence makes more dominant contribution to the compass sensitivity as compared with local electronic coherence [@SI]. The perspective that coherence is a resource for chemical magnetoreception offers a unique guide towards various design principles of radical pair molecules. For example, one can see that the reference-and-probe type of radical pairs (namely one radical of which is free from hyperfine coupling) [@Ritz10; @Maeda08] which tends to result in the highest sensitivity usually embody larger coherence. Considering the explicit example of the radical pair inspired by the flavin adenine dinucleotide in cryptochromes, we plot in Fig.\[fig:CRDS\](b) the magnetic sensitivity as a function of coherence when varying the nuclear spin polarization. The same feature is observed, namely the larger the global coherence is, the better the magnetic sensitivity. We remark that the nuclear spins are usually depolarized at ambient temperature. We assume here a larger state space to facilitate the explicit demonstration of how coherence (that changes with the nuclear polarization) can directly affect the function of chemical magnetoreception. In experiments, this may be achieved by chemically induced dynamic nuclear polarization or with the assistance of quantum control via, e.g. nitrogen-vacancy centers in diamond [@Cai12_njp; @Paz13; @Cai13]. In contrast, the prominent connection is absent between the compass sensitivity and local electronic coherence [@SI]. We also stress that the above measure of coherence in Eq.(\[eq:gcm\]) is determined by the hyperfine interaction Hamiltonian $\mathcal{H}_0$ (without an external magnetic field) and the joint system state $\rho_0$, yet it well predicts the property (namely the magnetic sensitivity) of the dynamics of a chemical compass when changing the magnetic field direction.
[*A unified picture of decoherence effects.—*]{} Following the present insight that quantum coherence of the global electron-nuclear state is a resource for chemical magnetoreception, it is possible to study the effects of different decoherence models on the functioning of a chemical compass [@Gauger11; @Cai12; @Band12] in a unified picture. More specifically, we can study how decoherence will destroy the global electron-nuclear quantum coherence in a chemical compass, as characterized by Eq.(\[eq:gcm\]), and thereby deteriorate its magnetic anisotropy of reaction yield. Since a chemical compass shall work under ambient conditions, the noise from the environment of the core system (i.e. the radical pair and the surrounding nuclear spins) will inevitably affect its function. The noise effects vary for different decoherence models, which can be described by the Lindblad type quantum master equation as [@Gauger11] =-i\[,\]-\[Q\_S,\]\_[+]{}-\[Q\_T,\]\_[+]{}+() with $
\mathcal{L}(\rho)=\sum_k\xi_k{\left[ \mathcal{L}_k \rho \mathcal{L}_k^{\dagger}-\frac{1}{2}{\left( \mathcal{L}_k^{\dagger} \mathcal{L}_k \rho-\rho \mathcal{L}_k^{\dagger} \mathcal{L}_k \right)} \right]}$, where $\mathcal{H}=\mathcal{H}_0+\mathrm{\hat{B}}(\theta,{\varphi})$ is the total Hamiltonian including the hyperfine interactions $\mathcal{H}_0$ and the external magnetic field $\mathrm{\hat{B}}(\theta,{\varphi})$, $Q_S$ and $Q_T$ are the projectors into the singlet and triplet subspace of the radical pair state individually, and $[x,y]_{+}=xy+yx$. The above master equation is based on the Haberkorn approach [@Steiner89] by adding the dissipator $\mathcal{L}(\rho)$ which represent the environmental noise. We remark that it is not trace preserving as it is restricted in the subspace of the active radical pair state, and the singlet yield can then be calculated as $\mbox{Y}_{\mbox{s}}=k_S\int {\left\langle \mathcal{S} \right\vert} \rho(t){\left\vert \mathcal{S} \right\rangle}dt$, which is equivalent to Eq.(\[singletyield\]) in the absence of decoherence (see Ref.[@Hore10]). We consider three typical classes of environmental noise (that are independent on the magnetic field by themselves [@Gauger11]), namely the local dephasing model $\mathcal{L}_{I}=\{\sigma_z^{D},\sigma_z^{A}\}$ with $\sigma_z={{\left\vert \alpha \right\rangle}{\left\langle \alpha \right\vert} }-{{\left\vert \beta \right\rangle}{\left\langle \beta \right\vert} }$, the spin relaxation $\mathcal{L}_{II}=\{\sigma_{\pm}^{D} ,\sigma_{\pm}^{A}\}$ with $\sigma_{+}={{\left\vert \alpha \right\rangle}{\left\langle \beta \right\vert} }$ and $\sigma_{-}={{\left\vert \beta \right\rangle}{\left\langle \alpha \right\vert} }$, and the singlet-triplet dephasing model [@Shu91] $\mathcal{L}_{III}=\{\sigma_{ST}^{DA}\}$ with $\sigma_{ST}=2{\left\vert \mathcal{S} \right\rangle}{\left\langle \mathcal{S} \right\vert}-\mathbb{I}$, where the superscript $D$ and $A$ represent two radicals in a pair. These decoherence models are most relevant and were included to fit the experiment data [@Maeda12] but behave very different from each other in the sense that they destroy varying aspects of radical pair coherence: they will respectively lead to the decay of local electron spin coherence, of electron spin longitudinal component, and of singlet-triplet coherence at a rate given by $\xi$.
![The noise effect on the magnetic sensitivity of a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$) for the local dephasing (red, square), the electron spin relaxation (blue, circle), and the singlet-triplet dephasing (violet, diamond). (a) The magnetic sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as a function of the decoherence rate $\xi$ (in the unit of $\mu s^{-1}$). (b) The magnetic sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as a function of the coherence $\mathcal{C}(\mathcal{H}_0,\rho_0)$. In both panels, the reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, and the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:DEC"}](FADH_DEC_NEW_v1.pdf "fig:"){width="4.56cm"} ![The noise effect on the magnetic sensitivity of a radical pair (in which one radical contains three $^1$H nuclei and two $^{14}$N nuclei as FADH$\cdot$) for the local dephasing (red, square), the electron spin relaxation (blue, circle), and the singlet-triplet dephasing (violet, diamond). (a) The magnetic sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as a function of the decoherence rate $\xi$ (in the unit of $\mu s^{-1}$). (b) The magnetic sensitivity $\mbox{D}_{\mbox{s}}(\mathcal{H}_0,\rho_0)$ as a function of the coherence $\mathcal{C}(\mathcal{H}_0,\rho_0)$. In both panels, the reaction rates are $k_S=k_T=0.5 \mu s^{-1}$, and the Earth’s magnetic field is $b=50\mu T$.[]{data-label="fig:DEC"}](FADH_DEC_NEW_v2.pdf "fig:"){width="4.58cm"}
In Fig.\[fig:DEC\](a), one can see that the effects of these noise models on the magnetic sensitivity of the radical pair are quantitatively very different as a function of the same decoherence rate $\xi$. Such a fact is natural but gives no general insight into the problem how and why different types of noise would affect chemical magnetoreception to different extent. Instead, we calculate the global coherence $\mathcal{C}$ (see Eq.\[eq:gcm\]) under the influence of environmental noise and find a universal relation between the decoherence effects on the magnetic sensitivity and the coherence, which holds for different noise models studied here, see Fig.\[fig:DEC\](b). We have checked that such a universal relation also holds for mutations of the radical pair through (partially) deuteration, which will be interesting to study in spin chemistry experiment. This supports the observation that the global coherence is an appropriate concept to quantify the role of coherence in chemical magnetoreception. This is demonstrated by a study of different noise sources whose effect on the magnetic sensitivity can be predicted from their effect on the global coherence.
[*Discussion and outlook.—*]{} We have introduced the viewpoint of chemical magnetoreception as a quantum interferometer. This perspective allows us to reveal the direct connection between the global electron-nuclear spin coherence and the magnetic sensitivity of chemical magnetoreception for general molecules, it thus evidences coherence as a resource in the chemical compass model of avian magnetoreception in a similar way as a coherence-based quantum device. The verification of the present observation is conceivable either in spin chemistry experiments or by quantum simulation with well controllable systems, for example nitrogen-vacancy defects in diamond [@Dolde13; @Cai13], to emulate the radical pair dynamics. We remark that coherent manipulations and readout of electron or nuclear spins in diamond have been well developed in experiments. The concept of global coherence offers a unified perspective to predict the magnetic sensitivity of a chemical compass. Therefore, it is possible to use it to advance our understanding of the design principles of a chemical compass, namely to exhibit a better sensitivity the molecule shall embody a larger global coherence. This will facilitate the construction of an artificial chemical compass which is sensitive to the weak geomagnetic field at ambient temperature [@Maeda08], the demonstration of which would serve as an important intermediate step to understand how nature might be able to design such a chemical compass. We expect that the present ideas may help to further transfer the concepts and methods developed in quantum information to the field of spin chemistry, and gain new insights into the other coherent phenomena in spin chemistry, such as low field effect, coherence transfer in spin correlated radical pairs [@Wasie11] and other variants of radical pair model of magnetoreception [@Lovett12; @Carrillo13].
[*Acknowledgements.—*]{} We are grateful for the valuable discussions with Susana F. Huelga, Till Biskup, Ulrich E. Steiner, Erik Gauger, Peter Hore and Ilia Solov’yov. The work was supported by the Alexander von Humboldt Foundation, EU Integrating Projects Q-ESSENCE and SIQS, EU STREP PAPETS, ERC Synergy grant BioQ. J.-M.C was supported also by a Marie-Curie Intra-European Fellowship (FP7). Computations were performed on the bwGRiD.\
[99]{} \[1\][`#1`]{} \[2\]\[\][[\#2](#2)]{}
G. S. Engel, T. R. Calhoun, E. L. Read, T.-K. Ahn, T. Manal, Y.-C. Cheng, R. E. Blankenship, G. R. Fleming, Nature [**446**]{}, 782-786 (2007).
E. Collini, C. Y. Wong, K. E. Wilk, P. M. G. Curmi, P. Brumer, G. D. Scholes, Nature 463, 644-647 (2010).
M. Mohseni, P. Rebentrost, S. Lloyd, and A. Aspuru-Guzik, J. Chem. Phys. **129**, 174106 (2008).
M. B. Plenio and S. F. Huelga, New J. Phys. **10**, 113019 (2008).
G. D. Scholes, G. R. Fleming, A. Olaya-Castro, and R. van Grondelle, Nature Chemistry [**3**]{}, 763-774 (2011).
A. W. Chin, J. Prior, R. Rosenbach, F. Caycedo-Soler, S. F. Huelga, and M. B. Plenio, Nature Physics. [**9**]{}, 113-118 (2013).
L. Turin, J. Theor. Biol. **216**(3), 367 (2002).
J. C. Brookes, F. Hartoutsiou, A. P. Horsfield, and A. M. Stoneham, Phys. Rev. Lett. **98**, 038101 (2007).
I. A. Solov’yov, P.-Y. Chang, and K. Schulten, Phys. Chem. Chem. Phys. [**14**]{}, 13861-13871 (2012) .
K. Schulten, C. E. Swenberg, and A. Weller, Z. Phys. Chem **NF111**, 1-5 (1978).
T. Ritz, S. Adem, and K. Schulten, Biophys. J. [**78**]{}, 707-718 (2000).
Ilia A. Solov’yov, D. Chandler, and K. Schulten, Biophys. J **92**, 2711 (2007).
C. T. Rodgers and P. J. Hore, Proc. Natl. Acad. Sci. U.S.A. [**106**]{}, 353 (2009).
T. Ritz, R. Wiltschko, P. J. Hore, C. T. Rodgers, K. Stapput, P. Thalau, C. R. Timmel, W. Wiltschko, Biophys. J. [**96**]{}, 3451 (2009).
T. Ritz, M. Ahmad, H. Mouritsen, R. Wiltschko, W. Wiltschko, J. Royal. Soc. Interface [**7**]{}, S135-146 (2010).
K. Maeda, K. B. Henbest, F. Cintolesi, I. Kuprov, C. T. Rodgers, P. A. Liddell, D. Gust, C. R. Timmel, P. J. Hore, Nature [**453**]{}, 387-390 (2008).
T. Ritz, Procedia Chemistry [**3**]{}, 262 (2011).
J. A. Jones, P. J. Hore, Chem. Phys. Lett. 488, 90 (2010).
I. K. Kominis, Phys. Rev. E [**80**]{}, 056115 (2009); I. K. Kominis, Phys. Rev. E [**83**]{}, 056118 (2011).
J.-M. Cai, G. G. Guerreschi, and H. J. Briegel, Phys. Rev. Lett. [**104**]{}, 220502 (2010).
E. M. Gauger, E. Rieper, J. J. L. Morton, S. C. Benjamin, and V. Vedral, Phys. Rev. Lett. [**106**]{}, 040503 (2011).
C. Y. Cai, Qing Ai, H. T. Quan, C. P. Sun, Phys. Rev. A [**85**]{}, 022315 (2012).
J.-M. Cai, F. Caruso, and M. B. Plenio, Phys. Rev. A [**85**]{}, 040304 (2012).
J. N. Bandyopadhyay, T. Paterek, and D. Kaszlikowski, Phys. Rev. Lett. [**109**]{}, 110502 (2012); E. M. Gauger, S. C. Benjamin, Phys. Rev. Lett. [**110**]{}, 178901 (2013); J. N. Bandyopadhyay, T. Paterek, and D. Kaszlikowski, Phys. Rev. Lett. [**110**]{}, 178902 (2013).
H. J. Hogben, T. Biskup, and P. J. Hore, Phys. Rev. Lett. [**109**]{}, 220501 (2012).
G. G. Guerreschi, M. Tiersch, U. Steiner, H. J. Briegel, Chem. Lett. [**572**]{}, 106 (2013).
S. Johnsen and K. J. Lohmann, Nature Rev. Neurosci. [**6**]{}, 703-712 (2005).
H.‐J. Werner, Z. Schulten, and K. Schulten, J. Chem. Phys. [**67**]{}, 646 (1977).
H. Mouritsen, U. Janssen-Bienhold, M. Liedvogel, G. Feenders, J. Stalleicken, P. Dirks, R. Weiler, Proc. Natl Acad. Sci. USA [**101**]{}, 14294-14299 (2004).
J. C. Weaver, T. E. Vaughan and R. D. Astumian, Nature [**405**]{}, 707-709 (2000).
F. Cintolesi, T. Ritz, C. W. M. Kay, C. R. Timmel and P. J. Hore, Chem. Phys. [**294**]{}, 385-399 (2003).
I. A. Solov’yov, D. E. Chandler and K. Schulten, Biophys. J. [**92**]{}, 2711-2726 (2007).
M. Ahmad, P. Galland, T. Ritz, R. Wiltschko and W. Wiltschko, Planta [**225**]{}, 615-624 (2007).
M. Zapka, D. Heyers, C. M. Hein, S. Engels, N.-L. Schneider, J. Hans, S. Weiler, D. Dreyer, D. Kishkinev, J. M. Wild, H. Mouritsen, Nature **461**, 1274-1277 (2009).
K. Maeda, A. J. Robinson, K. B. Henbest, H. J. Hogben, T. Biskup, M. Ahmad, E. Schleicher, S. Weber, C. R. Timmel, and P. J. Hore, Proc. Natl. Acad. Sci. U.S.A. [**109**]{}, 4774 (2012).
I. A. Solov’yov, H. Mouritsen, K. Schulten, Biophy. J. [**99**]{}, 40-49 (2010).
I. A. Solov’yov, T. Domratcheva, A. R. M. Shahi, and K. Schulten, Journal of the American Chemical Society. [**134**]{}, 8046-1805 (2012).
M. Liedvogel, H. Mouritsen, J. Roy. Soc. Interface. [**7**]{}, S147-S162 (2010).
J. B Phillips, and S. C. Borland, Nature [**359**]{}, 142-144 (1992).
W. Wiltschko, R. Wiltschko, J. Comp. Physiol. A. [**191**]{}, 675-693 (2005).
T. Ritz, P. Thalau, J. B. Phillips, R. Wiltschko and W. Wiltschko, Nature [**429**]{}, 177-180 (2004).
U. E. Steiner, T. Ulrich, Chem. Rev. [**89**]{}, 51 (1989).
D. J. Wineland, J. J. Bollinger, W. M. Itano, and D. J. Heinzen, Phys. Rev. A [**50**]{}, 67-88 (1994).
D. Budker, and M. Romalis, Nature Physics [**3**]{}, 227 - 234 (2007).
A. Shaji and C. M. Caves, Phys. Rev. A [**76**]{}, 032111 (2007).
O. Efimova and P. J. Hore, Biophys. J., **94**, 1565-1574 (2008).
See [Supplemental Material](http://link.aps.org/supplemental/10.1103/PhysRevLett.111.230503) for the calculation details and more discussions.
The radical pair that is responsible for the magnetic sense of birds has not been conclusively determined. In the present work, we thus explore the essential features of typical radical pair molecules (which in total consist of more than three up to seven nuclei with random hyperfine interaction strengths [@SI]) rather than restricting to the radical pairs with very specific hyperfine configurations, particularly we go beyond the simplified model with only one single nucleus [@Gauger11; @Cai12; @Hogben12; @Band12]. Besides the typical behaviors of general radical pair molecules, we also take the radical pair inspired by the flavin adenine dinucleotide (FADH$\cdot$) formed photochemically in cryptochromes as an explicit example to further elaborate our results.
J.-M. Cai, F. Jelezko, M. B. Plenio, A. Retzker, New J. Phys. [**15**]{}, 013020 (2013).
J.-M. Cai, A. Retzker, F. Jelezko, M. B. Plenio, Nature Physics [**9**]{}, 168 (2013).
P. London, J. Scheuer, J.-M. Cai, I. Schwarz, A. Retzker, M.B. Plenio, M. Katagiri, T. Teraji, S. Koizumi, J. Isoya, R. Fischer, L. P. McGuinness, B. Naydenov, and F. Jelezko, Phys. Rev. Lett. [**111**]{}, 067601 (2013).
A. I. Shushin, Chem. Phys. Lett. [**181**]{}, 274-278 (1991).
F. Dolde, I. Jakobi, B. Naydenov, N. Zhao, S. Pezzagna, C. Trautmann, J. Meijer, P. Neumann, F. Jelezko, J. Wrachtrup, Nature Physics [**9**]{}, 139 (2013).
M. T. Colvin, A. Butler Ricks, A. M. Scott, A. L. Smeigh, R. Carmieli, T. Miura, and M. R. Wasielewski, J. Am. Chem. Soc. [**133**]{}, 1240-1243 (2011).
B. W. Lovett, et al, Biophys. J., [**102**]{}, 961 (2012).
A. Carrillo, M. F. Cornelio, M. C. de Oliveira, arXiv:1304.3452.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We use the coupled cluster method in high orders of approximation to make a comprehensive study of the ground-state (GS) phase diagram of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on a two-dimensional honeycomb lattice with antiferromagnetic (AFM) interactions up to third-nearest neighbors. Results are presented for the GS energy and the average local on-site magnetization. With the nearest-neighbor coupling strength $J_{1} \equiv 1$ we find four magnetically ordered phases in the parameter window $J_{2},J_{3} \in [0,1]$, namely the Néel (N), striped (S), and anti-Néel (aN) collinear AFM phases, plus a spiral phase. The aN phase appears as a stable GS phase in the classical version of the model only for values $J_{3}<0$. Each of these four ordered phases shares a boundary with a disordered quantum paramagnetic (QP) phase, and at several widely separated points on the phase boundaries the QP phase has an infinite susceptibility to plaquette valence-bond crystalline order. We identify all of the phase boundaries with good precision in the parameter window studied, and we find three tricritical quantum critical points therein at: (a) $(J_{2}^{c_1},J_{3}^{c_1})=(0.51 \pm 0.01,0.69 \pm 0.01)$ between the N, S, and QP phases; (b) $(J_{2}^{c_2},J_{3}^{c_2})=(0.65
\pm 0.02,0.55 \pm 0.01)$ between the S, spiral, and QP phases; and (c) $(J_{2}^{c_3},J_{3}^{c_3})=(0.69 \pm 0.01,0.12 \pm 0.01)$ between the spiral, aN, and QP phases.
author:
- 'P. H. Y. Li and R. F. Bishop'
- 'D. J. J. Farnell'
- 'C. E. Campbell'
title: |
Phase diagram of a frustrated Heisenberg antiferromagnet on the honeycomb lattice:\
the $J_{1}$–$J_{2}$–$J_{3}$ model
---
INTRODUCTION
============
Phase transitions are common phenomena in nature and they have long been the subject of intense theoretical interest. Since they are driven by thermal and/or quantum fluctuations, an understanding of them at the microscopic level requires a good many-body description. Particular interest in recent years has centered on the so-called quantum phase transitions that occur in systems in their ground state at zero temperature ($T=0$) as some parameters in the Hamiltonian describing them are varied.
Spin-lattice systems with competing interactions provide a particularly rich field in which to study such quantum phase transitions. The exchange interactions that lead to collective magnetic behavior in spin systems are clearly quantum-mechanical in origin. One knows too that the interplay between reduced dimensionality, frustration (due to either competing interactions or the underlying lattice geometry), and strong quantum fluctuations generates a huge variety of new states of condensed matter beyond the usual states of quasiclassical long-range order (LRO). A particularly rich place to observe such exotic ground-state (GS) phases is in situations where they originate as a result of quantum fluctuations within a large set of configurations that are degenerate at the classical level.[@Ramirez:2008; @Balents:2010]
The search for such exotic phases that owe their existence purely to quantum effects is nowadays one of the primary reasons for the study of frustrated quantum spin-lattice systems. The interplay between magnetic frustration and quantum fluctuations provides a powerful mechanism for disturbing, destabilizing, or even completely destroying magnetic order. The search for a genuine quantum spin-liquid (QSL) phase,[@Ramirez:2008; @Balents:2010; @Anderson:1973; @Anderson:1987] which has no magnetic order and no LRO or long-range correlations of any kind, has itself attracted huge theoretical interest ever since its first proposal nearly 40 years ago by Anderson,[@Anderson:1973] and its subsequent pursuit by, for example, Shastry and Sutherland[@shastry1], and many others following them.
The GS wave function of a QSL is clearly a superposition of a large number of different configurations. An example could be the resonating valence-bond (RVB) state, which is a superposition of many short-range singlet valence bonds. The RVB state was itself first proposed by Fazekas and Anderson[@Fazekas:1974] as the GS wave function for the spin-$\frac{1}{2}$ isotropic Heisenberg antiferromagnet (HAFM) on the geometrically frustrated, two-dimensional, (2D) triangular lattice with only nearest-neighbor (NN) interactions, all of equal strength, although it is now known that this proposal is incorrect and that this system is magnetically ordered.
In a genuine QSL state no symmetry is broken and the quantum fluctuations are required to form a many-body singlet state that contains no long-range correlations with respect to [*any*]{} operator, although there may be present some form of topological order. Other exotic quantum paramagnetic (QP) states, which also have no magnetic LRO but which break some spatial symmetry with respect to short-range magnetic correlations, can also arise. The various QP valence-bond crystalline (VBC) solid phases fall into this category.
As we have noted, a combination of strong quantum fluctuations and strong frustration in a spin system provides an ideal scenario for the emergence of such novel quantum GS phases as the QSL and other QP phases discussed above, which do not possess the magnetic LRO that typifies the classical GS phases of the corresponding models taken in the limit $s \to \infty$ of the spin quantum number $s$ of the lattice spins. We know that quantum fluctuations tend to be largest for the smallest values of $s$, for lower dimensionality $D$ of the lattice, and for the smallest coordination number $z$ of the lattice. Thus, for spin-$\frac{1}{2}$ models the honeycomb lattice plays a special role since its coordination number ($z=3$) is the lowest possible for $D=2$. Frustration is easily incorporated by the inclusion of competing next-nearest-neighbor (NNN) and possibly also next-next-nearest-neighbor (NNNN) bonds. For these reasons such spin-$\frac{1}{2}$ frustrated Heisenberg models on the honeycomb lattice have engendered huge theoretical interest.[@Rastelli:1979; @Mattson:1994; @Fouet:2001; @Mulder:2010; @Cabra:2011; @Ganesh:2011; @Clark:2011; @Reuther:2011; @DJJF:2011_honeycomb; @Albuquerque:2011; @Mosadeq:2011; @Oitmaa:2011; @Mezzacapo:2012; @PHYLi:2012_honeycomb_J1neg; @PHYLi:2012_honeyJ1-J2; @PHYLi:2012_Honeycomb_J2neg]
Interest in the honeycomb lattice has been given further impetus by the discovery of a QSL phase in the exactly solvable Kitaev model,[@kitaev] in which the spin-$\frac{1}{2}$ particles are sited on a honeycomb lattice. Additional interest has also emanated from the recent synthesis of graphene monolayers[@graphene] and other magnetic materials with a honeycomb structure. For example, it is likely that Hubbard-like models on the honeycomb lattice may well describe many of the physical properties of graphene. In this context it is particularly interesting to note the clear evidence of Meng [*et al.*]{}[@meng] that quantum fluctuations are strong enough to trigger an insulating QSL phase between the non-magnetic metallic phase and the antiferromagnetic (AFM) Mott insulator for the Hubbard model on the honeycomb lattice at moderate values of the Coulomb repulsion $U$. This latter Mott insulator phase corresponds in the limit $U \to \infty$ to the pure HAFM on the bipartite honeycomb lattice, whose GS phase exhibits Néel LRO. However, higher-order terms in the $t/U$ expansion of the Hubbard model (where $t$ is the Hubbard hopping term strength parameter) lead to frustrating exchange couplings in the corresponding spin-lattice limiting model (and see, e.g., Ref. \[\]), in which the HAFM with NN exchange couplings is the leading term in the large-$U$ expansion.
The unexpected result of Meng [*et al.*]{},[@meng] together with other related work,[@Yang:2011_hcomb; @Vaezi:2010; @Vaezi:2011] has excited much interest in understanding the physics of frustrated quantum magnets on the honeycomb lattice. In particular, a growing consensus is emerging[@Mattson:1994; @Fouet:2001; @Cabra:2011; @Mosadeq:2011; @Albuquerque:2011; @Reuther:2011; @Mezzacapo:2012] that frustrated spin-$\frac{1}{2}$ HAFMs on the honeycomb lattice exhibit a frustration-induced QP phase. It is interesting to note in this context that recent experiments[@exp] on the layered compound Bi$_3$Mn$_4$O$_{12}$(NO$_3$) (BMNO) at temperatures below its Curie-Weiss temperature reveal QSL-like behavior. In BMNO the Mn$^{4+}$ ions are situated on the sites of (weakly-coupled) honeycomb lattices, although they have spin quantum number $s = \frac{3}{2}$. The successful substitution of the $s=\frac{3}{2}$ Mn$^{4+}$ ions in BMNO by V$^{4+}$ ions could lead to a corresponding experimental realization of a spin-$\frac{1}{2}$ HAFM on the honeycomb lattice.
Other realizations of quantum HAFMs which exhibit the honeycomb structure include magnetic compounds such as InNa$_{3}$Cu$_{2}$SbO$_{6}$[@Miura:2006] and InCu$_{2/3}$V$_{1/3}$O$_{3}$.[@Kataev:2005] In both of these materials the Cu$^{2+}$ ions in the copper oxide layers form a spin-$\frac{1}{2}$ HAFM on a (distorted) honeycomb lattice. Other similar honeycomb materials include the family of compounds BaM$_2$(XO$_4$)$_2$ (M=Co, Ni; X=P, As),[@Regnault:1990] in which the magnetic ions M are disposed in weakly-coupled layers where they are situated on the sites of a honeycomb lattice. The Co ions have spins $s = \frac{1}{2}$ and the Ni ions have spins $s = 1$. Recent calculations[@Tsirlin:2010] of the material $\beta$-Cu$_2$V$_2$O$_7$ have demonstrated that its properties can also be described in terms of a spin-$\frac{1}{2}$ model on an (anisotropic) honeycomb lattice.
Finally, we note that the very recent prospect of being able to realize spin-lattice models with ultracold atoms trapped in optical lattices[@Struck:2011] is likely to make even more data available about the quantum phase transitions in the models as the exciting possibility opens up in such trapped-atom experiments to tune the strengths of the competing magnetic bonds, and hence to drive the system from one phase to another.
Recently, we have made a series of studies of the frustrated spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice using the coupled cluster method (CCM) complemented in some cases with the Lanczos exact diagonalization (ED) of small lattices. We have studied various regimes in the full $(J_{1},J_{2},J_{3})$ parameter space of NN ($J_1$) bonds, NNN ($J_2$) bonds, and NNNN ($J_3$) bonds. These include (a) the AFM model (i.e., with $J_{1}>0$) in the special case where the NNN and NNNN bonds are also AFM and have equal strength ($J_{3}=J_{2} \equiv \kappa J_{1} > 0$);[@DJJF:2011_honeycomb] (b) the ferromagnetic (FM) model (i.e., with $J_{1}<0$) with frustrating NNN and NNNN bonds, again in the special case where they are both AFM and have equal strength ($J_{3}=J_{2}>0$);[@PHYLi:2012_honeycomb_J1neg] (c) the AFM model (i.e., with $J_{1}>0$) in the special case where the NNN and NNNN bonds are both FM and have equal strength ($J_{3} = J_{2} < 0$);[@PHYLi:2012_Honeycomb_J2neg] and (d) the AFM model (i.e., with $J_{1}>0$) in the special case where we have frustrating NNN bonds only (i.e., with $J_{2} \equiv x J_{1}>0$; $J_{3}=0$).[@PHYLi:2012_honeyJ1-J2] The aim of the present work is to extend the investigation of the phase diagram of the full spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice in the case where all of the NN, NNN, and NNNN bonds are AFM, but no further restriction is made except that we limit the parameter space window to $J_{2}/J_{1},J_{3}/J_{1} \in [0,1]$.
We briefly outline the structure of the rest of the paper. The model itself is first described in Sec. \[model\_section\], before briefly outlining the CCM formalism that we employ as our main calculational tool in Sec. \[CCM\]. To aid the reader we first give an overview of our main results in Sec. \[preview\], focussing on the phase diagram for the model, before we give a detailed presentation and discussion of our results in Sec. \[results\]. Finally, we conclude in Sec. \[summary\] with a summary and comparison of our results with the work of others.
THE HONEYCOMB MODEL {#model_section}
===================
The spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice, or special cases of it (e.g., when $J_{3}=J_{2}$ or $J_{3}=0$) have been intensively studied by many authors (see, e.g., Refs. \[\] and references cited therein). The Hamiltonian of the model is $$H = J_{1}\sum_{\langle i,j \rangle} \mathbf{s}_{i}\cdot\mathbf{s}_{j} +
J_{2}\sum_{\langle\langle i,k \rangle\rangle} \mathbf{s}_{i}\cdot\mathbf{s}_{k} +
J_{3}\sum_{\langle\langle\langle i,l \rangle\rangle\rangle} \mathbf{s}_{i}\cdot\mathbf{s}_{l}\,,
\label{eq1}$$ where $i$ runs over all lattice sites, and where $j$, $k$, and $l$ run over all NN sites, all NNN sites, and all NNNN sites to $i$, respectively, counting each bond once and once only. Each site $i$ of the lattice carries a spin-$s$ particle represented by an SU(2) spin operator ${\bf s}_{i}=(s^{x}_{i},s^{y}_{i},s^{z}_{i})$. We restrict ourselves here to the case $s=\frac{1}{2}$. The lattice and the exchange bonds are illustrated in Fig. \[model\].
Before discussing the $s=\frac{1}{2}$ version of the model that is the topic of the present paper it is useful to consider first the classical limit (i.e., $s \rightarrow \infty$). Thus the $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice has six classical GS phases in the case where $J_{1}>0$ (as considered here) and where $J_2$ and $J_{3}$ are arbitrary (i.e., can take either sign).[@Rastelli:1979; @Fouet:2001] The six phases comprise three collinear AFM phases, the FM phase, plus two different helical phases (and see, e.g., Fig. 2 of Ref. ). The three AFM phases are the Néel (N) phase, the striped (S) phase and the anti-Néel (aN) phase as shown in Figs. \[model\](a), (b), and (d) respectively. The S, aN, and N states have, correspondingly, 1, 2, and all 3 NN spins to a given spin antiparallel to it. Similarly, if we consider the sites of the honeycomb lattice as comprising a set of parallel sawtooth (or zigzag) chains (in any one of the three equivalent directions), the S state consists of alternating up-spin and down-spin FM chains, whereas the aN state consists of AFM chains in which NN spins on adjacent chains are parallel.
Although at $T=0$ there exists an infinite family of non-coplanar states degenerate in energy with respect to each of the S and aN states, both thermal and quantum fluctuations [@Fouet:2001] favor the collinear configurations. When $J_{3}>0$ the spiral state shown in Fig. \[model\](c) is the stable classical GS phase in some region of the parameter space. This state is characterized by a pitch vector perpendicular to one of the three equivalent $J_1$ bond directions, and a single spiral angle defined so that as we move along the parallel sawtooth chains \[drawn in the horizontal direction in Fig. \[model\](c)\] the spin angle increases by $\pi + \phi$ from one site to the next, and with NN spins on adjacent chains antiparallel. The classical GS energy for this spiral state is minimized when the pitch angle takes the value $\phi =
\cos ^{-1} [\frac{1}{4}(J_{1}-2J_{2})/(J_{2}-J_{3})]$. The corresponding minimum value for the GS energy per spin is then given as $$\frac{E^{\rm cl}_{\rm spiral}}{N} = \frac{s^{2}}{2}\left (-J_{1} - 2J_{2} + J_{3} - \frac{1}{4}
\frac{(J_{1} - 2J_{2})^{2}}{(J_{2} - J_{3})} \right ).$$ When $\phi \rightarrow 0$ this spiral state simply becomes the collinear N state with a corresponding energy per spin given by $$\frac{E^{\rm cl}_{\rm N}}{N} = \frac{s^{2}}{2}\left (-3J_{1} + 6J_{2} - 3J_{3}\right).$$ Clearly, the phase transition between this spiral state and the N state is of a continuous nature and the corresponding phase boundary is given by the equation $y=\frac{3}{2}x-\frac{1}{4}$, for $\frac{1}{6}<x<\frac{1}{2}$, where $y\equiv J_{3}/J_{1}$ and $x\equiv J_{2}/J_{1}$.
Similarly, when $\phi \rightarrow \pi$ this spiral state becomes the collinear S state with a corresponding GS energy per spin given by $$\frac{E^{\rm cl}_{\rm S}}{N} = \frac{s^{2}}{2}\left (J_{1} - 2J_{2} - 3J_{3}\right).$$ The classical spiral and S states undergo a continuous phase transition along their common phase boundary $y=\frac{1}{2}x+\frac{1}{4}$, for $x>\frac{1}{2}$. Furthermore there is a first-order phase transition between the collinear N and S states along the boundary line $x=\frac{1}{2}$, for $y>\frac{1}{2}$. These three phases (N, S, and spiral) meet at the tricritical point $(x,y) = (\frac{1}{2},\frac{1}{2})$. We note too that as $x
\rightarrow \infty$ (for a fixed finite value of $y$) the spiral pitch angle $\phi \rightarrow \frac{2}{3} \pi$. Thus in this limiting case the classical model simply becomes two HAFMs on weakly connected interpenetrating triangular lattices, with the classical triangular-lattice ordering of NN spins oriented at an angle $\frac{2}{3} \pi$ to each other on each sublattice.
When $y>0$ (and $J_{1}>0$) the above three states are the only classical GS phases. When $y<0$ the N state persists in a region bounded by the same boundary line as above, $y=\frac{3}{2}x-\frac{1}{4}$, for $-\frac{1}{2}<x<\frac{1}{6}$, on which it continuously meets a second spiral state, and by the boundary line $y=-1$, for $x<-\frac{1}{2}$, at which it undergoes a first-order transition to the FM state, which itself is the stable GS phase in the region $x<-\frac{1}{2}$ and $y<-1$. Another collinear AFM state, the aN state shown in Fig. \[model\](d), with a GS energy per spin given by $$\frac{E^{\rm cl}_{\rm aN}}{N} = \frac{s^{2}}{2}\left (-J_{1} - 2J_{2} + 3J_{3}\right),$$ becomes the stable GS phase in the region $x>\frac{1}{2}$, for $y < \frac {1}{2}\{x -[x^{2}+2(x-\frac{1}{2})^{2}]^{1/2} \}$. On the boundary it undergoes a first-order transition to the spiral state shown in Fig. \[model\](c).
Finally, for $\frac{1}{6}<x<\frac{1}{2}$ the spiral state shown in Fig. \[model\](c) meets a second spiral GS phase on the boundary line $y=0$, along which there is a first-order transition between the two spiral states. This second spiral phase is characterized by a pitch vector that is parallel to one of the $J_1$ bond directions and by two spiral angles $\alpha$ and $\beta$. Within a unit cell the spin directions deviate from those in the N state by an angle $\alpha$, and as one advances from one unit cell to the next there is a twist by an angle $\beta$. Both of the two pitch angles of this second spiral phase smoothly approach the value zero along the above boundary with the N state, and the value $\pi$ along a second boundary curve that joins the points $(x,y) = (-\frac{1}{2},-1)$ and $(\frac{1}{2},0)$, on which it meets the aN state. Both transitions are continuous in nature. This second spiral phase meets the three collinear states N, aN, and FM at the tetracritical point $(x,y) = (-\frac{1}{2},-1)$.
In this paper we further our study of the spin-$\frac{1}{2}$ model of Eq. (\[eq1\]) on the honeycomb lattice, restricting ourselves to the case where all of the bonds are antiferromagnetic in nature. Thus, henceforth we set $J_{1} \equiv 1$ to set the overall energy scale, and we work here within the parameter space window $J_{2},J_{3} \in [0,1]$.
Recently we have used the CCM to study the special $J_{1}$–$J_{2}$ case of this model (and in this parameter window with $J_{1}>0$ and $J_{2} \equiv xJ_{1}>0$) in which $J_{3}=0$.[@PHYLi:2012_honeyJ1-J2] We found a paramagnetic plaquette valence-bond crystalline (PVBC) phase for $x_{c_1}<x<x_{c_2}$, where $x_{c_1} \approx 0.207
\pm 0.003$ and $x_{c_2} \approx 0.385 \pm 0.010$. We found that the transition at $x_{c_1}$ to the N phase appeared to be of a continuous deconfined type (although we could not exclude a very narrow intermediate phase in the range $0.21 \lesssim x \lesssim 0.24$), while that at $x_{c_2}$ to the aN phase appeared to be of first-order type. As we noted above the aN phase exists in the classical version of the $J_{1}$–$J_{2}$ model only at the isolated and highly degenerate critical point $x = \frac{1}{2}$. The spiral phases that are present classically for all values $x > \frac{1}{6}$ were found to be absent for all $x \lesssim 1$.
We have also separately used the CCM to study the spin-$\frac{1}{2}$ model of Eq. (\[eq1\]) on the honeycomb lattice in the special case where $J_{3}=J_{2}\equiv \kappa J_{1}>0$ and $J_{1}>0$ (i.e. along the line $y=x \equiv \kappa$).[@DJJF:2011_honeycomb] We also found a PVBC phase in this case for $\kappa_{c_1}<\kappa<\kappa_{c_2}$, where $\kappa_{c_1} \approx 0.47$ and $\kappa_{c_2} \approx 0.60$. Once again, the evidence favored the transition at $\kappa_{c_1}$ to the N phase to be of a continuous deconfined type, while that at $\kappa_{c_2}$ to the S phase appeared to be first-order in nature.
In order to shed more light on the $J_{1}$–$J_{2}$–$J_{3}$ HAFM model (i.e., with $J_{1}>0$) on the honeycomb lattice, we now extend its study to map out its entire phase diagram in the parameter regime $x,y \in [0,1]$. In view of the proven success of using the CCM on the above special cases of this model, we continue to utilize it, and in Sec. \[CCM\] we accordingly first briefly outline the method as it is applied here.
THE CCM FORMALISM {#CCM}
=================
The CCM is a widely used microscopic many-body technique. It is has been demonstrated to be particularly efficient and very accurate in handling a wide variety of highly frustrated quantum magnets. Such frustrated systems are notoriously challenging at the theoretical level. Only a very limited number of established numerical methods exist for their accurate treatment, and other recent and very promising approaches such as those based on projected entangled pair states[@peps] have not yet been sufficiently widely tested to be able to evaluate properly their accuracy and efficacy.
Of the other well established and widely used techniques we note that exact diagonalization (ED) techniques, which involve the finite-size extrapolation of numerical exact data for finite-lattice systems, are much more challenging for the present honeycomb-lattice model than for comparable square-lattice models, to which they have been very efficiently and accurately applied (see, e.g., Refs. \[\]). The reasons include the facts that for the honeycomb lattice the unit cell now contains two sites, and that there exist relatively fewer finite-sized lattices (than in the square-lattice case) that are small enough for ED techniques to be used but which also contain the full point-group symmetry.[@Fouet:2001] Also, quantum Monte Carlo (QMC) methods are severely restricted in the presence of frustration by the well-known “minus-sign problem”.
By contrast the CCM, when evaluated to high orders in one of its systematic approximation hierarchies, as described below, has been proven through a huge variety of applications to provide a powerful tool both to determine with good accuracy the positions of quantum critical points, [@DJJF:2011_honeycomb; @PHYLi:2012_honeycomb_J1neg; @PHYLi:2012_honeyJ1-J2; @PHYLi:2012_Honeycomb_J2neg; @Reuther:2011_J1J2J3mod; @Kr:2000; @rachid05; @schmalfuss; @Bi:2008_PRB; @Bi:2008_JPCM; @darradi08; @Bishop:2009; @richter10; @UJack_ccm; @Bishop:2012_checkerboard] and to classify the nature of any QP phases in the system.[@DJJF:2011_honeycomb; @PHYLi:2012_honeyJ1-J2; @darradi08; @Bishop:2012_checkerboard] As noted above, we have also used the CCM very successfully in some previous applications to honeycomb-lattice models,[@DJJF:2011_honeycomb; @PHYLi:2012_honeycomb_J1neg; @PHYLi:2012_honeyJ1-J2; @PHYLi:2012_Honeycomb_J2neg] and for all these reasons we now employ it again here.
The CCM is a size-extensive method, in which the limit $N \rightarrow \infty$, where $N$ is the number of lattice spins, may automatically be imposed from the outset. The many-body system under study is assumed to have exact ket and bra GS energy eigenvectors, $|\Psi\rangle$ and $\langle \tilde{\Psi}|$ respectively, which satisfy the corresponding Schrödinger equations, $$H|\Psi\rangle=E|\Psi\rangle\,, \quad \langle \tilde{\Psi}|H=E\langle\tilde{\Psi}| \,,$$ and which are chosen to have the normalization $\langle \tilde{\Psi}|\Psi \rangle = 1$, i.e., $\langle \tilde{\Psi}| = \langle \Psi|/\sqrt{\langle \Psi|\Psi\rangle}$. The quantum correlations present in the exact ground state are expressed systematically in the CCM with respect to some suitable normalized model (or reference) state, $|\Phi\rangle$.[@ccm2; @j1j2_square_ccm1; @ccm3] It is common practice to choose simple quasiclassical states as CCM reference states, although other choices are certainly possible. In this study we choose various classical model states as our CCM model states, namely: (a) the Néel, (b) the striped, (c) the spiral, and (d) the anti-Néel states shown in Fig. \[model\], as we discuss below.
The model state $|\Phi\rangle$ is required to be a fiducial vector in the sense that all possible ket states in the many-body Hilbert space can be obtained by acting on it with an appropriate linear combination of mutually commuting many-body creation operators, $C^{+}_{I}$, which may be defined with respect to the model state. The operators $C^{+}_{I} \equiv (C^{-}_{I})^{\dagger}$, with $C_{0}^{\dagger} \equiv 1$, thus have the property that $\langle \Phi|C^{+}_{I} = 0 = C_{I}^{-}|\Phi\rangle = 0$; $\forall I \neq 0$. The CCM parametrizations of the exact ket and bra GS wave functions are given in terms of the usual exponentiated forms, $$|\Psi\rangle = {\rm e}^{S}|\Phi\rangle\,, \quad \langle\tilde{\Psi}| = \langle \Phi|\tilde{S}{\rm e}^{-S} \,,$$ where the CCM correlation operators $S$ and $\tilde{S}$ are themselves expressed as generalized multiconfigurational creation and destruction operators respectively, $$S=\sum_{i}{\cal S}_{I} C^{+}_{I}\,, \quad \tilde{S}_{I}=1 +\sum_{i} {\tilde{\cal S}}_{i} C^{-}_{I} \,\,,
\forall I \neq 0 \,.$$ Clearly these parametrizations satisfy the normalization relations $\langle \tilde{\Psi}|\Psi \rangle = \langle\Phi|\Psi\rangle = \langle\Phi|\Phi\rangle \equiv 1$.
The set of correlation coefficients (${\cal S}_{I}$, $\tilde{\cal S}_{I}$) is now determined by requiring the energy expectation value $\bar{H}\equiv\langle\tilde{\Psi}|H|\Psi\rangle$ to be a minimum with respect to each of the correlation coefficients themselves. This will result in the coupled sets of equations $\langle \Phi |C^{-}_{I}{\rm
e}^{-S}H{\rm e}^{S}|\Phi\rangle=0$ and $\langle\Phi|\tilde{S}({\rm e}^{-S}H{\rm e}^{S} -
E)C^{+}_{I}|\Phi\rangle=0$; $\forall I \neq 0$, which we normally solve for the correlation coefficients (${\cal S}_{I}$, $\tilde{\cal S}_{I}$) using parallel computing routines once the specific truncation scheme is specified, as described further below.
In order to treat each lattice site in the spin system on an equal basis it is extremely convenient to rotate the local spin-axes on each site in such a way that all the spins of each CCM reference state used point along the negative $z$-direction. Such rotations in spin space are obviously canonical transformations that have no effect on the fundamental SU(2) commutation relations. The spins of our system are then represented entirely by these locally defined spin coordinate frames. The multispin creation operators may be written as linear sums of products of the individual spin raising operators $s^{+}_{k} \equiv s^{x}_{k} + is^{y}_{k}$, i.e., $C^{+}_{I} \equiv
s^{+}_{k_{1}}s^{+}_{k_{2}} \cdots s^{+}_{k_{n}}$. After calculation of the correlation coefficients (${\cal S}_{I}$, $\tilde{\cal S}_{I}$), we can then calculate the GS energy using $E=\langle\Phi|{\rm
e}^{-S}H{\rm e}^{S}|\Phi\rangle$, and the magnetic order parameter, which is defined to be the average local on-site magnetization, $M \equiv -\frac{1}{N}\langle\tilde{\Psi}| \sum^{N}_{i=1} s^{z}_{i}|
\Psi\rangle$, with respect to the local rotated spin coordinates described above.
If we include all possible multispin configurations for the calculation of the correlation coefficients (${\cal S}_{I}$, $\tilde{\cal S}_{I}$), then the CCM formalism becomes exact. Of course in practice one needs to truncate the set, and there are several well-developed and systematically improvable truncation hierarchies that have been extremely widely tested by now. For spin-$\frac{1}{2}$ systems we usually use the well-established localized LSUB$m$ truncation scheme where we keep at a given truncation level specified by the truncation index $m$ only all of those multi-spin configurations which may be defined over all possible lattice animals (or polyominos) of size $m$ on the lattice. A lattice animal (or polyomino) of size $m$ is defined as a set of $m$ contiguous sites in the usual graph-theoretic sense where every site is adjacent (in the nearest-neighbor sense) to at least one other site. The method of solving for higher orders of LSUB$m$ approximations is well documented in Refs. \[\], to which the interested reader is referred for further details.
Table \[table\_FundConfig\] shows the number $N_f$ of fundamental configurations that are inequivalent after all space and point-group symmetries of both the Hamiltonian and the model state have been taken into account, for each of the Néel, striped, spiral, and anti-Néel model states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice.
0.2cm
-------- ------ --------- ----------- --------
Method
Néel striped anti-Néel spiral
LSUB4 5 9 9 66
LSUB6 40 113 85 1080
LSUB8 427 1750 1101 18986
LSUB10 6237 28805 17207 347287
-------- ------ --------- ----------- --------
: Number of fundamental configurations, $N_{f}$, for the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model ($J_{1}=1$) on the honeycomb lattice, using the Néel, striped, anti-Néel, and spiral states.
\[table\_FundConfig\]
We note that the number $N_f$ of such independent spin configurations taken into account in the CCM correlation operators $S$ and $\tilde{S}$ increases rapidly with the truncation index $m$. Clearly the number of independent configurations is smaller for states such as the Néel state that have a higher degree of point-group symmetry, and for which we can utilize conservation laws such as $s^{z}_{T}=0$, where ${\bf s}_{T} \equiv \sum_{i=1}^{N} {\bf s}_{i}$ is the total spin operator referred to the global spin coordinates.
Clearly the spiral state has the largest number $N_f$, for a given level of LSUB$m$ approximation, from among our four model states, and for this state we are limited to values $m \leq 10$ even with the use of massively parallel computing to derive and solve the corresponding coupled sets of CCM bra- and ket-state equations.[@ccm] For the spiral state we note too that we have the additional computational cost that the pitch angle $\phi$ at a given LSUB$m$ level must be chosen to minimize the corresponding estimate for the GS energy. For specified values of each of the exchange parameters $J_{2}$ and $J_{3}$ (with $J_{1} \equiv 1$) a typical computational run for the spiral phase at the LSUB$10$ level typically requires about 5$\,$h computing time using 3000 processors simultaneously.
Although the CCM works from the outset in the limit $N \to \infty$ of an infinite number of spins, and hence the need for any finite-size scaling is obviated, we do still need to extrapolate the LSUB$m$ data to reach results in the exact $m \rightarrow \infty$ limit. Although there are no known exact extrapolation rules, by now there exists a wealth of empirical experience in extrapolating the GS energy, $E$, and the magnetic order parameter (i.e., the average local on-site magnetization), $M$. For the GS energy per spin, $E/N$, a well-established and very accurate extrapolation ansatz (see, e.g., Refs. \[\]) is $$E(m)/N = a_{0}+a_{1}m^{-2}+a_{2}m^{-4}\,, \label{E_extrapo}$$ whereas for the magnetic order parameter, $M$, we use different schemes depending on different circumstances, specifically on whether the system is highly frustrated or not. Thus, for systems with a GS order-disorder transition or with a considerable degree of frustration, such as is the case for the present model, we use (see, e.g., Refs. \[\]) $$M(m) = c_{0}+c_{1}m^{-1/2}+c_{2}m^{-3/2}\,. \label{M_extrapo_frustrated}$$ When we have only three data points to fit to an extrapolation formula, such as will sometimes occur here, specifically for the spiral phase, a two-term extrapolation fit can easily be preferable in practice to a three-term fit. This is particularly the case when one of the data points is either far from the limiting case or when it does not represent all of the features of the system as well as the remaining, more accurate points. In such cases we sometimes use the alternative simpler forms, $$E(m)/N = b_{0}+b_{1}m^{-2}\,, \label{E_extrapo_linearFit}$$ and $$M(m) = d_{0}+d_{1}m^{-1/2}\,, \label{M_extrapo_frustrated_linearFit}$$ instead of their counterparts in Eqs. (\[E\_extrapo\]) and (\[M\_extrapo\_frustrated\]), respectively.
Finally we note that since the hexagon is an important structural element of the honeycomb lattice, it is preferable for the extrapolations to use only LSUB$m$ data with $m \geq 6$, wherever possible. However, especially for the spiral phase that is particularly costly of computational resource, as we explain below, we sometimes need to include LSUB4 results in the extrapolations. Under such circumstances, however, we always perform a sensitivity analysis, by doing some LSUB$m$ runs with higher values of $m$ for a few indicative points only, as we discuss in more detail in Sec. \[results\].
PREVIEW OF THE PHASE DIAGRAM {#preview}
============================
![ (Color online) Phase diagram of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$), in the parameter window $J_{2},J_{3} \in [0,1]$. The five regions correspond to four quasiclassical phases with (a) AFM Néel (N) order as shown in Fig. \[model\](a), (b) collinear AFM striped (S) order as shown in Fig. \[model\](b), (c) spiral order as shown in Fig. \[model\](c), (d) AFM anti-Néel (aN) order as shown in Fig. \[model\](d), plus (e) a magnetically disordered, or quantum paramagnetic (QP), phase that exhibits plaquette valence-bond crystalline (PVBC) order on at least part of the boundary region (and see below). The first-order phase transition boundary between the N and S phases, marked by the (grey) convolution (eight-pointed star, [$\times$]{}) symbols is found from points at which the curves for the magnetic order parameter $M$ of the two phases cross; the first-order phase transition boundary between the S and QP phases, marked by (green) plus ($+$) symbols, is found from points at which $M \to 0$ for the S phase; the first-order phase transition boundary between the S and spiral phases, marked by (cyan) open triangle ($\triangle$) symbols, is found from points at which the curves for the magnetic order parameter $M$ of the two phases cross; the phase transition boundary between the spiral and QP phases, marked by (orange) open circle ($\bigcirc$) symbols (of two sizes, see main text in Sec. \[results\_neel\_spiral\]), is found from points at which $M \to 0$ for the spiral phase; the first-order phase transition boundary between the spiral and aN states, marked by (magenta) times ($\times$) symbols, is found from points at which the curves for the magnetic order parameter $M$ of the two phases cross; the phase transition boundary between the aN and QP phases, marked by (blue) plus ($+$) symbols, is found from points at which $M \to 0$ for the aN phase; and the phase transition boundary between the N and QP states, marked by (red) times ($\times$) symbols, which is probably of continuous (second-order, and possibly of a deconfined) nature, is found from points at which $M \to 0$ for the N phase. Points marked by the larger (red) times ($\times$) and (green and blue) plus ($+$) symbols are found to be infinitely susceptible to PVBC order, and hence the QP state at these points is PVBC in nature.[]{data-label="phase"}](fig2.eps){width="6cm"}
Before discussing our results in detail it is perhaps useful to summarize our main findings first, and for that purpose we show in Fig. \[phase\] the phase diagram for the frustrated spin-$\frac{1}{2}$ model of Eq. (\[eq1\]) on the honeycomb lattice, in the case where all the bonds are antiferromagnetic in nature (i.e., $J_{n} > 0,\, n=1,2,3$). Furthermore, we set $J_{1} \equiv 1$ and restrict ourselves to the window $0 \leq J_m \leq 1, m=2,3$. Henceforth we denote $x \equiv J_{2}/J_{1}$, $y \equiv J_{3}/J_{2}$. The actual phase boundaries are determined from a variety of information that emerges from our CCM calculations, as we now describe briefly and with further details given in Sec. \[results\].
As we have already noted, we have previously studied this model for the two special cases with $J_{3}=J_{2}$ in Ref. \[\], and with $J_{3}=0$ in Ref. \[\], and the corresponding CCM results from those papers are included in Fig. \[phase\]. Firstly, along the line $J_{3}=J_{2}$ (i.e., when $y = x \equiv \kappa$) we found[@DJJF:2011_honeycomb] that the system has quasiclassical AFM Néel (N) order for $\kappa < \kappa_{c_1} \approx 0.47$, quasiclassical AFM striped (S) order for $\kappa > \kappa_{c_2} \approx 0.60$, and a quantum paramagnetic phase separating the N and S phases for $\kappa_{c_1} < \kappa < \kappa_{c_2}$. By studying the susceptibility of the N and S states to hexagonal plaquette valence-bond crystal (PVBC) ordering, we found that the most likely scenario was that the intervening state had PVBC order over the entire range $\kappa_{c_1} < \kappa < \kappa_{c_2}$. The transition at $\kappa = \kappa_{c_2}$ between the PVBC and S GS phases was seen to be of first-order type, while that at $\kappa = \kappa_{c_1}$ between the N and PVBC GS phases appeared to be a continuous one. Since the N and PVBC phases break different symmetries our results favored the transition point between them at $\kappa = \kappa_{c_1}$ to be a deconfined quantum critical point (QCP). The QCPs at $y = x = \kappa_{c_1}$ and at $y = x = \kappa_{c_2}$ are clearly shown in Fig. \[phase\] with the larger (red) times ($\times$) and the larger (green) plus ($+$) symbols respectively.
Secondly, in a separate study along the line $y=0$, we found[@PHYLi:2012_honeyJ1-J2] that the system has the quasiclassical N state as its GS phase for $x < x_{c_1} \approx 0.21$, the quasiclassical anti-Néel (aN) state as its GS phase for $x >
x_{c_2} \approx 0.39$, and again a quantum paramagnetic (QP) phase separating the N and aN phases for $x_{c_1} < x < x_{c_2}$. Similar CCM calculations of the susceptibility of the N and aN phases to PVBC order led again to the conclusion that the transition between the PVBC and aN phases was of first-order type, while the likely scenario for the transition between the N and PVBC phases is again that it is of the continuous deconfined type. Nevertheless, due to the difficulty in determining the lower critical value of $x$ at which PVBC order is established as accurately as we determined the value $x = x_{c_1}$ at which Néel order is destabilized, we could not exclude a second scenario in which the transition between the N and PVBC phases proceeds via an intervening phase (possibly even of an exotic spin-liquid variety) in the very narrow window $0.21 \lesssim x \lesssim 0.24$. Again, the QCPs at $(x,y) = (x_{c_1},0)$ and $(x_{c_2},0)$ are clearly shown in Fig. \[phase\] with the larger (red) times ($\times$) and the (blue) plus ($+$) symbols respectively.
We also showed previously,[@PHYLi:2012_honeyJ1-J2] by a comparison of the GS energies of the spiral and aN phases calculated separately with the CCM, that the spiral phases that are present classically (i.e., for the case where the spin quantum number $s \to \infty$) in the case $y=0$ for all values $x > \frac{1}{6}$ are absent for all values $x \lesssim 1$. The actual phase boundary between the spiral and aN phases shown in Fig. \[phase\] is now calculated in the present paper, as described below.
Based on our previous findings for the GS phases of these two special cases when (a) $J_{3} = J_{2}$ and (b) $J_{3} = 0$. we have now performed a series of CCM calculations based on the N, S, aN, and spiral states as model states, for a variety of cuts in the phase diagram at both constant values of $J_3$ and constant values of $J_2$. For example, the phase boundary between the N and the S phases is obtained, as explained more fully in Sec. \[results\_neel\_striped\], from our extrapolated ($m \to \infty$) LSUB$m$ results for the order parameter $M$ (namely, the average onsite magnetization) of the two phases, for a variety of constant $J_3$ cuts. We find that for values of $y \equiv J_{3}/J_{1} \gtrsim 0.69$ the two magnetization curves meet at a (positive) nonzero value, indicative of a direct first-order transition between the states. These points are shown in Fig. \[phase\] by the (grey) convolution (eight-pointed star, [$\times$]{}) symbols.
For the value $y \approx 0.69$ the two curves become zero at precisely the same point, $x \approx 0.51$. Conversely, when $y \lesssim 0.69$, the order parameters of the N and the S phases both become zero at respective critical values of $x$ before the curves cross (when solutions exist for both phases), indicating the emergence of a new phase separating them. The corresponding points where the magnetic order parameters for Néel order and striped order vanish are shown in Fig. \[phase\] by (red) times ($\times$) and (green) plus ($+$) symbols respectively. By continuity with our earlier results[@DJJF:2011_honeycomb] along the line $y=x$, we tentatively identify the intervening phase as the PVBC state. The tricritical QCP between the N, S, and PVBC phases is thus identified as being at $(x,y) \approx (0.51,0.69)$. We also note that for values of $y \lesssim 0.55$ no solution for the S phase exists with $M>0$ for any value of $x$, giving preliminary indications of a new phase boundary between the S state and another phase that we identify as a spiral phase.
By comparing the order parameters for the S and spiral phases at various constant $J_2$ cuts we find that for values of $x \gtrsim 0.66$ the two curves meet at a (positive) nonzero value, once again indicative of a direct transition between the states. These points are shown in Fig. \[phase\] as (cyan) open triangle ($\triangle$) symbols. For the value $x \approx 0.66$ the two curves become zero at the same point $y \approx 0.55$. Then, for values $x \lesssim 0.66$ the order parameters of the S and spiral phases both become zero at respective critical values of $y$ before the curves cross. Once again this indicates a phase separating the S and spiral phases for values of $x \lesssim 0.66$ (down to a lower value of $x \approx 0.635$ below which the spiral phase ceases to exist for any value of $y$), which we similarly identify tentatively as the PVBC phase.
In that very narrow window $0.635 \lesssim x \lesssim 0.66$, which is almost certainly an artifice of our approximations, we denote in Fig. \[phase\] the points where the magnetic order parameter vanishes ($M \to 0$) for the striped and spiral states by (cyan) open triangle ($\triangle$) and (orange) open circle ($\bigcirc$) symbols respectively. We argue in Sec. \[results\_spiral\_striped\] that these results are consistent with the existence of a second tricritical QCP at $(x,y) \approx (0.65,0.55)$ between the S, spiral, and (tentatively) PVBC phases. The remainder of the phase boundary between the spiral and PVBC states is similarly identified by the vanishing of the magnetic order parameter of the spiral phase, and these points are again shown in Fig. \[phase\] as (orange) open circle ($\bigcirc$) symbols.
Finally by comparing the energies of the aN and spiral phases we find that for all values of the parameter $J_{2} \leq 1$ where the spiral phase exists, the aN phase actually has a lower energy for values of the parameter $J_3$ below a certain critical value, which itself depends on $J_2$. Similarly, by comparing the order parameters of these two phases at various constant $J_2$ cuts, we find that for values of $x \gtrsim 0.69$ the two curves meet at a (positive) nonzero value, indicative once more of a direct phase transition between the aN and spiral phases. These points are shown in the phase diagram of Fig. \[phase\] by (magenta) times ($\times$) symbols.
For the value $x \approx 0.69$ the two curves become zero at the same point $y \approx 0.12$. Conversely, for values $x \lesssim 0.69$ the order parameters of the aN and spiral phases both become zero at respective critical values of $y$ before the curves cross. This is again indicative of a phase separating the aN and spiral phases for values of $x \lesssim 0.69$ (down to the lower value of $x \approx 0.635$ below which the spiral phase ceases to exist for any value of $y$), as noted above. This intermediate phase is again tentatively identified as having PVBC order. In this way we identify a third tricritical QCP at $(x,y) \approx (0.69,0.12)$ between the spiral, aN and (tentatively) PVBC phases. Remaining points on the phase boundary between the aN and PVBC phases are then identified as the points where the magnetic order parameter of the aN phase vanishes, and these are shown by (blue) plus ($+$) symbols on the the phase diagram of Fig. \[phase\].
In Sec. \[results\] we now describe in more detail how the various points in the phase diagram of Fig. \[phase\] are obtained. We also discuss the properties of the various phases that we have examined.
RESULTS
=======
In this section, we present and discuss our CCM results for the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ HAFM on the honeycomb lattice, with all of the bond strengths positive (i.e., antiferromagnetic in nature). To set the overall energy scale we put $J_{1} \equiv 1$, and we investigate the parameter space window $J_{2},J_{3} \in [0,1]$. We use each of the Néel (N), collinear striped (S), spiral, and anti-Néel (aN) states shown respectively in Figs. \[model\](a)-(d) as CCM model states.
Néel versus striped phases {#results_neel_striped}
--------------------------
Figures \[E\](a) and \[E\](b)
show the extrapolated ($m \to \infty)$ CCM LSUB$m$ values for the GS energy per spin for the Néel (N) and striped (S) states as functions of $J_{2}$ for various fixed values of $J_{3}$ in the range $0.5 \leq J_{3} \leq 1.0$. The extrapolations have been performed using Eq. (\[E\_extrapo\]) and the calculated LSUB$m$ results with $m=\{6,8,10\}$. We observe that the energy curves cross for all values of the parameter $J_{3} \gtrsim 0.68$, but for values $J_{3} \lesssim 0.68$ the curves do not cross. This gives us a first indication of the emergence of an intermediate phase between the N and S states, over a finite range of values of the $J_2$ parameter, below some critical value of the $J_3$ parameter.
We note that the extrapolations become more difficult in the vicinity of this critical point, and consequently the actual values of $J_2$ at which the curves cross for fixed values of $J_3$ near the critical value are more uncertain than those at larger values. Furthermore, at the actual energy crossing points very near the critical point the corresponding values of the magnetic order parameter (i.e., the average onsite magnetization) $M$ for one or both states becomes negative and hence unphysical. Indeed, for the S state, $M<0$ for the entire $J_{3}=0.5$ curve, which is why we have not shown it in Fig. \[E\](a).
In order to obtain more accurate values of the critical point we also show in Figs. \[M\_Neel\_striped\](a) and \[M\_Neel\_striped\](b) the
curves for the extrapolated order parameters $M$ of the N and S states, corresponding to values of $J_3$ shown in Figs. \[E\](a) and \[E\](b) for the GS energy per spin, $E/N$. The LSUB$\infty$ curves shown use the LSUB$m$ results with $m=\{6,8,10\}$, together with the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\]), which is appropriate for this highly frustrated regime.
We observe again that the curves intersect for values $J_{3} \gtrsim 0.69$, and that the corresponding values of $(J_{2},J_{3})$ are our best estimate for the phase boundary between the N and S states, as shown on the phase diagram of Fig. \[phase\] by the points denoted with (grey) convolution (eight-pointed star, [$\times$]{}) symbols. For values $J_{3} \lesssim 0.69$ the extrapolated order parameters of both the N and S phases become zero before the curves intersect, revealing the presence of an intermediate phase in that regime. The corresponding points in the case $J_{3} \lesssim 0.69$ where $M \to 0$ for the N and S phases are shown in the phase diagram of Fig. \[phase\] by (red) times ($\times$) and (green) plus ($+$) symbols respectively. Our best value for the corresponding tricritical QCP comes from the data shown in Fig. \[M\_Neel\_striped\](b), where it is seen to be at $(J_{2}^{c_1},J_{3}^{c_1})=(0.51 \pm 0.01,0.69 \pm 0.01)$, and where the error bars are estimates from a sensitivity analysis of the LSUB$m$ extrapolation scheme.
We note that the extrapolated order parameter $M$ becomes everywhere negative (i.e., for all values of $J_2$) for the S state for all values of $J_{3} \lesssim 0.55$, as may be seen from data similar to those shown in Fig. \[M\_Neel\_striped\](a). This is a clear first indication that the S state becomes unstable as the GS phase in this regime. From a comparison with the corresponding classical model (i.e., in the limit $s \to \infty$) discussed in Sec. \[model\_section\], we might expect the S state to yield to the spiral state, at least for sufficiently large values of $J_2$ in the present $s = \frac{1}{2}$ case. We investigate this further in Sec. \[results\_spiral\_striped\] below. It is clearly also expected that the Néel (N) phase will not survive for large enough frustrating values of $J_{2} > 0$, and again from a comparison with the classical model we expect that the spiral phase might exist in that case too. Hence, we first make a comparison in Sec. \[results\_neel\_spiral\] of the N and spiral phases.
Néel versus spiral phases {#results_neel_spiral}
-------------------------
We start by analyzing the GS energy per spin, $E/N$, for the spiral state as a function of the spiral pitch angle, $\phi$. In our CCM calculations we choose the angle $\phi$, for each point in the phase diagram where the spiral state exists, as the one that minimizes the energy estimate there. Clearly the minimizing angle in general also depends on the particular LSUB$m$ approximation being used with the spiral state as CCM model state. For example, we show in Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](a) the GS energy per spin, $E/N$, as a function of pitch angle $\phi$ in the LSUB6 approximation, for various illustrative values of $J_2$ at a fixed value of $J_{3}=0.4$ (and $J_{1} \equiv 1$).
We note first from Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](a) that for various fixed values of $J_{2}$ CCM solutions at a given LSUB$m$ level of approximation exist only for certain ranges of the spiral angle $\phi$. For example, for $J_{2}=0$ (and $J_{3}=0.4$), the CCM LSUB6 solutions based on the spiral state exist only for $0 \leq \phi \lesssim 0.12\pi$. In this case, where the Néel state (i.e., where $\phi=0$) is the stable GS phase that minimizes the energy, if we try to force the system too far away from Néel collinearity the CCM equations themselves become unstable in the sense that they no longer have a real solution. We note too that as $J_2$ is increased slowly (at fixed $J_{3}=0.4$), the minimum in the energy curve at $\phi=0$ becomes shallower, so that by the time $J_{2}=0.4$ it has almost disappeared. This is a first indication of the imminent instability of the Néel state as the GS phase if $J_2$ is increased slightly more.
Similarly, for $J_{2}=1$ (and $J_{3}=0.4$), the CCM LSUB6 solutions based on the spiral state exist only for $0.43 \lesssim \phi/\pi \leq 1$. In this case, a spiral state (i.e., with a value $\phi \neq 0,\pi$) is the stable GS phase that minimizes the energy, and if we now try to force the system too close to the Néel regime, the CCM solution collapses. We also observe from Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](a) that for the smaller value $J_{2}=0.8$ (and $J_{3}=0.4$), while the energy curve still shows a global minimum for a noncollinear spiral phase, it has now also developed a secondary minimum at a value $\phi = \pi$ (i.e., that of the collinear striped state), which indicates the proximity of the phase boundary between the spiral and striped states, as we examine more fully in Sec. \[results\_spiral\_striped\] below.
Conversely, as $J_3$ is increased further (for fixed $J_2$), the spiral minimum becomes more pronounced, and as $J_{2} \rightarrow \infty$ the pitch angle $\phi \rightarrow \frac{2}{3}\pi$. This is as expected, since in this limit the model becomes two weakly connected HAFMs on interpenetrating triangular lattices, with the classical ordering of NN spins oriented at angles $\frac{2}{3} \pi$ with respect to one another on each sublattice.
From data such as that shown in Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](a) we can calculate in a given LSUB$m$ approximation based on the spiral state as the CCM model state, the angle $\phi=\phi_{{\rm LSUB}m}$ that minimizes the energy, $E_{{\rm LSUB}m}(\phi)$ for given values of the exchange coupling strengths. For example, in Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](b) we show the angle $\phi=\phi_{{\rm LSUB}8}$ from the LSUB8 approximation, as a function of the parameter $J_2$ for several fixed values of the parameter $J_3$. There is clear preliminary evidence that for values of $J_3$ below some upper critical value there is no stable spiral solution for any value of $\phi \neq 0$ over a certain range of the parameter $J_2$, which itself depends on $J_3$.
Thus, we are led to expect a second tricritical QCP in the $(J_{2},J_{3}$) plane at $(J_{2}^{c_2},J_{3}^{c_2})$, with $J_{3}^{c_2}<J_{3}^{c_1}$, such that: (a) for values $J_{3}>J_{3}^{c_1}$ the N and S states meet at a common phase boundary discussed in Sec. \[results\_neel\_striped\] above, (b) for values $J_{3}^{c_1}>J_{3}>J_{3}^{c_2}$ there is an intermediate phase between the N and S states, and (c) for values $J_{3}<J_{3}^{c_2}$ there is an intermediate phase between the N state and the spiral state with $\phi \neq \pi$. Thus, the QCP at $(J_{2}^{c_2},J_{3}^{c_2})$ is a tricritical point between the N, S, and intermediate phases. From the results discussed in Sec. \[results\_neel\_striped\] we now expect that $J_{3}^{c_2} \approx 0.55$, and we discuss this further below.
Figure \[E\_Neel\_vs\_spiral\] shows our extrapolated CCM LSUB$\infty$ results for the GS energy per spin, $E/N$, as a function of $J_{2}$, for various fixed values of $J_{3}$ in the range $0.2 \leq J_{3} \leq
0.6$, for the Néel and the spiral states.
![ (Color online) Extrapolated CCM LSUB$\infty$ results for the GS energy per spin, $E/N$, as a function of $J_{2}$, for various fixed values of $J_{3}$ in the range $0.2 \leq J_{3} \leq 0.6$, for the Néel and the spiral states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$). The extrapolated LSUB$m$ ($m \rightarrow
\infty$) results are based on the extrapolation scheme of Eq. (\[E\_extrapo\]) and the calculated results with $m=\{6,8,10\}$ for the Néel state, and with $m=\{4,6,8\}$ for the spiral state. For the spiral state the results use the pitch angle $\phi =
\phi_{{\rm LSUB}m}$ that minimizes the energy $E=E_{{\rm LSUB}m}(\phi)$. []{data-label="E_Neel_vs_spiral"}](fig6.eps){width="6cm"}
Once again, the figure clearly illustrates the existence of an intermediate phase between the Néel and the spiral phases (including the striped state as a special case of the latter) for values $J_{3}<J_{3}^{c_1}$.
On a technical point we remark that for the spiral state the extrapolations are calculated using the LSUB$m$ calculated results with $m=\{4,6,8\}$, rather than with the set $m=\{6,8,10\}$ used for the Néel state. This is partly due to the very high number, $N_{f} = 347287$, of configurations needed for the spiral state at the LSUB10 level of approximation, compared with the corresponding much smaller number, $N_{f} = 6237$, for the Néel state, as seen from Table \[table\_FundConfig\]. This difference is compounded by the fact that for the spiral state we also need to do LSUB$m$ runs for each point in the phase space as a function of the pitch angle $\phi$, in order to determine the angle $\phi = \phi_{{\rm LSUB}m}$ that minimizes the corresponding estimate for the energy, $E=E_{{\rm LSUB}m}(\phi)$. This makes LSUB$m$ calculations for the spiral state with $m \geq 10$ particularly demanding of computational resources.
Nevertheless, we did perform LSUB10 calculations for the spiral state for the special case $J_{3}=0$ in our previous study of the $J_1$–$J_2$ model,[@PHYLi:2012_honeyJ1-J2] where we performed separate extrapolations using the LSUB$m$ results with $m=\{6,8,10\}$ and $m=\{4,6,8\}$. We found that both extrapolations were in very good agreement with one another, and hence now feel confident that the spiral-state extrapolations for the full $J_1$–$J_2$–$J_3$ model considered here with the limited set $m=\{4,6,8\}$ will be equally robust, since it is now prohibitively expensive of computational resource to perform LSUB10 calculations for the spiral state over the whole region of phase space where it is the stable GS phase.
We note that, as is usually the case, the CCM LSUB$m$ results for finite $m$ values for a given phase extend beyond the actual physical LSUB$\infty$ boundary for that phase. Thus, the energy curves shown in Fig. \[E\_Neel\_vs\_spiral\] for fixed values of $J_3$ terminate at certain values of $J_{2}$, which are determined by the termination points of the highest LSUB$m$ approximations used in the extrapolations, beyond which no real solution exists for the corresponding coupled CCM equations. We note from Fig. \[E\_Neel\_vs\_spiral\] that the maxima in the energy curves occur close to these LSUB$m$ termination points for the largest $m$ values employed, which in turn lie close to the physical (LSUB$\infty$) phase transition points.
It has been suggested[@Albuquerque:2011] that such an energy maximum approximately coincides with the avoided level crossing to a different phase, in which case it could be taken as an approximation for the phase transition point of either (the Néel or spiral) state to the intermediate (as yet unknown) state. However, we do not use this criterion here, since our results for the magnetic order parameter give us much more accurate estimates, as we now discuss.
Thus, we show in Fig. \[M\_Neel\_vs\_spiral\] our extrapolated CCM LSUB$\infty$ results for the GS magnetic order parameters, $M$, for both the Néel and the spiral states, as functions of $J_{2}$, for the same fixed values of $J_{3}$ shown in Fig. \[E\_Neel\_vs\_spiral\] for the GS energy.
![ (Color online) Extrapolated CCM LSUB$\infty$ results for the GS magnetic order parameter, $M$, as a function of $J_{2}$, for various fixed values of $J_{3}$ in the range $0.2 \leq J_{3} \leq
0.6$, for the Néel and the spiral states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$). The extrapolated LSUB$m$ ($m
\rightarrow \infty$) results are based on the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\_linearFit\]) and the calculated results with $m=\{6,8,10\}$ for the Néel state, and with $m=\{4,6,8\}$ for the spiral state. For the spiral state the results use the pitch angle $\phi = \phi_{{\rm LSUB}m}$ that minimizes the energy $E=E_{{\rm LSUB}m}(\phi)$. []{data-label="M_Neel_vs_spiral"}](fig7.eps){width="6cm"}
The extrapolated LSUB$m$ ($m \rightarrow \infty$) results for the Néel state are based on the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\]) with $m=\{6,8,10\}$, whereas the extrapolated results for the spiral state are based on the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\_linearFit\]) with $m=\{4,6,8\}$.
As we indicated above, LSUB10 calculations are prohibitively expensive for the spiral state, and hence we need for extrapolation purposes to include the LSUB4 results. When this point is included, and the data set $m=\{4,6,8\}$ is thus employed, it is clearly preferable to use the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\_linearFit\]) rather than that of Eq. (\[M\_extrapo\_frustrated\]), so as not to give the $m=4$ result too much weight. However, in order to check our results, we have performed LSUB10 calculations for the two values $J_{3}=0.2,0.4$. For these two values we have also made extrapolations using Eq. (\[M\_extrapo\_frustrated\_linearFit\]) with $m=\{6,8,10\}$, and these are indicated in Fig. \[phase\] by the larger (orange) open circles. We find, very gratifyingly, that the two extrapolations agree very well with one another at both values $J_{3}=0.2,0.4$, which gives credence to our results using the data set $m=\{4,6,8\}$ elsewhere for the spiral state.
We note too that in our earlier study[@PHYLi:2012_honeyJ1-J2] of this model with $J_{3}=0$ (and $J_{1} \equiv 1$), we also computed LSUB12 results for the Néel state and found that the Néel order vanished at a value $J_{2} \approx 0.207 \pm 0.003$ when we performed extrapolations including the $m=12$ point. It is this point that is shown in Fig. \[phase\] by the larger (red) times ($\times$) symbol, although the value obtained with the more limited data set $m=\{6,8,10\}$ used for the results in Fig. \[M\_Neel\_vs\_spiral\] is in good agreement with it. Similarly, in our earlier study of the model along the $J_{3}=J_{2}$ line[@DJJF:2011_honeycomb] we also used LSUB$m$ results with $m=\{6,8,10,12\}$ to perform the extrapolations, and found that in this case Néel order vanished at a value $J_{2}
\approx 0.466 \pm 0.005$, and this value is also shown in Fig. \[phase\] by a larger (red) times ($\times$) symbol.
From curves such as those shown in Fig. \[M\_Neel\_vs\_spiral\] we use the points where the extrapolated values of the order parameter $M$ vanish, for various fixed values of $J_3$, to plot the phase boundaries of the Néel and spiral phases denoted in Fig. \[phase\] by (red) times ($\times$) and (orange) open circle symbols respectively. As expected from our previous discussion in Sec. \[results\_neel\_striped\], a Néel-ordered phase exists for all values of $J_3$ up to some critical value of $J_2$ which marks its phase boundary. For values $J_{3}<J_{3}^{c_1}\approx 0.69 \pm 0.01$ this phase borders a quantum paramagnetic phase, whereas for $J_{3}>J_{3}^{c_1}$ it borders the striped state at a first-order phase transition boundary. We also see from curves such as those shown in Fig. \[M\_Neel\_vs\_spiral\] clear evidence for an intervening phase between the Néel and spiral phases (with pitch angle $\phi \neq \pi$) everywhere that the spiral phase exists. Instead the spiral phase meets the striped phase along a common boundary (on which $\phi=\pi$) for all values $J_{2} > J_{2}^{c_2} \gtrsim 0.65$. There is thus a second tricritical point at $(J_{2}^{c_2},J_{3}^{c_2})$, as we discuss more fully below in Sec. \[results\_spiral\_striped\], at which the striped, spiral and quantum paramagnetic phases meet.
Striped versus spiral phases {#results_spiral_striped}
----------------------------
We first recall that classically (i.e., when $s \to \infty$) we have for this model (with $J_{1}\equiv 1$) that for a fixed value of $J_{2}>\frac{1}{2}$ the GS phase is the striped phase for $J_{3}>\frac{1}{2}J_{2}+\frac{1}{4}$ and the spiral phase for $J_{3}<\frac{1}{2}J_{2}+\frac{1}{4}$. There is a continuous phase transition between the two classical states along the boundary line, $J_{3}=\frac{1}{2}J_{2}+\frac{1}{4}\,, J_{2}\geq\frac{1}{2}$, on which the spiral pitch angle $\phi=\pi$. Our results for the present $s=\frac{1}{2}$ model, as we shall see below, indicate that quantum fluctuations tend to stabilize the collinear order of the striped state to lower values of $J_3$, for fixed $J_2$, than the classical limit. Furthermore, as we shall see, the quantum fluctuations also seem to turn the classical second-order transition into a quantum first-order one.
Thus, we show in Fig. \[E\_spiral\_striped\](a) the angle, $\phi=\phi_{{\rm LSUB8}}$ that minimizes the energy, $E_{{\rm LSUB8}}(\phi)$, as a function of $J_{3}$, using the spiral state as our CCM model state, for various fixed values of $J_2$.
Very similar curves are found for other LSUB$m$ approximations. We observe that, unlike in the classical case, where $\phi \to \pi$ continuously at the critical value, there is now a discontinuous jump on the phase boundary. Its origin lies in the double-minimum structure of the corresponding energy curves (for fixed values of $J_2$ and $J_3$) as functions of the pitch angle $\phi$, comparable to that shown in Fig. \[E\_vs\_angle\_Neel\_vs\_spiral\](a) for the case $J_{2}=0.8$, $J_{3}=0.4$.
Clearly, if we consider the angle $\phi$ itself to be an order parameter (such that $\phi=\pi$ for striped order and $\phi\neq 0,\pi$ for spiral order) the typical scenario for a first-order transition, as now seen here, is the emergence of such a two-minimum structure for $E/N$ as a function of $\phi$ for fixed coupling parameter strengths, one at a value $\phi\neq\pi$ and the other precisely at $\phi=\pi$. For a fixed value of $J_2$ we find, at a given LSUB$m$ level of approximation, that when $J_3$ is above a certain critical value the global minimum in the $E=E(\phi)$ curve is at $\phi=\pi$, whereas when $J_3$ is below this value the global minimum is at the other minimum, $\phi\neq\pi$.
Figure \[E\_spiral\_striped\](b) shows the corresponding extrapolated CCM LSUB$\infty$ results for the GS energy per spin, $E/N$, of the spiral and striped states, as functions of the parameter $J_3$, for the same fixed values of $J_2$ shown in Fig. \[E\_spiral\_striped\](a). The first-order transition between the spiral and striped states can clearly be seen to occur close to, but not precisely at, the corresponding maxima in the energy curves.
As before, the actual phase boundary is most clearly seen from our similarly extrapolated CCM LSUB$\infty$ results for the magnetic order parameter, $M$, which are shown in Fig. \[M\_spiral\_vs\_Striped\].
![ (Color online) Extrapolated CCM LSUB$\infty$ results for the GS magnetic order parameter, $M$, as a function of $J_{3}$, for various fixed values of $J_{2}$ in the range $0.64 \leq J_{2} \leq
1.0$, for the spiral and the striped states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$). The extrapolated LSUB$m$ ($m
\rightarrow \infty$) results are based on the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\_linearFit\]) and the calculated results with $m=\{4,6,8\}$. For the spiral state the results use the pitch angle $\phi =
\phi_{{\rm LSUB}m}$ that minimizes the energy $E=E_{{\rm LSUB}m}(\phi)$. []{data-label="M_spiral_vs_Striped"}](fig9.eps){width="8cm"}
We note that for all values of $J_{2} \gtrsim 0.66$ there is a clear and sharp minimum in the magnetic order parameter at the phase transition point in the parameter $J_3$ where the striped and spiral phases meet. These points are indicated by the (cyan) open triangle ($\triangle$) symbols in the phase diagram shown in Fig. \[phase\].
At the value $J_{2} \approx 0.66$ the two curves meet at $M=0$. We note that for this value of $J_2$ the magnetic order parameter $M$ for the spiral state is very small (and positive) for all values of $J_3$, and that as $J_2$ is decreased further the spiral state rapidly disappears altogether for $J_{2} \lesssim 0.635$. In the very narrow regime $0.635 \lesssim J_{2} \lesssim 0.66$, we see from Fig. \[M\_spiral\_vs\_Striped\] that there appears to be an intrusion of the intermediate (quantum paramagnetic) phase, as shown in Fig. \[phase\] by the appearance of both (cyan) open triangle ($\triangle$) and (orange) open circle ($\bigcirc$) symbols at the striped-spiral phase boundary at the two values $J_{2}=0.64,0.65$. It seems almost sure, however, that this effect arises from our extrapolations, and is an indication of the (small) errors inherent in them. Our best estimate from the results shown in Fig. \[M\_spiral\_vs\_Striped\] is thus that the second tricritical QCP, where the spiral, striped and quantum paramagnetic phases meet, occurs at $(J_{2}^{c_2},J_{3}^{c_2})=(0.65 \pm 0.02,0.55 \pm 0.01)$.
We also note from Fig. \[M\_spiral\_vs\_Striped\] that for values $0.635 \lesssim J_{2} \lesssim 0.77$ and $J_{3}>0$ the magnetic order parameter $M$ of the striped state becomes zero at a lower critical value of $J_3$. These lower values in each case are shown in Fig. \[phase\] by the same (orange) open circle ($\bigcirc$) symbols as we discussed previously in Sec. \[results\_neel\_spiral\]. We note that for the special case $J_{3}=0$ that we investigated earlier,[@PHYLi:2012_honeyJ1-J2] the spiral state is actually unstable, since the anti-Néel state was seen to have lower energy for all values of $J_2$ in the range investigated, namely $J_{2} \leq 1$, where solutions for the spiral state could be found. From continuity, we expect that the anti-Néel state should remain the stable GS phase for small enough values of $J_3$ below some critical value for each fixed value of $J_2$, above which value the spiral phase then becomes the stable GS phase. Thus we are led to expect that there might exist a third tricritical QCP at $(J_{2}^{c_3},J_{3}^{c_3})$ between the spiral, quantum paramagnetic, and anti-Néel GS phases. We examine this further in Sec. \[results\_spiral\_aN\] below.
Spiral versus anti-Néel phases {#results_spiral_aN}
------------------------------
In Fig. \[E\_cross\_spiral\_vs\_aN\] we show the extrapolated CCM LSUB$\infty$ results for the GS energy per spin, $E/N$, of both the spiral and anti-Néel states, as functions of the parameter $J_{3}$, for various fixed values of the parameter $J_2$ in the range $0.7 \leq J_{2} \leq 1.0$.
![ (Color online) Extrapolated CCM LSUB$\infty$ results for the GS energy per spin, $E/N$, as a function of $J_{3}$, for various fixed values of $J_{2}$ in the range $0.7 \leq J_{2} \leq 1.0$, for the anti-Néel and the spiral states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$). The extrapolated LSUB$m$ ($m \rightarrow
\infty$) results are based on the extrapolation scheme of Eq. (\[E\_extrapo\_linearFit\]), and the calculated results with $m=\{4,6,8\}$ in both cases. For the spiral state the results use the pitch angle $\phi =
\phi_{{\rm LSUB}m}$ that minimizes the energy $E=E_{{\rm LSUB}m}(\phi)$. We note that in all cases curves without symbols attached refer to the anti-Néel state, whereas the corresponding curves with symbols refer to the spiral state. []{data-label="E_cross_spiral_vs_aN"}](fig10.eps){width="8cm"}
Although the energy differences are small for each fixed value of $J_2$, the results at each LSUB$m$ level, as well as the extrapolated results, clearly show an energy crossing point. These energy crossing points are thus our first estimates of the phase boundary points between the spiral and anti-Néel states.
On a technical point, we have noted previously that CCM LSUB$m$ calculations for the spiral state are computationally very expensive for values of the truncation index $m \geq 10$. Thus, we are for the most part restricted to the data set $m=\{4,6,8\}$ for the spiral state, although we have performed a very few calculations for a few select values in the parameter space with $m=10$. With only three data points to fit to an extrapolation formula, a two-term extrapolation fit (such as those in Eqs. (\[E\_extrapo\_linearFit\]) and (\[M\_extrapo\_frustrated\_linearFit\]), for example) can often be preferable in practice to a three-term fit (such as their counterparts in Eqs. (\[E\_extrapo\]) and (\[M\_extrapo\_frustrated\]), respectively). This is particularly the case when one of the data points is either far from the limiting case or when it does not represent all of the features of the system as well as the remaining, more accurate points, as is the case here for the $m=4$ points.
Thus, since the energy differences of the spiral and aN states are relatively small, we have found it preferable to use the extrapolation scheme of Eq. (\[E\_extrapo\_linearFit\]) in this case, and to employ the same data set with $m=\{4,6,8\}$ for both (aN and spiral) phases, even though results with $m=10$ are more readily available for the aN state. We have, however, demonstrated that the results so obtained are robust and reliable, by making further checks in some limited test cases using the extrapolation scheme of Eq. (\[E\_extrapo\]) fitted to the results $m=\{4,6,8,10\}$ or the extrapolation scheme of Eq. (\[E\_extrapo\_linearFit\]) fitted to the results $m=\{6,8,10\}$, for example.
We note that real CCM LSUB$m$ solutions based on the aN state as model state cease to exist, for a fixed value of $J_2$, above some termination value in the parameter $J_3$ that itself depends on the truncation index $m$, just as we have indicated above for other phases. These LSUB8 terminations are what cause our extrapolations for the aN state to be shown only up to certain values of $J_3$ for each curve shown in Fig. \[E\_cross\_spiral\_vs\_aN\]. In each case, the LSUB$m$ solution with a finite value of $m$ extends further into the region where the aN solution actually ceases to exist (i.e., to after the energy crossing point with the spiral phase). Presumably, in the $m \to \infty$ limit, the LSUB$m$ termination points for the aN phase would coincide with the phase boundary with the spiral phase. Simple heuristic extrapolations based on the results with $m=\{4,6,8\}$ agree well with the energy crossing points, and are hence entirely consistent with this hypothesis.
We note from Fig. \[E\_cross\_spiral\_vs\_aN\] that as the value of the parameter $J_2$ is decreased towards the lower limiting value $J_{2} \approx 0.635$, below which the spiral state ceases to exist, the energy curves for the aN and spiral phases lie increasingly close to one another, and hence the position of the crossing point becomes increasingly difficult to determine with high precision. Accordingly, we expect that a better indicator of the phase boundary might be obtained from a comparison of the magnetic order parameters of the two states, as now shown in Fig. \[M\_spiral\_vs\_aN\].
![ (Color online) Extrapolated CCM LSUB$\infty$ results for the GS magnetic order parameter, $M$, as a function of $J_{3}$, for various fixed values of $J_{2}$ in the range $0.4 \leq J_{3} \leq 1.0$, for the anti-Néel and the spiral states of the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ model on the honeycomb lattice (with $J_{1} \equiv 1$). The extrapolated LSUB$m$ ($m \rightarrow
\infty$) results are based on the extrapolation scheme of Eq. (\[M\_extrapo\_frustrated\_linearFit\]), and the calculated results with $m=\{4,6,8\}$ for both the aN and spiral phases for values $J_{2}=0.69,0.695,0.7,0.8,0.9,1.0$; and with $m=\{6,8,10\}$ for the aN phase for values $J_{2}=0.4,0.5,0.6$. For the spiral state the results use the pitch angle $\phi =
\phi_{{\rm LSUB}m}$ that minimizes the energy $E=E_{{\rm LSUB}m}(\phi)$. We note that in all cases curves without symbols attached refer to the anti-Néel state, whereas the corresponding curves with symbols refer to the spiral state.[]{data-label="M_spiral_vs_aN"}](fig11.eps){width="9cm"}
We see very clearly that for values of $J_{2} \gtrsim 0.69$ the curves for the magnetic order parameters $M$ of the aN and spiral phases cross in the physical regime (i.e., at a positive value of $M$). It is these crossing points that are our best estimates for the corresponding points on the boundary between the two phases, and these are shown in the phase diagram of Fig. \[phase\] by (magenta) times ($\times$) symbols. For values $J_{2} \gtrsim 0.8$, these values are in excellent quantitative agreement with the corresponding energy crossing points from Fig. \[E\_cross\_spiral\_vs\_aN\]. For smaller values of $J_2$, down to the value $J_{2} \approx 0.635$ below which the spiral state ceases to exist for any value of $J_3$, the energy crossing points become increasingly difficult to estimate accurately, as discussed above, and generally lie slightly below the much more accurate values obtained from Fig. \[M\_spiral\_vs\_aN\], although they are still in good qualitative agreement with them.
Our best estimate for the position of the third tricritical QCP, $(J_{2}^{c_3},J_{3}^{c_3}) = (0.69 \pm 0.01, 0.12 \pm 0.02)$, which marks the point where the spiral and aN phases meet the QP phase, comes from curves such as those shown in Fig. \[M\_spiral\_vs\_aN\].
For values $J_{2} < J_{2}^{c_3} \approx 0.69$, we use the corresponding values of $J_3$ at which the magnetic order parameter $M \to 0$ for the aN phase, as shown in Fig. \[M\_spiral\_vs\_aN\], to find the phase boundary between the aN and QP phases. The corresponding points are shown in the phase diagram of Fig. \[phase\] by (blue) plus ($+$) symbols.
The quantum paramagnetic (PVBC?) phase(s) {#results_paramagnet}
-----------------------------------------
From the results presented so far we have seen that in the parameter space window $J_{2},J_{3} \in [0,1]$ the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ HAFM on the honeycomb lattice with $J_{1} \equiv 1$ has regions of five different GS phases. Four of these (viz., the N, S, aN, and spiral phases) are quasiclassical in nature, and they almost completely surround the fifth QP phase, as shown in Fig. \[phase\], with each of them sharing a boundary with the (almost) enclosed region of the QP phase. (Indeed, it seems likely that if the diagram were extended slightly to negative values of $J_3$, the QP region would be seen to be entirely enclosed.) In the window $J_{2},J_{3} \in [0,1]$ there are three tricritical QCPs (and it seems likely that a fourth, which marks the meeting of the N, aN and QP phases, will occur just outside the window).
From our current results discussed here the question still remains open, however, as to the exact nature of this phase. What we know from previous work[@DJJF:2011_honeycomb; @PHYLi:2012_honeyJ1-J2] that employed the same CCM techniques as here is that the QP phase appears to be have PVBC ordering at least at four points along its boundary. These include the two points marked with the larger (red) times ($\times$) symbols in Fig. \[phase\] along its boundary with the N state where it crosses both the $J_{3}=0$ axis and the $J_{3}=J_{2}$ line, the point marked with the larger (green) plus ($+$) symbol along its boundary with the S state where it crosses the $J_{3}=J_{2}$ line, and the point marked with the larger (blue) plus ($+$) symbol along its boundary with the aN state where it crosses the $J_{3}=0$ axis.
Those points were identified as lying on a phase boundary with the PVBC state by calculating, within the same CCM LSUB$m$ approximations as used to calculate the phase transition points that marked the vanishing of the magnetic order parameter $M$ in each case (i.e., for the N, S, and aN phases respectively), the susceptibility of the respective phases against the formation of PVBC order. We showed, within the accuracy of our results, that each of the above four points where the respective magnetic order parameter of each quasiclassical phase goes to zero coincide with the points at which the corresponding susceptibility of the state to PVBC order becomes infinite.
In principle we could now repeat those calculations for the PVBC susceptibility parameter for all points along the phase boundary of the QP state with the four quasiclassical states. However, that would be particularly costly of computing resources for the spiral state. Even if we were to do so and hence show that the entire boundary region has PVBC order, we could still not be sure that the QP region was entirely PVBC-ordered, since there might still exists regions of other phases with other forms of order, possibly even of an exotic (spin-liquid) variety. The full characterization of the ordering within the QP phase(s) bounded by the four phases with quasiclassical ordering is an extremely challenging problem, and one that is essentially outside the scope of the present investigation.
SUMMARY AND DISCUSSION {#summary}
======================
In this paper we have studied the spin-$\frac{1}{2}$ HAFM on the honeycomb lattice with NN, NNN, and NNNN exchange interactions, namely the so-called $J_{1}$–$J_{2}$–$J_{3}$ model described by the Hamiltonian of Eq. (\[eq1\]). We have investigated the full phase diagram of the model, in the case where all the bonds are antiferromagnetic in nature (i.e., $J_{n}>0,\, n=1,2,3$), based on a combination of CCM techniques described in detail in Secs. \[CCM\] and \[results\]. In particular we have set $J_{1} \equiv 1$ to set the overall energy scale, and we have restricted attention here to the window $J_{2},J_{3} \in [0,1]$ for the remaining parameters.
Our results are summarized in the phase diagram of Fig. \[phase\]. In the window $J_{2},J_{3} \in [0,1]$ we find the five stable GS phases shown. Four of them are quasiclassical in nature, in the sense that they have counterparts in the classical version ($s \to \infty$) of the model. These comprise three phases showing collinear AFM order, viz., the Néel (N), striped (S), and anti-Néel states depicted in Figs. \[model\](a), (b), and (d) respectively, plus a noncollinear spiral phase depicted in Fig. \[model\](c). In the classical version of the model, however, only the three phases with N, S, and spiral order exist in the examined window $J_{2},J_{3} \in [0,1]$. In the classical model the phase with aN order exists only for a part of the phase space where $J_{3}<0$. We find that quantum fluctuations tend to stabilize this collinear aN phase at the expense of the spiral phase, in keeping with the very general observation that quantum fluctuations always seem to favor collinear phases over noncollinear ones. The fifth phase shown in Fig. \[phase\] is a magnetically disordered, or quantum paramagnetic (QP), phase that has no classical counterpart.
The QP phase has phase boundaries with each of the four quasiclassical phases, with three tricritical QCPs occurring in the window $J_{2},J_{3} \in [0,1]$, and a fourth presumably occurring just outside the window with a small negative value of $J_3$. The boundaries of the QP phase with each of the S, aN, and spiral phases appear to delimit first-order transitions, whereas there are strong indications that the boundary of the QP phase with the N phase delimits a continuous phase transition. At two points along this boundary, viz., along the lines $J_{3}=0$ and $J_{3}=J_{2}$, there are strong indications that the QP phase there has PVBC order. Since the N and PVBC phases break different symmetries we have argued that the transitions there favor the deconfinement scenario. Nevertheless, we cannot entirely exclude even at these points the possibilities of a very weak first-order transition or that the transition proceeds through a very narrow region of intervening phase (possibly even of an exotic spin-liquid variety).
We have found that all of the three remaining phase transition lines in the $J_{2},J_{3} \in [0,1]$ window between the pairs of quasiclassical states shown in Fig. \[phase\] (viz., between the N and S, S and spiral, and the spiral and aN phases) are first-order in nature. Rather strikingly, quantum effects in the $s=\frac{1}{2}$ model turn the continuous transition between the spiral and S states of the classical ($s \to \infty$) model into a first-order one.
Whereas in the classical model the only three phases present in the $J_{2},J_{3} \in [0,1]$ window (viz., the N, S, and spiral phases) meet at a single tricritical point at $(J_2^{c,{\rm cl}}, J_3^{c,{\rm cl}})=(0.5,0.5)$, there are now three tricritical QCPs in the same window for the $s=\frac{1}{2}$ model. The classical tricritical point separates into two tricritical QCPs at $(J_{2}^{c_1},J_{3}^{c_1})=(0.51 \pm 0.01,0.69 \pm 0.01)$ between the N, S, and QP phases, and at $(J_{2}^{c_2},J_{3}^{c_2})=(0.65 \pm 0.02,0.55 \pm 0.01)$ between the S, spiral, and QP phases. A third tricritical QCP at $(J_{2}^{c_3},J_{3}^{c_3}) = (0.69 \pm 0.01, 0.12 \pm 0.02)$, is identified for the spin-$\frac{1}{2}$ model between the spiral, aN, and QP phases.
In overall terms our results for the phase diagram are in good agreement with other very recent studies of this model. For example, a study using a combination of various exact diagonalization (ED) and self-consistent cluster mean-field (SCCMF) techniques[@Albuquerque:2011] finds a phase diagram with basically the same five phases as we identify in the window $J_{2},J_{3} \in [0,1]$. There is good agreement with the positions of the two tricritical QCPs at $(J_{2}^{c_1},J_{3}^{c_1})$ and $(J_{2}^{c_2},J_{3}^{c_2})$. The main difference seems to be in the position of the third tricritical point at $(J_{2}^{c_3},J_{3}^{c_3})$. Although both sets of calculations seem to be in good agreement about the boundary of the aN phase, the ED results generally place the boundary between the spiral and QP phases to lower values of $J_2$ such that the spiral phase occupies a larger region of phase space than our own calculations indicate. We note, however, that ED calculations are especially difficult for noncollinear phases, since the finite lattices used do not so easily sample such noncollinear phases.
A further recent study of the model, using an unbiased pseudo-fermion functional renormalization group (PFFRG) method,[@Reuther:2011] gives a phase diagram again in good overall agreement with ours, and with a phase boundary between the spiral and QP phases now in closer agreement with ours than from the ED results.[@Albuquerque:2011] Again, there is also good agreement with the phase boundaries involving the N and S states, and the positions of the two tricritical QCPs at $(J_{2}^{c_1},J_{3}^{c_1})$ and $(J_{2}^{c_2},J_{3}^{c_2})$. The only qualitative disagreement is that the PFFRG study finds no evidence in the $J_{2},J_{3} \in [0,1]$ window for the aN phase. Nevertheless, this study did find that for larger values of $J_2$ and small values of $J_3$ there were large incommensurability shifts from the spiral phase and, furthermore, that there was evidence for $J_{2} \gtrsim 0.4$ of sizeable staggered dimer order. Such staggered dimer order is often difficult in practice to distinguish from the striped AFM order.
Series expansion (SE) techniques have also been applied to this model recently.[@Oitmaa:2011] In particular, expansions were performed around our N, S, and spiral states, as well as about the second classical spiral phase and the staggered dimer valence-bond crystal (SDVBC) state, which is also called the lattice nematic state. The results from the SE analysis are much more qualitative than ours or those of the ED+SCCMF[@Albuquerque:2011] and PFFRG[@Reuther:2011] studies. Nevertheless, the SE study is also in broad agreement, with the exception again of finding no evidence for the aN state (which might have been located for the only value $J_{3}=0$ that was studied by those authors with the second spiral phase that exists classically in the range $\frac{1}{6}<J_{2}<\frac{1}{2}$ for $J_{3}=0$, and for which the aN phase is a limiting collinear form as discussed in Sec. \[model\_section\]).
It is clear that the existence or not of the aN phase in the $J_{2},J_{3} \in [0,1]$ window is one of the points on which there is still disagreement between various studies. It seems clear that, if it does indeed exist, as we argue here, it becomes unstable at very small values of $J_3$ for all values of $J_{2} < 1$ for which it exists. We note that a quite different recent ED study[@Mosadeq:2011] of the model (along the $J_{3}=0$ line only), which used a NN singlet valence-bond basis, found very similar critical points to ours for the boundaries of the QP phase, but found that while the QP phase was bounded on one side (for smaller values of $J_2$) by the N phase, it was bounded on the other side by a SDVBC phase. However, these results would again be equally consistent with the identification of this phase as our aN phase, since their interpretation is strongly biased by their choice of basis.
A quite separate study of the case $J_{3}=0$ has also been performed recently based on an entangled-plaquette variational (EPV) ansatz.[@Mezzacapo:2012] This EPV study uses a single very broad class of entangled-plaquette states as trial wave functions, and finds in this very unbiased way that along the $J_{3}=0$ line the model has N, QP, and aN phases with critical points very close to ours. In the same region studied by us (viz., $J_{2}<1$) the EPV study finds no evidence of spiral order along the $J_{3}=0$ line.
It is clear that the spin-$\frac{1}{2}$ $J_{1}$–$J_{2}$–$J_{3}$ HAFM on the honeycomb lattice is a challenging model, but one in which there seems now to be a growing consensus on its overall phase structure. There is very good agreement over the regions in which the N and S phases exist, and we believe our own CCM results now give perhaps the best quantitative results in these cases for the positions of the phase boundaries and the positions of the two tricritical QCPs at $(J_{2}^{c_1},J_{3}^{c_1})$ and $(J_{2}^{c_2},J_{3}^{c_2})$.
An uncertainty remains over the precise extent of the phase with spiral order, and whether or not there is an aN phase along the $J_{3}=0$ line for values of $J_2$ beyond the point where the QP phase disappears, and hence also for small positive values of $J_3$ up to the point where spiral order sets in. The other main uncertainty is the nature of the QP phase itself. We have argued here that over at least some widely separated points on the boundaries with the N, S, and aN phases the QP phase has PVBC order. Two quite separate ED calculations[@Albuquerque:2011; @Mosadeq:2011] also give clear evidence that much of the QP phase has PVBC order, although the latter calculations[@Mosadeq:2011] are only done along the $J_{3}=0$ line.
By contrast the EPV calculations[@Mezzacapo:2012] along the $J_{3}=0$ line seem to favor a disordered (spin-liquid) phase, while spin-wave calculations[@Mulder:2010] favor SDVBC order along the same line in the QP regime. The PFFRG study,[@Reuther:2011] also done over the entire $J_{2},J_{3} \in [0,1]$ window, finds evidence too that a large part of the QP regime has strong SDVBC order, while the part with smaller values of $J_2$ has only weak PVBC order. On the other hand the SE study[@Oitmaa:2011] finds that SDVBC order is not favored, at least for low values of $J_3$.
We should note, however, that the SE study is in broad disagreement with most other studies along the $J_{3}=0$ line, in that it finds no evidence at all for a magnetically disordered phase there, but instead finds that the N phase first gives way to the second classical spiral phase, and then later to the spiral phase considered here, as $J_2$ is increased. On the other hand, the SE results are consistent with the finding from the ED+SCCMF analysis[@Albuquerque:2011] that at least for some parameter ranges the SDVBC state might be very difficult to distinguish from magnetically ordered states such as our S state.
Finally we note that the ED+SCCMF study also presents evidence for the N to PVBC transition being a strong candidate for a deconfined transition, just as we have found in our earlier CCM studies of the model along the $J_{3}=J_{2}$ line[@DJJF:2011_honeycomb] and along the $J_{3}=0$ line.[@PHYLi:2012_honeyJ1-J2] Clearly this model still has open questions, but we believe that the CCM results presented here have furthered our understanding of it.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
We thank J. Richter for fruitful discussions. We are also grateful to J. Schulenburg for his assistance in the updating and maintenance of the CCM computer code. We thank the University of Minnesota Supercomputing Institute for the grant of supercomputing facilities.
[200]{}
A. P. Ramirez, Nat. Phys. [**4**]{}, 442 (2008).
L. Balents, Nature (London) [**464**]{}, 199 (2010).
P. W. Anderson, Mater. Res. Soc. Bull. [**8**]{}, 153 (1973).
P. W. Anderson, Science [**235**]{}, 1196 (1987).
B. S. Shastry and B. Sutherland, Physica B [**108**]{}, 1069 (1981).
P. Fazekas and P. W. Anderson, Phil. Mag. [**30**]{}, 423 (1974).
E. Rastelli, A. Tassi, and L. Reatto, Physica B & C [**97**]{}, 1 (1979).
A. Mattsson, P. Fröjdh, and T. Einarsson, Phys. Rev. B [**49**]{}, 3997 (1994).
J. B. Fouet, P. Sindzingre, and C. Lhuillier, Eur. Phys. J. B [**20**]{}, 241 (2001). A. Mulder, R. Ganesh, L. Capriotti, and A. Paramekanti, Phys. Rev. B [**81**]{}, 214419 (2010).
D. C. Cabra, C. A. Lamas, and H. D. Rosales, Phys. Rev. B [**83**]{}, 094506 (2011).
R. Ganesh, D. N. Sheng, Y.-J. Kim, and A. Paramekanti, Phys. Rev. B [**83**]{}, 144414 (2011).
B. K. Clark, D. A. Abanin, and S. L. Sondhi, Phys. Rev. Lett. [**107**]{}, 087204 (2011).
J. Reuther, D. A. Abanin, and R. Thomale, Phys. Rev. B [**84**]{}, 014417 (2011).
D. J. J. Farnell, R. F. Bishop, P. H. Y. Li, J. Richter, and C. E. Campbell, Phys. Rev. B [**84**]{}, 012403 (2011).
A. F. Albuquerque, D. Schwandt, B. Hetényi, S. Capponi, M. Mambrini, A. M. Läuchli, Phys. Rev. B [**84**]{}, 024406 (2011).
H. Mosadeq, F. Shahbazi, and S. A. Jafari, J. Phys.: Condens. Matter [**23**]{}, 226006 (2011).
J. Oitmaa and R. R. P. Singh, Phys. Rev. B [**84**]{}, 094424 (2011).
F. Mezzacapo and M. Boninsegni, Phys. Rev. B [**85**]{}, 060402(R) (2012).
P. H. Y. Li, R. F. Bishop, D. J. J. Farnell, J. Richter, and C. E. Campbell, Phys. Rev. B [**85**]{}, 085115 (2012).
P. H. Y. Li, R. F. Bishop, D. J. J. Farnell, and C. E. Campbell, arXiv:1201.3512v1 \[cond-mat.str-el\] (2012).
P. H. Y. Li and R. F. Bishop, arXiv: 1202.6249v1 \[cond-mat.str-el\] (2012).
A. Kitaev, Ann. Phys. (N.Y.) [**321**]{}, 2 (2006); G. Baskaran, S. Mandal, and R. Shankar, Phys. Rev. Lett. [**98**]{}, 247201 (2007); J. Chaloupka, G. Jackeli, and G. Khaliullin, [*ibid.*]{} [**105**]{}, 027204 (2010).
A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, and A. K. Geim, Rev. Mod. Phys. [**81**]{}, 109 (2009).
Z. Y. Meng, T. C. Lang, S. Wessel, F. F. Assaad, and A. Muramatsu, Nature (London) [**464**]{}, 847 (2010).
H. Y. Yang and K. P. Schmidt, Europhys. Lett. [**94**]{}, 17004 (2011).
A. Vaezi and X. G. Wen, arXiv:1010.5744v1 \[cond-mat.str-el\] (2010).
A. Vaezi, M. Mashkoori, and M. Hosseini, arXiv:1110.0116v2 \[cond-mat.str-el\] (2011).
S. Okubo, F. Elmasry, W. Zhang, M. Fujisawa, T. Sakurai, H. Ohta, M. Azuma, O. A. Sumirnova, and N. Kumada, J. Phys.: Conf. Series [**200**]{}, 022042 (2010).
Y. Miura, R. Hiari, Y. Kobayashi, and M. Sato, J. Phys. Soc. Jpn. [**75**]{}, 084707 (2006).
V. Kataev, A. Möller, U. Löw, W. Jung, N. Schittner, M. Kriener, and A. Freimuth, J. Magn. Magn. Mater. [**290/291**]{}, 310 (2005).
L. P. Regnault and J. Rossat-Mignod, in [*Phase Transitions in Quasi-Two-Dimensional Planar Magnets*]{}, edited by L. J. De Jongh (Kluwer Academic Publishers, Dordrecht, 1990), p. 271.
A. A. Tsirlin, O. Janson, and H. Rosner, Phys. Rev. B [**82**]{}, 144416 (2010).
J. Struck, C. [Ö]{}lsch[ä]{}ger, R. Le Targat, P. Soltan-Panahi, A. Eckardt, M. Lewenstein, P. Windpassinger, and K. Sengstock, Science [**333**]{}, 996 (2011).
V. Murg, F. Verstraete, and J. I. Cirac, Phys. Rev. B [**79**]{}, 195119 (2009).
H. J. Schulz, T. A. L. Ziman, and D. Poilblanc, J. Phys. I [**6**]{}, 675 (1996).
J. Richter and J. Schulenberg, Eur. Phys. J. B [**73**]{}, 117 (2010).
J. Reuther, P. Wölfle, R. Darradi, W. Brenig, M. Arlego, and J. Richter, Phys. Rev. B [**83**]{}, 064416 (2011).
S. E. Krüger, J. Richter, J. Schulenburg, D. J. J. Farnell, and R. F. Bishop, Phys. Rev. B [**61**]{}, 14607 (2000).
R. Darradi, J. Richter, and D. J. J. Farnell, Phys. Rev. B [**72**]{}, 104425 (2005). D. Schmalfu[ß]{}, R. Darradi, J. Richter, J. Schulenburg, and D. Ihle, Phys. Rev. Lett. [**97**]{}, 157201 (2006).
R. F. Bishop, P. H. Y. Li, R. Darradi, J. Schulenburg, and J. Richter, Phys. Rev. B [**78**]{}, 054412 (2008).
R. F. Bishop, P. H. Y. Li, R. Darradi, and J. Richter, J. Phys.: Condens. Matter [**20**]{}, 255251 (2008).
R. Darradi, O. Derzhko, R. Zinke, J. Schulenburg, S. E. Krüger, and J. Richter, Phys. Rev. B [**78**]{}, 214415 (2008).
R. F. Bishop, P. H. Y. Li, D. J. J. Farnell, and C. E. Campbell, Phys. Rev. B [**79**]{}, 174405 (2009).
J. Richter, R. Darradi, J. Schulenburg, D. J. J. Farnell, and H. Rosner, Phys. Rev. B [**81**]{}, 174429 (2010).
R. F. Bishop, P. H. Y. Li, D. J. J. Farnell, and C. E. Campbell, Phys. Rev. B [**82**]{}, 024416 (2010).
R. F. Bishop, P. H. Y. Li, D. J. J. Farnell, J. Richter, and C. E. Campbell, arXiv:1202.2722v1 \[cond-mat.str-el\] (2012).
C. Zeng, D. J. J. Farnell, and R. F. Bishop, J. Stat. Phys. [**90**]{}, 327 (1998).
R. F. Bishop, D. J. J. Farnell, and J. B. Parkinson, Phys. Rev. B [**58**]{}, 6394 (1998).
R. F. Bishop, D. J. J. Farnell, S. E. Krüger, J. B. Parkinson, J. Richter, and C. Zeng, J. Phys.: Condens. Matter [**12**]{}, 6887 (2000).
We use the program package CCCM of D. J. J. Farnell and J. Schulenburg, see http://www-e.uni-magdeburg.de/jschulen/ccm/index.html.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Much of our understanding of the internal structure of the Sun derives from so-called standard theoretical solar models. Unfortunately, none of those models agrees completely with observation. The discrepancy is commonly associated with chemical abundance, and has led to what is now called the solar abundance problem, the resolution of which has previously been out of sight. But now the Borexino Collaboration, who recently announced measurements of the pp-chain solar neutrinos, are optimistic that they will be able to measure the flux $\Phi_{\rm CNO}$ of the neutrinos emitted by the relatively weak CNO cycle. Since C, N and O constitute the majority of the heavy elements, that measurement will permit a crucial determination of the heavy-element abundance $Z_{\rm c}$ in the Sun’s energy-generating core, thereby shedding important light on the problem. To accomplish that determination, a robust relation between $Z_{\rm c}$ and $\Phi_{\rm CNO}$ will be required. That relation is $Z_{\rm c}= 0.400\, \Phi_{\rm CNO}$, where $\Phi_{\rm CNO}$ is in units of $10^{10} {\rm cm}^{-2}{\rm s}^{-1}$.'
bibliography:
- 'refs.bib'
title: 'Anticipating the Sun’s heavy-element abundance'
---
\[firstpage\]
Sun: neutrinos – Sun: helioseismology – Sun: abundances – opacity
Introduction {#sec:introduction}
============
The recent report by @borexino2018 of the impressive measurement of pp-chain solar neutrinos has provided the first observational confirmation of the agreement (to within about 10%) between the nuclear energy production rate in the Sun and the radiant luminosity, confirming that the Sun is more-or-less in thermal balance, as it is normally assumed to be. The result is a significant step towards deepening our understanding of the inner workings of a star
The collaboration relate their results to two modern sets of standard solar models (SSMs) published by @2017ApJ...835..202V, one with the commonly adopted GS98 [@1998SSRv...85..161G] abundances, the other with the more recent, and lower, AGSS09 values, which appear to be better representations of photospheric values [see also @2011SoPh..268..255C]. By comparing $^{7}{\rm Be}$ and $^{8}{\rm B}$ fluxes, suitably adjusted for flavour transitions, with theoretical values from the models, the collaboration report a preference for the former composition, which is in conflict with the standard assumption that, aside from gravitational settling and radiative levitation, abundances in the radiative envelope are the same as those in the photosphere. However, the preference is at least comforting because it has long been known that solar models with low heavy-element abundance, $Z$, are ruled out by helioseismology [@DOGBridgeingap1983Natur; @duvallharveybridgeingap1983Natur; @dirtysunobs1998Obs.118.25C; @basuantia2008PhR...457..217B]. @2017ApJ...835..202V present a careful demonstration of this issue in the specific case of the GS98 and AGSS09 abundances. The conflict has been named the solar abundance problem.
Circumventing opacity {#sec:circumvention}
=====================
Opacity, which depends directly on $Z$, is the principal cause of the model differences, via its control of the flow of heat in the Sun’s radiative envelope. A lower opacity requires a smaller temperature gradient, and hence a lower temperature, leading to a lower sound speed in the radiative envelope, although the latter is offset in the energy-generating core by the higher hydrogen abundance, $X$, required to fuel the otherwise slower nuclear reactions. A reliable knowledge of opacity is therefore a crucial ingredient for understanding the ‘so-called’ standard structure and evolution of the Sun. Its calculation involves very complicated physics, and the outcome has commonly been questioned.
It should be appreciated, however, that nuclear reactions do not themselves depend directly on opacity. Moreover, the pertinent properties of the internal structure of the Sun can be determined, subject to relatively minor additional assumptions, by seismological analysis of acoustic modes of oscillation. Acoustic propagation depends on relatively simple physics, rendering correctly interpreted, yet admittedly more limited, inferences from helioseismology more reliable than those from SMMs.
Setting aside an acoustic glitch that is present immediately beneath the convection zone, whose origin at least in part is material redistribution in the tachocline [@elliottdogtachthickness1999ApJ...516..475E; @2018MNRAS.477.3845C], ignored in SSMs but having a relatively minor impact on the overall stratification of the deep interior of the star [@2017ApJ...835..202V], the principal flaw in the construction of SSMs is likely to be either an error in the opacity calculation, a possibility that is now not wholly accepted (but there is, in particular, a degree of acceptance resulting from recent laboratory measurements by @baileyetalironopacity2015Natur), or it is the standard assumption that photospheric abundances directly reflect the abundances in the radiative envelope [@guzik_mussack_2010ApJ...713.1108G]. A potential resolution could be, for example, that mechanical waves generated off-resonance at the base of the convection zone [e.g. @press1981ApJ...245..286P] have amplitudes enough to carry a significant, yet uncertain, fraction of the total luminosity. If that were so, the role of opacity would be substantially diminished. The associated wave momentum flux would have negligible influence on the hydrostatic balance, which alone (aside from the adiabatic exponent $\gamma_1$, the theory of which is relatively robust in the radiative interior) determines the Sun’s seismic structure. However, the thermal structure would then depend on the uncertain details of the wave spectrum. Without knowing the thermal structure, no SMM can be trusted.
Determining $Z$ via seismology {#sec:seismology}
==============================
As the Borexino Collaboration point out, their anticipated measurement of the neutrino flux $\Phi_{\rm CNO}$ produced by the CNO cycle will provide a direct evaluation of the CNO abundances, the dominant contributors to $Z$, irrespective of opacity. To accomplish that, a relation between $Z$ and $\Phi_{\rm CNO}$ will be needed. One might be tempted to interpolate between SMMs such as those provided by @2017ApJ...835..202V, but the outcome would be subject to their reliability, which, as I pointed out above, is in doubt. However, an adequate estimate of pertinent conditions in the core can be obtained more reliably from helioseismology, provided one adopts an assumption such as a core that was initially homogeneous and which suffered no material redistribution of the products of the nuclear reactions during the subsequent main-sequence evolution, an assumption that is adopted also in creating SMMs. For the purpose of estimating $Z$, it is adequate to linearize the difference between the seismic structure of the Sun and an appropriate accurately computed SSM which, in the case of the representation I use here [@DOGLorentz2004AIPC], was Model S of @1996Sci...272.1286C. The outcome is an estimate of the sound speed $c$ and the density $\rho$, which are subject to an uncertainty of a few parts in $10^3$ [cf. @mtdog2003ESASP.517..397T; @basuantia2008PhR...457..217B], from which the pressure $p$ can be determined from the constraint of hydrostatic support. It is then necessary to estimate the helium abundance $Y$ and the temperature $T$, neither of which is seismically accessible. To this end a shell representing the tachocline in Model S was homogenized with the convection zone, and then a constant $\delta Y$ was added to $Y$, enabling $T$ to be determined implicitly from $c$, $p$ and $\rho$ with the help of the equation of state: $\delta Y$ and $T$ were determined simultaneously by requiring that the nuclear energy generation rate in the core is equal to the observed luminosity at the surface. The resulting structure, Model Ss, is presented by @DOGLorentz2004AIPC. To the precision required here, that structure is independent of the reference SSM adopted [cf. @basupinsonneaultbahcall2000ApJ...529.1084B], and of the imperfect [cf. @2018MNRAS.477.3845C] representation of the tachocline.
The relation between the central heavy-element abundance $Z_{\rm c}$ and the CNO neutrino fluxes was obtained using the cross-sections adopted by @2017ApJ...835..202V. It was achieved by scaling model B16-GS98 of @2017ApJ...835..202V to the seismic structure of the Sun, in the form of Model Ss, using the functional forms of the CNO reaction rates, presumed to be in equilibrium. The relative abundances of C, N and O are thereby determined, irrespective of their initial values [e.g. @claytonbook1983]. The outcome is that the Sun produces neutrino fluxes $\Phi_{13} = 1.41$, $\Phi_{15} = 1.06$ and $\Phi_{17} = 0.027$ per $Z_{\rm c}$, all in units of $10^{10} {\rm cm}^{-2}{\rm s}^{-1}$, from the beta decays of $^{13}$N, $^{15}$O and $^{17}$F respectively. The central heavy-element abundance is related to the the total CNO neutrino flux according to $$Z_{\rm c}= \alpha \Phi_{\rm CNO}\,,
\label{eq:ZPhi}$$ where $\alpha = 0.400$ and $\Phi_{\rm CNO}$ is also in units of $10^{10} {\rm cm}^{-2}{\rm s}^{-1}$. The uncertainty in $\alpha$ is dominated by uncertainties in the nuclear reaction rates, which are detailed by @borexino2018 and @2017ApJ...835..202V.
Unlike hydrogen and helium abundances, the total CNO abundance is unaltered by the nuclear reactions, so the relation between $Z_{\rm c}$ and the heavy-element abundance Z throughout the radiative envelope is relatively secure, depending only on the weak variation resulting from gravitational settling. Hence one can be assured that the analysis leading to equation (\[eq:ZPhi\]) is reliable. We now await the promised future measurement by the Borexino Collaboration to resolve the abundance issue.
I thank T. Sekii and the referee for useful suggestions.
\[lastpage\]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Random K-out graphs are used in several applications including modeling by sensor networks secured by the [*random pairwise*]{} key predistribution scheme, and payment channel networks. The random K-out graph with $n$ nodes is constructed as follows. Each node draws an edge towards $K$ distinct nodes selected uniformly at random. The orientation of the edges is then ignored, yielding an [*undirected*]{} graph. An interesting property of random K-out graphs is that they are [*connected*]{} almost surely in the limit of large $n$ for any $K \geq2$. This means that they attain the property of being connected very [*easily*]{}, i.e., with far fewer edges ($O(n)$) as compared to classical random graph models including Erdős-Rényi graphs ($O(n \log n)$). This work aims to reveal to what extent the asymptotic behavior of random K-out graphs being connected easily extends to cases where the number $n$ of nodes is [*small*]{}. We establish upper and lower bounds on the probability of connectivity when $n$ is finite. Our lower bounds improve significantly upon the existing results, and indicate that random K-out graphs can attain a given probability of connectivity at much smaller network sizes than previously known. We also show that the established upper and lower bounds *match* order-wise; i.e., further improvement on the order of $n$ in the lower bound is not possible. In particular, we prove that the probability of connectivity is $1-\Theta({1}/{n^{K^2-1}})$ for all $K \geq 2$. Through numerical simulations, we show that our bounds closely mirror the empirically observed probability of connectivity.'
author:
-
bibliography:
- 'IEEEabrv.bib'
- 'references.bib'
title: |
Tight Bounds for the Probability of Connectivity\
in Random K-out Graphs
---
Random Graphs, Connectivity, Wireless Sensor Networks, Security
Introduction
============
Random graphs constitute an important framework for analyzing the underlying structural characteristics of complex real-world networks such as communication networks, social networks and biological networks[@boccaletti2006complex; @goldenberg2010survey; @newman2002random]. A class of random graphs called the random K-out graphs is one of the earliest known models of random graphs [@FennerFrieze1982; @Bollobas]. The random K-out graph comprising $n$ nodes, denoted by $\mathbb{H}(n;K)$, is constructed as follows. Each node draws $K$ edges towards $K$ distinct nodes chosen uniformly at random from all other nodes. The orientation of the edges is then ignored, yielding an [*undirected*]{} graph. Due to their unique connectivity properties, random K-out graphs have received renewed interest for analyzing secure wireless sensor networks and routing in cryptocurrency networks.
In the context of wireless sensor networks (WSNs), random K-out graphs have been used extensively for evaluating strategies for secure communication. The limited computation and communication capabilities of WSNs precludes the use of traditional key exchange protocols for establishing secure connectivity [@Gligor_2002; @perrig2004security; @XiaoSurvey]. Moreover, WSNs deployed for applications such as battlefield surveillance and environmental monitoring are vulnerable to adversarial attacks and operational failures. For facilitating secure connectivity in WSNs, Eschenauer and Gligor [@Gligor_2002] proposed the [*random*]{} predistribution of symmetric cryptographic keys. Subsequently, several variants of random key predistribution schemes have been studied; see [@security_survey; @XiaoSurvey] and the references therein. A widely adopted approach is the random *pairwise* key predistribution introduced by Chan et al. [@Haowen_2003]. The random pairwise scheme is implemented in two phases. In the first phase, each sensor node is paired [*offline*]{} with $K$ distinct nodes chosen uniformly at random among all other sensor nodes. Next, a *unique* pairwise key is inserted in the memory of each of the paired sensors. After deployment, two sensor nodes can communicate securely only if they have at least one key in common. ; see Figure \[fig:0\]. In Section \[sec:Model\], we provide more details about the implementation of this scheme. The deployment of unique, pairwise keys brings several advantages including resilience against node capture and replication attacks, and quorum-based key revocation [@Haowen_2003].
In the context of cryptocurrency networks, a growing body of work is investigating the efficacy of routing protocols over different network topologies [@sivaraman2018high; @pcn_spider2018; @tang2019privacy]. A structure analogous to random K-out graphs have been proposed to make message propagation robust to [*de-anonymization*]{} attacks [@FantiDandelion2018 Algorithm 1]. In order to make cryptocurrency networks more scalable, payment channel networks (PCNs) such as the Lightning network have been introduced. A key challenge in the design of PCNs is the trade-off between the number of edges in the network (which is constrained since each edge corresponds to funds escrowed in the PCN) and connectivity (which is desirable to facilitate transactions between participating nodes). Given their ability to get connected with a relatively smaller number of edges, random K-out graphs offer a promising potential for informing the topological properties of such networks.
![*A WSN with $6$ nodes secured by the random pairwise key predistribution scheme with scheme parameter $K=2$. Each node randomly selects $K=2$ nodes and unique pairwise cryptographic keys are given to node pairs per selection. Two nodes can communicate securely if they have at least one key in common. This induces a graph with edges corresponding to node pairs that share a key.*[]{data-label="fig:0"}](kout.pdf)
In several networked applications, [*connectivity*]{} is a fundamental determinant of the system performance. For instance, connectivity enables any pair of nodes to exchange messages in a communication network, or exchange funds in a cryptocurrency network. However, establishing links can be costly and often the goal is to obtain a connected network as [*efficiently*]{} as possible, i.e., by using the least amount of resources (links). The connectivity of random K-out graphs and their heterogeneous variants have been extensively studied [@FennerFrieze1982; @Yagan2013Pairwise; @eletrebycdc2018; @mansi_globecom]. It is known [@Yagan2013Pairwise; @FennerFrieze1982] that random K-out graphs are connected with probability tending to one (as $n \rightarrow \infty$) if and only if $K \geq 2$. In particular, the following zero-one law holds: $$\lim_{n \to \infty} \mathbb{P}\left[ \mathbb{H}(n;K) \text{ is connected}\right] =
\begin{cases}
1 & \mathrm{if} ~~ K\geq 2, \\
0 & \mathrm{if} ~~ K=1.
\end{cases}
\label{eq:homogeneous_zero_one_law}$$ A key advantage of ${\mathbb{H}(n;K)}$ is its ability to get connected very easily. With $K=2$, ${\mathbb{H}(n;K)}$ contains at most 2n edges meaning that on average, each node has a degree of less than 4. On the other hand, the classical Erdős-Rényi (ER) random graph [@erdos61conn] requires an average degree of the order of $\log n$ for connectivity; other models with similar connectivity behavior to ER graphs include random key graphs [@yagan2012zero] and random geometric graphs [@PenroseBook]. Most existing results for random K-out graphs describe the behavior of the network when the number of nodes $n$ approaches $\infty$ in the form of asymptotic zero-one laws. However, in practical scenarios, the number of nodes in the network are often constrained to be finite. This raises the need to go beyond the asymptotic results (valid for $n \rightarrow \infty$) and obtain as tight bounds as possible for the case when $n$ is small.
Let $P(n;K)$ denote the probability of connectivity of ${\mathbb{H}(n;K)}$ as a function of the number of nodes ($n$) and the number of selections ($K$) per node. Connectivity is a monotonic increasing property in the number of edges and as a consequence $P(n;K)$ increases as $K$ increases. The case where $K=2$ corresponds to the critical threshold (\[eq:homogeneous\_zero\_one\_law\]) for connectivity. Therefore, we first focus on deriving bounds on $P(n;K)$ for the case $K=2$, and then generalize them to all $K \geq 2$. First, we derive the best known lower bound for $P(n;2)$, i.e., the probability of connectivity for $K=2$. Next, by deriving an upper bound on $P(n;2)$, we show that the lower bound *matches* the upper bound order-wise, implying that further improvement on the order of $n$ is not possible. While our key focus is on the $K=2$ threshold, we also derive the lower and upper bounds for all $K \geq 2$, with our lower bound beating the existing bounds for all $K \geq 2$; see Section \[sec:results\] for a detailed comparison of the bounds and the empirical probability of connectivity. Moreover, to the best of our knowledge, our work is the first to derive an upper bound on $P(n;K)$, which shows that the lower bound is order-wise optimal. The *matching* upper and lower bounds for the general case $K=2$ derived in this paper establish that the probability of connectivity is $1-\Theta({1}/{n^{K^2-1}})$, i.e, the probability of [*not*]{} being connected decays as $\Theta({1}/{n^{K^2-1}})$. Our results significantly improve the probabilistic guarantees for network designs that induce random K-out graphs. For example, we show that $n=30$ (resp. $n=60$) is sufficient to have a probability of connectivity of $1-10^{-4}$ (resp. $1-10^{-5}$), while the best known previous result would indicate that $n \geq 72$ (resp. $n \geq 150$) is necessary.
**Organization:** In Section \[sec:Model\] we describe the random pairwise scheme and the resulting random K-out graphs. In Section \[sec:results\] we present our bounds for connectivity in random K-out graphs and compare them with existing results. We present the proofs of the lower and upper bounds, respectively in Sections \[sec:upperbound\] and \[sec:lowerbound\], and conclude in Section \[sec:conc\].
**Notation:** All limits are understood with the number of nodes $n$ going to infinity. While comparing asymptotic behavior of a pair of sequences $\{a_n\},\{b_n\}$, we use $a_n = \oo(b_n)$, $a_n=\omega(b_n)$, $a_n = \OO(b_n)$, $a_n=\Theta(b_n)$, [and $a_n = \Omega(b_n)$]{} with their meaning in the standard Landau notation. All random variables are defined on the same probability triple $(\Omega, {\mathcal{F}}, \mathbb{P})$. Probabilistic statements are made with respect to this probability measure $\mathbb{P}$, and we denote the corresponding expectation operator by $\mathbb{E}$. The cardinality of a discrete set $A$ is denoted by $|A|$ and the set of all positive integers by ${\mathbb{N}}_0$.
Model: Random K-out Graphs {#sec:Model}
==========================
The random pairwise key predistribution scheme of Chan et al. is parametrized by two positive integers $n$ and $K$ such that $K <
n$. This scheme is implemented as follows. Consider a network comprising of $n$ nodes [indexed by labels]{} $i=1,2,\dots n$ with unique IDs: ${\rm Id}_1, \ldots , {\rm Id}_n$. Each of the $n$ nodes draws $K$ edges towards $K$ distinct nodes chosen uniformly at random from among all other nodes. Nodes $v_i$ and $v_j$ are deemed to be [*paired*]{} if at least one of them selected the other; i.e., either $v_i$ selects $v_j$, or $v_j$ selects $v_i$, or both. Once the offline pairing process is complete, the set of keys to be inserted to nodes are determined as follows. For any $v_i, v_j$ that are [*paired*]{} with each other as described above, a unique pairwise key $\omega_{ij}$ is generated and inserted in the memory modules of both nodes $v_i$ and $v_j$ along with the corresponding node IDs. It is important to note that $\omega_{ij}$ is assigned *exclusively* to nodes $v_i$ and $v_j$ to be used solely in securing the communication between them. In the post-deployment *key-setup* phase, nodes first broadcast their IDs to their neighbors following which each node searches for the corresponding IDs in their key rings. Finally, nodes that have been paired verify each others’ identities through a cryptographic handshake [@Haowen_2003].
Let ${\mathcal{N}}:=\{1,2,\dots,n\}$ denote the set of node labels. For each $i \in {\mathcal{N}}$, let $\Gamma_{n,i} \subseteq {\mathcal{N}}_{-i}$ denote the labels selected by node $v_i$ (uniformly at random from ${\mathcal{N}}_{-i}$). Specifically, for any subset $A \subseteq {\cal N}_{-i}$, we have $${\mathbb{P}}{[\Gamma_{n,i} = A ]} = \left \{
\begin{array}{cl}
{{n-1}\choose{K}}^{-1} & \mbox{if $|A|=K$} \\
0 & \mbox{otherwise.}
\end{array}
\right .
\label{eq:main_eqn_for_gamma}$$ Thus, the selection of $\Gamma_{n,i}$ is done [*uniformly*]{} amongst all subsets of ${\cal N}_{-i}$ which are of size exactly $K$. Under the *full-visibility* assumption, i.e., when one-hop secure communication between a pair of sensors hinges solely on them having a common key, a WSN comprising of $n$ sensors secured by the pairwise key predistribution scheme can be modeled by a random K-out graph defined as follows. With $n=2,3, \ldots $ and positive integer $K < n$, we say that two distinct nodes $v_i$ and $v_j$ are adjacent, denoted by $v_i \sim v_j$ if they have at least one common key in their respective key rings. More formally, $$\begin{aligned}
v_i \sim v_j ~~\quad \mbox{if} ~~~\quad j \in \Gamma_{n,i} \vee i \in \Gamma_{n,j}.
\label{eq:Adjacency}\end{aligned}$$ Let $\mathbb{H}(n;K)$ denote the undirected random graph on the vertex set $\{ v_1, \ldots , v_n \}$ induced by the adjacency notion (\[eq:Adjacency\]). In the literature on random graphs, $\mathbb{H}(n;K)$ is often referred to as a random $K$-out graph and have been widely studied [@Bollobas; @FennerFrieze1982; @philips1990diameter; @Yagan2013Pairwise; @yavuz2017k; @yagan2013scalability].
Results and Discussion {#sec:results}
======================
In this section, we present our main results, upper and lower bounds for the probability of connectivity of ${\mathbb{H}(n;K)}$, and compare them with existing results. Throughout, we write $$P(n;K)
:=
{\mathbb{P}}[{~ \mathbb{H}(n;K) ~\mbox{is connected}~}] .$$
Main results
------------
We provide our first technical result– an upper bound for the probability of connectivity $P(n;K)$.
\[thm:UpperBoundForConnectivity\] For any fixed positive integer $K\geq 2$, we have $$\begin{aligned}
P(n;K)
\leq 1- \frac{({K!})^K e^{-K(K+1)}}{{K+1}}\cdot \frac{1}{n^{K^2-1}} (1+\oo(1))
\label{eqn:UpperBoundForConnectivity}\end{aligned}$$
We present the asymptotic version of the upper bound in Theorem \[thm:UpperBoundForConnectivity\] to make it easier to interpret; see Appendix for the more detailed bound with an explicit expression replacing the $(1+o(1))$ term in (\[eqn:UpperBoundForConnectivity\]). The dependence of the upper bound on the scheme parameter $K$ can be succinctly captured as follows.
For a fixed positive integer $K\geq 2$, $$\begin{aligned}
P(n;K)=1 - \Omega\left(\frac{1}{n^{K^2-1}}\right).
\label{eq:upperbound-asymptotic}\end{aligned}$$
Given a fixed value of the parameter $K$ ($K \geq 2$), we derive the upper bound on the probability of connectivity by computing the likelihood of existence of isolated components comprising $K+1$ nodes. Due to space constraints, we outline the proof for the case of $K=2$ in Section \[sec:upperbound\] and present the full proof ($K\geq2$) in the Appendix. In our second main result, we derive an order-wise [*matching*]{} lower bound and show that the probability of connectivity is also $1-O\left(\frac{1}{n^{K^2-1}}\right)$.
[*For any fixed positive integer $K\geq 2$, for all $ n \geq 4(K+2)$, we have $$\begin{aligned}
P(n;K)
\geq 1 - c(n;K) Q(n;K)
\label{eq:LowerBoundForConnectivity}\end{aligned}$$ where,* ]{} \[thm:LowerBoundForConnectivity\]$$\begin{aligned}
c(n;K)&= \frac{e^{- (K^2-1)(1- \frac{K+1}{n})}}{\sqrt{2 \pi (K+1)}}\sqrt{\frac{n}{(n-K-1)}},
\label{eq:c(K)}\\
Q(n;K)
&=
\left ( \frac{K+1}{n} \right )^{K^2-1}
+ \frac{n}{2}
\left ( \frac{K+2}{n} \right )^{(K+2)(K-1)}
\label{eq:q(K)}\end{aligned}$$
For any fixed positive integer $K\geq 2$ we have $$\begin{aligned}
P(n;K)
= 1 - O\left(\frac{1}{n^{K^2-1}}\right).\label{eq:lowerbound-asymptotic}\end{aligned}$$
This shows that our lower bound (\[eq:LowerBoundForConnectivity\]) for connectivity [*matches*]{} our upper bound (\[eqn:UpperBoundForConnectivity\]), and is therefore order-wise optimal. Combining (\[eq:upperbound-asymptotic\]) and (\[eq:lowerbound-asymptotic\]), we obtain the following result.
[*For any positive integer $K\geq 2$, for all $ n \geq 4(K+2)$, we have $$\begin{aligned}
P(n;K)
= 1 - \Theta\left(\frac{1}{n^{K^2-1}}\right)\label{eq:corollary-asymptotic}\end{aligned}$$* ]{}
The above equation indicates how rapidly $P(n;K)$ converges to one as $n$ grows large.
Previous results in [@Yagan2013Pairwise; @FennerFrieze1982]
-----------------------------------------------------------
We present a summary of the related lower bounds [@Yagan2013Pairwise; @FennerFrieze1982] on the probability of connectivity. To the best of our knowledge, our work is the first to compute an upper bound on the probability of connectivity for random K-out graphs.
### Earlier results by Yağan and Makowski [@Yagan2013Pairwise]
It was established [@Yagan2013Pairwise Theorem 1] that for $K \geq 2$, [ $$P(n;K)
\geq 1 - a(K) Q(n;K)
\label{eq:oyam_lb}$$ holds for all $n \geq n(K)$ with $n(K) = 4(K+2)$, where ]{} $$a(K)= e^{-\frac{1}{2} (K+1)(K-2)}.
\label{eq:a(K)}$$
### Earlier results by Fenner and Frieze [@FennerFrieze1982]
A lower bound for probability of connectivity can be inferred from the proof of [@FennerFrieze1982 Theorem 2.1, p. 348]. Upon inspecting Eqn. 2.2 in [@FennerFrieze1982 p. 349] with $p=0$; it can be inferred that $$P(n;K) \geq 1 - b(n;K) Q(n;K)
\label{eq:BoundByFF}$$ holds for all $n$ and $K$ such that $K < n$, where $$\begin{aligned}
b(n;K) = \frac{12 n}{12 n -1}\sqrt{ \frac{1}{ 2\pi (K+1) } }
\sqrt{\frac{n}{n-K-1}}.
\label{eq:b(K)}\end{aligned}$$ Observe from (\[eq:LowerBoundForConnectivity\]), (\[eq:oyam\_lb\]) and (\[eq:BoundByFF\]), that the smaller the values of $c(n;K), a(K)$ and $b(n;K)$, the better is the corresponding lower bound. As discussed in [@Yagan2013Pairwise], the bound (\[eq:BoundByFF\]) by Fenner and Frieze is tighter than (\[eq:oyam\_lb\]) when $K=2$, while (\[eq:oyam\_lb\]) is tighter than (\[eq:BoundByFF\]) for all $K \geq 3$. Upon examining (\[eq:c(K)\]), (\[eq:a(K)\]) and (\[eq:b(K)\]), we can see that our bound given in Theorem \[thm:UpperBoundForConnectivity\] is tighter than both (\[eq:oyam\_lb\]) and (\[eq:BoundByFF\]) for all $K \geq 2$. We illustrate the performance of these bounds in the succeeding discussion.
Discussion
----------
Through simulations, we study how our upper and lower bounds compare with the empirically observed probability of connectivity. We consider a network secured by the pairwise scheme with parameter $K=2$ and compute the empirical probability of connectivity as we vary the number of nodes $n$. For each parameter pair $(n,K)$, we generate $10^6$ independent realizations of ${\mathbb{H}(n;K)}$. To obtain the empirical probability of connectivity, we divide the number of instances for which the generated graph is connected by the total number ($10^6$) of instances generated; see Figure \[fig:2\]. Next, we compare the lower bound for $P(n;K)$ presented in Theorem \[thm:LowerBoundForConnectivity\] with the corresponding bounds in [@Yagan2013Pairwise; @FennerFrieze1982]. Recall from (\[eq:homogeneous\_zero\_one\_law\]) that $K=2$ is the critical threshold for connectivity of ${\mathbb{H}(n;K)}$ in the limit of large network size; thus, we focus on the case $K=2$ throughout the simulations. Substituting $K=2$ in (\[eq:q(K)\]), (\[eq:c(K)\]), (\[eq:a(K)\]) and (\[eq:b(K)\]), we obtain the following lower bounds on $P(n;2)$, $$\begin{aligned}
& \textrm{YM \cite{Yagan2013Pairwise}}: P(n;2) \geq 1- \frac{155}{n^3} \label{eq:YM2}\\
& \textrm{FF \cite{FennerFrieze1982}}: P(n;2) \geq 1- \frac{155}{n^3}\cdot
\frac{12 n/(12 n -1)}{\sqrt{ 6 \pi }}
\sqrt{\frac{n}{n-3}}
\label{eq:ff2}
\\
& \textrm{This work} : P(n;2) \geq 1- \frac{155}{n^3}\cdot \frac{e^{-(3-\frac{9}{n})}}{\sqrt{6 \pi }}
\sqrt{\frac{n}{n-3}}
\label{eq:ourlb2}\end{aligned}$$
![*A zoomed in view of our results and empirical probability of connectivity (computed by averaging $10^6$ independent experiments for each data point) for $K=2$ as a function of $n$ for $n\geq 16$. The lower bound corresponds to Theorem \[thm:LowerBoundForConnectivity\] and the upper bound corresponds to Theorem \[thm:UpperBoundForConnectivity\].* []{data-label="fig:2"}](lb_ub_emp.pdf)
![*Lower bounds and empirical probability of connectivity (computed by averaging $10^6$ independent experiments for each data point) for $K=2$ as a function of $n$ for $n\geq 16$. Our lower bound given in Theorem \[thm:LowerBoundForConnectivity\] significantly improves the existing lower bounds by Yağan and Makowski [@Yagan2013Pairwise], and Fenner and Frieze [@FennerFrieze1982].* []{data-label="fig:3"}](plot_compare.pdf)
With $K=2$, we plot the lower bounds (\[eq:LowerBoundForConnectivity\]), (\[eq:oyam\_lb\]) and (\[eq:BoundByFF\]) for comparison in Figure \[fig:3\]. In Table \[tab:tab1\], we compare the mean number of realizations of ${\mathbb{H}(n;K)}$ generated until one disconnected realization is observed corresponding to the lower bounds (\[eq:YM2\]), (\[eq:ff2\]) and (\[eq:ourlb2\]) for $K=2$.
Our results show that ${\mathbb{H}(n;K)}$ gets connected with probabilistic guarantees as high as $99.92\%$ even when $K=2$ and network consisting of as few as $16$ nodes. These results complete and complement the existing asymptotic zero-one laws for random K-out graphs.
----- ------------------------------------------- ------------------------- ------------------------
$n$ Theorem \[thm:LowerBoundForConnectivity\] YM [@Yagan2013Pairwise] FF [@FennerFrieze1982]
16 1 in [**1183**]{} 1 in [**26**]{} 1 in [**102**]{}
20 1 in [**2645**]{} 1 in [**51**]{} 1 in [**205**]{}
25 1 in [**5753**]{} 1 in [**100**]{} 1 in [**409**]{}
35 1 in [**17834** ]{} 1 in [**276**]{} 1 in [**1145** ]{}
----- ------------------------------------------- ------------------------- ------------------------
: *Comparison of the lower bound (\[eq:LowerBoundForConnectivity\]) with existing lower bounds (\[eq:oyam\_lb\]) and (\[eq:BoundByFF\]) from [@Yagan2013Pairwise] and [@FennerFrieze1982], respectively for $K=2$. The entries in the table corresponds to the mean number of realizations of ${\mathbb{H}(n;K)}$ generated until one disconnected realization is observed.*[]{data-label="tab:tab1"}
Upper bound on probability of connectivity {#sec:upperbound}
==========================================
For easier exposition, we give a proof of Theorem \[thm:UpperBoundForConnectivity\] here for $K=2$. Due to space constraints, the general version of our proof for $K\geq 2$ is given in the Appendix. For $K=2$, each node selects at least two other nodes and there can be no isolated nodes or node pairs in ${\mathbb{H}(n;K)}$. Thus, for $K=2$, the smallest possible isolated component is a [*triangle*]{}, i.e., a complete sub-networks over [*three*]{} nodes such that each node selects the other two nodes. To derive the upper bound on connectivity, we first derive a lower bound on the probability of existence of isolated triangles in ${\mathbb{H}(n;K)}$. In the proof for the general case $(K \geq 2)$ presented in the Appendix, we investigate the existence of isolated components of size $K+1$.
Let $\Delta_{ijk}$ denote the event that nodes $v_i,v_j$ and $v_k$ form an isolated triangle in ${\mathbb{H}(n;K)}$. The number of isolated triangles in ${\mathbb{H}(n;K)}$, denoted by $Z_n$ is given by $$\begin{aligned}
Z_n = \sum_{1\leq i<j<k \leq n} {\mathds{1}}\{ \Delta_{ijk}\}\end{aligned}$$Note that the existence of one or more isolated triangles ($Z_n \geq 1$), implies that ${\mathbb{H}(n;K)}$ is *not* connected. Thus, we can upper bound the probability of connectivity of ${\mathbb{H}(n;K)}$ as $$\begin{aligned}
&{\mathbb{P}}[~{\mathbb{H}(n;K)}\text{~is connected}~]\nonumber \\
&=1- {\mathbb{P}}[~{\mathbb{H}(n;K)}\text{~is \emph{not} connected}~]\nonumber \\
&=1-{\mathbb{P}}[~\exists \text{~at least one isolated \emph{sub-network} in~} {\mathbb{H}(n;K)}~]\nonumber \\
&\leq 1-{\mathbb{P}}[~ \exists \text{~at least one {isolated} \emph{triangle} in~} {\mathbb{H}(n;K)}~] \nonumber \\
&= 1-{\mathbb{P}}[~ Z_n \geq 1 ~]. \label{eq:p_con_upper_bound}\end{aligned}$$ where, $$\begin{aligned}
\vspace{-1mm}
[Z_n \geq 1] = \cup_{1\leq i<j<k \leq n} {\mathds{1}}\{\Delta_{ijk}\}.\end{aligned}$$
In the succeeding discussion, we assume $K=2$ and use the Bonferroni inequality [@bonferroni] to lower bound the union of the events ${\mathds{1}}\{\Delta_{ijk}\}$, where $1\leq i<j<k \leq n$. $$\begin{aligned}
&{\mathbb{P}}[~ Z_n \geq 1 ~] \nonumber\\
& \geq \sum_{ i<j<k } {\mathbb{P}}[~\Delta_{ijk}~]-\sum_{ i<j<k }\sum_{ x<y<z } {\mathbb{P}}[~\Delta_{ijk}\cap\Delta_{xyz}~]
\label{eq:2out-union-lb}\end{aligned}$$For all $1\leq i<j<k\leq n$ and $1 \leq x<y<z \leq n$, we have $$\begin{aligned}
{\mathbb{P}}[~\Delta_{ijk}~]= \left(\frac{{1}}{{n-1 \choose 2}}\right)^3\left(\frac{{n-4 \choose 2}}{{n-1 \choose 2}}\right)^{n-3}
\label{eq:2out-singlesum}\end{aligned}$$ Moreover, note that if the sets $\{i,j,k\}$ and $\{x,y,z\}$ have one or more nodes in common, then these sets cannot simultaneously constitute isolated triangles; i.e., the events $\Delta_{ijk}$, $\Delta_{xyz}$ are mutually exclusive if $\{i, j, k\} \cap \{x, y, z\} \neq \emptyset $. Thus, $$\begin{aligned}
{\mathbb{P}}[~\Delta_{ijk} \cap\Delta_{x y z}~]=
\begin{cases}
& 0 \text{~~~if~} \{i,j,k\}\cap\{x,y,z\} \neq \phi, \\
& \left(\frac{{1}}{{n-1 \choose 2}}\right)^6\left(\frac{{n-7 \choose 2}}{{n-1 \choose 2}}\right)^{n-6} \text{otherwise}.
\label{eq:2out-doublesum}
\end{cases}\end{aligned}$$We now calculate the term appearing in (\[eq:2out-union-lb\]) in turn. We have $$\begin{aligned}
\sum_{ i<j<k } {\mathbb{P}}[~\Delta_{ijk}~]
&= {n \choose 3} {\mathbb{P}}[~\Delta_{ijk}~]\nonumber\\
&= {n \choose 3}\left(\frac{{1}}{{n-1 \choose 2}}\right)^3\left(\frac{{n-4 \choose 2}}{{n-1 \choose 2}}\right)^{n-3}\nonumber\\
&= \frac{{4n}}{{3(n-1)^2 (n-2)^2}}
\prod_{\ell=1}^2\left( 1-\frac{3}{n-\ell}\right)^{n-3}
\nonumber\\
&\geq \frac{{4}}{{3n^3}}\left( 1-\frac{3}{n-2}\right)^{2n-6}
\label{eq:2out-bound-singlesum},\end{aligned}$$ and $$\begin{aligned}
&\sum_{ i<j<k }\sum_{ x<y<z } {\mathbb{P}}[~\Delta_{ijk}\cap\Delta_{xyz}~] \nonumber\\
&= {n \choose 3}{n-3 \choose 3}\left(\frac{{1}}{{n-1 \choose 2}}\right)^6\left(\frac{{n-7 \choose 2}}{{n-1 \choose 2}}\right)^{n-6}\nonumber\\
&= \frac{{16n(n-3)(n-4)(n-5)}}{{9 (n-1)^5(n-2)^5}}
\prod_{\ell=1}^2\left( 1-\frac{6}{n-\ell}\right)^{n-6}
\nonumber\\
&\leq \frac{{16n^4}}{{9 (n-2)^{10}}}\left( 1-\frac{6}{n-1}\right)^{2n-12}
\label{eq:2out-bound-doublesum}.\end{aligned}$$ Substituting (\[eq:2out-bound-singlesum\]) and (\[eq:2out-bound-doublesum\]) in (\[eq:2out-union-lb\]), we obtain $$\begin{aligned}
&{\mathbb{P}}[~ Z_n \geq 1 ~] \nonumber\\
& \geq \sum_{ i<j<k } {\mathbb{P}}[~\Delta_{ijk}~]-\sum_{ i<j<k }\sum_{ x<y<z } {\mathbb{P}}[~\Delta_{ijk}\cap\Delta_{xyz}~]\nonumber\\
& \geq \frac{{4}}{3{n^3}}\hspace{-.5mm}\left( \hspace{-.5mm}1\hspace{-.5mm}-\hspace{-.5mm}\frac{3}{n-2}\hspace{-.5mm}\right)^{2n\hspace{-.5mm}-6} \hspace{-1.5mm}-\frac{{16n^4}}{{9 (n-2)^{10}}}\left(\hspace{-.5mm} 1\hspace{-.5mm}-\hspace{-.5mm}\frac{6}{n-1}\hspace{-.5mm}\right)^{2n-12}\nonumber\\
& =\frac{{4 e^{-6}}}{3{n^3}}(1+\oo(1)) \label{eq:upperrbound}$$ Reporting this into (\[eq:p\_con\_upper\_bound\]) leads to establishing Theorem \[thm:UpperBoundForConnectivity\] for $K=2$. More compactly, this result can be stated as $P(n;2)=1-\Omega\left(\frac{1}{n^3}\right)$. We prove the more general result for $K \geq 2$ in the Appendix. The next Section is devoted establishing a [*matching*]{} lower bound on the probability of connectivity.
Lower bound on probability of connectivity
==========================================
Fix $n=2,3, \ldots $ and consider a fixed positive integer $K$. The conditions $$2 \leq K \quad \mbox{and} \quad e(K+2) < n
\label{eq:OneLawConditions}$$ are enforced throughout. Note that the condition $e(K+2) < n$ automatically implies $K<n$.
\[sec:lowerbound\]
Preliminaries
-------------
Before proceeding with the proof, we discuss one of the key steps which distinguishes our proof and improves upon existing[@Yagan2013Pairwise; @FennerFrieze1982] bounds. In contrast to the standard bound ${n \choose r} \leq \left(\frac{ne}{r} \right)^r$ used in [@Yagan2013Pairwise], we upper bound ${n \choose r}$ using a variant [@stirlingremark] of Stirling formula. For all $x=1,2,\dots$, we have $$\begin{aligned}
\sqrt{2 \pi} x^{x+0.5} e^{-x} e^{\frac{1}{12x+1}} < x! < \sqrt{2 \pi} x^{x+0.5} e^{-x} e^{\frac{1}{12x}}, \label{eq:stirling}\end{aligned}$$ which gives $$\begin{aligned}
{n \choose{ r}}& \leq \frac{1}{\sqrt{2 \pi}} \left(\frac{n}{n-r}\right)^{n-r} \left( \frac{n}{r}\right)^r \frac{\sqrt{n}}{\sqrt{n-r}\sqrt{r}} \nonumber\\
&\qquad \cdot \exp\left\{ \frac{1}{12n}-\frac{1}{12(n-r)+1}-\frac{1}{12r+1}\right\}\nonumber\\
& \leq \frac{1}{\sqrt{2 \pi}} \left(\frac{n}{n-r}\right)^{n-r} \left( \frac{n}{r}\right)^r \frac{\sqrt{n}}{\sqrt{n-r}\sqrt{r}},
\label{eq:stirlingtakeaway}\end{aligned}$$ since $$\frac{1}{12n}-\frac{1}{12(n-r)+1}-\frac{1}{12r+1} < 0$$ Using the upper bound for ${n \choose r}$ as presented in (\[eq:stirlingtakeaway\]) eventually leads to the factor $e^{- (K^2-1)(1- \frac{K+1}{n})}$ improvement in the lower bound on probability of connectivity in Theorem \[thm:LowerBoundForConnectivity\]. Next, we note that for $0\leq K \leq x \leq y$, $$\begin{aligned}
\frac{ {x \choose K} }{ {y \choose K} }
= \prod_{\ell=0}^{K-1}
\left (
\frac{x-\ell}{y-\ell} \right ) \leq \left ( \frac{x}{y}
\right )^K \label{eq:ratio} \end{aligned}$$ since $\frac{x-\ell}{y-\ell}$ decreases as $\ell$ increases from $\ell = 0$ to $\ell=K-1$. Lastly, for all $x \in {\mathbb{R}}$, we have $$\begin{aligned}
1 \pm x &\leq e^{\pm x}. \label{eq:kcon_1pmx}\end{aligned}$$
Proof of Theorem \[thm:LowerBoundForConnectivity\]
--------------------------------------------------
If $\mathbb{H} (n;K)$ is [*not*]{} connected, then there exists a non-empty subset $S$ of nodes that is isolated. Further, since each node is paired with at least $K$ neighbors, $|S|\geq K+1$. Let $C_n (K)$ denote the event that $\mathbb{H}(n;K)$ is connected. We have $$C_n(K)^c
\subseteq
\bigcup_{S \in \mathcal{P}_n: ~ |S| \geq K+1} ~ B_n (K ; S)
\label{eq:BasicIdea}$$ where $\mathcal{P}_n$ stands for the collection of all non-empty subsets of ${\cal N}$. Let $\mathcal{P}_{n,r} $ denotes the collection of all subsets of ${\cal N}$ with exactly $r$ elements. A standard union bound argument yields $$\begin{aligned}
{{\mathbb{P}}\left[{ C_n(K)^c }\right]}
&\leq &
\sum_{ S \in \mathcal{P}_n: K+1 \leq |S| \leq \lfloor \frac{n}{2} \rfloor }
{{\mathbb{P}}\left[{ B_n (K ; S) }\right]}
\nonumber \\
&=&
\sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }
\left ( \sum_{S \in \mathcal{P}_{n,r} } {{\mathbb{P}}\left[{ B_n (K; S) }\right]} \right ).
\label{eq:BasicIdea+UnionBound}\end{aligned}$$ For each $r=1, \ldots , n$, let $B_{n,r} (K) = B_n (K ; \{ 1, \ldots , r \} )$. Under the enforced assumptions, exchangeability implies $${{\mathbb{P}}\left[{ B_n (K ; S) }\right]} = {{\mathbb{P}}\left[{ B_{n,r} (K ) }\right]}, \quad S \in
\mathcal{P}_{n,r}$$ and since $|\mathcal{P}_{n,r} | = {n \choose r}$, we have $$\sum_{S \in \mathcal{P}_{n,r} } {{\mathbb{P}}\left[{ B_n (K ; S) }\right]}
= {n \choose r} ~ {{\mathbb{P}}\left[{ B_{n,r}(K ) }\right]}
\label{eq:ForEach=r}$$ Substituting (\[eq:ForEach=r\]) into (\[eq:BasicIdea+UnionBound\]) we obtain $$\begin{aligned}
{{\mathbb{P}}\left[{ C_n(K)^c }\right]}& \leq \sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }{n \choose r} ~ {{\mathbb{P}}\left[{ B_{n,r}(K ) }\right]} .
\nonumber\\
&\leq
\sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor } {n \choose r}\hspace{-1mm}
\left({ {r-1}\choose K }\over { {n-1} \choose K } \right)^{\hspace{-.5mm}r}\hspace{-1mm}
\left(\hspace{-1mm}{ {n-r-1}\choose K }\over { {n-1} \choose K } \hspace{-1mm}\right)^{n-r}.\label{eq:BasicIdea+UnionBound2}\end{aligned}$$
Using (\[eq:ratio\]) in (\[eq:BasicIdea+UnionBound2\]) together with (\[eq:stirlingtakeaway\]), we conclude that $$\begin{aligned}
& {{\mathbb{P}}\left[{ C_n(K)^c }\right]}
\nonumber \\
& \leq \sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor } {n \choose r} \left(\frac{r-1}{n-1}\right)^{rK}
\left(1-\frac{r}{n-1}\right)^{(n-r)K}\\
& \leq \sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor } {n \choose r} \left(\frac{r}{n}\right)^{rK}
\left(1-\frac{r}{n}\right)^{(n-r)K}\nonumber\\
& \leq \sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }
\frac{1}{\sqrt{2 \pi}} \left(\frac{n}{n-r}\right)^{n-r} \left( \frac{n}{r}\right)^r \frac{\sqrt{n}}{\sqrt{n-r}\sqrt{r}}\nonumber\\ &\qquad \cdot \left(\frac{r-1}{n-1}\right)^{rK}
\left(1-\frac{r}{n-1}\right)^{(n-r)K}
\nonumber \\
& =
\sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }\frac{\sqrt{n}}{\sqrt{2 \pi}\sqrt{n-r}\sqrt{r}}
\left(\frac{r}{n}\right)^{r(K-1)}
\left(1-\frac{r}{n}\right)^{(n-r)(K-1)}
\nonumber \\
&\leq
\hspace{-1mm}\sum_{\hspace{-.5mm}r=K+1}^{ \lfloor \frac{n}{2} \rfloor }\hspace{-.5mm}\frac{\sqrt{n}}{\sqrt{2 \pi}\sqrt{n-r}\sqrt{r}}
\left(\hspace{-.5mm}\frac{r}{n}\hspace{-.5mm}\right)^{r(K-1)}
\hspace{-1.5mm}e^{-\left(\frac{r}{n}\right){(n-r)(K-1)}}
\label{eq:eapply},\end{aligned}$$ where (\[eq:eapply\]) follows from (\[eq:kcon\_1pmx\]). For $K+1 \leq r \leq \lfloor \frac{n}{2} \rfloor,$ we have $$\begin{aligned}
r(n-r) \geq (K+1)(n-K-1) \end{aligned}$$ Substituting in (\[eq:eapply\]), $$\begin{aligned}
&{{{\mathbb{P}}\left[{ C_n(K)^c }\right]}}\nonumber \\
& \leq \sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }\frac{\sqrt{n}}{\sqrt{2 \pi}\sqrt{n-K-1}\sqrt{K+1}}
\left(\frac{r}{n}\right)^{r(K-1)}\nonumber\\ &\qquad \cdot
e^{-\left(\frac{K+1}{n}\right){(n-K-1)(K-1)}}
\nonumber \\
& =
\sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }
\left(\frac{r}{n}\right)^{r(K-1)}
\frac{e^{- (K^2-1)(1- \frac{K+1}{n})}}{\sqrt{2 \pi (K+1)}}\sqrt{\frac{n}{(n-K-1)}}.
\nonumber \\
& = c(n;K)
\sum_{r=K+1}^{ \lfloor \frac{n}{2} \rfloor }
\left(\frac{r}{n}\right)^{r(K-1)}
\nonumber \\
& = c(n;K)\hspace{-1mm}
\left(\hspace{-0.5mm}\frac{K+1}{n}\hspace{-0.5mm}\right)^{K^2-1}
\hspace{-1.5mm} +
c(n;K)\hspace{-1.5mm}
\sum_{r=K+2}^{ \lfloor \frac{n}{2} \rfloor }
\left(\frac{r}{n}\right)^{r(K-1)}
\label{eq:con_1_law_last_step} \end{aligned}$$ with $c(n;K)$ given by (\[eq:c(K)\]). Due to space constraints we in present the sequence of steps leading to the final bound in Theorem \[thm:LowerBoundForConnectivity\] in the Appendix.
Conclusions
===========
In this work we derive upper and lower bounds for connectivity for random K-out graphs when the number of nodes is finite. Our [matching]{} upper and lower bounds prove that the probability of connectivity is $1-\Theta({1}/{n^{K^2-1}})$ for all $K \geq 2$. Our lower bound is shown to significantly improve the existing ones. In particular, our results further strengthen the applicability of random K-out graphs as an efficient way to construct a connected network topology even when the number of nodes is [*small*]{}. It would be interesting to pursue further applications of K-out graphs in the context of cryptographic payment channel networks. \[sec:conc\]
Acknowledgements {#acknowledgements .unnumbered}
================
This work has been supported in part by the National Science Foundation through grant CCF \#1617934.
Appendix {#appendix .unnumbered}
========
Upper bound on probability of connectivity for $K\geq2$ {#sec:genupperbound}
-------------------------------------------------------
In Section \[sec:upperbound\] we proved Theorem \[thm:UpperBoundForConnectivity\] for the case $K=2$. In this section, we prove the upperbound for the general case $K \geq 2$. Let $K$ be a fixed positive integer such that $ K \geq 2$. Let $\Delta_{i_1 \dots i_{K+1}}$ denote the event that nodes $v_i,v_j,\dots,v_{K+1}$ form an isolated component in ${\mathbb{H}(n;K)}$. The number of such isolated components of size $K+1$ in ${\mathbb{H}(n;K)}$, denoted by $Z_n$ is given by $$\begin{aligned}
Z_n = \sum_{1\leq i_1<i_2 \dots < i_{K+1} \leq n} {\mathds{1}}\{ \Delta_{i_1 \dots i_{K+1}}\}\end{aligned}$$ Note that the existence of one or more isolated components of size $K+1$ ($Z_n \geq 1$), implies that ${\mathbb{H}(n;K)}$ is *not* connected. We can upper bound the probability of connectivity of ${\mathbb{H}(n;K)}$ as $$\begin{aligned}
&{\mathbb{P}}[~{\mathbb{H}(n;K)}\text{~is connected}~]\nonumber \\
&=1- {\mathbb{P}}[~{\mathbb{H}(n;K)}\text{~is \emph{not} connected}~]\nonumber \\
&=1-{\mathbb{P}}[~\exists \text{~at least one isolated \emph{sub-network}} ~]\nonumber \\
&\leq 1-{\mathbb{P}}[~ \exists \text{~at least one {isolated} component of size $K+1$} ~] \nonumber \\
&= 1-{\mathbb{P}}[~ Z_n \geq 1 ~]. \label{eq:gen_p_con_upper_bound}\end{aligned}$$ where, $$\begin{aligned}
\{Z_n \geq 1\} = \bigcup_{1\leq i_1<i_2 \dots < i_{K+1} \leq n} {\mathds{1}}\{\Delta_{i_1 \dots i_{K+1}}\}.\end{aligned}$$
In the succeeding discussion, we use the Bonferroni inequality [@bonferroni] to lower bound the union of the events given in (\[eq:gen\_p\_con\_upper\_bound\]). $$\begin{aligned}
&{\mathbb{P}}[~ Z_n \geq 1 ~] \nonumber\\
& \geq \sum_{ i_1<i_2 \dots < i_{K+1} } {\mathbb{P}}[~\Delta_{i_1 \dots i_{K+1}}~]\\
&\qquad -\sum_{i_1<i_2 \dots < i_{K+1} }\sum_{ j_1<j_2 \dots < j_{K+1}} {\mathbb{P}}[~\Delta_{i_1 \dots i_{K+1}}\cap\Delta_{j_1 \dots j_{K+1}}~]
\label{eq:gen_2out-union-lb}\end{aligned}$$ For all $1\leq i_1<i_2 \dots < i_{K+1}\leq n$ and $1 \leq j_1<j_2 \dots < j_{K+1}\leq n$, we have $$\begin{aligned}
{\mathbb{P}}[~\Delta_{i_1 \dots i_{K+1}}~]= \left(\frac{{1}}{{n-1 \choose K}}\right)^{K+1}\left(\frac{{n-K-2 \choose K}}{{n-1 \choose K}}\right)^{n-K-1}
\label{eq:gen_2out-singlesum}\end{aligned}$$ Moreover, note that if the sets $\{i_1,\dots,i_{K+1}\}$ and $\{j_1,\dots,j_{K+1}\}$ have one or more nodes in common, then these sets cannot simultaneously constitute isolated components. Thus, ${\mathbb{P}}[~\Delta_{i_1 \dots i_{K+1}} \cap\Delta_{j_1 \dots j_{K+1}}~]=$ $$\begin{aligned}
\begin{cases}
& 0 \text{~~~if~~~} \{i_1,\dots,i_{K+1}\}\cap\{j_1,\dots,j_{K+1}\} \neq \phi, \\
& \left(\frac{{1}}{{n-1 \choose K}}\right)^{2(K+1)}\left(\frac{{n-2K-3 \choose K}}{{n-1 \choose K}}\right)^{n-2(K+1)} \text{otherwise}. \label{eq:gen_2out-doublesum}
\end{cases}\end{aligned}$$ We now calculate the term appearing in (\[eq:gen\_2out-union-lb\]) in turn. We have $$\begin{aligned}
&\sum_{ i_1<i_2 \dots < i_{K+1}} {\mathbb{P}}[~\Delta_{i_1\dots i_{K+1}}~]\nonumber\\
& = {n \choose {K+1}} {\mathbb{P}}[~\Delta_{i_1 \dots i_{K+1}}~]\nonumber\\
& = {n \choose {K+1}}\left(\frac{{1}}{{n-1 \choose K}}\right)^{K+1}\left(\frac{{n-K-2 \choose K}}{{n-1 \choose K}}\right)^{n-K-1}\nonumber\\
& = \frac{({K!})^K n}{{K+1}}\cdot \left(\frac{(n-K-1)!}{(n-1)!}\right)^{K} \prod_{\ell=1}^{K}\cdot\left(1-\frac{K+1}{n-\ell}\right)^{n-K-1} \nonumber\\
& \geq \frac{({K!})^K}{{K+1}}\cdot \frac{1}{n^{(K^2-1)}} \prod_{\ell=1}^{K}\cdot\left(1-\frac{K+1}{n-\ell}\right)^{n-K-1} \nonumber\\
& \geq \frac{({K!})^K}{{K+1}}\cdot \frac{1}{n^{(K^2-1)}} \cdot\left(1-\frac{K+1}{n-K}\right)^{K(n-K-1)} \label{eq:gen_2out-bound-singlesum}\\
& = \frac{({K!})^K e^{-K(K+1)}}{{K+1}}\cdot \frac{1}{n^{(K^2-1)}} (1+\oo(1)) \nonumber\end{aligned}$$ where (\[eq:gen\_2out-bound-singlesum\]) is plain from the observation that for all $\ell$ in $1, \dots, K$, $$1-\frac{K+1}{n-\ell} \geq 1-\frac{K+1}{n-K}.$$ Next, $$\begin{aligned}
&\sum_{ i_1< \dots < i_{K+1}}\sum_{ j_1< \dots < j_{K+1} } {\mathbb{P}}[~\Delta_{ i_1, \dots, i_{K+1}}\cap\Delta_{j_1, \dots , j_{K+1}}~] \nonumber\\
&= {n \choose {K+1}}{n-K-1 \choose {K+1}}\left(\frac{{1}}{{n-1 \choose K}}\right)^{2(K+1)}\nonumber\\
&\qquad \cdot \left(\frac{{n-2K-3 \choose K}}{{n-1 \choose K}}\right)^{n-2(K+1)}\nonumber\\
&\leq
{n \choose {K+1}}{n-K-1 \choose {K+1}}\left(\frac{{1}}{{n-1 \choose K}}\right)^{2(K+1)}\nonumber\\
&\qquad \cdot \left(\frac{n-2K-3}{n-1}\right)^{K(n-2(K+1))}\label{eq:gen_2out-bound-doublesum-2}\\
&=
\frac{n!}{(n-2(k+1))!((K+1)!)^2}\left(\frac{{K!}}{{n-1 \choose K}}\right)^{2(K+1)}\nonumber\\
&\qquad \cdot \left(\frac{n-2K-3}{n-1}\right)^{K(n-2(K+1))}\nonumber\\
&= \frac{(K!)^{2(K+1)}}{((K+1)!)^2}
\frac{n(n-1)\dots (n-2K-3)}{(n (n-1) \dots (n-K))^{2(K+1)}}\nonumber\\
&\qquad \cdot \left(\frac{n-2K-3}{n-1}\right)^{K(n-2(K+1))}\nonumber\\
&\leq \frac{(K!)^{2(K+1)}}{((K+1)!)^2} \cdot \frac{n^{2(K+1)}}{(n-K)^{2K(K+1)}}\nonumber\\
&\qquad \cdot \left(1-\frac{2(K+1)}{n-1}\right)^{K(n-2(K+1))}
\label{eq:gen_2out-bound-doublesum},\end{aligned}$$ where (\[eq:gen\_2out-bound-doublesum-2\]) follows from (\[eq:ratio\]).Substituting (\[eq:gen\_2out-bound-singlesum\]) and (\[eq:gen\_2out-bound-doublesum\]) in (\[eq:gen\_2out-union-lb\]), we obtain $$\begin{aligned}
&{\mathbb{P}}[~ Z_n \geq 1 ~] \nonumber\\
& \geq \sum_{ i_1<i_2\dots<i_{K+1}} {\mathbb{P}}[~\Delta_{i_1\dotsi_{K+1}}~]\nonumber\\
&-\sum_{ i_1<i_2\dots<i_{K+1}}\sum_{ j_1<j_2\dots<j_{K+1}} {\mathbb{P}}[~\Delta_{i_1\dotsi_{K+1}}\cap\Delta_{j_1,\dots,j_{K+1}}~]\nonumber\\
& \geq \frac{({K!})^K}{{K+1}}\cdot \frac{1}{n^{(K^2-1)}} \left(1-\frac{K+1}{n-K}\right)^{K(n-K-1)}\nonumber\\
&\qquad -\frac{(K!)^{2(K+1)}}{((K+1)!)^2} \cdot \frac{n^{2(K+1)}}{(n-K)^{2K(K+1)}}\nonumber\\
&\qquad \cdot \left(1-\frac{2(K+1)}{n-1}\right)^{K(n-2(K+1))}\nonumber\\
& = \frac{({K!})^K e^{-K(K+1)}}{{K+1}}\cdot \frac{1}{n^{(K^2-1)}} (1+\oo(1)) \label{eq:gen_upperrbound}\\
&= \Omega\left(\frac{1}{n^{K^2-1}} \right).\nonumber\end{aligned}$$ [\
]{}In view of (\[eq:gen\_p\_con\_upper\_bound\]), we then obtain for $K\geq2$ that $${\mathbb{P}}[~{\mathbb{H}(n;K)}\text{~is connected}~] = 1-\Omega\left(\frac{1}{n^{K^2-1}} \right).$$
Bounding the sum in (\[eq:con\_1\_law\_last\_step\])
----------------------------------------------------
Recall that $K$ is a fixed positive integer $\geq 2$ and under the constraint (\[eq:OneLawConditions\]) we have $K+2 \leq \lfloor \frac{n}{2} \rfloor$, and therefore the sum in (\[eq:con\_1\_law\_last\_step\]) is not empty. Let $$\left(\frac{x}{n}\right)^{x(K-1)} = e^{(K-1) f_n (x) },
\quad x \geq 1
\label{eq:Exponentiation}$$ with $$f_n (x) = x \log \left ( \frac{x}{n} \right )
= x \left ( \log x - \log n \right ).$$ Observe that $r \rightarrow f_n(r)$ decreases monotonically on the interval $r=1, \ldots, \lfloor \frac{n}{e} \rfloor$ and increases monotonically thereafter on the interval $r= \lfloor \frac{n}{e} \rfloor + 1, \ldots , \lfloor \frac{n}{2} \rfloor$. Therefore, $$\begin{aligned}
\lefteqn{
\max \left (
f_n(r), \ r=K+2, \ldots , \left \lfloor \frac{n}{2} \right \rfloor
\right )
} & &
\nonumber \\
&=&
\max \left (
f_n(K+2),
f_n \left ( \left \lfloor \frac{n}{2} \right \rfloor \right )
\right ) .
\label{eq:IntermediaryBound}\end{aligned}$$ From (\[eq:OneLawConditions\]), we have $K+2 \leq \lfloor \frac{n}{e} \rfloor $. Next, we show that $$f_n \left ( \left \lfloor \frac{n}{2} \right \rfloor \right )
\leq
f_n(K+2)
\label{eq:WhichIsLarger}$$ for all $n$ large enough, say $n \geq n(K)$ for some finite integer $n(K)$ which depends on $K$. (\[eq:WhichIsLarger\]) is equivalent to $$\left \lfloor \frac{n}{2} \right \rfloor
\log
\left( \frac{ \left \lfloor \frac{n}{2} \right \rfloor }{n} \right )
\leq
(K+2) \left ( \log (K+2) - \log n \right ),$$ a condition can be expressed as $$n \left (
\frac{ \left \lfloor \frac{n}{2} \right \rfloor }{n}
\right )
\log
\left( \frac{ \left \lfloor \frac{n}{2} \right \rfloor }{n} \right )
\leq
(K+2) \left ( \log (K+2) - \log n \right ).$$ The mapping $t \rightarrow t \log t $ is monotone increasing on the interval $( e^{-1}, \infty)$. Since $\left \lfloor \frac{n}{2} \right \rfloor \leq \frac{n}{2}$, for the inequality (\[eq:WhichIsLarger\]) to hold, if suffices to show $$- \left ( \frac{n}{2} \right ) \log 2
\leq
(K+2) \left ( \log (K+2) - \log n \right )
\label{eq:WhichIsLarger2}$$ for all $n$ satisfying the constraint $$\frac{1}{e} < \frac{1}{n} \left \lfloor \frac{n}{2} \right \rfloor .$$ It is easy to see that this occurs for all $n > 4$, which is in fact automatically guaranteed under (\[eq:OneLawConditions\]). Condition (\[eq:WhichIsLarger2\]) can be simplified to yield $$\log n
\leq
\left ( \frac{\log 2}{2(K+2)} \right ) \cdot n
+ \log (K+2).
\label{eq:WhichIsLarger3}$$ It can be verified that (\[eq:WhichIsLarger3\]) holds as an equality for $n=4(K+2)$ and a strict inequality for all $n > 4(K+2)$. Setting $n(K) = 4(K+2) $ is therefore sufficient for (\[eq:WhichIsLarger2\]) (hence (\[eq:WhichIsLarger\])) to hold. Using (\[eq:Exponentiation\]), (\[eq:IntermediaryBound\]) and (\[eq:WhichIsLarger\]) we get $$\begin{aligned}
\lefteqn{\max \left ( \left(\frac{r}{n}\right)^{r(K-1)} : \ r = K+2 ,
\ldots, \left \lfloor \frac{n}{2} \right \rfloor \right )} &&
\nonumber \\
&=&
\left(\frac{K+2}{n}\right)^{(K+2)(K-1)} \hspace{ 1cm}
\nonumber\end{aligned}$$ for all $n \geq n(K)$ yielding $$\sum_{r=K+2}^{ \lfloor \frac{n}{2} \rfloor }
\left(\frac{r}{n}\right)^{r(K-1)}
\leq
\left \lfloor \frac{n}{2} \right \rfloor
\cdot
\left(\frac{K+2}{n}\right)^{(K+2)(K-1)} .$$ Substituting in (\[eq:con\_1\_law\_last\_step\]) and noting that $P(n;K)=1-{{\mathbb{P}}\left[{ C_n(K)^c }\right]}$, we obtain (\[eq:LowerBoundForConnectivity\]. [\
]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Denoting by ${\mathcal L}_d(m_0,m_1,\ldots,m_r)$ the linear system of plane curves passing through $r+1$ generic points $p_0,p_1,\ldots,p_r$ of the projective plane with multiplicity $m_i$ (or larger) at each $p_i$, we prove the Harbourne-Hirschowitz Conjecture for linear systems ${\mathcal L}_d(m_0,m_1,\ldots,m_r)$ determined by a wide family of systems of multiplicities $\bold{m}=(m_i)_{i=0}^r$ and arbitrary degree $d$. Moreover, we provide an algorithm for computing a bound of the regularity of an arbitrary system $\bold{m}$ and we give its exact value when $\bold{m}$ is in the above family. To do that, we prove an $H^1$-vanishing theorem for line bundles on surfaces associated with some pencils “at infinity”.'
author:
- 'F. Monserrat'
title: Curves having one place at infinity and linear systems on rational surfaces
---
Introduction
============
This paper deals with the problem of computing the dimension of linear systems on smooth projective surfaces. The main result provides, for any arbitrary number of generic points in the projective plane over the field of complex numbers, ${\mathbb{P}}^2$, a wide family of systems of multiplicities for which the Harbourne-Hirschowitz Conjecture holds. Moreover, we show an algorithm for computing upper bounds of the regularity of a system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_r)$.
Our proofs of these results are based on Section \[infinity\], where we give an $H^1$-vanishing theorem for line bundles on those surfaces $X$ obtained from ${\mathbb{P}}^2$ eliminating (by means of successive blowing-ups) the indeterminacies of the rational map $f:{\mathbb{P}}^2 \cdots \rightarrow {\mathbb{P}}^1$ given by certain pencils of plane curves. These are the pencils “at infinity” associated with rational projective curves of ${\mathbb{P}}^2$ that have one place at infinity and are smooth in their affine parts. The set formed by the centers of the blowing-ups used to obtain such a surface $X$ turns out to be a [*P-sufficient configuration*]{} (this type of configurations has been introduced and studied in [@g-m-bis], [@g-m] and [@g-m-2]). This fact, together with the simplicity and good properties of the effective semigroup of $X$, leads up to the above mentioned vanishing theorem. Then, semicontinuity arguments will allow to deduce our main result.
Fixing $r+1$ points $p_0,p_1,\ldots,p_r$ of ${\mathbb{P}}^2$ in generic position and given $r+1$ non-negative integers $m_0,m_1,\ldots,m_r$, the linear system ${\mathcal L}_d(m_0,m_1,\ldots,m_r)$ of plane projective curves of fixed degree $d$ having multiplicity $m_i$ (or larger) at $p_i$ for each $i$, has an expected dimension (attained when all the conditions being imposed are independent). Those systems whose dimension is larger than the expected one are called [*special*]{}. The Harbourne-Hirschowitz Conjecture intends to give a description of all special linear systems. Basically, it asserts that a linear system is special if and only if it has a multiple fixed component such that its strict transform on the surface obtained by blowing-up the points $p_0,p_1\ldots,p_r$ is a $(-1)$-curve (that is, an integral curve with self-intersection equal to $-1$ and genus zero). This conjecture goes back to B. Segre [@seg], being reformulated by several authors (see [@harb1], [@gim1], [@hir], [@harb4], [@cil], [@cil2], and [@cil3] for a survey).
Different approaches have been applied to obtain partial results on the Harbourne-Hirschowitz Conjecture. It has been proved for $r+1\leq 9$ points (Castelnuovo was the first to deal with these cases [@castel], although modern proofs are due to Nagata [@nag], Gimigliano [@gim1] and Harbourne [@harb4]). Arbarello and Cornalba [@a-c] treated the homogeneous case with multiplicity $2$ (that is, $m_0=m_1=\ldots=m_r=2)$ using infinitesimal deformation theory, and Hirschowitz [@hir1] proved the conjecture for the homogeneous case with multiplicity not greater than 3, by using a specialization technique (the so-called [*Horace method*]{}). This result has been generalized by Ciliberto and Miranda ([@cil] and [@cil2]) applying a different degeneration technique, showing that the Harbourne-Hirschowitz Conjecture is true for the quasihomogeneous case $m_1=m_2=\ldots=m_r\leq 3$ and $m_0$ arbitrary, and for the homogeneous case with multiplicity $m$ up to 12 (the cases $13\leq
m\leq 20$ are treated in [@cil4] with the same technique and the help of a computer program). Using a similar approach, Seibert [@sei] proved the conjecture for the quasihomogeneous case with $m_1=m_2=\ldots=m_r=4$ and, recently, Laface [@laf] has done it for $m_1=m_2=\ldots=m_r=5$. Other advances have been done by Mignon [@mig2] after proving the conjecture when $m_i\leq 4$ for all $i$, and Évain [@ev], who proves it for the homogeneous case when the number of points $r+1$ is a power of 4. Also, using a refinement of the Horace method (the so-called [*differential Horace method*]{}), Alexander and Hirschowitz [@a-h-1] obtained a bound $d_0=d_0(m)$ (only depending on $m$) such that, for any $d\geq
d_0$ and any system of multiplicities $(m_0,m_1,\ldots,m_r)$ with $m_i\leq m$ for all $i$, the linear system ${\mathcal
L}_d(m_0,m_1,\ldots,m_r)$ is non-special. A result which shows that the conjecture holds whenever there exist sufficiently many small multiplicities $m_i\leq 4$, at least one of them being $1$, is the one recently proved by Bunke and Lossen in [@los] by applying the differential Horace method. More recent advances on the subject are the papers of S. Yang [@yang], who proves the conjecture when $m_i\leq 7$ for all $i$, and M. Dumnicki and W. Jarnicki [@dum1], who do so when $m_i\leq 11$ for all $i$; also, in [@dum2] the conjecture is proved for the homogeneous cases with multiplicity bounded by 42.
Our contribution to the study of linear systems ${\mathcal
L}_d(m_0,m_1,\ldots,m_r)$ is made in Section \[s4\]. Using the iterated blowing-ups (introduced by Kleiman in [@kle1] and [@kle2], and also studied in [@harb2], [@roe] and [@fdb2]) and results developed in Section \[infinity\], we deduce a sufficient condition for the non-speciality of a linear system of that type (Theorem \[super\]). As a consequence, we determine, for any arbitrary number of points $r+1\geq 2$, a wide family of systems of multiplicities $(m_0,m_1,\ldots,m_r)$ for which the special linear systems of the form ${\mathcal
L}_d(m_0,m_1,\ldots,m_r)$ are completely characterized, proving that the Harbourne-Hirschowitz Conjecture is true for them. This result has the particularity of providing, for each arbitrary integer $r\geq 1$, a large set of systems of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_{r})$ satisfying the Harbourne-Hirschowitz Conjecture for which the possible $m_i$ are unbounded. Moreover, we also determine, for whichever $\bold{m}$ in the above set, the least degree $d$ such that these multiplicities impose independent conditions to curves of degree $d$ (that is, the [*regularity*]{} of $\bold{m}$).
There are many results giving upper bounds of the regularity of a system of multiplicities (see [@gim], [@cat], [@hir], [@bal], [@b-c], [@xu], [@harbc], [@hhf], [@hr], [@dum1] or [@harb3] for a survey). In Section \[s43\] we introduce a generalization of the algorithm given in [@roe3], based on our results in Section \[infinity\], providing bounds of the regularity which, in many cases, are better than the existing ones (as far as the author knows).
Every variety $X$ in this article will be considered over the field of complex numbers ${\mathbb{C}}$. Moreover, $K_X$ will denote a canonical divisor on $X$.
I thank Javier Fernández de Bobadilla for pointing out to me his geometric-combinatorial proof of Jung’s Theorem on factorization of automorphisms of the plane.
Preliminaries {#pre}
=============
Configurations {#s1}
--------------
In this section we summarize some concepts and notations that will be used throughout the paper. We start with the definition of configuration.
An [*ordered configuration*]{} over ${\mathbb{P}}^2$ (a [*configuration*]{} in the sequel) will be a finite sequence ${{\mathcal K}}=(p_0,p_1,\ldots,p_n)$ of closed points such that $p_0$ belongs to $X_0:={\mathbb{P}}^2$ and, inductively, if $i\geq 1$ then $p_i$ belongs to the blowing-up $X_i$ of $X_{i-1}$ at $p_{i-1}$. Among the points of ${\mathcal K}$ there is a natural partial ordering: $p_i\leq
p_j$ whenever $p_i=p_j$ or the composition of blowing-ups $X_j\rightarrow X_i$ maps $p_j$ to $p_i$. We will say that ${\mathcal K}$ is a [*chain configuration*]{} when $\leq$ be a total ordering.
Denote by $\pi_{{\mathcal K}}:Z_{{\mathcal K}} \rightarrow {\mathbb{P}}^2$ the morphism given by the composition of all the successive blowing-ups centered at the points of ${\mathcal K}$. Each blowing-up at $p_i$ gives rise to an exceptional divisor $E_i$ whose total (resp., strict) transform on $Z_{{\mathcal K}}$ will be denoted by $E_i^{\mathcal K}$ (resp., $\tilde{E}^{\mathcal K}_i$). In the same way, for each effective divisor $C$ on $X$, $C^{\mathcal K}$ (resp., $\tilde{C}^{\mathcal K}$) will be the total (resp., strict) transform of $C$ on $Z_{{\mathcal K}}$. Also, for each divisor $D$ on $Z_{{\mathcal K}}$, $[D]$ will denote its class in ${{\rm Pic}}(Z_{{\mathcal K}})$. The system $\{[L^{\mathcal K}], [E_0^{\mathcal K}],[E_1^{\mathcal K}],\ldots, [E_n^{\mathcal K}]\}$ is a ${\mathbb{Z}}$-basis of ${{\rm Pic}}(Z_{{\mathcal K}})$, $L$ denoting a general line on ${\mathbb{P}}^2$.
A point $p_i\in {\mathcal K}$ is said to be [*proximate*]{} to another point $p_j\in {\mathcal K}$ (in short, $i\rightarrow j$ or $p_i \rightarrow p_j$) if either $i=j+1$ and $p_i$ belongs to the exceptional divisor $E_j$, or $i>j+1$ and $p_i$ belongs to the strict transform on $X_i$ of the exceptional divisor $E_j$. The point $p_i$ is said to be a [*free*]{} point if it is proximate to, at most, one point of ${\mathcal K}$; otherwise, $p_i$ is said to be a [*satellite*]{} point. The proximity relation among the points of ${\mathcal K}$ is an equivalent datum to a matrix $\bold{P}_{{\mathcal K}}=(q_{ij})_{0\leq i,j\leq n}$, called [*proximity matrix*]{} of ${\mathcal K}$, and defined as follows: $q_{ij}=1$ if $i=j$, $q_{ij}=-1$ if $p_i$ is proximate to $p_j$, and $q_{ij}=0$ otherwise. For each $j=0,1,\ldots,n$, the entries of its $j$th column are the coefficients of the expression of the divisor $\tilde{E}_j^{{\mathcal K}}$ as linear combination of the divisors $E_0^{{\mathcal K}}, E_1^{{\mathcal K}}\ldots, E_n^{{\mathcal K}}$. The proximity relations can also be represented by means of a combinatorial object, the [*proximity graph*]{}. It will be denoted by ${\mathbb{G}}({\mathcal K})$ and it is a labelled graph whose vertices represent the points of ${\mathcal K}$ and whose edges join vertices associated with proximate points. Each vertex is labelled with the subindex $i$ of its associated point $p_i$. An edge joining $p_j$ and $p_i$ ($i>j$) is a continuous straight line whenever $p_i$ is a minimal point of ${\mathcal K}$ (with respect to the ordering $\leq $) which is proximate to $p_j$, and it is a dotted curved line otherwise (the label of an edge is determined by its property of being continuous-straight or curved-dotted). For the sake of simplicity, when we will depict a proximity graph, we will not draw those edges which can be deduced from others. Notice that the subgraph consisting of the vertices and the continuous edges has a forest structure whose trees are rooted on the vertices corresponding to those points in the configuration which lie in ${\mathbb{P}}^2$. A proximity graph will be called [*unibranched*]{} if it is associated to a chain configuration. The proximity graph is an equivalent datum either to the Enriques diagram or the dual graph of ${\mathcal K}$. By a [*system of multiplicities*]{} we mean a finite sequence $(m_0,m_1,\ldots,m_n)$ of non-negative integers. A [*weighted configuration*]{} (resp., [*weighted proximity graph*]{}) will be a pair $({\mathcal K},\bold{m})$ (resp., (${\mathbb{G}}({\mathcal K}),\bold{m})$), where ${\mathcal K}=(p_i)_{i=0}^n$ is a configuration and $\bold{m}=(m_0,m_1,\ldots,m_n)$ is a system of multiplicities. It can be seen as a map that assigns, to each point $p_j$ of ${\mathcal K}$ (resp., to the corresponding vertex of ${\mathbb{G}}({\mathcal K})$), the non-negative integer (multiplicity) $m_j$. The [*excesses*]{} of the weighted configuration $({\mathcal K},\bold{m})$ are defined to be the integers $\rho_j({\mathcal K}, \bold{m}):=m_j-\sum_{k\rightarrow j} m_k$, $0\leq j \leq n$. Since the excesses only depend on the proximity relations among the points of the configuration, we can define the [*excesses of a given weighted proximity graph*]{} $({\mathbb{G}}, \bold{m})$ as those associated with every weighted configuration $({\mathcal K},\bold{m})$ such that ${\mathbb{G}}({\mathcal K})={\mathbb{G}}$; they will be denoted by $\rho_j({\mathbb{G}},\bold{m})$. Similarly, the [*proximity matrix associated with a proximity graph*]{} ${\mathbb{G}}$, which will be denoted $\bold{P}_{{\mathbb{G}}}$, can be defined in an obvious way. If ${\mathcal K}$ is a configuration of $n+1$ points and $\bold{v}=(v_0,v_1,\ldots,v_t)$ is a system of multiplicities such that $t<n$, the pair $({\mathcal K},
\bold{v})$ (resp., (${\mathbb{G}}({\mathcal K}),\bold{v})$) will also be considered a weighted configuration (resp., weighted proximity graph), identifying $\bold{v}$ with the sequence of multiplicities of length $n+1$ obtained adding $n-t$ zero components to $\bold{v}$, that is, $( v_0 ,v_1 , \ldots ,v_t ,0,0, \ldots ,0)$.
P-sufficient configurations {#psuf}
---------------------------
Consider a configuration ${\mathcal K}=(p_0,p_1,\ldots,p_n)$ and take the notations of Section \[s1\]. Denote by $(b_{ij})_{0\leq i,j\leq
n}$ the entries of the matrix $\bold{P}_{{\mathcal K}}^{-1}$, whose columns contain the coefficients of the expressions of each divisor $E_j^{{\mathcal K}}$ as linear combinations of the divisors $\tilde{E}_0^{\mathcal K},\tilde{E}_1^{\mathcal K},\ldots,\tilde{E}_n^{\mathcal K}$. For each integer $i$ such that $0\leq i\leq n$ we consider the divisor on $Z_{{\mathcal K}}$ defined by $D_i:=\sum_{j=0}^i b_{ij} E_{j}^{{\mathcal K}}$, which has the following property: $D_i\cdot \tilde{E}_j^{{\mathcal K}}$ equals $-1$ if $i=j$ and $0$ otherwise. We define the square symmetric matrix $G_{{\mathcal K}}=(g_{ij})_{0\leq i,j\leq n}$ by $$g_{ij}=-9D_i\cdot D_j -(K_{Z_{{\mathcal K}}}\cdot D_i)(K_{Z_{{\mathcal K}}}\cdot
D_j).$$
Given an element $\bold{x}\in {\mathbb{R}}^{n+1}$, we set $\bold{x}> 0$ when all the coordinates of $\bold{x}$ are non-negative and at least one of them is positive. Recall [@minors] that an ($n+1$)-dimensional square symmetric matrix $A$ is called to be [*conditionally positive definite*]{} if $\bold{x} A \bold{x}^t>0$ for all vector $\bold{x}\in {\mathbb{R}}^{n+1}$ such that $\bold{x}>0$.
[A configuration ${\mathcal K}$ is called to be [*P-sufficient*]{} if the matrix $G_{{\mathcal K}}$ is conditionally positive definite. ]{}
This type of configurations has been recently introduced in [@g-m-bis] and [@g-m] and, in them, it is proved that the cone of curves of $Z_{{\mathcal K}}$ is (finite) polyhedral whenever ${\mathcal K}$ is a P-sufficient configuration. Recall that the [*cone of curves*]{} of a projective surface $X$, which we will denote by $NE(X)_{{\mathbb{R}}}$, is the convex cone of the real vector space ${{\rm Pic}}(X)\otimes_{{\mathbb{Z}}}{\mathbb{R}}$ spanned by the classes of the effective divisors on $X$.
When ${\mathcal K}$ is a chain configuration, checking whether it is P-sufficient or not is equivalent to checking a single condition [@g-m Cor. 2]: ${\mathcal K}$ is P-sufficient if and only if $-9D_n^2-(K_{Z_{{\mathcal K}}}\cdot D_n)^2>0$.
The following result, whose proof can be found in [@g-m-2], provides a property of the surfaces obtained from P-sufficient configurations which will be useful in Section \[infinity\].
\[2\] Let ${\mathcal K}=(p_0,p_1,\ldots,p_n)$ be a P-sufficient configuration and $D$ an effective divisor on $Z_{{\mathcal K}}$ such that $D^2\geq 0$ and $D\cdot \tilde{E}_i^{{\mathcal K}}\geq 0$ for all $i=0,1,\ldots,n$. Then, $K_{Z_{{\mathcal K}}}\cdot D<0$.
Plane curves having one place at infinity {#plane}
-----------------------------------------
With the exception of Proposition \[villa\] and Corollary \[pernia\], this section is expository and its aim is to summarize some facts related to plane curves having one place at infinity. This type of curves has been extensively studied by several authors (see, for instance, [@abh2], [@moh], [@abh], [@sat], [@pink], [@suzuki] or [@fujimoto]).
\[definition\] [A projective curve $C\hookrightarrow {\mathbb{P}}^2$ (which we will assume that is not a line) is said [*to have one place along a line $H \hookrightarrow {\mathbb{P}}^2$*]{} if the intersection $C\cap H$ is a single point $p$ and $C$ is reduced and has only one analytic branch at $p$. If $H$ is viewed as the line of infinity in the compactification of the affine plane to ${\mathbb{P}}^2$, we say that [*$C$ has one place at infinity.*]{} ]{}
Throughout this section, we fix a projective curve $C$ having one place at infinity, $p$ being the intersection point of $C$ with the line of infinity $H$. Consider the infinite sequence of morphisms $$\cdots \rightarrow X_{i+1} \rightarrow X_i\rightarrow \cdots
\rightarrow X_1\rightarrow X_0:={\mathbb{P}}^2,$$ where $X_1\rightarrow
X_0={\mathbb{P}}^2$ is the blowing-up of ${\mathbb{P}}^2$ at $p_0:=p$ and, for each $i\geq 1$, $X_{i+1}\rightarrow X_i$ denotes the blowing-up of $X_i$ at the unique point $p_i$ which lies on the strict transform of $C$ and on the exceptional divisor $E_{i-1}$ created by the preceding blowing-up.
### $\delta$-sequences {#deltasec}
The unique branch at infinity of $C$ corresponds to a normalized discrete valuation $v$ of the field of rational functions of $C$ over ${\mathbb{C}}$. We define the [*semigroup at infinity*]{} (resp., [*Weierstrass semigroup*]{}) associated with $C$, and we denote it by $\Gamma_C$ (resp., $S_C$), as the subsemigroup of ${\mathbb{N}}$ generated by all the integers of the form $-v(g)$, where $g$ belongs to the affine ${\mathbb{C}}$-algebra ${{\mathcal O}}_C(C\setminus \{p\})$ (resp., to the normalization of ${{\mathcal O}}_C(C\setminus \{p\})$). Obviously, $\Gamma_C$ is contained in $S_C$ and they are equal if and only if $C\setminus \{p\}$ is an smooth affine curve. Abhyankar and Moh proved in [@abh2] the existence of a positive integer $s$ and a sequence of positive generators $\delta_0,\delta_1,\ldots,\delta_s$ of $\Gamma_C$ such that:
- If $d_i=\gcd (\delta_0,\delta_1,\ldots,\delta_{i-1})$, for $1\leq i\leq s+1$ and $n_i=d_i/d_{i+1}$, $1\leq i\leq s$, then $d_{s+1}=1$ and $n_i>1$ for $1\leq i\leq s$.
- For $1\leq i\leq s$, $n_i\delta_i$ belongs to the semigroup generated by $\delta_0,\delta_1,\ldots, \delta_{i-1}$.
- $\delta_0>\delta_1$ and $\delta_i<\delta_{i-1}
n_{i-1}$ for $i=2,3,\ldots,s$.
The sequence $\{\delta_i\}_{i=0}^s$ can be obtained from an equation of the curve $C$ using approximate roots [@abh2 Chapter II, Sections 6,7]. We will refer to it as a [*$\delta$-sequence associated with $C$*]{}.
Moreover, it turns out that any sequence $(\delta_0,\delta_1,\ldots,\delta_s)$ satisfying the above conditions (I), (II) and (III) is a $\delta$-sequence associated with some curve having one place at infinity, which can be chosen of degree $\delta_0$ (see, for instance, [@sat] or [@pink]).
Associated with the branch at infinity of a curve having one place at infinity, there is a sequence of Newton polygons $P_0,P_1,\ldots,P_{g-1}$ which determines the equisingularity class of that branch [@cam 3.4]. Assume that each $P_i$ is the segment which joins the points $(0,e_i)$ and $(m_i,0)$, $e_i,m_i\in {\mathbb{N}}$. These Newton polygons can be explicitly recovered from a $\delta$-sequence $(\delta_0,\delta_1,\ldots,\delta_s)$ associated with the curve:
If $\delta_0-\delta_1$ does not divide $\delta_0$, then $s=g$ and $$e_0=\delta_0-\delta_1, \;\;\; e_i=d_{i+1}$$ $$m_0=\delta_0, \;\;\; m_i=n_i\delta_i-\delta_{i+1}$$ for $1\leq i\leq s-1$. Otherwise, $s=g+1$ and $$e_0=d_2=\delta_0-\delta_1, \;\;\; e_i=d_{i+2}$$ $$m_0=\delta_0+n_1\delta_1-\delta_2, \;\;\;
m_i=n_{i+1}\delta_{i+1}-\delta_{i+2}$$ for $1\leq i\leq s-2$.
The above equalities are considered and used in [@reg] and the proximity relations among the infinitely near points $p_0,p_1,\ldots$ can be easily deduced from the $\delta$-sequence, as we will describe next (see [@cam] for complete details):
Define $s_0=k_0=0$ and let $h_i$, $k_t$ and $s_t$ (with $0\leq
i\leq s_g-1$ and $1\leq t\leq g$) be the positive integers obtained from the following continued fractions:
$$\frac{m_{j-1}}{e_{j-1}}+k_{j-1}= h_{s_{j-1}} +
\frac{1}{{h_{s_{j-1}+1} + _{ \ddots + \frac{1}{{h_{s_j-1} +
\frac{1}{{k_j }}}}} }},$$ for $j=1,2,\ldots,g$. Also, for each $n\in
\{1,2,\ldots,s_g\}$, define $f(n):=k_t-1$ whenever $n=s_t$ for some $t\in \{1,2,\ldots,g\}$, and $f(n):=h_n$ otherwise. Then, the proximity relations are the following: $l\rightarrow l-1$ for each positive integer $l$, and $l\rightarrow \sum_{i=0}^{n-1} h_i-1$ for each pair $(n,l)$ such that $1\leq n\leq s_g$ and $\sum_{i=0}^{n-1}
h_i<l\leq \sum_{i=0}^{n-1} h_i+f(n)$.
Thus, a $\delta$-sequence associated with a plane curve $C$ having one place at infinity determines the equisingularity class of the branch of $C$ at $p$ and, therefore, the proximity graph of whichever configuration of the form $(p_0,p_1,\ldots,p_l)$ with $l\in {\mathbb{N}}$ (in particular, that of the minimal embedded resolution of the branch).
### Curves of Abhyankar-Moh-Suzuki type {#ams}
In this paper we are mainly interested in a certain class of curves having one place at infinity: the so-called curves of Abhyankar-Moh-Suzuki type, which we define next.
[A plane curve $C$ having one place at infinity is said to be of [*Abhyankar-Moh-Suzuki*]{} type ($AMS$ type for short) if it is rational and smooth in its affine part, that is, $C\setminus H$ is isomorphic to ${\mathbb{C}}$, $H$ being the line of infinity. ]{}
Let $H$ be the line of infinity in ${\mathbb{P}}^2$ and identify ${\mathbb{C}}^2$ with ${\mathbb{P}}^2\setminus H$. Recall that, by [@abh1], a curve $C$ is of AMS type if and only if it is the compactification in ${\mathbb{P}}^2$ of the zero locus of a component of a certain automorphism $\phi:
{\mathbb{C}}^2\rightarrow {\mathbb{C}}^2$. The embedding of ${\mathbb{C}}^2$ in ${\mathbb{P}}^2$ allows to extend $\phi$ to a birational transformation $\tilde{\phi}:{\mathbb{P}}^2\rightarrow {\mathbb{P}}^2$. The minimal embedded resolution of the singularity of $C$ is closely related to the minimal resolution of the indeterminacy of $\tilde{\phi}$, and the combinatorics of the last one can be described precisely, as we will show next. For details see [@fdb1] or [@fdb2].
First, we will define an associative operation $\uparrow$ in the set of unibranched proximity graphs with two or more vertices (see Figure 1 for an example).
Let ${\mathbb{F}}_1$ and ${\mathbb{F}}_2$ be two proximity graphs of this type and assume that $V_1=\{v_0,v_1,\ldots,v_n\}$ (resp., $V_2=\{w_0,w_1,\ldots,w_m\}$) is the set of vertices of ${\mathbb{F}}_1$ (resp., ${\mathbb{F}}_2$) where, if $\leq$ denotes the ordering induced in $V_{{\mathbb{F}}_1}$ (resp., $V_{{\mathbb{F}}_2}$) by the natural ordering among the points of a configuration whose proximity graph is ${\mathbb{F}}_1$ (resp., ${\mathbb{F}}_2$), it holds that $v_0<v_1\ldots <v_n$ (resp., $w_0<w_1\ldots <w_m$). The graph ${\mathbb{F}}_1 \uparrow {\mathbb{F}}_2$ is the unibranched proximity graph such that:
- its set of vertices is $V_{{\mathbb{F}}_1}\cup V_{{\mathbb{F}}_2}$;
- its set of edges is $A\cup \{e_1,e_2 \}$, where $A$ is the union of the sets of edges of ${\mathbb{F}}_1$ and ${\mathbb{F}}_2$ and $e_1,e_2$ are two new edges such that $e_1$ is a continuous straight line joining $v_n$ and $w_0$, and $e_2$ is a curved dotted line joining $v_n$ and $w_1$;
- the vertex $v_i$ (resp., $w_i$) is labelled with $i$ (resp., $n+i+1$) for each $i$ such that $0\leq i\leq n$ (resp., $0\leq i\leq m$).
\[figu\]
(55,50) (10,5)(5,12.5)(10,20)
(10,5) (10,5)[(0,1)[5]{}]{} (10,10) (10,10)[(0,1)[5]{}]{} (10,15) (10,15)[(0,1)[5]{}]{} (10,20) (10,20)[(0,1)[5]{}]{} (10,25)
(12,5)[$0$]{} (12,10)[$1$]{} (12,15)[$2$]{} (12,20)[$3$]{} (12,25)[$4$]{}
(20,10)(15,15)(20,20)
(20,5) (20,5)[(0,1)[5]{}]{} (20,10) (20,10)[(0,1)[5]{}]{} (20,15) (20,15)[(0,1)[5]{}]{} (20,20)
(22,5)[$0$]{} (22,10)[$1$]{} (22,15)[$2$]{} (22,20)[$3$]{}
(40,5)(35,12.5)(40,20)
(40,5) (40,5)[(0,1)[5]{}]{} (40,10) (40,10)[(0,1)[5]{}]{} (40,15) (40,15)[(0,1)[5]{}]{} (40,20) (40,20)[(0,1)[5]{}]{} (40,25)
(40,35)(35,40)(40,45)
(40,30) (40,30)[(0,1)[5]{}]{} (40,35) (40,35)[(0,1)[5]{}]{} (40,40) (40,40)[(0,1)[5]{}]{} (40,45)
(40,25)[(0,1)[5]{}]{}
(40,25)(35,30)(40,35)
(42,5)[$0$]{} (42,10)[$1$]{} (42,15)[$2$]{} (42,20)[$3$]{} (42,25)[$4$]{}
(42,30)[$5$]{} (42,35)[$6$]{} (42,40)[$7$]{} (42,45)[$8$]{}
(9,0)[${\mathbb{F}}_1$]{} (19,0)[${\mathbb{F}}_2$]{} (35,0)[${\mathbb{F}}_1 \uparrow
{\mathbb{F}}_2$]{}
For each integer $n\geq 2$ consider a chain configuration $(p_0,p_1,\ldots,p_{2n-2})$ such that $p_i$ is proximate to $p_0$ for all $i$ such that $1\leq i\leq n-1$ and the remaining points $p_{n}, p_{n+1}\ldots p_{2n-2}$ are free. Define ${\mathbb{G}}(n)$ to be the proximity graph of this configuration (see Figure 2).
\[figu\]
(25,50) (15,0)(5,12.5)(15,25)
(15,0) (15,0)[(0,1)[5]{}]{} (15,5) (15,5)[(0,1)[5]{}]{} (15,10) (15,10)[(0,1)[2.5]{}]{}
(15,25) (15,25)[(0,1)[2.5]{}]{} (15,25)[(0,-1)[2.5]{}]{}
(15,40)[(0,-1)[2.5]{}]{}
(15,40) (15,40)[(0,1)[5]{}]{} (15,45)
(14.7,16)[$\vdots$]{}
(14.7,31)[$\vdots$]{}
(16,0)[ 0]{}(16,5)[ 1]{}(16,10)[ 2]{}(16,40)[ $2n-3$]{}(16,45)[ $2n-2$]{} (17,25)[$n-1$]{}
For each finite ordered sequence $(n_1,n_2,\ldots,n_r)$ of integers such that $r\geq 1$ and $n_i\geq 2$ for all $i$, we define a proximity graph, depending only on that sequence, by using the above considered associative operation: $${\mathbb{G}}(n_1,n_2,\ldots,n_r):={\mathbb{G}}(n_1) \uparrow {\mathbb{G}}(n_2)
\uparrow \cdots \uparrow {\mathbb{G}}(n_r).$$ Also, we denote by ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^{-}$ (resp., ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^{+}$) the proximity graph obtained from ${\mathbb{G}}(n_1,n_2,\ldots,n_r)$ by deleting (resp., adding) the last $n_r-1$ vertices and the edges which are adjacent to them (resp., a new vertex with label $2\sum_{i=1}^r n_i -r$ and a new edge joining it with the vertex with label $2\sum_{i=1}^r n_i-r-1$).
Now consider, as above, a curve $C$ of AMS type and an affine automorphism $\phi: {\mathbb{C}}^2\rightarrow {\mathbb{C}}^2$ such that $C$ is the zero locus of a component of it. Let $\pi:X\rightarrow {\mathbb{P}}^2$ be the minimal resolution of the indeterminacy of $\tilde{\phi}:{\mathbb{P}}^2\rightarrow {\mathbb{P}}^2$ and let ${\mathcal K}$ be the configuration of centers of the blowing-ups involved in $\pi$. Then, there exists a sequence of integers $(n_1,n_2,\ldots,n_r)$ (with $n_i \geq 2$ for all $i$) such that ${\mathbb{G}}({\mathcal
K})={\mathbb{G}}(n_1,n_2,\ldots,n_r)$. Moreover, the strict transform on $X$ of the line of infinity, $\tilde{H}^{{\mathcal K}}$, is a $(-1)$-curve, that is, a smooth rational curve with self-intersection $-1$.
If ${\mathcal C}$ is the configuration such that $\pi_{\mathcal
C}:X_{\mathcal C}\rightarrow {\mathbb{P}}^2$ induces the minimal embedded resolution of the singularity of $C$ at infinity, there are two possibilities: either $\pi_{\mathcal C}$ is the composition of all the blowing-ups of $\pi$ except the last $n_r-1$ of them (in this case, ${\mathbb{G}}({\mathcal C})={\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$), or it is the composition of the first $2\sum_{i=1}^{r-2}n_i+n_{r-1}-r+2$ blowing-ups of $\pi$ (in this case, ${\mathbb{G}}({\mathcal
C})={\mathbb{G}}(n_1,n_2,\ldots,n_{r-1})^-$). The following proposition shows that, for each proximity graph of the form ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$, there exists a curve of AMS type such that the proximity graph associated to its minimal embedded resolution is this one. Then, the proximity graphs associated to minimal resolutions of curves of AMS type are exactly those of the form ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$.
\[villa\] Let $(n_1,n_2,\ldots,n_r)$ be an ordered sequence of integers such that $n_i\geq 2$ for all $i=1,2,\ldots,r$. Then, there exists a curve $C$ of AMS type such that its degree is $n_1n_2\cdots n_r$ and the proximity graph associated with its minimal embedded resolution is ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$.
[*Proof*]{}. Define the integers $\delta_k=n_{k+1}
n_{k+2}\cdots n_r$ for $k=0,1,\ldots,r-1$ and $\delta_r=1$. It is obvious that the sequence $(\delta_0,\delta_1,\ldots,\delta_r)$ satisfies the conditions (I), (II) and (III) which characterize the $\delta$-sequences and, therefore, there exists a curve $C$ of degree $\delta_0=n_1 n_2 \cdots n_r$ having one place at infinity with associated $\delta$-sequence $(\delta_0,\delta_1,\ldots,\delta_r)$. From this sequence one can compute, using the formulae given in Section \[deltasec\], the proximity relations among the points of the configuration which provides the minimal embedded resolution of the singularity of $C$ and check that the proximity graph associated with this configuration is ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$. Finally, since $\delta_r=1$, the Weierstrass semigroup of $C$ and its semigroup at infinity are both equal to ${\mathbb{N}}$ and, therefore, $C$ is rational and smooth in its affine part.[$\ \ \square$]{}\
A direct consequence of the above proposition and the genus formula is the following
\[pernia\] Let $C$ be a curve of AMS type and $n_1,n_2,\ldots,n_r\geq 2$ integers such that the proximity graph associated with its minimal embedded resolution is ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$. Then, the degree of $C$ is $n_1 n_2\cdots n_r$.
Surfaces associated with pencils “at infinity” {#infinity}
==============================================
Let $C$ be a projective curve of ${\mathbb{P}}^2$ having one place at infinity and consider the notations of Section \[pre\]. Take projective coordinates $(X:Y:Z)$ on ${\mathbb{P}}^2$ such that $Z=0$ be the equation of the line of infinity $H$ and let $F(X,Y,Z)$ be an homogeneous polynomial in $k[X,Y,Z]$ such that $F(X,Y,Z)=0$ is an equation of $C$. The [*pencil “at infinity” associated with $C$*]{}, which we will denote by ${{\mathcal P}}(C)$, will be the linear subspace of $H^0({\mathbb{P}}^2, {{\mathcal O}}_{{\mathbb{P}}^2}(d))$ spanned by $F$ and $Z^d$, $d$ being the degree of $F$. Let $n$ be the smallest integer such that the composition of morphisms $X_{n+1}\rightarrow
X_n\rightarrow \cdots \rightarrow X_0={\mathbb{P}}^2$ eliminates the indeterminacies of the rational map ${\mathbb{P}}^2 \cdots \rightarrow
{\mathbb{P}}^1$ defined by ${{\mathcal P}}(C)$. We will denote by ${\mathcal K}_C$ the chain configuration $(p_0,p_1,\ldots,p_n)$ and by $X_C$ the surface $Z_{{\mathcal K}_C}=X_{n+1}$. It turns out that all the curves of ${{\mathcal P}}(C)$, except the non-reduced one with equation $Z^d=0$, are integral curves having one place at infinity, and the above morphism $X_{C}\rightarrow {\mathbb{P}}^2$ induces a simultaneous embedded resolution of all of them (see [@moh]).
The objective of this section is to give a vanishing theorem for line bundles on the surface $X_C$, when $C$ is a curve of AMS type. This result will allow to determine the dimension of whichever complete linear system on $X_C$.
The following proposition provides two characterizations of the curves of AMS type depending on the associated configuration ${\mathcal K}_C$.
\[rat\] Let $C$ be a curve having one place at infinity. Then, the following conditions are equivalent:
- The configuration ${\mathcal K}_C$ is P-sufficient.
- $K_{X_C}\cdot \tilde{C}^{{\mathcal K}_C}<0$.
- $C$ is a curve of AMS type.
[*Proof*]{}. First, observe that the class of $\tilde{C}^{{\mathcal K}_C}$ in ${{\rm Pic}}(X_C)$ coincides with $[dL^{{\mathcal K}_C}-D_n]$, $d$ being the degree of $C$ and $D_n$ the divisor defined in Section \[psuf\] (associated with the configuration ${\mathcal K}_C$). The reason is that $D_n=\sum_{i=0}^n
u_iE_i^{{\mathcal K}_C}$, $u_i$ being the multiplicity of the strict transform of $C$ at the point $p_i$, $0\leq i\leq n$, because $D_n\cdot \tilde{E}_i^{{\mathcal K}_C}$ equals $-1$ if $i=n$ and $0$ otherwise.
The equivalence between $(a)$ and $(b)$ is consequence of the equalities $$-9D_n^2-\left(K_{X_C}\cdot D_n\right)^2=9d^2-\left(\sum_{i=0}^n u_i\right)^2=\left(3d+\sum_{i=0}^n
u_i\right)\left(-K_{X_C}\cdot \tilde{C}^{{\mathcal K}_C}\right),$$ where the first one follows by Bézout’s Theorem.
The equivalence between $(b)$ and $(c)$ follows from the expression of the arithmetic genus of $\tilde{C}^{{\mathcal K}_C}$, $$p_a(\tilde{C}^{{\mathcal K}_C})=1+\frac{1}{2}K_{X_C}\cdot
\tilde{C}^{{\mathcal K}_C},$$ and the following fact: if $C$ were not smooth in its affine part, then its geometric genus would be less than $p_a(\tilde{C}^{{\mathcal K}_C})$. [$\ \ \square$]{}\
A direct consequence of this proposition is the following
\[www\] If $C$ is a curve of AMS type, then all the curves of the pencil ${{\mathcal P}}(C)$, except the non-reduced one, are also of AMS type.
\[rectarat2\] [If $C$ is a curve of AMS type, then $\tilde{H}^{{\mathcal K}_C}$ is a $(-1)$-curve of $X_C$. This fact is trivial from what is said in Section \[ams\].]{}
The following result provides a characterization of the proximity graphs of the form ${\mathbb{G}}({\mathcal K}_C)$, where $C$ is a curve of AMS type.
\[sss\] Let ${\mathbb{G}}$ be a proximity graph. Then, there exists a curve $C$ of AMS type such that ${\mathbb{G}}={\mathbb{G}}({\mathcal K}_C)$ if and only if there is a sequence of integers $n_1,n_2,\ldots,n_r\geq 2$ such that ${\mathbb{G}}={\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$.
[*Proof*]{}. Assume the existence of a curve $C$ of AMS type such that ${\mathbb{G}}={\mathbb{G}}({\mathcal K}_C)$. The degree of $C$ is $n_1
n_2\cdots n_r$ by Corollary \[pernia\]. Suppose that the configuration ${\mathcal K}_C$ is $(p_0,p_1,\ldots,p_n)$ and let ${\mathcal C}$ be the configuration provided by the minimal embedded resolution of $C$. Since the morphism $\pi_{{\mathcal K}_C}: X_C
\rightarrow {\mathbb{P}}^2$ induces an embedded resolution of $C$, one has that ${\mathcal C}=(p_0,p_1,\ldots,p_m)$ for some $m\leq n$. Then, applying Bézout’s Theorem to two curves of the pencil ${{\mathcal P}}(C)$, the following equality holds: $$\label{b}
(n_1 n_2\cdots n_r)^2=\sum_{i=0}^m u_i^2 +n-m,$$ where $u_i$ denotes the multiplicity of the strict transform of $C$ at $p_i$. Taking into account that ${\mathbb{G}}({\mathcal
C})={\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$ and $u_i=\sum_{p_j\rightarrow p_i}
u_j$ for all $i=1,2,\ldots,m$, the multiplicities $u_i$ can be easily computed in terms of $n_1,n_2,\ldots,n_r$ and, from the equality $(\ref{b})$, it holds that $n-m=n_r$. Therefore, ${\mathbb{G}}({\mathcal K}_C)={\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$.
Conversely, assume that ${\mathbb{G}}={\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$ for certain integers $n_1,n_2,\ldots,n_r\geq 2$. By Proposition \[villa\], there exists a curve $C$ of AMS type such that the proximity graph associated with its minimal embedded resolution is ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^-$. Similar arguments to those used in the above paragraph show that ${\mathbb{G}}({\mathcal K}_C)={\mathbb{G}}$.[$\ \ \square$]{}
[For each curve $C$ having one place at infinity, the [*effective*]{} (resp., [*nef*]{}) [*semigroup*]{} of $X_C$, denoted by $NE(X_C)$ (resp., $P(X_C)$), is defined as the subsemigroup of ${{\rm Pic}}(X_C)$ generated by the classes of all effective (resp., numerically effective) divisors on $X_C$. ]{}
Campillo, Piltant and Reguera described, in [@c-p-r1], the effective semigroup of $X_C$. They proved the following equality: $$NE(X_C)={\mathbb{N}}[\tilde{H}^{{\mathcal K}_C}]\oplus \bigoplus_{i=0}^n {\mathbb{N}}[\tilde{E}_i^{{\mathcal K}_C}],$$ where $H$ denotes, as above, the line of infinity. Moreover, as a consequence of [@c-p-r1 Cor. 7 and Prop. 6] and Corollary \[www\], we have the following result:
\[sup\] If $C$ is a curve of AMS type, then $P(X_C)$ coincides with the semigroup of classes in ${{\rm Pic}}(X_C)$ of the form $[\tilde{D}^{{\mathcal K}_C}]$, where $D\hookrightarrow {\mathbb{P}}^2$ is a projective curve whose support does not contain the line of infinity.
Now, we will state and prove the announced $H^1$-vanishing result for line bundles on $X_C$.
\[gordo\] Let $C$ be a curve of AMS type.
- If $D$ is a numerically effective divisor on $X_C$, then $h^1(X_C, {{\mathcal O}}_{X_C}(D))=0$.
- Let $D$ be an effective divisor on $X_C$ such that $D\cdot \tilde{E}_i^{{\mathcal K}_C}\geq 0$ for all $i=0,1,\ldots,n$ and $D\cdot \tilde{E}_1^{{\mathcal K}_C}\geq 1$ . Then, $h^1(X_C,
{{\mathcal O}}_{X_C}(D))=0$ if and only if $D\cdot \tilde{H}^{{\mathcal K}_C}\geq -1$.
[*Proof*]{}. Firstly, notice that the configuration ${\mathcal K}_C$ is P-sufficient, by Proposition \[rat\]. In order to prove $(a)$, we will reason by contradiction. So, we assume the existence of a numerically effective divisor $D$ such that $h^1(X_C, {{\mathcal O}}_{X_C}(D))>0$. Since $[D]$ is an effective class (by Proposition \[sup\]) we can apply Proposition \[2\] and [@harb Lem. II.7], deducing that the complete linear system $|D|$ has fixed part. Moreover, again by Proposition \[sup\], this fixed part has not exceptional components (that is, it has no divisor $\tilde{E}_i^{{\mathcal K}_C}$ as a component).
Now, the integral fixed components of $|D|$ have negative self-intersection. Indeed, if we assume the existence of an integral fixed component $R$ such that $R^2\geq 0$, then $K_{X_C}\cdot R<0$ (by Proposition \[2\]) and, applying the Riemann-Roch Formula, we get $$h^0(X_C,{{\mathcal O}}_{X_C}(R))\geq 1+(R^2-K_{X_C}\cdot R)/2\geq 2,$$ which is false, since $h^0(X_C,{{\mathcal O}}_{X_C}(R))=1$.
Finally Proposition \[sup\] provides a contradiction, because the unique non-exceptional integral curve on $X_C$ with negative self-intersection is $\tilde{H}^{{\mathcal K}_C}$.
To prove $(b)$, we consider a divisor $D$ satisfying the hypotheses. First, we will assume the inequality $h^1(X_C, {{\mathcal O}}_{X_C}(D))>0$ and we will show that this implies $D\cdot \tilde{H}^{{\mathcal K}_C}\leq -2$. By applying $(a)$ we have $D\cdot \tilde{H}^{{\mathcal K}_C}\leq -1$ and, therefore, it only remains to prove that the equality $D\cdot
\tilde{H}^{{\mathcal K}_C}=-1$ leads to a contradiction. Using the hypotheses and Remark \[rectarat2\], it can be deduced that $D-\tilde{H}^{{\mathcal K}_C}$ is a numerically effective divisor. A similar reasoning to that given in the proof of $(a)$ shows that $\tilde{H}^{{\mathcal K}_C}$ is the unique possible integral fixed component of the linear system $|D-\tilde{H}^{{\mathcal K}_C}|$ and, then, it must be fixed part free by Proposition \[sup\]. So, we have a decomposition $[D]=[\tilde{H}^{{\mathcal K}_C}]+[T]$, where $T$ is an effective divisor such that $|T|$ is fixed part free. Applying Part $(a)$ to the divisor $T$, we deduce that $h^1(X_C,{{\mathcal O}}_{X_C}(T))=0$. Taking into account this fact, Riemann-Roch Theorem and the equality $h^0(X_C,{{\mathcal O}}_{X_C}(D))=h^0(X_C,{{\mathcal O}}_{X_C}(T))$, the following chain of equalities and inequalities holds: $$0<h^1(X_C,{{\mathcal O}}_{X_C}(D))=h^0(X_C,{{\mathcal O}}_{X_C}(D))-1-\frac{1}{2}(D^2-K_{X_C}\cdot
D)=$$ $$=h^0(X_C,{{\mathcal O}}_{X_C}(T))-1-\frac{1}{2}(D^2-K_{X_C}\cdot
D)=$$ $$=1+\frac{1}{2}(T^2-K_{X_C}\cdot
T)-1-\frac{1}{2}(D^2-K_{X_C}\cdot D)=$$ $$=\frac{1}{2}(-(\tilde{H}^{{\mathcal K}_C})^2-2 \tilde{H}^{{\mathcal K}_C}\cdot T+K_{X_C}\cdot \tilde{H}^{{\mathcal K}_C}).$$ Hence, $(\tilde{H}^{{\mathcal K}_C})^2-K_{X_C}\cdot
\tilde{H}^{{\mathcal K}_C}<-2\tilde{H}^{{\mathcal K}_C}\cdot T\leq 0$. But this is a contradiction, since $\tilde{H}^{{\mathcal K}_C}$ is a $(-1)$-curve.
It only remains to prove that, if $D\cdot \tilde{H}^{{\mathcal K}_C}\leq -2$, then $h^1(X_C, {{\mathcal O}}_{X_C}(D))>0$. But the inequality $D\cdot
\tilde{H}^{{\mathcal K}_C}\leq -2$ implies that the $(-1)$-curve $\tilde{H}^{{\mathcal K}_C}$ is a multiple fixed component of the linear system $|D|$, and it is easy to see that this fact implies that $h^1(X_C, {{\mathcal O}}_{X_C}(D))$ is positive (by a similar reasoning to that given in [@mir pag. 197]).[$\ \ \square$]{}
\[dim\]
With the hypotheses of Theorem \[gordo\], Part $(a)$ allows us to determine the dimension of all complete linear systems on $X_C$. Indeed, let $D$ be a divisor on $X_C$ and consider the set $S=\{[\tilde{H}^{{\mathcal K}_C}],
[\tilde{E}_0^{{\mathcal K}_C}],[\tilde{E}_1^{{\mathcal K}_C}],\ldots,[\tilde{E}_n^{{\mathcal K}_C}]\}\subseteq
{{\rm Pic}}(X_C)$. If for some $F\in S$ we have $D\cdot F<0$, then it is obvious that $h^0(X_C,{{\mathcal O}}_{X_C}(D))=h^0(X_C,{{\mathcal O}}_{X_C}(D-F))$. Therefore, we can perform the following process: check $D\cdot F$ for each $F\in S$, replace $D$ by $D-F$ whenever $D\cdot F<0$ and continue with the new $D$. The process ends when it gives rise to a divisor $D'$ such that either it is obviously not effective (because either $D'\cdot L^{{\mathcal K}_C}<0$ or $D'\cdot(L^{{\mathcal K}_C}-E_1^{{\mathcal K}_C})<0$) or $D'$ is numerically effective. Since $h^0(X_C,{{\mathcal O}}_{X_C}(D))=h^0(X_C,{{\mathcal O}}_{X_C}(D'))$, in the first case the linear system $|D|$ is empty and in the second case its dimension is $(D'^2-K_{X_C}\cdot D')/2$, by Part $(a)$ of Theorem \[gordo\].
Linear systems of curves through generic points on ${\mathbb{P}}^2$ {#s4}
===================================================================
In this section we will use Theorem \[gordo\] and a specialization process to deduce some results about the dimension of linear systems of curves passing through a finite set of points of the plane in generic position.
Special linear systems and the Harbourne-Hirschowitz Conjecture {#s41}
---------------------------------------------------------------
Given a projective curve $C$ of ${\mathbb{P}}^2$, we will say that $C$ [*goes through*]{} a weighted configuration $({\mathcal K}=(p_i)_{i=0}^n,\bold{m}=(m_i)_{i=0}^n)$ if and only if the divisor $C^{{\mathcal K}}-\sum_{i=0}^n m_i E_i^{{\mathcal K}}$ is effective. For any degree $d$, denote by ${{\mathcal L}}_d({\mathcal K},\bold{m})$ the set of projective curves on ${\mathbb{P}}^2$ of degree $d$ going through $({\mathcal K},\bold{m})$. This is a linear system of ${\mathbb{P}}^2$ that is projectively isomorphic to the complete linear system $|D_{d,{\mathcal K},\bold{m}}|$ of $Z_{{\mathcal K}}$, $D_{d,{\mathcal K},\bold{m}}$ being the divisor $d L^{{\mathcal K}}-\sum_{i=0}^n m_i
E_i^{{\mathcal K}}$. From Riemann-Roch Theorem one gets $$\dim {{\mathcal L}}_d({\mathcal K},\bold{m})-h^1({{\mathcal L}}_d({\mathcal K},
\bold{m}))=\frac{d(d+3)}{2}-\sum_{i=0}^n \frac{m_i(m_i+1)}{2},$$ where $\dim {{\mathcal L}}_d({\mathcal K},\bold{m})$ is the dimension of ${{\mathcal L}}_d({\mathcal K},\bold{m})$ as projective space and $h^1({{\mathcal L}}_d({\mathcal K},
\bold{m})):=h^1(Z_{{\mathcal K}},{{\mathcal O}}_{Z_{{\mathcal K}}}(D_{d,{\mathcal K},\bold{m}}))$ will be called the [*superabundance*]{} of ${{\mathcal L}}_d({\mathcal K},\bold{m})$. The independence of the linear conditions imposed by the weighted configuration $({\mathcal K},\bold{m})$ is equivalent to the vanishing of this superabundance.
If ${\mathcal K}$ is a configuration whose points lie all in ${\mathbb{P}}^2$, the dimension and the superabundance of a linear system ${{\mathcal L}}_d({\mathcal K},\bold{m})$ depend on the position of the points of ${\mathcal K}$, and they reach their minimal values for a generic set of points. We will denote by ${\mathcal K}_0(n)$ a configuration consisting of $n+1$ closed points of ${\mathbb{P}}^2$ in generic position. For each integer $d\geq 1$ and for each system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_n)$ we will denote by ${{\mathcal L}}_d(\bold{m})$ the linear system ${{\mathcal L}}_d({\mathcal K}_0(n),\bold{m})$. Also, we define the [*expected dimension*]{} of ${{\mathcal L}}_d(\bold{m})$ to be the following number: $${{\rm edim\;}}{{\mathcal L}}_d(\bold{m}):=\max \left\{\frac{d(d+3)}{2}-\sum_{i=0}^n
\frac{m_i(m_i+1)}{2},-1\right\}.$$
[We will say that a linear system ${{\mathcal L}}_d(\bold{m})$ is [*special*]{} if and only if $\dim {{\mathcal L}}_d(\bold{m})>{{\rm edim\;}}{{\mathcal L}}_d(\bold{m})$, that is, ${{\mathcal L}}_d(\bold{m})$ is non-empty and the superabundance $h^1({{\mathcal L}}_d(\bold{m}))$ is positive.]{}
Given a positive integer $d$ and a system of multiplicities $\bold{m}=(m_i)_{i=0}^n$, it is easy to prove that, if there exists a curve $C$ on ${\mathbb{P}}^2$ such that its strict transform on $Z_{{\mathcal K}_0(n)}$ is a $(-1)$-curve and $D_{d,{\mathcal K}_0(n),\bold{m}}\cdot
\tilde{C}^{{\mathcal K}_0(n)}\leq -2$, then the linear system ${\mathcal
L}_d(\bold{m})$ is special (see, for instance, [@mir pag. 197]). One of the equivalent statements of the Harbourne-Hirschowitz Conjecture is just the converse assertion:\
[**Conjecture**]{}. (Harbourne-Hirschowitz) If a linear system ${{\mathcal L}}_d(\bold{m})$ is special, then there exists a curve $C$ on ${\mathbb{P}}^2$ such that its strict transform on $Z_{{\mathcal K}_0(n)}$ is a $(-1)$-curve and $D_{d,{\mathcal K}_0(n),\bold{m}}\cdot
\tilde{C}^{{\mathcal K}_0(n)}\leq -2$.
A non-speciality result and some consequences {#s42}
---------------------------------------------
For each positive integer $n$, there exists a variety $Y_{n}$ whose points are naturally identified with the configurations over ${\mathbb{P}}^2$ with $n+1$ points. These varieties, known as [*iterated blowing-ups*]{}, were introduced by Kleiman in [@kle1] and [@kle2] and they have also been treated in [@harb2], [@roe] and [@fdb2] (see also [@roe2]). There is a family of projective morphisms $Y_{n+1}\rightarrow Y_{n}$ and relative divisors $F_{-1},F_0,F_1,\ldots,F_n$ on $Y_{n+1}$ such that the fiber over a given configuration ${\mathcal K}=(p_0,p_1,\ldots,p_n)$ (viewed as a point of $Y_{n}$) is isomorphic to the surface $Z_{{\mathcal K}}$ obtained by blowing-up the points in ${\mathcal K}$ and, if $i\geq 0$ (resp., $i=-1$), the restriction of $F_i$ to this fiber corresponds to the total transform $E_i^{{\mathcal K}}$ of the exceptional divisor appearing in the blowing-up centered at $p_i$ (resp., the total transform of a general line of ${\mathbb{P}}^2$).
For each positive integer $d$ and for each sequence of multiplicities ${\bold m}=(m_0,m_1,\ldots,m_n)$ we apply the Semicontinuity Theorem [@har III, 12.8] to the invertible sheaf ${{\mathcal O}}_{Y_{n+1}}(dF_{-1}-m_0F_0-m_1F_1-\ldots-m_nF_n)$, obtaining that the functions $Y_n\rightarrow {\mathbb{Z}}$ given by $$\label{semic}
{{\mathcal K}} \mapsto h^i(Z_{{\mathcal K}}, {{\mathcal O}}_{Z_{{\mathcal K}}}(D_{d,{\mathcal K}, \bold{m}})),$$ for $i\in \{0,1\}$, are upper-semicontinuous.
For each proximity graph ${\mathbb{G}}$ with $n+1$ vertices, we define $U({\mathbb{G}})$ as the subset of $Y_n$ containing exactly the configurations ${\mathcal K}$ whose proximity graph is ${\mathbb{G}}$. This is an irreducible smooth locally closed subvariety ([@roe] and [@fdb2]). As a consequence of the upper-semicontinuity of the functions given in (\[semic\]), for any positive integer $d$ and any system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_n)$, the dimension and the superabundance of the linear systems ${\mathcal
L}_d({\mathcal K},\bold{m})$, for ${\mathcal K}$ varying in $U({\mathbb{G}})$, achieve the minimum value in a dense open subset of $U({\mathbb{G}})$.
[We will say that a weighted configuration $({\mathcal K},
\bold{m}=(m_i)_{i=0}^n)$ (resp., a weighted proximity graph $({\mathbb{G}},\bold{m})$) is [*consistent*]{} if all the excesses $\rho_j({\mathcal K},\bold{m})$ (resp., $\rho_j({\mathbb{G}},\bold{m})$)) are non-negative (see Section \[s1\] for the definition of excesses). In this case, and provided that $n\geq 1$, we associate with $({\mathcal K},
\bold{m})$ an integer, denoted by $\epsilon({\mathcal K}, \bold{m})$ (or $\epsilon({\mathbb{G}}({\mathcal K}), \bold{m})$, since it depends only on the weighted proximity graph) and defined to be either $1$, if $\rho_1({\mathcal K}, \bold{m})\geq 1$, or $0$, if $\rho_1({\mathcal K}, \bold{m})=0$. ]{}
Given a weighted proximity graph $({\mathbb{G}}={\mathbb{G}}({\mathcal K}),\bold{m}=(m_i)_{i=0}^n)$, it is possible to obtain a unique system of multiplicities, which will be denoted by $\bold{m}^{{\mathbb{G}}}=(m_0^{{\mathbb{G}}},m_1^{{\mathbb{G}}},\ldots,m_n^{{\mathbb{G}}})$, such that $({\mathbb{G}}, \bold{m}^{{\mathbb{G}}})$ is consistent and the ideal sheaves given by ${\pi_{{\mathcal K}}}_*{{\mathcal O}}_{Z_{{\mathcal K}}}(-\sum_{i=0}^n m_i E_i^{{\mathcal K}})$ and ${\pi_{{\mathcal K}}}_*{{\mathcal O}}_{Z_{{\mathcal K}}}(-\sum_{i=0}^n m_i^{{\mathbb{G}}} E_i^{{\mathcal K}})$ coincide. So, there exists a canonical bijection between the linear systems $|D_{d,{\mathcal K},\bold{m}}|$ and $|D_{d,{\mathcal K},\bold{m}^{{\mathbb{G}}}}|$ for all integers $d\geq 1$. The procedure used to obtain $\bold{m}^{{\mathbb{G}}}$ is called [*unloading*]{} [@cas 4.6] and it depends only on the proximity graph ${\mathbb{G}}$, and not on a special election of the configuration ${\mathcal K}$ associated with ${\mathbb{G}}$. In each step of the unloading procedure ([*unloading step*]{}) one must detect a point $p_i$ of ${\mathcal K}$ such that its associated excess $\rho_i({\mathcal K},\bold{m})$ is negative; then, one replaces the system of multiplicities $\bold{m}$ by the system $\bold{m}'=(m_0',m_1',\ldots,m_n')$ where $m_i'=m_i+1$, $m_j'=m_j-1$ for those indexes $j$ such that $p_j$ is proximate to $p_i$, and $m_j'=m_j$ otherwise (if some multiplicity in $\bold{m}'$ is negative, it must be replaced by $0$). Now, we must perform another unloading step from the new system $\bold{m}'$, and so on. A finite number of unloading steps lead to the desired system of multiplicities $\bold{m}^{{\mathbb{G}}}$. An unloading step applied to a point $p_i$ whose associated excess equals $-1$ is called [*tame*]{}. Tame unloadings will be very useful for us, since they preserve independence of conditions, that is, if $\bold{m}'$ is obtained from $\bold{m}$ performing a tame unloading step, then $h^1(Z_{{\mathcal K}},{{\mathcal O}}_{Z_{{\mathcal K}}}(D_{d,{\mathcal K},\bold{m}}))=h^1(Z_{{\mathcal K}},{{\mathcal O}}_{Z_{{\mathcal K}}}(D_{d,{\mathcal K},\bold{m}'}))$ for all positive integer $d$ (this fact can be easily deduced from [@cas 4.7.1] and [@cas 4.7.3]).
[We will say that a weighted configuration $({\mathcal K}, \bold{m})$ (resp., a weighted proximity graph $({\mathbb{G}}, \bold{m})$) is [*almost consistent*]{} if either it is consistent or there exists a sequence of tame unloading steps leading from $\bold{m}$ to $\bold{m}^{{\mathbb{G}}({\mathcal K})}$ (resp., $\bold{m}^{{\mathbb{G}}}$). ]{}
The following theorem gives a sufficient condition for the non-speciality of a linear system ${\mathcal L}_d(\bold{m})$, when $\bold{m}$ is a system of multiplicities such that $({\mathcal K}_C,\bold{m})$ is almost consistent, $C$ being a curve of AMS type. Moreover, when this weighted configuration is consistent and the excess $\rho_1({\mathcal K}_C,\bold{m})$ is positive, it provides a characterization of such non-special linear systems which are not empty.
\[super\] Let $d$ be a positive integer and $\bold{m}=(m_0,m_1,\ldots,m_n)$ a system of multiplicities, with $n\geq 1$. Assume the existence of a curve $C$ of AMS type such that $({\mathcal K}_C,\bold{m})$ is almost consistent. Then, the linear system ${\mathcal L}_d(\bold{m})$ is non-special whenever $d\geq m_0^{{\mathbb{G}}}+m_1^{{\mathbb{G}}}-\epsilon({\mathbb{G}},
\bold{m}^{{\mathbb{G}}})$, where ${\mathbb{G}}:={\mathbb{G}}({\mathcal K}_C)$. Moreover, if ${\mathcal
L}_d(\bold{m})$ is not empty, $({\mathcal K}_C,\bold{m})$ is consistent and $\rho_1({\mathbb{G}}, \bold{m})\geq 1$, then the following equivalence holds: ${\mathcal L}_d(\bold{m})$ is non-special if and only if $d\geq
m_0+m_1-1$.
[*Proof*]{}. Set ${\mathcal K}_C=(p_0,p_1,\ldots,p_n)$ (adding null multiplicities to $\bold{m}$, if it is necessary, we can assume that the cardinality of ${\mathcal K}_C$ is $n+1$). In order to prove the first assertion of the statement, we will reason by contradiction. So, assume that $d\geq m_0^{{\mathbb{G}}}+m_1^{{\mathbb{G}}}-\epsilon({\mathbb{G}},
\bold{m}^{{\mathbb{G}}})$ and ${\mathcal L}_d(\bold{m})$ is special. The subset $U({\mathbb{G}}({\mathcal K}_0(n)))$ is dense in $Y_n$ (see [@kle1]) and then, as a consequence of the upper-semicontinuity of the functions given in (\[semic\]), the following inequalities hold: $\dim
{\mathcal L}_d(\bold{m})\leq
h^0(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}}))-1$ and $h^1({\mathcal
L}_d(\bold{m}))\leq h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}}))$. Thus, the complete linear system on $X_C$ given by $|D_{d,{\mathcal K}_C,\bold{m}}|$ is not empty and $h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}}))$ is positive. But, since $({\mathcal K}_C,\bold{m})$ is almost consistent, we have $$h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}^{{\mathbb{G}}}}))=h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}})).$$ The consistency of $({\mathcal K}_C,\bold{m}^{{\mathbb{G}}})$ implies that $D_{d,{\mathcal K}_C,\bold{m}^{{\mathbb{G}}}}\cdot \tilde{E}_i^{{\mathcal K}_C}\geq 0$ for all $i=0,1,\ldots,n$ and, therefore, we can apply Theorem \[gordo\] to deduce the inequality $D_{d,{\mathcal K}_C,\bold{m}^{{\mathbb{G}}}}\cdot \tilde{H}^{{\mathcal K}_C}\leq
-1-\epsilon({\mathbb{G}}, \bold{m}^{{\mathbb{G}}})$. But, taking into account that $\tilde{H}^{{\mathcal K}_C}$ is a $(-1)$-curve, this is equivalent to the condition $d\leq m_0^{{\mathbb{G}}}+m_1^{{\mathbb{G}}}-1-\epsilon({\mathbb{G}},
\bold{m}^{{\mathbb{G}}})$, a contradiction.
For the last assertion, it only remains to prove that, assuming the consistency of $({\mathbb{G}},\bold{m})$ and the inequality $\rho_1({\mathbb{G}},
\bold{m})\geq 1$, the non-speciality of the linear system ${\mathcal
L}_d(\bold{m})$ implies the inequality $d\geq m_1+m_2-1$. We will reason by contradiction. So, assume that ${\mathcal L}_d(\bold{m})$ is non-special and $d\leq m_1+m_2-2$. Again taking into account that $\tilde{H}^{{\mathcal K}_C}$ is a $(-1)$-curve, we have $D_{d,{\mathcal K}_C,\bold{m}}\cdot \tilde{H}^{{\mathcal K}_C}\leq -2$. If $N$ denotes the line of ${\mathbb{P}}^2$ joining the two first points of the configuration ${\mathcal K}_0(n)$, then $$D_{d,{\mathcal K}_0(n),\bold{m}}\cdot \tilde{N}^{{\mathcal K}_0(n)}=d-m_0-m_1=D_{d,{\mathcal K}_C,\bold{m}}\cdot
\tilde{H}^{{\mathcal K}_C}\leq -2,$$ which is a contradiction with the non-speciality of ${\mathcal L}_d(\bold{m})$.[$\ \ \square$]{}\
As a consequence of Proposition \[sss\], there is a bijection between the set of ordered sequences $(n_1,n_2,\ldots,n_r)\in
({\mathbb{N}}\setminus \{0,1\})^r$ (with $r\in {\mathbb{N}}\setminus \{0\}$) and the set of proximity graphs of the form ${\mathbb{G}}({\mathcal K}_C)$, $C$ being a curve of AMS type. Taking this fact into account, we obtain the following reformulation of Theorem \[super\], expressed in purely arithmetical terms:
\[super2\] Let $d$ be a positive integer and $\bold{m}=(m_0,m_1,\ldots,m_n)$ a system of multiplicities with $n\geq 1$. Assume that there exist integers $n_1,n_2,\ldots,n_r\geq 2$ such that the weighted proximity graph $({\mathbb{G}}={\mathbb{G}}(n_1,n_2,\ldots,n_r)^+,\bold{m})$ is almost consistent. Then, the linear system ${\mathcal L}_d(\bold{m})$ is non-special whenever $d\geq
m_0^{{\mathbb{G}}}+m_1^{{\mathbb{G}}}-\epsilon({\mathbb{G}},\bold{m})$. Moreover, if ${\mathcal
L}_d(\bold{m})$ is not empty, $({\mathbb{G}},\bold{m})$ is consistent and $\rho_1({\mathbb{G}},\bold{m})\geq 1$, then the following equivalence holds: ${\mathcal L}_d(\bold{m})$ is non-special if and only if $d\geq
m_0+m_1-1$.
Next, we will give two examples by applying Corollary \[super2\] to two specific sequences of integers.
\[ex1\]
Let $n\geq 2$ be an integer and consider the proximity graph ${\mathbb{G}}={\mathbb{G}}(t+1)^+$, where $t:=\lfloor n/2 \rfloor$. The number of points $s+1$ of whichever configuration whose proximity graph be ${\mathbb{G}}$ is $n+1$ (resp., $n+2$) if $n$ is odd (resp., if $n$ is even) and, moreover, the complete list of proximity relations among the points of such a configuration $(p_0,p_1,\ldots,p_s)$ are the following: $p_i\rightarrow p_{i-1}$ for all $i=1,2,\ldots,s$ and $p_j\rightarrow p_0$ for all $j=2,3,\ldots,t$. Then, applying Corollary \[super2\] to the graph ${\mathbb{G}}$, we get the following result:
Let $\bold{m}=(m_0,m_1,\ldots,m_n)$ be a system of multiplicities such that $m_1\geq m_2\geq \ldots \geq m_n$ and $m_0\geq
\sum_{i=1}^t m_i$. A linear system ${\mathcal L}_d(\bold{m})$ is non-special whenever $d\geq m_0+m_1-\epsilon$, where $\epsilon=\min\{1,m_1-m_2\}$. Moreover, if ${\mathcal
L}_d(\bold{m})$ is not empty and $m_1>m_2$, then ${\mathcal
L}_d(\bold{m})$ is non-special if and only if $d\geq m_0+m_1-1$.
\[ex2\]
Let $k\geq 2$ be an integer. If $(p_0,p_1,\ldots,p_s)$ is whichever configuration whose proximity graph is ${\mathbb{G}}(2,2,\ldots,2)^+$ (where the number $2$ appears $k$ times), one gets that $s=3k$ and the proximity relations among the points are the following: $p_i\rightarrow p_{i-1}$ for all $i=1,2,\ldots,s$ and $p_{3j+1}\rightarrow p_{3j-1}$ for all $j=1,2,\ldots,k-1$. Applying again Corollary \[super2\] to this graph one gets the following result:
Let $\bold{m}=(m_0,m_1,\ldots,m_{3k})$ be a system of multiplicities such that $m_0\geq m_1\geq \ldots \geq m_{3k}$ and $m_{3i-1}\geq
m_{3i}+m_{3i+1}$ for all $i=1,2,\ldots,k-1$. Then, a linear system ${\mathcal L}_d(\bold{m})$ is non-special if $d\geq m_0+m_1$. If, in addition, ${\mathcal L}_d(\bold{m})$ is not empty and $m_1>m_2$, then ${\mathcal L}_d(\bold{m})$ is non-special if and only if $d\geq
m_0+m_1-1$.
The following direct consequence of Theorem \[super\] exhibits a wide range of cases in which the Harbourne-Hirschowitz Conjecture is satisfied.
\[conj\] Let $\bold{m}=(m_0,m_1,\ldots,m_n)$ be a system of multiplicities such that $({\mathcal K}_C, \bold{m})$ is consistent and $\rho_1({\mathcal K}_C,\bold{m})\geq 1$, $C$ being a curve of AMS type. Denote ${\mathcal K}_0(n)=(p_0,p_1,\ldots,p_n)$ and set $N$ the line joining $p_0$ and $p_1$. If a linear system of the form ${\mathcal
L}_d(\bold{m})$ is special, then $D_{d,{\mathcal K}_0(n),\bold{m}}\cdot
\tilde{N}^{{\mathcal K}_0(n)}\leq -2$.
\[nequivalence\] [For a fixed positive integer $n$, let us denote by ${\mathcal
S}_n$ the set of proximity graphs of the form ${\mathbb{G}}({\mathcal K}_C)$, where $C$ is a curve of AMS type, whose number of vertices is greater than or equal to $n+1$. Each proximity graph ${\mathbb{G}}={\mathbb{G}}({\mathcal K}_C)\in {\mathcal S}_n$ provides, by Corollary \[conj\], an infinite family of systems of multiplicities $\bold{m}=(m_i)_{i=0}^n$ for which the Harbourne-Hirschowitz Conjecture is true. This family is given by the non-negative integer solutions of the following system of linear inequalities in $m_0,m_1,\ldots,m_n$: $$\left\{ \begin{array}{l}
m_i - \sum\limits_{j\leq n;\; p_j \to p_i } {m_j \ge 0,\quad i = 0,2,3,4, \ldots ,n} \\
m_1 - \sum\limits_{j\leq n;\; p_j \to p_1 } {m_j \ge 1} \\
\end{array} \right.$$ where ${\mathcal K}_C=(p_0,p_1,\ldots,p_n)$. Two graphs ${\mathbb{G}}$ and ${\mathbb{G}}'$ in ${\mathcal S}_n$ give rise to the same system of inequalities if the proximity relations involving the first $n+1$ vertices are the same for both graphs; in this case, we will say that ${\mathbb{G}}$ and ${\mathbb{G}}'$ are [*$n$-equivalent*]{}. Taking into account Proposition \[sss\], a complete system of representants of the quotient set of ${\mathcal S}_n$ by this equivalence relation is given by the proximity graphs ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$ (with $n_i\geq 2$ for all $i$) such that either $r=1$ and $(n+1)/2\leq n_1\leq n+1$, or $r>1$, $t:=n-2\sum_{i=1}^{r-1} n_i+r>0$ and $t/2\leq n_r\leq t$. Thus, for a fixed positive integer $n$, the set of distinct systems of linear inequalities in $n+1$ variables provided by Corollary \[conj\] (each of them satisfying that the Harbourne-Hirschowitz Conjecture is true for the solutions) is finite. In fact, they are in one-to-one correspondence with the $n$-equivalence classes. ]{}
[We define the [*regularity*]{} of a system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_n)$ as the minimum integer $d$ such that $({\mathcal K}_0(n),\bold{m})$ imposes independent conditions to the curves of degree $d$, and we will denote it by $\tau(\bold{m})$. ]{}
The following result is another consequence of Theorem \[super\] and it allows to compute the exact value of the regularity of a wide range of systems of multiplicities:
\[regu\] Let $\bold{m}=(m_0,m_1,\ldots,m_n)$ be a system of multiplicities, with $n\geq 1$. Assume the existence of a curve $C$ of AMS type such that $({\mathcal K}_C,\bold{m})$ is consistent and $\rho_1({\mathcal K}_C,\bold{m})\geq 1$. Then, $\tau(\bold{m})=m_0+m_1-1$ if $(m_0+m_1-1)(m_0+m_1+2)-\sum_{i=0}^n m_i(m_i+1)\geq -2$, and $\tau(\bold{m})=m_0+m_1$ otherwise.
[*Proof*]{}. First, we will show that $[D_{d,{\mathcal K}_0(n),\bold{m}}]$ is an effective class, where $d=m_0+m_1$. Indeed, the class $[D_{d,{\mathcal K}_C,\bold{m}}]$ is numerically effective, due to the hypotheses and the fact that $\tilde{H}^{{\mathcal K}_C}$ is a $(-1)$-curve. So, it is an effective class of ${{\rm Pic}}(X_C)$ by Proposition \[sup\], and $h^1({\mathcal
L}_d({\mathcal K}_C,\bold{m}))=h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}}))=0$, by Theorem \[gordo\]. Therefore, we get $${{\rm edim\;}}{\mathcal L}_d(\bold{m})=\dim {\mathcal L}_d({\mathcal K}_C,\bold{m})\geq
0.$$ Hence, ${\mathcal L}_d(\bold{m})$ is non-empty and then, by Theorem \[super\], $\tau(\bold{m})\leq d$.
If $(d-1)(d+2)-\sum_{i=0}^n m_i(m_i+1)\geq -2$, then the inequality ${{\rm edim\;}}{\mathcal L}_{d-1}(\bold{m})\geq -1$ holds. So, the superabundance $h^1({\mathcal L}_{d-1}(\bold{m}))$ must be zero, in virtue of Theorem \[super\], and then $\tau(\bold{m})=d-1$ by [@hir 2].
Finally, if $(d-1)(d+2)-\sum_{i=0}^n m_i(m_i+1)< -2$, then $$\dim
{\mathcal L}_{d-1}(\bold{m})-h^1({\mathcal L}_{d-1}(\bold{m}))<-1$$ and this implies that the superabundace $h^1({\mathcal
L}_{d-1}(\bold{m}))$ is positive. Therefore, in this case, $\tau(\bold{m})=d$.[$\ \ \square$]{}
Bounding the regularity {#s43}
-----------------------
In [@roe3] it is described an algorithm, based on the unloading method, which provides an upper bound of the regularity of whichever system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_n)$. Although only the case of homogeneous multiplicities is explicitly treated (i.e., $m_0=m_1=\cdots=m_n$) this algorithm can be adapted without difficulty to the case of arbitrary multiplicities. In this section we introduce a generalization of this algorithm, based on our results in Section \[infinity\].
Let $\bold{m}=(m_0,m_1,\ldots,m_n)$ be a sequence of multiplicities (with $n\geq 1$) such that $m_0\geq m_1\geq \cdots
\geq m_{n}$. Take a sequence of integers $(n_1, n_2, \ldots,n_r)$ such that $n_i\geq 2$ for all $i=1,2,\ldots,r$ and $n+1$ is not greater than the number of vertices of the graph ${\mathbb{G}}:={\mathbb{G}}(n_1,
n_2,\ldots,n_r)^+$. By completing with zero multiplicities, if it is necessary, we will assume that this number of vertices coincides with $n+1$. Denote by ${\mathbb{G}}_i$ the proximity graph obtained from ${\mathbb{G}}$ by deleting all curved-dotted edges involving some vertex with label greater than $i$. Let $(i_1,
i_2,\ldots,i_w)$ be an increasing sequence of integers such that ${\mathbb{G}}_{i_1}, {\mathbb{G}}_{i_2},\ldots, {\mathbb{G}}_{i_w}$ are the distinct elements of the set $\{{\mathbb{G}}_i \mid 1\leq i < n\}$ and let $a$ be the maximum integer such that $0\leq a< n$ and, if $(p_0,p_1,\ldots,p_n)$ denotes a configuration with proximity graph ${\mathbb{G}}$, the cardinality of the set $\{ j\mid a\leq j\leq n, \;\;p_j\rightarrow
p_a\;\;\mbox{and}\;\; m_j>0\}$ is greater than $1$ (if that integer does not exist, we will take $a=0$).
[Let ${\mathbb{G}}$ be the proximity graph ${\mathbb{G}}(s)^+$, where $s>1$ is an integer. The above described sequence of graphs ${\mathbb{G}}_{i_1},
{\mathbb{G}}_{i_2},\ldots, {\mathbb{G}}_{i_w}$ is, in this case, the sequence ${\mathbb{G}}^1, {\mathbb{G}}^2,\ldots,{\mathbb{G}}^{s-1}$ where ${\mathbb{G}}^1$ denotes the proximity graph of a chain of $2s$ free points and, for each $k=2,3,\ldots,s-1$, ${\mathbb{G}}^k$ stands for the graph depicted in Figure 3 of page taking $n=2s-1$; the integer $a$ is $0$.]{}
Set $\bold{m}_1:=\bold{m}$, which is consistent for ${\mathbb{G}}_{i_1}$, and define recursively the systems of multiplicities $\bold{m}_2,\bold{m}_3,\ldots,\bold{m}_w$ as follows. Suppose we have defined $\bold{m}_k$ and perform the following two-steps algorithm applied to $\bold{v}:=\bold{m}_k$, which will give rise to $\bold{m}_{k+1}$:\
[*Step 1*]{}. If $({\mathbb{G}}_{i_{k+1}},\bold{v})$ is consistent, then define $\bold{m}_{k+1}:=\bold{v}$. Otherwise, there exists a unique $j$ such that the excess $\rho_j({\mathbb{G}}_{i_{k+1}},\bold{v})$ is negative. In this case, if $j=a$ and this excess equals $-1$, define also $\bold{m}_{k+1}:=\bold{v}$; else, perform an unloading step to $({\mathbb{G}}_{i_k},\bold{v})$ at the vertex which corresponds to that excess, replace $\bold{v}$ by the obtained new system of multiplicities and go to Step 2.\
[*Step 2*]{}. Replace $\bold{v}$ by $\bold{v}^{{\mathbb{G}}_{i_k}}$ and return to Step 1.\
Once we have computed $\bold{m}_w$, we must consider the system of multiplicities $\bold{m}':=\bold{m}_w^{{\mathbb{G}}}=(m_0',m_1',\ldots,
m_n')$. Let $C$ be a curve having one place at infinity whose associated proximity graph is ${\mathbb{G}}_{i_w}={\mathbb{G}}$. Notice that $h^1(X_C,{{\mathcal O}}_{X_C}(D_{m_0'+m_1', {\mathcal K}_C, \bold{m}'}))=0$ (by Theorem \[gordo\]), since $D_{m_0'+m_1', {\mathcal K}_C, \bold{m}'}$ is a numerically effective divisor of $X_C$. We will compute the successive dimensions $h^1(X_C,{{\mathcal O}}_{X_C}(D_{m_0'+m_1'-j, {\mathcal K}_C,
\bold{m}'}))$ for $j=1,2,\ldots$ (using the process described in Remark \[dim\]) until finding the minimum $j$ such that the mentioned dimension is positive. Finally, we will define $\beta(\bold{m}):=m_0'+m_1'-j+1$. Note that this process is independent of the chosen curve $C$; in fact, it only depends on the proximity graph ${\mathbb{G}}$.
Now, we will justify that the obtained value $\beta({\bold m})$ is an upper bound of the regularity $\tau(\bold{m})$. We start with a lemma whose proof is an adaptation of that of [@roe3 Lem. 2.1] and we will omit it.
\[a\] Let $d$ be a positive integer, ${\mathcal K}=\{p_0,p_1,\ldots, p_n\}$ a configuration and $\bold{m}=(m_i)_{i=0}^n$ a system of multiplicities. Let $i\in \{0,1,\ldots,n\}$ be such that $\rho_i({\mathcal K},\bold{m})\geq -1$ and let $\bold{m}'=(m_0',m_1',\ldots,m_n')$ be the sequence of multiplicities obtained from $\bold{m}$ by performing an unloading step at the point $p_i$. Then, $h^1({\mathcal L}_d({\mathcal K},\bold{m}))=0$ whenever $h^1 ({\mathcal L}_d({\mathcal K},\bold{m}'))=0$.
\[just\] Let $\bold{m}=(m_0,m_1,\ldots,m_n)$ be a sequence of multiplicities, let $(n_1,n_2,\ldots,n_r)$ be a sequence of integers such that $n_i\geq 2$ for all $i=1,2,\ldots,r$ and set $\beta(\bold{m})$ defined as above. Then, $\beta(\bold{m})$ is an upper bound of $\tau(\bold{m})$.
[*Proof*]{}. Recall the notations of Section \[s42\]. Taking into account that the proximity graphs ${\mathbb{G}}_{i_1},{\mathbb{G}}_{i_2},\ldots,{\mathbb{G}}_{i_w}$ are associated with chain configurations and the matrix ${\bf P}_{{\mathbb{G}}_{i_{k-1}}}^{-1}\cdot
{\bf P}_{{\mathbb{G}}_{i_k}}$ has no negative entries for each $k=2,3,\ldots,w$, it can be deduced, from [@roe], the existence of a chain of inclusions $$\label{seq}
U({\mathbb{G}})=U({\mathbb{G}}_{i_w})\subseteq \overline{U({\mathbb{G}}_{i_{w-1}})}\subseteq
\ldots \subseteq \overline{U({\mathbb{G}}_{i_1})}.$$
Let $d= \beta(\bold{m})$ and, for each system of multiplicities $\bold{v}$, set $h^1(d,{\mathbb{G}}_{i_k},\bold{v})$ the minimum of the superabundances $h^1({\mathcal L}_d({\mathcal K},\bold{v}))$ when ${\mathcal K}$ varies in $U({\mathbb{G}}_{i_k})$.
Take a plane curve $C$ of AMS type such that ${\mathbb{G}}({\mathcal
K}_C)={\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$. From the above description of the algorithm, it follows that $$h^1({\mathcal
L}_d({\mathcal K}_C,\bold{m}'))=h^1(X_C,{{\mathcal O}}_{X_C}(D_{d,{\mathcal K}_C,\bold{m}'}))=0$$ and, since the weighted proximity graph $({\mathbb{G}},\bold{m}')$ is obtained from $({\mathbb{G}},\bold{m}_w)$ by tame unloading steps, we get that the integer $h^1({\mathcal L}_d({\mathcal K}_C,\bold{m}_w))$ vanishes and, hence, $h^1(d,{\mathbb{G}}_{i_w},\bold {m}_w)=0$.
Finally, for $2\leq k \leq w$, we will show that the vanishing of $h^1(d,{\mathbb{G}}_{i_k},\bold{m}_k)$ implies that of $h^1(d,{\mathbb{G}}_{i_{k-1}},\bold{m}_{k-1})$. In order to prove this assertion observe firstly that, if we assume that $h^1(d,{\mathbb{G}}_{i_k},\bold{m}_k)=0$, then $h^1(d,{\mathbb{G}}_{i_{k-1}},\bold{m}_k)=0$ by (\[seq\]) and the upper-semicontinuity of the functions given in (\[semic\]). Choose a configuration ${\mathcal K}\in U({\mathbb{G}}_{i_{k-1}})$ such that $h^1({\mathcal
L}_d({\mathcal K}, \bold{m}_k))=0$. It is not hard to see that the unloading procedure of the Step 2 of the algorithm to obtain the sequence $\bold{m}_1,\bold{m}_2,\ldots,\bold{m}_w$ can be performed by means of tame unloading steps. From this fact and Lemma \[a\], the equality $h^1({\mathcal L}_d({\mathcal K}, \bold{m}_{k-1}))=0$ is obtained and, therefore, $h^1(d,{\mathbb{G}}_{i_{k-1}},\bold{m}_{k-1})=0$.
Now, it follows, by induction, that $h^1(d,{\mathbb{G}}_{i_1},\bold{m})=0$. Finally, using again semicontinuity and taking into account the density of $U({\mathbb{G}}({\mathcal K}_0(n)))$ in $Y_n$, we get $h^1({{\mathcal L}}_d(\bold{m}))=0$. Hence, $d$ is an upper bound of $\tau(\bold{m})$.[$\ \ \square$]{}\
We conclude the paper with some remarks on the above described algorithmic bound.
First observe that, given a system of multiplicities $\bold{m}$, there is a bound $\beta(\bold{m})$ for each election of a proximity graph ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$ with, at least, $n+1$ vertices (that is, $2\sum_{i=1}^r n_i-r\geq n$). It is clear that $n$-equivalent proximity graphs give rise to the same bound (see Remark \[nequivalence\]). Thus, one can apply the algorithm to all the graphs ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$ such that either $r=1$ and $(n+1)/2\leq n_1\leq n+1$, or $r>1$, $t:=n-2\sum_{i=1}^{r-1}
n_i+r>0$ and $t/2\leq n_r\leq t$, and then pick the best bound.
We will show that the algorithm given by Roé in [@roe3] can be obtained as a particular case of the one we have described (essentially, it corresponds to a specific type of proximity graph ${\mathbb{G}}(n_1,n_2,\ldots,n_r)^+$). To apply his algorithm to a system of multiplicities $\bold{m}=(m_0,m_1,\ldots,m_n)$, he uses successive specializations, starting from a configuration of $n+1$ general points of the plane and following with configurations corresponding to the sequence of proximity graphs ${\mathbb{G}}^1, \ldots, {\mathbb{G}}^n$ where, for each $k=1,2,\ldots,n$, ${\mathbb{G}}^k$ is the one shown in Figure 3 (assuming that ${\mathbb{G}}^1$ has no curved-dotted edge). The knowledge of the dimensions of all complete linear systems on the surfaces obtained by blowing-up at the points of whichever configuration whose associated proximity graph is ${\mathbb{G}}^n$ allows him to deduce, using a similar reasoning to the one explained in the algorithm we present here (but adapted to the above mentioned specific sequence of specializations), an upper bound of the regularity of $\bold{m}$. From this explanation, it is easy to deduce that applying Roé’s algorithm to $\bold{m}$ is equivalent to computing our bound $\beta(\bold{m})$ taking the graph ${\mathbb{G}}(n+1)^+$, adding previously to $\bold{m}$ the suitable number of zeros. This proximity graph corresponds, for instance, to the curve whose equation in projective coordinates $(X:Y:Z)$ is $X Z^n+Y^{n+1}=0$, $Z=0$ being the line of infinity.
\[figure2\]
(25,60) (15,5)(5,17.5)(15,30)
(15,5) (15,5)[(0,1)[5]{}]{} (15,10) (15,10)[(0,1)[5]{}]{} (15,15) (15,15)[(0,1)[2.5]{}]{}
(15,30) (15,30)[(0,1)[2.5]{}]{} (15,30)[(0,-1)[2.5]{}]{}
(15,45)[(0,-1)[2.5]{}]{}
(15,45) (15,45)[(0,1)[5]{}]{} (15,50) (15,50)[(0,1)[5]{}]{} (15,55)
(14.7,21)[$\vdots$]{}
(14.7,36)[$\vdots$]{}
(16,5)[ 0]{}(16,10)[ 1]{}(16,15)[ 2]{}(16,45)[ $n-2$]{}(16,50)[ $n-1$]{}(17,55)[$n$]{} (17,30)[$k$]{}
(15,-2)[${\mathbb{G}}^k$]{}
Note that, for each integer $n\geq 2$, the proximity graph ${\mathbb{G}}(\lfloor n/2 \rfloor+1)^+$ (considered in Example \[ex1\]) is the graph ${\mathbb{G}}^k$ of Figure 3 for $k=\lfloor n/2 \rfloor$, if $n$ is odd, and the one obtained from the same graph adding a new vertex corresponding to a free point at the top, if $n$ is even. The fact that this is one of the intermediate proximity graphs which appear in the sequence of specializations used in [@roe3] and easy reasonings concerning semicontinuity imply that our bounds $\beta(\bold{m})$, taking the above proximity graph, are either equal or lower than those obtained from [@roe3]. For homogeneous systems of multiplicities $\bold{m}=(m,m,\ldots,m)$ and for a fixed value of $n$, examples show that the difference between both bounds increases when the multiplicity does so. For instance, this is the behavior for $n+1=1000$ and $m$ taking values between 1 and 100. In fact, when $m\leq 38$ the two bounds coincide, $\beta(\bold{m})$ is sometimes better when $39\leq m\leq 68$ (in which case, the difference is 1) and it is always better when $69\leq m\leq 100$ (the difference is 1 in all cases except for $m=98$, where it equals 2). Also, for $m=500$ (resp., $m=800$) (resp., $m=1200$), $\beta(\bold{m})=16014$ (resp., $\beta(\bold{m})=25617$) (resp., $\beta(\bold{m})=38417$) and Roé’s bound is $16021$ (resp., 25629) (resp., $38436$). However, [@hr] gives better values in all the checked cases where our bound is less than Roé’s one.
For quasihomogeneous systems of multiplicities, there are cases in which our bound seems to improve the existing ones (as far as the author knows). As an example, consider the system of multiplicities $\bold{m}=(4000,1000_{19})$ (where the subindex is the number of occurrences). Taking the graph ${\mathbb{G}}(10)$, it is obtained the bound $\beta(\bold{m})=6009$. The Harbourne-Hirschowitz Conjecture predicts that the regularity of $\bold{m}$ is 5917 and the bounds provided in [@hir], [@gim], [@cat], [@roe3], [@bal] and [@xu] are 8367, 8000, 8000, 6183, 11140 and 6238, respectively. Also, the bound 6015 is obtained by an algorithm based on the reduction method described in [@dum1] and Cremona transformations (it has been computed by using the computer program provided in [@dumweb]), and the bound 7667 is obtained by using the algorithm given in [@harbc] for computing the dimension of line bundles on an smooth rational surface $X$ with anticanonical bundle having an irreducible and reduced global section $D$, with the further assumption that the morphism ${{\rm Pic}}(X)\rightarrow
{{\rm Pic}}(D)$ induced by the inclusion $D\subseteq X$ has trivial kernel. The bound provided by the algorithm of [@hr], using the parameters $r=9$ and $d=2$, is 7667 (this algorithmic bound depends on the choice of two parameters, but it is not clear how to obtain the optimal values).
Now, consider the family of systems of multiplicities $\bold{m}(m):=(m,1000_{19})$ for $m\geq 1000$. By applying Corollary \[super2\] to the graph ${\mathbb{G}}(10)$ (see also Example \[ex1\]) it can be deduced that $\tau(\bold{m}(m))=m+1000$ when $m\geq 9000$. Computing the above mentioned bounds of the regularity for the remaining values of $m$ it holds that, when either $m\in \{1619, 1622, 1623\}$ or $1625\leq m\leq 7765$, the value $\beta(\bold{m}(m))$ (taking the graph ${\mathbb{G}}(10)$) is less than all the non-parametric bounds given in [@hir], [@gim], [@cat], [@roe3], [@bal] and [@xu]. When either $3935\leq m\leq 3939$, $3944\leq m\leq 4081$, $4083\leq m\leq 4085$ or $m\in \{3941,3942,4087,4089,4090,4092\}$ it holds that the bound $\beta(\bold{m}(m))$ is also better than the one provided in [@dum1] and [@dumweb]; moreover, in these cases, we have not found any pair of parameters $(r,d)$ for which the bound given in [@hr] improves $\beta(\bold{m}(m))$. It is worth adding that, by looking at systems of multiplicities of the type $(m,h_{19})$ with $h\in \{1100,1200,1300,1400,1500\}$, we have observed an increasing tendency (when $h$ grows) on the number of values of $m$ for which $\beta(m,h_{19})$ seems to be the best bound.
Although, in order to establish comparisons, it is natural to look at homogeneous and quasihomogeneous cases, our algorithm can be applied to arbitrary systems of multiplicities. Finally we notice that, when the system of multiplicities $\bold{m}$ is either homogeneous or quasihomogeneous, examples suggest that the bound $\beta(\bold{m})$ is better when the graph ${\mathbb{G}}(\lfloor n/2
\rfloor+1)^+$ is taken (where $n+1$ is the length of $\bold{m}$).
[The results proved in Section \[infinity\] and the explanations given in the current section suggest that the algorithm provided in [@roe2] for giving a lower bound of the least degree $d$ such that a linear system ${\mathcal L}_d(\bold{m})$ is not empty can also be generalized. However, we have not found evidences of any significant improvement of this generalization with respect to the existing bounds. ]{}
[99]{} S. S. Abhyankar, [*Lectures on expansion techniques in Algebraic Geometry*]{}, Tata Institute of Fundamental Research Lectures on Mathematics and Physics [**57**]{}, Tata Institute of Fundamental Research, Bombay (1977). S. S. Abhyankar, T. T. Moh, Newton-Puiseux expansion and generalized Tschirnhausen transformation, [*J. Reine Angew. Math.*]{} [**260**]{} (1973), 47—83 and [**261**]{} (1973), 29—54. S. S. Abhyankar, T. T. Moh, Embeddings of the line in the plane, [*J. Reine Angew. Math.*]{} [**276**]{} (1975), 148—166. J. Alexander, A. Hirschowitz, An asymptotic vanishing theorm for generic unions of multiple points. [*Invent. Math.*]{} [**140**]{}, no. 2 (2000), 303—325. E. Arbarello, M. Cornalba, Footnotes to a paper of B. Segre, [*Math. Ann.*]{}, [**256**]{} (1981), 341—362. E. Ballico, Curves of minimal degree with prescribed singularities, [*Illinois J. Math.*]{} [**45**]{} (1999), 672—676.
E. Ballico, L. Chiantini, Nodal curves and postulation of generic fat points on ${\mathbb{P}}^2$, [*Arch. Math.*]{} [**71**]{} (1998) no. 6, 501—504. F. Bunke, C. Lossen, An $H^1$-vanishing theorem for generic fat points in ${\mathbb{P}}^2$, Preprint Univ. Kaiserslautern (2002). A. Campillo, [*Algebroid curves in positive characteristic*]{}, Lecture Notes in Math. [**813**]{}, Springer (1980). A. Campillo, O. Piltant, A. J. Reguera, Cones of curves and of line bundles on surfaces associated with curves having one place at infinity [*Proc. London Math. Soc.*]{} [**84**]{} (2002), 559—580.
E. Casas-Alvero, [*Singularities of plane curves*]{}, London Math. Soc. Lecture Note Ser. [**276**]{}, Cambridge University Press (2000).
G. Castelnuovo, Ricerche generali sopra i sistemi lineari di curve piane, [*Mem. Accad. Sci. Torino, II*]{} [**42**]{} (1891). M. V. Catalisano, Linear systems of plane curves through fixed fat points of ${\mathbb{P}}^2$, [*J. Algebra*]{} [**142**]{} (1991), no. 1, 81—100. C. Ciliberto, Geometric aspects of polynomial interpolation in more variables and of Waring’s Problem, European Congress of Mathematics, Vol. I (Barcelona, 2000), 289–316, [*Progr. Math.*]{} [**201**]{}, Birkhäuser, Basel (2001). C. Ciliberto, F. Cioffi, R. Miranda, F. Orecchia, Bivariate Hermite interpolation and linear systems of plane curves with base fat points, Lecture Notes Series on Computing, vol. 10 (2003), 87—102, World Scientific. C. Ciliberto, R. Miranda, Degenerations of planar linear systems, [*J. Reine Angew. Math.*]{} [**501**]{} (1998), 191—220. C. Ciliberto, R. Miranda, Linear systems of plane curves with base points of equal multiplicity, [*Trans. Amer. Math. Soc.*]{} [**352**]{} (2000), no. 9, 4037—4050. M. Dumnicki, W. Jarnicki, New effective bounds on the dimension of a linear system in ${\mathbb{P}}^2$, math.AG/0505183. M. Dumnicki, Reduction method for linear systems of plane curves with base fat points, arXiv: math.AG/0606716. M. Dumnicki, `http://gamma.im.uj.edu.pl/dumnicki/interpol.htm`, June 1 (2006). L. Évain, La fonction de Hilbert de la réunion de $4\sp h$ gros points génériques de $ {\mathbb{P}}^2$ de même multiplicité, [*J. Algebraic Geom.*]{} [**8**]{} (1999), no. 4, 787—796. J. Fernández de Bobadilla, A new geometric proof of Jung’s theorem on factorisation of automorphisms of ${\mathbb{C}}^2$, [*Proc. Amer. Math. Soc.*]{} [**133**]{} (2005), 15—19. J. Fernández de Bobadilla, Moduli spaces of polynomials in two variables, [*Mem. Amer. Math. Soc.*]{} [**173**]{} (2005), no. 817. M. Fujimoto, M. Suzuki, Construction of affine plane curves with one place at infinity, [*Osaka J. Math.*]{}, [**39**]{} (2002), no. 4, 1005—1027. J. W. Gaddum, Linear inequalities and quadratic forms, [*Pacific J. Math.*]{} [**8**]{} (1958), 411—414.
C. Galindo, F. Monserrat, The cone of curves associated to a plane configuration, [*Comment. Math. Helv.*]{} [**80**]{} (2005), 75—93. C. Galindo, F. Monserrat, On the cone of curves and of line bundles of a rational surface, [*Internat. J. Math.*]{} [**15**]{} (2004), no. 4, 393—407. C. Galindo, F. Monserrat, The total coordinate ring of a smooth projective surface, [*J. Algebra*]{} [**284**]{} (2005), 91—101. A. Gimigliano, [*On linear systems of plane curves*]{}, Thesis, Queen’s University, Kingston (1987). A. Gimigliano, Regularity of linear systems of plane curves, [*J. Algebra*]{} [**124**]{} (1989), 447—460.
B. Harbourne, Complete linear systems on rational surfaces, [*Trans. Amer. Math. Soc.*]{} [**289**]{} (1985), no. 1, 213—226. B. Harbourne, The geometry of rational surfaces and Hilbert functions of points in the plane, [*Can. Math. Soc. Conf. Proc.*]{} [**6**]{} (1986), 95—111. B. Harbourne, Points in good position in ${\mathbb{P}}^2$, Zero-dimensional schemes (Ravello, 1992), 213—229, de Gruyter, Berlin (1994). B. Harbourne, Anticanonical rational surfaces, [*Trans. Amer. Math. Soc.*]{} [**349**]{} (1997), no. 3, 1191—1208.
B. Harbourne, Iterated blow-ups and moduli for rational surfaces, Algebraic geometry (Sundance, UT, 1986), 101—117, Lecture Notes in Math. [**1311**]{}, Springer, Berlin (1988). B. Harbourne, Problems and progress: a survey on fat points in ${\mathbb{P}}2$, Zero-dimensional schemes and applications (Naples, 2000), Queen’s Papers in Pure and Appl. Math. [**123**]{} (2002), 85–132. B. Harbourne, S. Holay, S. Fitchett, Resolutions of ideals of quasiuniform fat point subschemes of ${\mathbb{P}}^2$, [*Trans. Amer. Math. Soc.*]{} [**355**]{} (2003), no. 2, 593—608. B. Harbourne, J. Roé, Linear systems with multiple base points in ${\mathbb{P}}^2$, [*Adv. Geom.*]{} [**4**]{} (2004), 41—59. R. Hartshorne, [*Algebraic geometry*]{}, Graduate Texts in Mathematics, vol. 52, Springer (1977). A. Hirschowitz, La méthode d’Horace pour l’interpolation à plusieurs variables, [*Manuscripta Math.*]{} [**50**]{} (1985), 337—388. A. Hirschowitz, Une conjecture pour la cohomologie des diviseurs sur les surfaces rationnelles génériques, [*J. Reine Angew. Math.*]{} [**397**]{} (1989), 208—213. S. Kleiman, [*The enumerative theory of singularities*]{}, Real and Complex Singularities, Oslo 1976 (P. Holm, ed.), Sijthoff & Noordhoof (1977), 297–396. S. Kleiman, Multiple point formulas I: Iteration, [*Acta Math.*]{} [**147**]{} (1981), 13—49. A. Laface, Quasi-homogeneous linear systems on ${\mathbb{P}}^2$ with base points of multiplicity 5, [*Canad. J. Math.*]{} [**55**]{} (2003), no. 3, 561—575. T. Mignon, Systèmes linéaires de courbes planes à singularités ordinaires imposées, [*C. R. Acad. Sci. Paris Sér. I Math.*]{} [**327**]{} (1998), no. 7, 651—654. R. Miranda, Linear systems of plane curves, [*Notices Amer. Math. Soc.*]{} [**46**]{} (1999), no. 2, 192–201. T. T. Moh, On analytic irreducibility at $\infty$ of a pencil of curves, [*Proc. Amer. Math. Soc.*]{} [**44**]{} (1974), 22—23. M. Nagata, On rational surfaces II, [*Coll. Sci. Univ. Kyoto, Ser. A, Math.*]{} [**33**]{} (1960), no. 2, 271—293.
H. Pinkham, Séminaire sur les singularités des surfaces (Demazure-Pinkham-Teissier), Course donné au Centre de Math. de l’Ecole Polytechnique (1977-1978). A. J. Reguera, Semigroups and clusters at infinity, Algebraic geometry and singularities (La Rábida, 1991), 339–374, [*Progr. Math.*]{}, [**134**]{}, Birkhäuser, Basel (1996). J. Roé, Varieties of clusters and Enriques diagrams, [*Math. Proc. Cambridge Philos. Soc.*]{}, [**137**]{} (2004), no. 1, 69—94. J. Roé, On the existence of plane curves with imposed multiple points, [*J. Pure App. Alg.*]{} [**156**]{} (2001), 115—126. J. Roé, Linear systems of plane curves with imposed multiple points, [*Illinois J. Math.*]{} [**45**]{} (2001), no. 3, 895—906. A. Sathaye, On planar curves, [*Amer. J. Math.*]{} [**99**]{} (1977), no. 5, 1105—1135. B. Segre, Alcune questioni su insiemi finiti di punti in geometria algebrica, [*Atti Convegno Intern. di Geom. Alg. di Torino*]{} (1961), 15—33. J. Seibert, The dimension of quasihomogeneous planar linear systems with multiplicity four, [*Comm. Algebra*]{} [**29**]{} (2001), no. 3, 1111—1130. M. Suzuki, Affine plane curves with one place at infinity, [*Ann. Inst. Fourier*]{} [**49**]{} (1999), no. 2, 375—404. S. Yang, Linear systems of plane curves with base points of bounded multiplicity, math.AG/0406591. G. Xu, Ample line bundles on smooth surfaces, [*J. Reine Angew. Math.*]{} [**469**]{} (1995), 199—209.
Dept. de Matemàtiques (ESTCE), UJI, Campus Riu Sec.\
12071 Castelló. SPAIN.\
[email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, there has been a growing interest in automating the process of neural architecture design, and the Differentiable Architecture Search (DARTS) method makes the process available within a few GPU days. In particular, a hyper-network called one-shot model is introduced, over which the architecture can be searched continuously with gradient descent. However, the performance of DARTS is often observed to collapse when the number of search epochs becomes large. Meanwhile, lots of “[*skip-connect*]{}s” are found in the selected architectures. In this paper, we claim that the cause of the collapse is that there exist cooperation and competition in the bi-level optimization in DARTS, where the architecture parameters and model weights are updated alternatively. Therefore, we propose a simple and effective algorithm, named “DARTS+”, to avoid the collapse and improve the original DARTS, by “early stopping” the search procedure when meeting a certain criterion. We demonstrate that the proposed early stopping criterion is effective in avoiding the collapse issue. We also conduct experiments on benchmark datasets and show the effectiveness of our DARTS+ algorithm, where DARTS+ achieves $2.32\%$ test error on CIFAR10, $14.87\%$ on CIFAR100, and $23.7\%$ on ImageNet. We further remark that the idea of “early stopping” is implicitly included in some existing DARTS variants by manually setting a small number of search epochs, while we give an [*explicit*]{} criterion for “early stopping”.'
author:
- 'Paper ID: 10453'
- |
Hanwen Liang$^{1*}$Shifeng Zhang$^{2}\thanks{Equal contribution. This work was done when the first two authors were interns at Huawei Noah's Ark Lab.}$Jiacheng Sun$^{1}\thanks{Corresponding email: [email protected]}$Xingqiu He$^1$\
**Weiran Huang$^1$Kechen Zhuang$^1$Zhenguo Li$^1$\
\
$^1$Huawei Noah’s Ark Lab $^2$TNList, Tsinghua University\
**
bibliography:
- 'mybibfile.bib'
title: '[DARTS+: Improved Differentiable Architecture Search with Early Stopping]{}'
---
Introduction
============
Neural Architecture Search (NAS) plays an important role in Automatic Machine Learning (AutoML), which has attracted lots of attention recently. The neural architectures searched by NAS have achieved the state-of-the-art results over handcrafted neural architectures in various tasks, including object classification [@regular2016neural; @pham2018efficient; @real2019regularized; @liu2018darts; @chen2019progressive; @luo2018neural], object detection [@hang2019auto-fpn; @ghiasi2019fpn], semantic segmentation [@liu2019auto], recommender systems [@joglekar2019neural], etc.
The common practice of NAS first derives an architecture search space, and then finds the best architecture in the search space with a specific search method. Early works of NAS usually adopt the REINFORCE method [@regular2016neural; @pham2018efficient] and evolutionary algorithms [@real2019regularized] for searching effective architectures. However, obtaining state-of-the-art architectures with such methods involves huge computational cost (e.g., thousands of GPU days [@regular2016neural]), because a large number of architectures need to be trained and evaluated. Recently, one-shot methods [@brock2017smash; @cai2018proxylessnas; @pham2018efficient; @liu2018darts; @xie2018snas] are proposed to reduce the search cost. Among the one-shot methods, @liu2018darts propose the Differentiable Architecture Search (DARTS) method to relax the search space to be continuous, so that one can search architectures and learn model weights directly with gradient descent. In particular, DARTS encodes the architecture search space with continuous parameters and performs searching with bi-level optimization, where the model weights and architecture parameters are optimized with training data and validation data alternatively. DARTS can reduce the search cost from thousands of GPU days to a few GPU days while keeping comparable performance. Despite the efficiency of DARTS, a severe issue underlying DARTS has been found [@chen2019progressive]. Namely, after a certain searching epochs, the number of [*skip-connect*]{}s increases dramatically in the selected architecture, which results in poor performance of the selected architecture. We call the phenomenon of performance drop after a certain number of epochs the “collapse” of DARTS. To tackle such issue, @chen2019progressive propose a search space regularization in their P-DARTS, where dropout [@srivastava2014dropout] is used to alleviate the dominance of [*skip-connect*]{}s during the search procedure, and the number of [*skip-connect*]{}s is manually controlled after the search procedure. However, this approach involves more hyper-parameters like dropout rate and the number of [*skip-connect*]{}s, which need to be carefully tuned by human experts. Moreover, @stamoulis2019single propose Single-Path NAS using the one-level optimization instead of the bi-level optimization in DARTS, where the architecture parameters and model weights are updated simultaneously. However, the search space of Single-Path NAS is carefully designed. If searching is done in the original search space of DARTS, one-level optimization will perform worse than bi-level ones [@liu2018darts]. In summary, the mechanism of the collapse of DARTS is still unknown and needs to be addressed.
In this paper, we first show the cause of the collapse of DARTS is that there exist [*cooperation*]{} and [*competition*]{} in the bi-level optimization in DARTS, where the architecture parameters and model weights are updated alternatively. In particular, we give an explanation of why lots of [*skip-connect*]{}s are involved in the selected architectures in DARTS and why they hurt the performance. To avoid the collapse of DARTS, we add the simple and effective “early stopping” paradigm to the original DARTS, named “DARTS+”, where the search procedure stops by a certain criterion, illustrated in Fig. \[visulization\](a). We remark that recent progresses of DARTS, including P-DARTS [@chen2019progressive], Auto-DeepLab [@liu2019auto] and PC-DARTS [@xu2019pc], also adopt the early stopping idea implicitly where fewer search epochs are manually set in their methods.
{width="7in"}
Moreover, we conduct sufficient experiments on benchmark datasets to demonstrate the effectiveness of the proposed DARTS+ algorithm. Specifically, DARTS+ achieves the state-of-the-art $2.32\%$ test error on CIFAR10 and $14.87\%$ test error on CIFAR100, while the search time is less than $0.4$ GPU days. When transferring to ImageNet, DARTS+ achieves the state-of-the-art $23.7\%$ top-1 error and impressive $22.5\%$ top-1 error if SE-Module [@hu2018squeeze] is introduced. DARTS+ is also able to search on ImageNet directly and achieves $23.9\%$ top-1 error.
In summary, our main contributions are listed as follows:
- We study the collapse issue of the DARTS method, and point out the underlying reason is the cooperation and competition in the bi-level optimization.
- We introduce an efficient “early stopping” paradigm to DARTS to avoid the collapse, and propose an effective criterion for early stopping.
- we conduct extensive experiments on benchmark datasets to demonstrate the effectiveness of the proposed algorithm, which achieves the state-of-the-art results on all of them.
Collapse of DARTS
=================
There is a severe issue underlying DARTS [@liu2018darts], that lots of [*skip-connect*]{}s tend to appear in the selected architecture when the number of searching epoch is large, making the performance poor. The phenomenon of performance drop is called the “collapse” of DARTS in our paper. In this section, we first give a quick review of the original DARTS, and then point out the collapse issue of DARTS. Moreover, we will discuss the cause of the collapse issue.
Preliminary: DARTS
------------------
The goal of DARTS is to search for a cell, which can be stacked to form a convolutional network or recursively connected to form a recurrent network. Each cell can be regraded as a directed acyclic graph (DAG) of $N$ nodes $\{x_i\}_{i=0}^{N-1}$, where each node represents a network layer. We denote the operation space as $\mathcal{O}$, and each element is a candidate operation, e.g., [*zero, skip-connect, convolution, max-pool*]{}, etc. Each edge $(i,j)$ of DAG represents the information flow from node $x_i$ to $x_j$, which consists of the candidate operations weighted by the architecture parameter $\alpha^{(i,j)}$. In particular, each edge $(i,j)$ can be formulated by a function $\bar{o}^{(i,j)}$ where $\bar{o}^{(i,j)} (x_i) = \sum_{o \in \mathcal{O}} p_o^{(i,j)} \cdot o(x_i),$ and the weight of each operation $o\in\mathcal{O}$ is a softmax of the architecture parameter $\alpha^{(i,j)}$, that is $p_o^{(i,j)} = \frac{\exp (\alpha_o^{(i,j)})}{\sum_{o' \in \mathcal{O}} \exp (\alpha_{o'}^{(i,j)})}$. An intermediate node is $x_j=\sum_{i<j} \bar{o}^{(i,j)} (x_i)$, and the output node $x_{N-1}$ is depth-wise concatenation of all the intermediate nodes excluding input nodes. The above hyper-network is called one-shot model, and we denote $w$ as the weights of the hyper-network.
For the search procedure, we denote $\mathcal{L}_{train}$ and $\mathcal{L}_{val}$ as the training and validation loss respectively. Then the architecture parameters are learned with the following bi-level optimization problem: $$\begin{aligned}
\min_\alpha\quad &\mathcal{L}_{val} (w^*(\alpha), \alpha), \\
\mathrm{s.t.}\quad &w^*(\alpha) = \arg \min_w \mathcal{L}_{train} (w, \alpha).
\end{aligned}$$
After obtaining architecture parameters $\alpha$, the final discrete architecture is derived by: 1) setting $o^{(i,j)} = \arg \max_{o \in \mathcal{O}, o\neq zero} p_o^{(i,j)}$, and 2) for each intermediate node, choosing two incoming edges with the two largest values of $\max_{o \in \mathcal{O}, o\neq zero} p_o^{(i,j)}$. More technical details can be found in the original DARTS paper [@liu2018darts].
Collapse Issue {#sec:collapse}
--------------
It has been observed in DARTS that lots of [*skip-connect*]{}s are involved in the selected architecture, which makes the architecture shallow and the performance poor. As an example, let us consider searching on CIFAR100. The alpha value of [*skip-connect*]{}s (green line in Fig. \[all\_exp\](c)) becomes large when the number of search epochs is large, and thus the number of [*skip-connect*]{}s increases in the selected architecture as shown in the green line in Fig. \[all\_exp\](a). Such shallow network has less learnable parameters than deep ones, thus it has weaker expressive power. As a result, architectures with lots of [*skip-connect*]{}s have poor performance, i.e., [*collapsed*]{}, indicated as the blue line in Fig. \[all\_exp\](a). To be more intuitive, we draw the selected architectures from different search epochs on CIFAR100 in Fig. \[visulization\](b). When the number of search epochs increases, the number of [*skip-connect*]{}s in the selected architecture also increases. Such phenomenon can also be observed on other datasets, such as CIFAR10 and ImageNet.
To avoid the collapse, one might propose to adjust search hyper-parameters, such as 1) adjusting learning rates, 2) changing the portion of training and validation data, and 3) adding regularization on [*skip-connect*]{}s like [*dropout*]{}. Unfortunately, such methods would only delay the collapse as the choice of hyper-parameters is not the essential cause of collapse.
![The collapse issue of DARTS. (a) The performance of architectures at different epochs on CIFAR10, CIFAR100 and Tiny-ImageNet-200, respectively (in blue line), and the number of [*skip-connect*]{}s in the normal cell (in green line). (b) The change of $\alpha$ in the deepest edge (connecting the last two nodes) of one-shot model. We omit the $\alpha$ of [*none*]{} operation as it increases while $\alpha$ of other operations drops. (c) The change of architecture parameters $\alpha$ in the shallowest edge. The dashed line denotes the early-stopping paradigm introduced in Sec. \[sec:early\_stopping\], and the circle denotes the point that the $\alpha$ ranking of learnable parameters becomes stable.[]{data-label="all_exp"}](imgs/all_in_three33.pdf){width="3.3in"}
The underlying reason of the collapse issue is that there exist [*cooperation*]{} and [*competition*]{} in the bi-level optimization in DARTS, where the architecture parameters and model weights are updated alternatively. Intuitively, the architecture parameters and model weights are well optimized at the beginning, and then gradually turn to compete against each other after a while. Since the model weights have more advantages than the architecture parameters in the competition, (e.g., the number of the model weights is far more than the number of architecture parameters, the architecture parameters are not sensitive to the final loss in the bi-level optimization, etc.), the architecture parameters cannot beat the model weights in the competition. As a result, the performance of selected architecture will first increase and then decrease (See blue lines in Fig. \[all\_exp\](a)).
In particular, at the initial state of the search procedure, the one-shot model underfits the training data. Thus, architecture parameters $\alpha$ and model weights $w$, which are learnable parameters of the one-shot model, will get better together at the beginning of the search procedure. This is the cooperation period. Note that the first cells in the whole one-shot model can touch the fresh data information, while the data that feed to the last cells are much noisier. If we allow different cells to have distinct architectures in the one-shot model, the first cells will learn features more quickly than the last cells. Since the feature representation learned by the last cells is relatively worse than the one learned by the first cells, the last cells are more likely to select more [*skip-connect*]{}s to obtain the good feature representation directly from the first cells.
Fig. \[fig:diffcell\] shows the learned normal cell architectures at different layers if we allow different architectures at different stages[^1]. It can be seen that the algorithm tends to select deep architectures with learnable operations (namely, operations with parameters to be learned such as [*convolutions*]{}) in the first cells (Fig. \[fig:diffcell\](a)), while architectures with many [*skip-connect*]{}s are preferred in the last cells (Fig. \[fig:diffcell\](c)). If different cells are forced to have the same architecture, as DARTS does, [*skip-connect*]{}s will be broadcasted from the last cells to the first cells. With the increase of searching epochs, the number of [*skip-connect*]{}s in the selected architecture will largely grow. During this period, the architecture becomes bad, thus the competition of architecture parameters $\alpha$ and model weights $w$ occurs, making the performance collapse.
![The selected architecture of normal cells at different layers when searching distinct cell architectures in different stages (stages are split with reduction cells). The searched dataset is CIFAR100. The first cells contain mostly convolutions, while the last cells are shallow with numerous [*skip-connect*]{}s.[]{data-label="fig:diffcell"}](imgs/diffcell_structure.pdf){width="3.3in"}
Moreover, the cooperation and competition phenomenon can also be observed in other bi-level optimization problems (e.g., GAN, meta-learning, etc.). Take GAN as an example, it is proved that a good learned discriminator is essential for training the generator [@goodfellow2014generative], which is the cooperation between generator and discriminator. However, if the input data (fake or real) lies in low-dimensional manifold and the discriminator is over-parameterized, the discriminator will easily separate the generated fake data from the real, and the generator will suffer from gradient vanishment and fail to generate real data [@arjovsky1701towards], which is the competition.
The Early Stopping Methodology {#sec:early_stopping}
==============================
Since the collapse issue of DARTS is caused by the cooperation and competition in the bi-level optimization as pointed out in Sec. \[sec:collapse\], we propose a simple and effective “early stopping” paradigm based on DARTS to avoid the collapse. In particular, the search procedure should be early stopped at a certain epoch, when DARTS starts to collapse. Such paradigm leads to both better performance and less search cost than the original DARTS. We remark that in this paper, we still follow the architecture sharing mechanism among cells used by DARTS[^2]. We use DARTS+ to denote the DARTS algorithm with our early stopping criterion, which is stated as follows.
**Criterion 1** [*The search procedure stops when there are two or more than two [*skip-connect*]{}s in one cell.* ]{}
The major advantage of the proposed stopping criterion is its simplicity. Compared with other DARTS variants, DARTS+ only needs a few modifications based on DARTS, and can significantly increase the performance with less search time. As mentioned in Sec. \[sec:collapse\], too many [*skip-connect*]{}s will hurt the performance of DARTS. On the other hand, an appropriate number of [*skip-connect*]{}s is helpful for transferring the information from first layers to last layers and stabilizing the training process, e.g., ResNet [@he2016deep], which makes the architectures achieve better performance. Therefore, stopping by Criterion 1 is a reasonable choice.
Criterion 1 is motivated by P-DARTS [@chen2019progressive], where the number of [*skip-connect*]{}s in the cell of final architecture is manually cut down to two. Although both DARTS+ and keep two [*skip-connect*]{}s in the cell of their final architectures, DARTS+ is essentially different from P-DARTS in dealing with the [*skip-connect*]{}s. does not intervene the number of [*skip-connect*]{}s during the search procedure, but only replaces the redundant [*skip-connect*]{}s with other operations as a post-processing after the search procedure finishes. In contrast, our DARTS+ ends up with desired architectures with a proper number of [*skip-connect*]{}s to avoid the collapse of DARTS. It controls the number of [*skip-connect*]{}s more directly and also more effectively (See Table \[tab:cifar\_results\] for a performance comparison between DARTS+ and P-DARTS).
Now we give some intuition for Criterion 1 in Fig. \[all\_exp\]. The red circles in Fig. \[all\_exp\](a-b) denote the points that the ranking of architecture parameters $\alpha$ for learnable operations (e.g., [*convolution*]{}s) becomes stable on CIFAR10, CIFAR100 and Tiny-ImageNet-200, respectively. Note that only the operation with the maximum $\alpha$ value is chosen in the selected architecture. When the ranking of $\alpha$ becomes stable, the operations to be selected are nearly determined, which implies that the search is almost saturated. The experiments also verify that after the point that the ranking of $\alpha$ becomes stable (red circles in Fig. \[all\_exp\](a)), the validation accuracies of selected architectures on all datasets (blue lines) tend to decrease, i.e., collapse. It can be seen that this point is close to the stopping point when Criterion 1 holds (the red dash line in Fig. \[all\_exp\](a)), thus stopping by Criterion 1 can prevent the collapse of DARTS.
Since the stable ranking of architecture parameters $\alpha$ for learnable operations indicates the saturated search procedure in DARTS, we can also use the following stopping criterion:
**Criterion 1\*** [*The search procedure stops when the ranking of architecture parameters $\alpha$ for learnable operations becomes stable for a determined number of epochs (e.g., 10 epochs).* ]{}
We remark that Criterion 1 is much easier to operate, but if one needs stopping more precisely or other search spaces are involved, Criterion 1\* could be used instead. We further remark that our early stopping paradigm solves an intrinsic issue of DARTS and is orthogonal to other tricks, thus it has potential to be used in other DARTS-based algorithms to achieve better performance.
We note that recent state-of-the-art differentiable architecture search methods also introduce the early stopping idea in an ad hoc manner. To avoid the collapse, P-DARTS [@chen2019progressive] uses 1) searching for 25 epochs instead of 50 epochs, 2) adopting [*dropout*]{} after [*skip-connect*]{}s, and 3) manually reducing the number of [*skip-connect*]{}s to two. Auto-DeepLab [@liu2019auto] starts to update architecture parameters for a few epochs after updating weight parameters. PC-DARTS [@xu2019pc] uses partial-channel connections to reduce search time, and therefore more epochs are needed for convergence of searching. Thus, setting 50 training epochs is also an implicit early stopping paradigm.
Experiments and Analysis {#Exp_set}
========================
----------------------------------------- ---------- --------------------------------- ----------------------------------- ------------- ------------ -----------
Search Param Search Cost Search
Dataset CIFAR10 CIFAR100 (M) (GPU days)
DenseNet-BC [@huang2017densely]$^1$ - $3.46$ $17.18$ $25.6$ - manual
NASNet-A [@regular2016neural] CIFAR10 $2.65$ - $3.3$ $1800$ RL
AmoebaNet-B [@real2019regularized] CIFAR10 $2.55 \pm 0.05$ - $2.8$ $3150$ evolution
PNAS [@liu2018progressive]$^1$ CIFAR10 $3.41 \pm 0.09$ - $3.2$ $225$ SMBO
ENAS [@pham2018efficient] CIFAR10 $2.89$ - $4.6$ $0.5$ RL
NAONet [@luo2018neural] CIFAR10 $3.18^{1}$ $15.67$ $10.6$ $200$ NAO
DARTS [@liu2018darts] CIFAR10 $3.00$ $17.76$ $3.3$ $1.5$ gradient
SNAS (moderate) [@xie2018snas] CIFAR10 $2.85$ - $2.8$ $1.5$ gradient
ProxylessNAS [@cai2018proxylessnas]$^2$ CIFAR10 $2.08$ - $5.7$ $4$ gradient
P-DARTS [@chen2019progressive] CIFAR10 $2.50$ $16.55$ $3.4$ $0.3$ gradient
P-DARTS [@chen2019progressive] CIFAR100 $2.62$ $15.92$ $3.6$ $0.3$ gradient
ASAP [@noy2019asap] CIFAR10 $2.49 \pm 0.04$ $15.6$ $2.5$ $0.2$ gradient
PC-DARTS [@xu2019pc] CIFAR10 $2.57 \pm 0.07$ - $3.6$ $0.1$ gradient
**DARTS+** CIFAR10 $\mathbf{2.32} (2.50 \pm 0.11)$ $16.28$ $3.7$ $0.4$ gradient
**DARTS+** CIFAR100 $2.46$ $\mathbf{14.87} (15.42 \pm 0.30)$ $3.8$ $0.2$ gradient
**DARTS+\*** CIFAR10 $\mathbf{2.20} (2.37 \pm 0.13)$ $15.04$ $4.3$ $0.6$ gradient
**DARTS+\*** CIFAR100 $2.46$ $\mathbf{14.87} (15.45 \pm 0.30)$ $3.9$ $0.5$ gradient
**DARTS+ (Large)**[$^3$]{} - $\mathbf{1.68}$ $\mathbf{13.03}$ $7.2$ - gradient
----------------------------------------- ---------- --------------------------------- ----------------------------------- ------------- ------------ -----------
Datasets
--------
In this section, we conduct extensive experiments on benchmark classification datasets to evaluate the effectiveness of the proposed DARTS+ algorithm. We use four popular datasets including CIFAR10 [@krizhevsky2009cifar], CIFAR100 [@krizhevsky2009cifar], Tiny-ImageNet-200[^3] and ImageNet [@deng2009ImageNet]. CIFAR10/100 consists of 50K training images and 10K testing images and the resolution is $32 \times 32$. Tiny-ImageNet-200 contains 100K $64 \times 64$ training images and 10K testing images. ImageNet is obtained from ILSVRC2012 [@russakovsky2015ImageNet], which contains more than 1.2M training images and 50K validation images. We follow the general setting on the ImageNet dataset where the images are resized to $224 \times 224$ for training and testing.
Architecture Search {#sec:search_exp}
-------------------
Unless specified, we use Criterion 1 as the stopping condition for DARTS+ in the experiments. Note that the stopping points by Criterion 1 and 1\* are almost the same in the proposed search space, and we will discuss the two criteria in detail in Sec. \[sec:diffstop\].
[ccc]{} Architecture & Test Err. (%) & Params (M)\
ResNet18 [@yao2019differentiable] & $47.3$ & $11.7$\
DenseNet-BC [@lan2018self]
-----------------
DenseNet-BC
[lan2018self]{}
-----------------
: Results of different architectures on Tiny-ImageNet-200. $^\dag$ denotes directly searching with Tiny-ImageNet-200, otherwise transferred from CIFAR10.[]{data-label="tab:tiny_imagenet_resutls"}
& $37.1$ & -\
NASNet & $29.8$ & $4.5$\
DARTS & $30.4$ & $3.8$\
DARTS$^\dag$ & $46.1$ & $2.1$\
SNAS & $30.6$ & $3.4$\
ASAP & $30.0$ & $3.3$\
**DARTS+** & $\mathbf{29.1}$ & $4.2$\
**DARTS+**$^\dag$ & $\mathbf{28.3}$ & $3.8$\
------------------------------------------------------ ----------------- ---------------- ------------ ------------- ------------ ----------
Params $\times +$ Search Cost
Top-1 Top-5 (M) (M) (GPU days)
MobileNet [@howard2017mobilenets] $29.4$ $10.5$ $4.2$ $569$ - manual
MobileNet-V2 ($1.4\times$) [@sandler2018mobilenetv2] $25.3$ - $6.9$ $585$ - manual
ShuffleNet-V2 ($2\times$) [@ma2018shufflenet] $25.1$ - $7.4$ $591$ - manual
NASNet-A [@regular2016neural] $26.0$ $8.4$ $5.3$ $564$ $1800$ RL
AmoebaNet-C [@real2019regularized] $24.3$ $7.6$ $6.4$ $570$ $3150$ RL
PNAS [@liu2018progressive] $25.8$ $8.1$ $5.1$ $588$ $225$ SMBO
MnasNet-92 [@tan2019mnasnet] $25.2$ $8.0$ $4.4$ $388$ - RL
EfficientNet-B0 [@tan2019efficientnet] $23.7$ $6.8$ $5.3$ $390$ - RL
DARTS [@liu2018darts] $26.7$ $8.7$ $4.7$ $574$ $4.0$ gradient
SNAS (mild) [@xie2018snas] $27.3$ $9.2$ $4.3$ $522$ $1.5$ gradient
ProxylessNAS [@cai2018proxylessnas]$^\dag$ $24.9$ $7.5$ $7.1$ $465$ $8.3$ gradient
P-DARTS (CIFAR10) [@chen2019progressive] $24.4$ $7.4$ $4.9$ $557$ $0.3$ gradient
ASAP [@noy2019asap] $26.7$ - - - $0.2$ gradient
XNAS [@nayman2019xnas] $24.0$ - $5.2$ $600$ $0.3$ gradient
PC-DARTS [@xu2019pc]$^\dag$ $24.2$ $7.3$ $5.3$ $597$ $3.8$ gradient
PC-DARTS$^{*\dag}$ $23.8$ $7.3$ $5.3$ $597$ $3.8$ gradient
**DARTS+ (CIFAR100)** $\mathbf{23.7}$ $\mathbf{7.2}$ $5.1$ $591$ $0.2$ gradient
**DARTS+$^{\dag}$** $23.9$ $7.4$ $5.1$ $582$ $6.8$ gradient
**SE-DARTS+ (CIFAR100)**$^\ddag$ $\mathbf{22.5}$ $\mathbf{6.4}$ $6.1$ $594$ $0.2$ gradient
------------------------------------------------------ ----------------- ---------------- ------------ ------------- ------------ ----------
{width="7in"}
### Implementation Details.
We have similar experimental settings as DARTS. The experiments are carried out in two stages: architecture search and architecture evaluation. The search space is the same as DARTS which has 8 candidate operations including [*skip-connect, max-pool-3x3, avg-pool-3x3, sep-conv-3x3, sep-conv-5x5, dil-conv-3x3, dil-conv-5x5, zero*]{}, and the structure of each operation is exactly the same as DARTS.
For CIFAR10 and CIFAR100, we use the same one-shot model as the original DARTS in which 8 cells (i.e. 6 normal cells and 2 reduction cells) with 16 channels are trained for searching. We use a half of the training data to train the model weights and the other half to update the architecture parameters. We search for a maximum of $60$ epochs with batch size $64$. We use SGD to optimize the model weights with initial learning rate $0.025$, momentum $0.9$ and weight decay $3\times10^{-4}$. Adam [@kingma2014adam] is used to optimize architecture parameters with initial learning rate $3\times10^{-4}$, momentum $(0.5,0.999)$ and weight decay $10^{-3}$. Early stopping is applied at certain epoch when Criterion 1 introduced in Sec. \[sec:early\_stopping\] is met.
For Tiny-ImageNet-200, the one-shot model is almost the same as CIFAR10/100 except that a $3\times3$ convolution layer with stride 2 is added on the first layer to reduce the input resolution from $64 \times 64$ to $32 \times 32$. Other settings are the same as those used in CIFAR10/100, including the “early stopping” criterion.
For ImageNet, following [@xu2019pc], the one-shot model starts with three $3\times 3$ convolution layers with stride 2 to reduce the resolution from $224 \times 224$ to $28 \times 28$, and the rest of the network consists of $8$ cells. We select $10\%$ data from the training set for updating model weights, and another $10\%$ for updating architecture parameters. We search with batch size $512$ for both training and validation sets. SGD is used for model weights training with initial learning rate $0.2$ (cosine learning rate decay), momentum $0.9$, and weight decay $3\times 10^{-4}$. The architecture parameters are trained with Adam with learning rate $3\times 10^{-3}$, momentum $(0.5, 0.999)$ and weight decay $10^{-3}$.
For all the datasets, the one-shot model weights and architecture parameters are optimized alternatively. The cell structure is determined by architecture parameters, following DARTS [@liu2018darts].
### Search Results and Analysis.
The proposed DARTS+ needs less searching time as “early stopping” is adopted. For CIFAR10, the search procedure requires 0.4 GPU days with a single Tesla V100 GPU, and stops at about 35 epochs. For CIFAR100, the searching time is 0.2 GPU days and the search procedure stops at about 18 epochs. For Tiny-ImageNet-200, searching stops at about 10 epochs. For ImageNet, the search procedure involves 200 epochs, and it requires 6.8 GPU days on Tesla P100 GPU.
The selected architectures are shown in Fig. \[best\_struc\]. We observe that the cells searched by DARTS+ contain most [*convolution*]{}s and a few [*skip-connect*]{}s. It should be noticed that DARTS+ succeeds in searching with all three datasets including CIFAR10/100, Tiny-ImageNet-200 and ImageNet. However, the original DARTS fails to search on CIFAR100 as the selected architecture is full of [*skip-connect*]{}s [@chen2019progressive], and most previous works on differentiable search [@liu2018darts; @xie2018snas; @chen2019progressive] do not search on ImageNet.
Architecture Evaluation {#arch_eval}
-----------------------
For each selected architecture, we follow the configurations and hyper-parameters of the previous works [@liu2018darts; @chen2019progressive] for evaluation on different datasets.
### Results on CIFAR10 and CIFAR100.
We use network of 20 cells and 36 initial channels for evaluation to ensure a comparable model size as other baseline models. We use the whole training set to train the model for 2000 epochs with batch size 96 to ensure convergence. Other hyper-parameters are set the same as the ones in the search stage. Following existing works [@pham2018efficient; @zoph2018learning; @liu2018progressive; @real2019regularized], we also add some enhancements including cutout, path dropout with probability 0.2 and auxiliary towers with weight 0.4.
The evaluation results are summarized in Table \[tab:cifar\_results\]. For each selected cell from either CIFAR10 or CIFAR100, we report the performance on both datasets. With the simple “early stopping” paradigm, we achieve the best results with $2.32\%$ test error on CIFAR10 and $14.87\%$ test error on CIFAR100. The results are much better than the original DARTS, which gets $3\%$ test error on CIFAR10 and $19.5\%$ test error on CIFAR100 (See Fig. \[all\_exp\](a) with large epochs). The proposed DARTS+ is much simpler and better than other modified DARTS algorithms like P-DARTS and PC-DARTS. ProxylessNAS uses a different search space, and it involves more search time. Moreover, DARTS+ is much easier to implement than other modified DARTS variants including ASAP. We point out that “early stopping” can be used in many other differentiable search algorithms and search spaces to obtain better architectures.
We further increase the initial channel number from 36 to 50, and add more augmentation tricks including AutoAugment [@cubuk2018autoaugment] and mixup [@zhang2017mixup] to achieve better results. Table \[tab:cifar\_results\] shows that DARTS+ achieves impressive $1.68\%$ test error on CIFAR10 and $13.03\%$ test error on CIFAR100, demonstrating the effectiveness of DARTS+.
### Results on Tiny-ImageNet-200.
The network is similar as CIFAR10/100 where 20 cells and 36 channels are involved, except that an additional $3 \times 3$ convolution layer with stride 2 is inserted in the first layer. We transfer the architectures searched from other algorithms to Tiny-ImageNet-200 and evaluate the performance for fair comparison. Other experimental settings are the same as CIFAR10/100.
The results are shown in Table \[tab:tiny\_imagenet\_resutls\]. DARTS+ achieves the state-of-the-art $28.3\%$ test error, which is much better than other baselines. Note that architecture searched on Tiny-ImageNet-200 with DARTS+ performs much better than DARTS and its parameter size is much larger than DARTS, because DARTS suffers from collapse and the architecture searched with DARTS contains lots of [*skip-connect*]{}s. Desired architectures are more likely to be generated with the “early stopping” paradigm.
### Results on ImageNet.
We use the architecture searched directly from ImageNet for evaluation, and the architecture from CIFAR100 to test the tranferability of the selected architecture. We follow DARTS such that the number of cells is 14 and the initial number of channels is 48. We train the model for 800 epochs with batch size 2048 on 8 Nvidia Tesla V100 GPUs as more epochs can achieve better convergence. The model is optimized with the SGD optimizer with an initial learning rate 0.8 (cosine decayed to 0), momentum of 0.9 and weight decay $3\times10^{-5}$. We use learning rate warmup [@goyal2017accurate] for the first 5 epochs and other training enhancements including label smoothing [@szegedy2016rethinking] and auxiliary loss tower.
The experimental results are shown in Table \[tab:imagenet\_results\]. Note that we re-implement PC-DARTS and the results are reported. When searching on ImageNet with the proposed DARTS+, the selected architecture achieves impressive $23.9\%/7.4\%$ top-1/top-5 error, and the architecture transferred from CIFAR100 achieves state-of-the-art $23.7\%/7.2\%$ error. The results imply that DARTS with “early stopping” succeeds in searching a good architecture with impressive performance on large-scale datasets with limited time.
We also adopt SE-module [@hu2018squeeze] in the architecture transferred from CIFAR100, and introduce AutoAugment [@cubuk2018autoaugment] and mixup [@zhang2017mixup] for training to obtain better model. The results are shown in Table \[tab:imagenet\_results\], and we achieve $22.5\%/6.4\%$ top-1/top-5 error with only additional 3M flops, showing the effectiveness of the selected architecture.
Effectiveness of Early Stopping {#sec:diffstop}
-------------------------------
To further verify the effectiveness of DARTS+, we conduct extensive experiments on selected architectures at different epochs. The classification results on CIFAR10, CIFAR100 and Tiny-ImageNet-200 are shown in Fig. \[all\_exp\]. We also point out the time to “early stop” under two criteria, marked as “red dashed line” and “red circle” respectively. We observe that the selected architecture performs worse with larger epochs, implying that the original DARTS suffers from the collapse issue. In contrast, “early stopping” is able to generate good architectures at both stopping criteria, regardless of the type of datasets.
We also compare “early stopping” Criterion 1 and 1\* in Table \[tab:cifar\_results\] and Fig. \[all\_exp\]. We observe that both criteria achieve comparable performance on all datasets as the stopping points are very close.
Conclusion
==========
In this paper, we conduct comprehensive analysis and extensive experiments to show that DARTS suffers from the collapse problem, which is mainly caused by the cooperation and competition problem in the bi-level optimization in DARTS. We propose the “DARTS+” algorithm, in which the “early stopping” paradigm is introduced to avoid the collapse of DARTS. The experiments show that we succeed in searching on various benchmark datasets including large-scale ImageNet with limited GPU days, and the resulting architectures achieve the state-of-the-art performances on all benchmark datasets. Moreover, it should be noticed that many recent progresses of DARTS could use “early stopping” to achieve better results, and the proposed “early stopping” criteria could be applied to many other types of search spaces, including the RandWire search space [@xie2019exploring], mobile convnets search space [@stamoulis2019single], etc.
[^1]: Stages are split with reduction cells, and each stage consists of a number of stacked cells.
[^2]: In future work, we may relax this constraint and allow different cells or stages to have different structures.
[^3]: <https://tiny-ImageNet.herokuapp.com/>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a dual-pump third-order nonlinear scheme for producing pairs of correlated photons that is less susceptible to Raman noise than typical spontaneous four wave mixing methods (SFWM). Beginning with the full multimode Hamiltonian we derive a general expression for the joint spectral amplitude, from which the probability of producing a pair of photons can be calculated. As an example, we demonstrate that a probability of $0.028$ pairs per pulse can be achieved in an appropriately designed fused silica microfiber. As compared with single pump SFWM in standard fiber, we calculate that our process shows significant suppression of the spontaneous Raman scattering and an improvement in the signal to noise ratio.'
author:
- 'Daniel R. Blay'
- 'L. G. Helt'
- 'M. J. Steel'
bibliography:
- 'sstpdc3.bib'
title: 'Proposal for an Integrated Raman-free Correlated Photon Source'
---
The on-demand generation of single photons is keenly sought in quantum optics. There are a multitude of photon generation schemes, including atom-like sources [@Somaschi2016; @Darquie2005; @Hennessy2007; @Albrecht2014], and heralded photon pair sources based on nonlinear optics. Of the latter, the two most common schemes are spontaneous parametric down conversion (SPDC)[@Kwiat1995; @Bonfrate1999; @Tanzilli2001] and spontaneous four wave mixing (SFWM)[@Fiorentino2002; @Inoue2004; @Sharping2006].
Due to the strength of the ${\chi^{(2)}}$ nonlinearity as compared with ${\chi^{(3)}}$, SPDC sources typically require lower pump powers than SFWM sources, and consequently exhibit negligible noise from competing ${\chi^{(3)}}$ processes. However, most optical materials do not possess the symmetries required to have a ${\chi^{(2)}}$ response. Utilising the universal ${\chi^{(3)}}$ response allows for a greater number of materials and platforms to be used in heralded single photon sources, and integrates well with current telecommunication networks. In particular, silica and silicon allow for near infra-red generation and thus efficient coupling to standard SMF-28 fiber[@Fiorentino2002; @Inoue2004].
Many of these ${\chi^{(3)}}$ materials, such as fused silica, are amorphous and therefore exhibit broadband spontaneous Raman scattering (SpRS) due to inhomogenous broadening of the Raman transitions. In the quantum regime, this corresponds to emission of uncorrelated single photons. In a typical amorphous degenerate SFWM source, the strong pump field produces these uncorrelated Raman photons over a broad energy range. This noise often overlaps with the desired frequency range of the generated pairs, [@Lin2007] (see \[subfig:sfwm\_diagram\]). Without due care, one may generate many more Raman photons than correlated pairs [@Collins2012a].
Attempts have been made to mitigate Raman noise in amorphous SFWM sources [@Clark2011], including dispersion engineering waveguides such that the produced pairs lie in a window of low Raman photon production [@Collins2012a; @Li2006; @He2012]. However, these are subject to material and engineering constraints that may be challenging to implement.
[0.5]{} {width=".9\linewidth"}
[0.5]{} {width=".9\linewidth"}
Here we propose a new method which sidesteps the issue of SpRS in amorphous materials. Pumping a ${\chi^{(3)}}$ material strongly at or near the third harmonic, $\omega_p \approx 3\omega_s$, three photons can be spontaneously generated at the third sub-harmonic $\omega_s$, which we call the fundamental, with low probability. When unstimulated this process is inefficient [@Bencheikh2007], and so authors seeking three photon generation tend to use cascaded ${\chi^{(2)}}$ processes [@Hubel2010; @Dot2012; @Hamel2014]. As we are aiming for pair generation, here we add a weak coherent field near the fundamental frequency $\omega_s$ to seed the process. This leaves the desired photon pair accessible, with energy conservation dictated by $\omega_p = \omega_1 + \omega_2 + \omega_s$. For a seed exactly at the fundamental, the pair of photons lie on either side of the seed, as seen in \[subfig:sstpdc\_raman\]. We may consider this process “stimulated spontaneous three photon down conversion” (SSTPDC). The principal advantage of this scheme is that the Stokes band of the Raman spectrum lies between the generated pairs and the pump frequency. The bandwidth of the spontaneous Raman response is typically of order 10 THz, whereas the fundamental and third harmonic fields involved at optical frequencies are separated by around 400 THz. This large spectral separation ensures low Raman noise in the signal band.
To describe this process and all of the multimode physics involved, we follow the formalism outlined by Yang, Liscidini and Sipe [@Yang2008]. We look for solutions to the first order Schrödinger equation $$\label{eq:first_order_schro}
\ket{\psi_\text{out}} \approx \left[1 - \frac{i}{\hbar} \int_{t_0}^{t_1} \Diff{t} \hat{H}_\text{I}(t)\right] \ket{\psi_\text{in}},$$ where $\ket{\psi_\text{in}}$ describes the coherent state input in the pump and seed modes, and is vacuum in the photon pair bands. As the nonlinearity outside the interaction length is zero, we are free to extend the integration limits to infinity, $t_0 \to -\infty$ and $t_1 \to \infty$. The relevant interaction Hamiltonian is given by $$\label{eq:full_3rdo_hamil}
\hat{H}_\text{I}(t)= -\frac{1}{4\epsilon_0}\int \Diff{^3\mathbf{r}} \Gamma^{(3)}_{ijmn} \hat{D}^i(\mathbf{r},t) \hat{D}^j(\mathbf{r},t) \hat{D}^m(\mathbf{r},t) \hat{D}^n(\mathbf{r},t).$$ Here $\hat{D}^{(i)}(\mathbf{r},t)$ are the components of the vector displacement operator, written in the interaction picture associated with the linear Hamiltonian ${\hat{H}_L = \sum_{\gamma} \int \Diff{k} \hbar \omega_{\gamma,k} \aopd{\gamma,k} \aop{\gamma,k}}$, (the summation is over the different modes $\gamma$, and the vacuum term is disregarded). The symbol $\Gamma^{(3)}$ is a rank four tensor, related to the standard third order susceptibility tensor[@Yang2008] (see below).
To treat the SSTPDC process, we express the displacement field operator as a sum over modes $\gamma$, $$\begin{aligned}
\label{eq:dop_exp}
\hat{\mathbf{D}}(\mathbf{r},t) = &\hphantom{+}\sum_{\gamma} \int_0^{k_B} \Diff{k} \sqrt{\frac{\hbar \omega_k}{2}} \mathbf{D}_{\gamma,k}(\mathbf{r}) \aop{\gamma,k} {\mathrm{e}}^{-i \omega_k t}+ \sum_{\gamma'} \int_{k_B}^{\infty} \Diff{k} \sqrt{\frac{\hbar \omega_k}{2}} \mathbf{F}_{\gamma',k}(\mathbf{r}) \bop{\gamma',k}{\mathrm{e}}^{-i \omega_k t} + \text{h.c.}\end{aligned}$$ where $\aop{\gamma,k}$ and $\bop{\gamma,k}$ are the usual bosonic annihilation operators for mode $\gamma$ and wavenumber $k$, $\mathbf{D}_{\gamma,k}$ and $\mathbf{F}_{\gamma,k}$ are mode functions, and h.c. denotes the Hermitian conjugate. We separate the expansion into low ($\aop{\gamma,k}$) and high ($\bop{\gamma,k}$) frequency bands, with $k_B$ a wavenumber between the two. For uniform waveguides, the field mode functions may be decomposed into a transverse mode function and a longitudinal plane wave, $$\begin{aligned}
\mathbf{D}_{\gamma,k}(\mathbf{r}) = \frac{\mathbf{d}_{\gamma,k}(x,y) {\mathrm{e}}^{i k z}}{\sqrt{2\pi}}, \qquad
\mathbf{F}_{\gamma,k}(\mathbf{r}) = \frac{\mathbf{f}_{\gamma,k}(x,y) {\mathrm{e}}^{i k z}}{\sqrt{2\pi}}.\end{aligned}$$ On substituting \[eq:dop\_exp\] into \[eq:full\_3rdo\_hamil\], and considering only one mode per band, we keep only terms involving the annihilation of one photon in the high frequency band and the creation of three photons in the low frequency band, and their conjugates. We neglect the SPM and XPM terms, with the expectation that this photon source will operate in a regime where there is not enough power for these effects to be significant [@Helt2013]. The interaction Hamiltonian can then be expressed as $$\label{eq:interaction_hamiltonian}
\begin{aligned}
\hat{H}_\text{I}(t) =& - \frac{3 \alpha^* \beta \hbar^2}{16 \pi^2 \epsilon_0} \int_0^{\infty} \int_0^{\infty}\int_0^{\infty}\int_0^{\infty}
\Diff{k_1} \Diff{k_2} \Diff{k_3} \Diff{k_4}
\sqrt{\omega_{k_1} \omega_{k_2} \omega_{k_3} \omega_{k_4}} \frac{\mathfrak{s}^*(\Delta k) {\overline{\chi}^{(3)}}}{{\overline{n}}^4}
\frac{\phi^*_s(k_4) \phi_p(k_4) {\mathrm{e}}^{-i\Delta \omega_k t}}{\mathcal{A}(k_1,k_2,k_3,k_4)} \aopd{k_1}\aopd{k_2}
+ \text{h.c.},
\end{aligned}$$ where $\Delta k = k_4 - k_3 - k_2 - k_1$, $\Delta \omega_k = \omega_{k_4} - \omega_{k_3} - \omega_{k_2} - \omega_{k_1}$, $\left|\alpha\right|^2$ and $\left|\beta\right|^2$ are the average numbers of photons in the input classical seed and pump pulses respectively, and $\phi_s(k)$, $\phi_p(k)$ are their spectral profiles, localised in $k$. The effective mode coupling area satisfies
$$\label{eq:effective_area}
\begin{aligned}
\frac{1}{\mathcal{A}(k_1,k_2,k_3,k_4)} \equiv& \int_{-\infty}^{\infty} \Diff{x} \int_{-\infty}^{\infty} \Diff{y} \frac{{\overline{n}}^4 {\chi^{(3)}}_{ijmn}(x,y)}{4{\overline{\chi}^{(3)}}\epsilon_0^2 n_{k_1}^2 n_{k_2}^2 n_{k_3}^2 n_{k_4}^2}\\
&\times \left[\left(\dop{i}{k_1}\dop{j}{k_2}\dop{m}{k_3}\right)^*\fop{n}{k_4}+
\left(\dop{i}{k_1}\dop{j}{k_2}\dop{n}{k_3}\right)^*\fop{m}{k_4}+
\left(\dop{i}{k_1}\dop{n}{k_2}\dop{m}{k_3}\right)^*\fop{j}{k_4}+
\left(\dop{n}{k_1}\dop{j}{k_2}\dop{m}{k_3}\right)^*\fop{i}{k_4}\right]\\
=& \int_{-\infty}^{\infty} \Diff{x} \int_{-\infty}^{\infty} \Diff{y} \frac{{\overline{n}}^4}{4{\overline{\chi}^{(3)}}\epsilon_0^2 n_{k_1}^2 n_{k_2}^2 n_{k_3}^2 n_{k_4}^2}\\ &\times\bigg\{\left(2{\chi^{(3)}}_{1122}(x,y)+{\chi^{(3)}}_{1212}(x,y)+{\chi^{(3)}}_{1221}(x,y)\right) \left(\mathbf{d}_{k_1}\cdot\mathbf{d}_{k_2}\right)^* \left(\mathbf{d}_{k_3}\right)^*\cdot\mathbf{f}_{k_4}\\
&+ \left({\chi^{(3)}}_{1122}(x,y)+2{\chi^{(3)}}_{1212}(x,y)+{\chi^{(3)}}_{1221}(x,y)\right) \left(\mathbf{d}_{k_1}\cdot\mathbf{d}_{k_3}\right)^* \left(\mathbf{d}_{k_2}\right)^*\cdot\mathbf{f}_{k_4}\\
&+ \left({\chi^{(3)}}_{1122}(x,y)+{\chi^{(3)}}_{1212}(x,y)+2{\chi^{(3)}}_{1221}(x,y)\right) \left(\mathbf{d}_{k_2}\cdot\mathbf{d}_{k_3}\right)^* \left(\mathbf{d}_{k_1}\right)^*\cdot\mathbf{f}_{k_4}\bigg\},
\end{aligned}$$
where $\Gamma^{(3)}_{ijmn} = {\chi^{(3)}}_{ijmn}/({\epsilon_0 n_{k_1}^2 n_{k_2}^2 n_{k_3}^2 n_{k_4}^2})$, the refractive index is abbreviated as ${n(x,y;\omega_{k_i}) = n_{k_i}}$, and the nonlinear susceptibility has been decomposed into a transverse and longitudinal part ${{\chi^{(3)}}_{ijmn}(\mathbf{r}) = {\chi^{(3)}}_{ijmn}(x,y) s(z)}$. Additionally, the typical size of a nonvanishing component of ${\chi^{(3)}}(x,y)$ is denoted ${\overline{\chi}^{(3)}}$, and ${\overline{n}}$ represents a typical value of the local refractive index, both introduced solely for convenience[@Yang2008]. We associate $k_1$ and $k_2$ with the generated pairs, $k_3$ with the seed and $k_4$ with the pump. Note that the second form of \[eq:effective\_area\] in vector notation holds if the material is isotropic, and both definitions for the effective area account for fields of arbitrary polarisation. The phasematching condition for SSTPDC in \[eq:interaction\_hamiltonian\] is captured in the spatial Fourier transform of the longitudinal nonlinearity profile $s(z)$: $$\mathfrak{s}(k) \equiv \int_{-\infty}^{\infty} \Diff{z} s(z) {\mathrm{e}}^{-ikz}.$$
We use \[eq:interaction\_hamiltonian\] to find the first order solution as given by \[eq:first\_order\_schro\] and transform from $k$ to $\omega$. This introduces factors in the group velocity $v_g$ which account for the density of states in frequency[@Yang2008]. The integration over all time yields $\int_{-\infty}^{\infty} {\mathrm{e}}^{-i \Delta \omega_k t} = 2\pi \delta(\Delta \omega_k)$, allowing the further integration over one frequency. Now the state can be described as $$\label{eq:biphoton_state}
\ket{\psi_\text{out}} \approx \frac{{\ket{\text{vac}}}+ \eta \ket{\text{II}}}{\sqrt{1+\left|\eta\right|^2}},$$ where $\eta$ is a normalisation factor, and the biphoton state is described by $$\label{eq:biphoton_state_star}
\ket{\text{II}} = \frac{1}{\sqrt{2}} \int_0^{\infty} \Diff{\omega_1} \Diff{\omega_2} \Phi(\omega_1,\omega_2) \aopd{\omega_1} \aopd{\omega_2} {\ket{\text{vac}}}.$$
The joint spectral amplitude (JSA) is $$\begin{gathered}
\label{eq:jsa}
\Phi(\omega_1,\omega_2) = \frac{1}{\eta}\Bigg(\frac{3\sqrt{2} i \alpha^* \beta \hbar}{8 \pi \epsilon_0} \int_0^{\infty} \Diff{\omega} \sqrt{\frac{\omega_1 \omega_2 \omega(\omega_1 + \omega_2 + \omega)}{v_g(\omega_1) v_g(\omega_2) v_g(\omega) v_g(\omega_1+\omega_2+\omega)}}\\
\times \frac{\mathfrak{s}^*(\Delta k){\overline{\chi}^{(3)}}}{{\overline{n}}^4} \frac{\bar{\phi}_s^*(\omega) \bar{\phi}_p(\omega_1+\omega_2+\omega)}{\mathcal{A}\left[k(\omega_1),k(\omega_2),k(\omega),k(\omega_1+\omega_2+\omega)\right]}\Bigg).\end{gathered}$$ This is our main result. It fully describes the biphoton state for SSTPDC, and from \[eq:jsa\] one can calculate the rate of photon pair production as well as arbitrary expectation values. In particular, normalising the biphoton state $\left<\text{II}|\text{II}\right>=1$ imposes the normalisation of the JSA ${\int \Diff{\omega_1}\Diff{\omega_2} \left|\Phi(\omega_1,\omega_2)\right|^2 = 1}$. From \[eq:biphoton\_state\_star\], this allows the physical interpretation of $\left|\eta\right|^2$ as the probability of pair production per pump pulse.
Approximating the seed and pump fields as Gaussians, $\bar{\phi}_j(\omega) = \left(\sqrt{\tau}_j/\pi^{1/4}\right) {\mathrm{e}}^{-\tau(\omega-\omega_j)^2/2}$, with long pulse durations $\tau_s$ and $\tau_p$ respectively, it is possible to arrive at a coarse approximation to the pair production probability $$\label{eq:rough_rate}
\left|\eta\right|^2 \approx \frac{4 \gamma^2 L^2}{3\pi} \sqrt{\frac{2 \tau_s^2 \tau_p^2}{\left|\beta_2(\omega_s)\right|L\left(\tau_s^2+\tau_p^2\right)}} P_p P_s,$$ where $L$ is the interaction length, the nonlinear parameter is $$\label{eq:gamma}
\gamma = \dfrac{3 {\chi^{(3)}}\omega_s}{4 \epsilon_0 \sqrt{v_g^3(\omega_s) v_g(\omega_p)} {\overline{n}}^4 \mathcal{A}},$$ $P_s = \hbar \omega_s \left|\alpha\right|^2/\tau_s$, and $P_p = \hbar \omega_p \left|\beta\right|^2/\tau_p$. These are nominal average *pulse* powers, related to the time averaged power by $\bar{P}_j = P_j \sigma$, where $\sigma$ is the duty cycle of a high repetition rate laser.
As a proof-of-principle, we consider a system where we expect to be able to phasematch this process. The phasematching and efficient conversion of one-third harmonic generation (OTHG) or backwards third harmonic generation has been studied by Grubsky [*et al*. ]{}[@Grubsky2005] and further refined by Zhang [*et al*. ]{}[@Zhang2015a], demonstrating how to phasematch the process in fused silica microfiber. By tuning the width of the fiber, the high frequency HE$_{21}$ mode can be phasematched with the low frequency HE$_{11}$ mode.
Setting the pump to the common frequency-doubled laser wavelength , we solve for these modes exactly and find that $n_\text{eff}^{\text{HE}_{21}}(\omega_p)=n_\text{eff}^{\text{HE}_{11}}(\omega_s)$ when the diameter of the fiber is . The effective area then is $\mathcal{A} = \SI{4.9}{\micro\meter^2}$, with the third order susceptibility ${\chi^{(3)}}=2.5\times10^{-22}$. We take an interaction length $L=\SI{10}{\milli\meter}$ typical of fiber tapers. We envisage modest pump configurations, typical of current mode-locked green sources, as shown in \[tab:sstpdc\_values\]. Note that to manage the fast walk-off, the seed pulse duration is on the order of nanoseconds. With this set of parameters, the pair production probability as given by \[eq:rough\_rate\] is $\left|\eta\right|^2=0.029$ per pulse. Without relying on these coarse approximations and assuming Gaussian pulses, the equivalent numerical result derived from \[eq:jsa\] yields $\left|\eta\right|^2=0.028$ per pulse, across a bandwidth of . This is sufficient for producing an effective photon source, for a pump laser with a typical 100 MHz repetition rate. The resulting joint spectral intensity (JSI) is shown in \[fig:norm\_jsi\]. It has a Schmidt number[@Law2004] of $K=106.3$, found from the singular value decomposition of the JSA. This indicates that the state is highly correlated, with an unheralded second order correlation function[@Christ2011] of $g^{(2)}(0)=1.0094$.
**Quantity** ** Symbol ** ** Seed value $(\omega_s)$** ** Pump value $(\omega_p)$**
--------------------------- ------------------ -------------------------------------------- -------------------------------------------
Wavelength $\lambda$ $\SI{1596}{\nano\meter}$ $\SI{532}{\nano\meter}$
Refractive index $n$ 1.46 1.44
Group index $n_g$ 1.396 1.695
Group velocity dispersion $\beta_2$ $\SI{2344}{\pico\second^2\per\kilo\meter}$ $\SI{-10}{\pico\second^2\per\kilo\meter}$
Pulse Power $P$ $\SI{1}{\watt}$ $\SI{10}{\kilo\watt}$
Pulse duration $\tau$ $\SI{1.0}{\nano\second}$ $\SI{10.0}{\pico\second}$
: Material and pulse parameters for the SSTPDC simulation. The group velocity and group velocity dispersion were acquired from the exact solutions in microfiber.
\[tab:sstpdc\_values\]
Being a four wave mixing process with four distinct fields, the parameter space for SSTPDC is quite large, and there is room to engineer desirable JSAs. For example, decreasing the pump duration to $\tau_p = \SI{1}{\pico\second}$ decreases the Schmidt number to $K=66.9$ ($g^{(2)}(0)=1.015$), decreases the rate to $\left|\eta\right|^2 = 0.0016$ pairs per pulse, and increases the generation bandwidth to . Doubling the interaction length $L=\SI{20}{\milli\meter}$ decreases the Schmidt number to $K=90.4$ ($g^{(2)}(0)=1.011$), decreases the generation bandwidth to , but increases the rate to $\left|\eta\right|^2=0.076$ pairs per pulse. As a final example, a four fold increase in the group velocity dispersion decreases the rate to $\left|\eta\right|^2 = 0.014$ pairs per pulse (as we would expect from \[eq:rough\_rate\]), decreases the generation bandwidth to , and also decreases the Schmidt number to $K=53.6$ ($g^{(2)}(0)=1.019$).
![The normalised joint spectral intensity for SSTPDC in fused silica microfiber, with a Schmidt number of $K=106.3$, and a pair production probability of $\left|\eta\right|^2=0.028$ per pulse. Here $\omega_s/(2\pi) = \SI{187.84}{\tera\hertz}$.\[fig:norm\_jsi\]](norm_jsi_inset.eps){width=".5\linewidth"}
What about the Raman problem? As a first approximation we follow Agrawal [@Agrawal2006] and Lin [*et al*. ]{}[@Lin2007], working in the quasi-CW limit and expressing the Raman photon flux as ${I_u^R = \Delta \nu_u P_p L \left|g_R(\Delta)\right| [\rho+1/({\mathrm{e}}^{h \Delta/(k_B T)}-1) ]/\mathcal{A}}$, where $u=s,i$, the detuning from the seed is $2\pi\Delta = \left|\omega - \omega_s\right|$, $\Delta \nu_u$ is a filter bandwidth. The factor $\rho=1$ for the Stokes process and is zero for the anti-Stokes process. The Stokes channel coincides with the idler photon band ($\omega<\omega_s$) and the anti-Stokes channel coincides with the signal photon ($\omega>\omega_s$). The Raman gain $g_R(\Delta)$ is taken directly from Agrawal [@Agrawal2006]. In this limit, the SFWM and SSTPDC spectral densities of the desired photon pairs are expressed as $$\begin{aligned}
\frac{I^\text{\scriptsize{SFWM}}_u}{\Delta\nu_u} &= \left(\gamma_\text{\scriptsize{SFWM}} P_\text{sp} L\right)^2 \text{sinc}^2\left(\pi\beta_2^{\text{\scriptsize{SFWM}}}(\omega_s)L \Delta \right),\label{eq:sfwm_flux}\\
\frac{I^\text{\scriptsize{SSTPDC}}_u}{\Delta\nu_u} &= 4\left(\gamma \sqrt{P_p P_s} L\right)^2 \text{sinc}^2\left(\pi\beta_2(\omega_s)L \Delta \right),\label{eq:sstpdc_flux}\end{aligned}$$ where $P_\text{sp}$ is the single-pump power, and similarly to \[eq:gamma\], $$\label{eq:gamma_sfwm}
\gamma_\text{\scriptsize{SFWM}} = \dfrac{3 {\chi^{(3)}}\omega_s}{4 \epsilon_0 v_g^2(\omega_s) {\overline{n}}^4 \mathcal{A}_\text{\scriptsize{SFWM}}},$$ with $\mathcal{A}_\text{\scriptsize{SFWM}}=\SI{84.0}{\micro\meter}$, $n_g^{\text{\scriptsize{SFWM}}}(\omega_s)=1.463$, and $\beta_2^\text{\scriptsize{SFWM}}(\omega_s) = -\SI{26.18}{\square\pico\second\per\kilo\meter}$. As the spontaneous Raman bandwidth is approximately , and the pump and seed fields are separated by roughly , we do not expect there to be any measurable contribution from the pump to the Raman noise in the pair generation band.
Defining the signal to noise ratio (SNR) as $\text{SNR} = I_u/I^R_u$, we define a figure of merit (FOM) to be the improvement in the SNR for SSTPDC in microfiber over the SNR of SFWM in SMF-28. Taking the single pump power to be $P_\text{sp} = \sqrt{P_s P_p}$, the FOM is expressed as $$\label{eq:fom}
\mathcal{F} \equiv \frac{\text{SNR}_\text{\scriptsize{SSTPDC}}}{\text{SNR}_\text{\scriptsize{SFWM}}} \approx 4\frac{\mathcal{A}_\text{\scriptsize{SFWM}}}{\mathcal{A}} \sqrt{\frac{P_p}{P_s}}.$$ To keep $\left|\eta\right|^2 \ll 1$ (recall \[eq:rough\_rate\]), we fix $P_p$ at $\SI{10}{\kilo\watt}$ with a pulse duration of $\SI{1}{\nano\second}$ ( average power at a repetition rate of ), and plot the peak spectral densities (\[eq:sstpdc\_flux,eq:sfwm\_flux\]), and FOM (\[eq:fom\]) as functions of $P_s$. The spectral densities themselves are plotted for a representative seed power of (the vertical line in \[fig:fom\]) in \[fig:spectral\_densities\]. shows a considerably improved SNR as compared with SFWM for a range of pumping configurations. demonstrates a reduction in the production of Raman photons and, critically, that the SSTPDC signal lies well above the SpRS of the SSTPDC seed, whereas the SFWM signal is buried beneath the SpRS from the SFWM pump.
[0.5]{} {width=".9\linewidth"}
[0.5]{} {width=".9\linewidth"}
Evaluating the two processes for a common fiber, the ratio between $P_\text{sp}$ and $P_s$ predicts an 100-fold decrease in Raman noise. If we compare SSTPDC in microfiber with SFWM in SMF-28, taking into account that the Raman gain increases in magnitude for smaller areas, we find that the gain in microfiber is larger by a factor of $\mathcal{A}_\text{SFWM}/\mathcal{A} = 17.0$, and so the ratio that determines the suppression is approximately $P_\text{sp}\mathcal{A}/(P_s\mathcal{A}_\text{\scriptsize{SFWM}}) = 5.9$. In other systems that support third harmonic generation, for example in photonic crystal fiber[@Tarnowski2011], one might expect to achieve or exceed the suppression given by $P_\text{sp}/P_s$. The freedom to separate the strong pump from the pair generation, inherent to SSTPDC, is what drives this improvement.
We have proposed and demonstrated the feasibility of the third order photon pair generation method we call stimulated spontaneous three photon down conversion (SSTPDC). The process can be phasematched, and with realistic pump requirements produce pairs at rates sufficiently high for an effective heralded single photon source. With the strongest pump field spectrally distinct from the generated pairs, it sidesteps the issue of noise from spontaneous Raman scattering. In microfiber, we calculate an improvement in the ratio of photon pairs to uncorrelated Raman photons as compared with standard SFWM in SMF-28, for various pumping configurations.
This work was supported in part by the ARC Centre for Ultrahigh bandwidth Devices for Optical Systems (CUDOS) (Project No. CE110001018). We thank M. J. Collins for useful discussions.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Hillclimbing is an essential part of any optimization algorithm. An important benchmark for hillclimbing algorithms on pseudo-Boolean functions $f: \{0,1\}^n \to {\ensuremath{\mathbbm{R}}}$ are (strictly) *montone* functions, on which a surprising number of hillclimbers fail to be efficient. For example, the $(1+1)$-Evolutionary Algorithm is a standard hillclimber which flips each bit independently with probability $c/n$ in each round. Perhaps surprisingly, this algorithm shows a phase transition: it optimizes any monotone pseudo-boolean function in quasilinear time if $c<1$, but there are monotone functions for which the algorithm needs exponential time if $c>2.2$. But so far it was unclear whether the threshold is at $c=1$.
In this paper we show how Moser’s entropy compression argument can be adapted to this situation, that is, we show that a long runtime would allow us to encode the random steps of the algorithm with less bits than their entropy. Thus there exists a $c_0 > 1$ such that for all $0<c\le c_0$ the $(1+1)$-Evolutionary Algorithm with rate $c/n$ finds the optimum in $O(n \log^2 n)$ steps in expectation.
address: '$^*$Department of Computer Science, ETH Zürich, Switzerland'
author:
- 'Johannes Lengler$^*$'
- 'Anders Martinsson$^*$'
- 'Angelika Steger$^*$'
bibliography:
- 'references.bib'
title: 'When Does Hillclimbing Fail on Monotone Functions: An entropy compression argument'
---
Introduction
============
Hillclimbing is an essential part of any optimization algorithm. The *$(1+1)$-Evolutionary Algorithm* or $(1+1)$-EA is a simple greedy hillclimbing scheme for maximizing a pseudo-Boolean objective function $f:\{0, 1\}^n\rightarrow \mathbb{R}$. We start with a search point $X_0 \in \{0, 1\}^n$ uniformly at random. In the $t$-th round we create an *offspring* $X'$ from the *parent* $X_t$ by flipping each bit of $X_t$ independently with probability $c/n$, where $c$ is the *mutation parameter*. Then we replace the current search point by $X'$ if it has at least the same objective, i.e., we set $X_{t+1} := X'$ if $f(X') \geq f(X_{t})$, and $X_{t+1} := X_t$ otherwise. The phrase $(1+1)$ reflects that in each round the next search point is chosen from one parent plus one offspring. It is clear that, on any function $f$ with a unique global maximum, the $(1+1)$-EA will eventually fixate at at this maximum of $f$.
Here we study the performance of this algorithm on (strictly) monotone functions. A function $f:\{0, 1\}^n\rightarrow\mathbb{R}$ is said to be *monotone*[^1] if $f(x)<f(y)$ whenever $x\neq y$ and $x^i\leq y^i$ for all $i\in[n]$, where $x^i$ denote the $i$-th coordinate of $x$. For any such function, the unique global maximum is the all-ones string. Monotone functions are an important class of benchmark functions for hillclimbing schemes, since there exist a large variety of hillclimbing schemes that optimize all monotone functions efficiently. For example, the $(1+1)$ algorithm that creates the offspring by flipping exactly one random bit in each round resembles a coupon collector process, and thus finds the optimum in time $O(n\log n)$.[^2] Nevertheless, a surprising number of hillclimbing schemes fail on some monotone functions, see [@lengler2018general] for an overview.
For the [$(1 + 1)$-EA]{}, while for any constant $c<1$ it is easy to see that the algorithm needs time $O(n\log n)$ to find the optimum of any monotone function [@DoerrJSWZ10], it was shown in a sequence of papers [@DoerrJSWZ10; @doerr2013mutation; @lengler2016drift] that for $c > 2.13\ldots$ there are monotone functions (dubbed [<span style="font-variant:small-caps;">HotTopic</span>]{}functions in [@lengler2018general]) on which the algorithm needs exponential time. The standard proof techniques for upper runtime bounds fail precisely at $c=1$, and there were split opinions in the community on whether there should be a phase transition from polynomial to exponential at $c=1$ [@kotzingpersonal]. On the presumed threshold $c=1$ it follows from a more general model of Jansen [@jansen2007brittleness] that the runtime is $O(n^{3/2})$, but it remained unclear whether the runtime is quasilinear or not.
The value $c=1$ is of special interest, for several reasons. From a practical perspective, it is considered the standard choice for the mutation parameter and explicitly recommended by textbooks on the subject [@back1996evolutionary; @back1997handbook]. From a theoretical perspective, $c=1$ is known to be the optimal parameter choice for linear functions, i.e., for functions of the form $f(X) = \sum_{i=1}^n w_iX^i$, where the $w_i$ are fixed weights. More precisely, the choice $c=1$ gives runtime $(1+o(1)) en\log n$ on any linear function, while any other (constant) choice of $c$ gives a strictly worse leading constant on any linear function [@witt2013tight].
In this paper we use an entropy compression argument to show that for $c = 1+{\varepsilon}$ the runtime remains quasilinear. More precisely, we show that a long runtime would allow us to encode the random trajectory of the algorithm with fewer bits than its entropy, which is an information theoretic contradiction. This type of argument is attributed to Moser, who used the technique in his celebrated algorithmic proof of the Lovász local lemma [@moser2009constructive; @fortnow2009kolmogorow]. Since then, the method has been used to extend and apply the local lemma [@moser2010constructive; @achlioptas2016random], and used to some extent for colouring problems [@achlioptas2016focused; @przybylo2016facial; @esperet2013acyclic; @grytczuk2013new; @dujmovic2016nonrepetitive]. The same idea has been used for analysing optimal data structures, e.g., for cell probing [@brody2015adapt; @larsen2012cell] and for sampling [@bringmann2013succinct]. Despite these results, and despite popular blogposts [@tao2009moser; @fortnow2009kolmogorow], the method still does not seem to be widely known outside of these communities.
We use this technique to prove that no phase transition occurs at $c=1$. More precisely, we show the following.
\[thm:main\] There exists an $\varepsilon>0$ such that, for any (strictly) monotone function $f$ and any constant $0<c\leq 1+\varepsilon$, the [$(1 + 1)$-EA]{}with mutation parameter $c$ requires $O(n \log^2 n)$ steps until it finds the maximum of $f$, and it visits an expected number of $O(n)$ search points. The same remains true if the initial search point of the algorithm is chosen by an adversary.
To be more precise, we show that there are constants ${\varepsilon}, C>0$ such that for all $0<c\leq 1+{\varepsilon}$ the runtime is at most $C/c \cdot n\log^2 n$, and the number of search points is at most $C \cdot n$.
For the proof, it turns out to be natural to not measure performance in the number of time steps, but in the number of *updates*, i.e. the number of times $X_{t+1} \neq X_t$. Note that because of the greedy nature of the algorithm, the number of updates coincides with the number of visited search points (minus one).
We first give some intuition on the behaviour of the algorithm and on our proof. For a general monotone function $f$, it is natural to measure the progress of the $(1+1)$-EA is terms of number of one-bits in the current search point. In order to make an update, it is necessary to flip at least one zero-bit into a one-bit, since otherwise the offspring would be rejected due to monotonicity (unless it is identical to the parent, in which case there is no update either). Thus, if the average update does not flip too many ones to zeros, the number of ones in the current search point will tend to $n$ efficiently. For a small mutation parameter (specifically for $c<1$), this is indeed true as the average number of ones flipped to zeros is at most $c$, and this remains true for update steps. For larger $c$, one might expect this to still hold as, intuitively, any offspring with more ones flipped to zeros than zeros flipped to ones should be unlikely to be fitter than its parent. However, the reason this intuition fails for sufficiently large $c$ is that one can “trick” the algorithm by weighing a fraction of the remaining zeros much higher than most ones already in the search point. Then, whenever one of these zeroes is flipped, the algorithm will happily keep the offspring, regardless of how many ones are flipped to zeros in the process.
Based on this intuition, we define *good* and *bad* updates. The bad updates capture cases in which a zero-bit with a disproportionally high weight is flipped. Then we study the entropy of the update steps of the algorithms. On the one hand, we will analyze the algorithm in a forward manner to give a lower bound on the entropy of each update. On the other hand, we give a backwards encoding (from last step to first) of the updates steps, and this encoding saves some bits in bad update steps. Since the expected number of bits needed for the encoding is lower bounded by the entropy, we get an upper bound for the expected number of bad update steps. This, in turn, gives us a linear upper bound for the expected number of update steps. Finally, the runtime bound follows by a slight refinement of the calculation, in which we compute how many steps we need to decrease the number of zero-bits from $2^k$ to $2^{k-1}$, for $k=\log n, \dots, 1$.
Preliminaries: properties of single updates
===========================================
The idea of this section is to collect properties of single update steps. Throughout this section we will use the following notation. We assume that we are in an (arbitrary, but fixed) state $y\in\{0,1\}^n$. We also assume that $k$ denotes the number of ones in $y$. We denote by $Y'$ the string obtained by flipping each coordinate independently with probability $c/n$. With ${\ensuremath{{\mathcal E}_{keep}}}$ we denote the event that $f(Y') \ge f(y)$, corresponding to the event that $y'$ is accepted as the new state. We also denote by $U$ the number of bits that are zero in $y$, but one in $Y'$ (upflips) and by $D$ the number of bits that are one in $y$, but zero in $Y'$ (downflips).
In this section we will repeatedly make use of the following fact. Let $X$ be a random variable, and ${\ensuremath{\mathcal{E}}}$ be some event. With ${\ensuremath{\mathds 1}}_{\ensuremath{\mathcal{E}}}$ we denote the indicator variable for the event ${\ensuremath{\mathcal{E}}}$. Then by the law of conditional expectation, $$\label{eq:conditional:expectation}
{\ensuremath{\mathbbm{E}}}[X|{\ensuremath{\mathcal{E}}}] = {{\ensuremath{\mathbbm{E}}}[X\cdot {\ensuremath{\mathds 1}}_{\ensuremath{\mathcal{E}}}]} / {\Pr[{\ensuremath{\mathcal{E}}}]}.$$
Expected number of bits flips
-----------------------------
${\ensuremath{\mathbbm{E}}}[U+D]$ is easily computed by linearity of expectation to be equal to $c$. However, we are interested in ${\ensuremath{\mathbbm{E}}}[U+D\mid {\ensuremath{{\mathcal E}_{keep}}}]$. Observe that $\Pr[{\ensuremath{{\mathcal E}_{keep}}}]$ is at least the probability that we flip exactly one zero-bit to a one and no other bit. Thus, $\Pr[{\ensuremath{{\mathcal E}_{keep}}}] \ge (n-k)\frac{c}n(1-\frac{c}n)^{n-1}\geq \frac{n-k}n ce^{-c}$, where the latter step holds for all $0<c<2$ and $n$ sufficiently large, and follows from the expansion $1-c/n = e^{-c/n + 2c^2/n^2 +O(1/n^3)}$. Observe also that ${\ensuremath{\mathds 1}}_{{\ensuremath{{\mathcal E}_{keep}}}} \le {\ensuremath{\mathds 1}}_{{\ensuremath{\mathcal{A}}}_1}+\ldots+ {\ensuremath{\mathds 1}}_{{\ensuremath{\mathcal{A}}}_{n-k}}$, where ${\ensuremath{\mathcal{A}}}_i$ denotes the event that we flip the $i$-th zero-bit to a one, where $1\le i\le n-k$. Thus implies that for $0<c<2$, and $n$ sufficiently large, $$\label{eq:UplusD}
{\ensuremath{\mathbbm{E}}}[U+D\mid {\ensuremath{{\mathcal E}_{keep}}}]
\le \frac{\sum_{i=1}^{n-k}{\ensuremath{\mathbbm{E}}}[(U+D)\cdot {\ensuremath{\mathds 1}}_{{\ensuremath{\mathcal{A}}}_i}]}{\Pr[{\ensuremath{{\mathcal E}_{keep}}}]}
\le \frac{(n-k)\frac{c}n(1+c)}{ \frac{n-k}n ce^{-c}} = (1+c)e^c.$$
Change in the number of ones {#sec:number:ones}
----------------------------
Our goal in this section is to (lower) bound the change in the number of ones, i.e., to bound ${\ensuremath{\mathbbm{E}}}[U-D\mid {\ensuremath{{\mathcal E}_{keep}}}]$. Clearly, ${\ensuremath{\mathbbm{E}}}[U \mid {\ensuremath{{\mathcal E}_{keep}}}] \ge 1$ by monotonicity. To bound $D$, note that it is intuitively clear that ${\ensuremath{\mathbbm{E}}}[D \mid {\ensuremath{{\mathcal E}_{keep}}}] \leq {\ensuremath{\mathbbm{E}}}[D] \leq c$. A full proof can be found in the appendix. For $c<1$ we thus get from that ${\ensuremath{\mathbbm{E}}}[U - D | {\ensuremath{{\mathcal E}_{keep}}}] \ge 1-c$. Standard drift arguments thus imply that the expected number of updates till EA reaches the all-ones string is $O(n)$.
In order to also be able to apply a similar argument for $c > 1$, we need to be more careful. What we will do is to partition updates into good and bad ones, i.e., we let ${\ensuremath{{\mathcal E}_{keep}}}= {\ensuremath{{\mathcal E}_{good}}}\uplus {\ensuremath{{\mathcal E}_{bad}}}$ and define ${\ensuremath{{\mathcal E}_{bad}}}$ in such a way that ${\ensuremath{{\mathcal E}_{bad}}}$ happens “rarely” [and]{} ${\ensuremath{\mathbbm{E}}}[U - D | {\ensuremath{{\mathcal E}_{good}}}]$ is positive.
To do so, observe that whenever $U=1$, say bit $i$ is flipped from a zero to a one, we can attribute a *value* to each 1-bit in $y+e_i$ (that is, to the indices $j\in \{i\}\cup \{ a\in[n] : y^a=1\}$) according to $$val_{y+e_i}(j) := f(y+ e_i) - f(y + e_i - e_j),$$ where $e_1, e_2, \dots, e_n$ denotes the standard basis vectors, and $+$ and $-$ denotes vector addition and subtraction respectively. It is natural to think of this value as the “cost” of flipping bit $j$ to a zero. Indeed if $Y'$ is obtained from $y$ by flipping bit $i$ from a zero to a one, and bits $j_1, j_2, \dots$ from ones to zeros, then if at least one of the $j$-bits, say $j_1$, has a strictly higher value than $i$, we have $$f(Y') \leq f(y+ e_i - e_{j_1}) = val_{y+e_i}(i)-val_{y+e_i}(j)<0,$$ which means such a $Y'$ will never be kept. Similarly, if $j_1$ has equal value to $i$, and this is not the only $1$-bit flipped to a zero, then $Y'$ will not be kept.
With this notion at hand we say that an update from $y$ to $Y'$ belongs to [${\mathcal E}_{bad}$]{} iff
1. $Y' \neq y$ and $f(Y') \leq f(y)$, i.e., we actually make an update,
2. there is exactly one zero-bit $i$ that is flipped, i.e, $U=1$, and
3. there exist at least $(1-\alpha) n$ one-bits in $y + e_i$ whose value is strictly smaller than the value of bit $i$.
To get some intuition behind this definition, observe that it indeed captures cases in which the number of ones may likely [*decrease*]{}: we only flip one bit from zero to one [*and*]{} there are many candidate bits for which we may be able to flip two or more of them back to zero. This intuition is formalized by the following proposition.
\[prop:excex\] For any $0 \leq \alpha \leq \frac{1}{2}$ and $0 \leq c\leq 1/(1-\alpha)$, we have for all $n \geq 3$ $$\begin{aligned}
\mathbb{E}[U-D \mid {\ensuremath{{\mathcal E}_{bad}}}] &\geq 1-c,\\
\mathbb{E}[U-D \mid{\ensuremath{{\mathcal E}_{good}}}] &\geq \left(1-\frac{c}{n}\right)^{\alpha n} \left( 1-(1-\alpha)c\right) \geq e^{-2\alpha c} \left( 1-(1-\alpha)c\right).\end{aligned}$$
Let $\mathcal{B}_i$ denote the event that the $i$-th one-bit in $y$ gets flipped in $Y'$. Then, by linearity of expectation, we have $\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{bad}}}] \geq 1-\sum_{i=1}^k\Pr[\mathcal{B}_i\mid {\ensuremath{{\mathcal E}_{bad}}}].$ By Bayes’ Theorem, we have $\Pr[\mathcal{B}_i\mid {\ensuremath{{\mathcal E}_{bad}}}] = \frac{c}{n}\cdot\frac{\Pr[{\ensuremath{{\mathcal E}_{bad}}}\mid \mathcal{B}_i]}{\Pr[{\ensuremath{{\mathcal E}_{bad}}}]}\leq \frac{c}{n},$ where the last step follows by a simple coupling argument, similar as in Section \[sec:number:ones\]. Hence $\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{bad}}}] \geq 1-c\frac{k}{n}\geq 1-c,$ as desired.
As for ${\ensuremath{{\mathcal E}_{good}}}$, the second inequality follows from $1-(1-\alpha)c \geq 0$, since $1-x \geq e^{-2x}$ for all $0\leq x \leq 2/3$. For the first inequality, let ${\ensuremath{\mathcal{U}}}$ denote the set of indices of zero-bits in $y$ that get flipped to one-bits in $Y'$. Then $$\mathbb{E}[U-D\mid{\ensuremath{{\mathcal E}_{good}}}] = \sum_A \Pr[\mathcal{U}=A\mid{\ensuremath{{\mathcal E}_{good}}}] \cdot \mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}\cap\{\mathcal{U}=A\}].$$ Thus, it suffices to estimate $\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}\cap\{\mathcal{U}=A\}]$ for any set $A\subset[n]$ such that $\Pr[\mathcal{U}=A\vert{\ensuremath{{\mathcal E}_{good}}}]$ is non-zero.
If ${\ensuremath{\left| A \right|}}\geq 2$, the same argument as for ${\ensuremath{{\mathcal E}_{bad}}}$ gives $$\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}\cap\{\mathcal{U}=A\}] \geq 2-c.$$ It remains to consider the case of ${\ensuremath{\left| A \right|}}=1$, say $A=\{i\}$. In this case, let $k'=\min(k, \lfloor (1-\alpha) n\rfloor)$, and order the one-bits in $y$, $j_1, j_2, \dots j_k$, in descending order with respect to $val_{y+e_i}(j)$ with ties broken arbitrarily. In order for $\mathcal{U}=A$ to be compatible with a good update, we can assume that the values of $j_1, j_2, \dots j_{k-k'}$ must be greater than or equal to the value of $i$.
With these definitions at hand, we write $$\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}\cap\{{\ensuremath{\mathcal{U}}}=A\}] = \frac{1}{\Pr[{\ensuremath{{\mathcal E}_{good}}}\mid{\ensuremath{\mathcal{U}}}=A]}\mathbb{E}[\mathbbm{1}_{{\ensuremath{{\mathcal E}_{good}}}}(U-D)\mid{\ensuremath{\mathcal{U}}}=A].$$ Note that, conditioned on ${\ensuremath{\mathcal{U}}}=A$, $\mathbbm{1}_{{\ensuremath{{\mathcal E}_{good}}}}(U-D)=0$ whenever one of the bits $j_1, \dots j_{k-k'}$ are flipped. This is because this is either the only bit flipped to a zero, in which case $U-D=1-1=0$, or one additional bit is flipped to a zero, in which case $f(Y')<f(y)$. Whenever the bits $j_1, \dots j_{k-k'}$ remain ones, on the other hand, we can lower bound $\mathbbm{1}_{{\ensuremath{{\mathcal E}_{good}}}}(U-D)$ by one minus the number of bits among $j_{k-k'+1}, \dots j_{k}$ that get flipped. Thus $$\mathbb{E}[\mathbbm{1}_{{\ensuremath{{\mathcal E}_{good}}}}(U-D)\mid{\ensuremath{\mathcal{U}}}=A] \geq (1-\frac{c}{n})^{k-k'}(1-\frac{c}{n}k').$$ By assumption, we have $1-\frac{c}{n}k' \geq 1-(1-\alpha)c \geq 0$ which means that $$\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}\cap\{{\ensuremath{\mathcal{U}}}=A\}] \geq \mathbb{E}[\mathbbm{1}_{{\ensuremath{{\mathcal E}_{good}}}}(U-D)\mid{\ensuremath{\mathcal{U}}}=A] \geq (1-\frac{c}{n})^{\alpha n}(1-(1-\alpha)c),$$ as desired. The proposition follows by observing that $2-c \geq 1-(1-\alpha)c$ as $2-c-1+(1-\alpha)c = 1-\alpha c \geq 1-\frac{\alpha}{1-\alpha} \geq 0,$ where in the second to last step we used $c \leq 1/(1-\alpha)$.
Note that for any fixed $0< \alpha \leq 1/2$ the considered range in Proposition \[prop:excex\] also contains some values $c>1$. Moreover, for the considered range the lower bound for $\mathbb{E}[U-D\mid {\ensuremath{{\mathcal E}_{good}}}] $ is positive. If we could thus show that ${\ensuremath{{\mathcal E}_{bad}}}$ occurs only sufficiently rarely, then we might hope to be able to bound the number of updates. This is what we will do with the entropy compression argument.
Entropy of an update step
-------------------------
In this section we study the entropy of a single update starting from a fixed state $y$. We refer the reader who is not familiar with information theory to the introduction in [@cover2012elements]. Naturally, the entropy will depend on $y$. Recall that the random variables $U$ and $D$ denote the number of upflips and downflips, respectively.
\[prop:entropy:update\] For any $c < 4/3$, any $0\le k < n$ and any $y\in\{0, 1\}^n$ with exactly $k$ ones we have $$\mathbb{H}(Y'\mid {\ensuremath{{\mathcal E}_{keep}}}) \geq {\ensuremath{\mathbbm{E}}}[\log_2\left({n-k\choose U}{k+U \choose D}\right) \mid {\ensuremath{{\mathcal E}_{keep}}}],$$ where $\mathbb{H}(Y'\mid {\ensuremath{{\mathcal E}_{keep}}})$ denotes the binary entropy of the conditional distribution of $Y'$ given ${\ensuremath{{\mathcal E}_{keep}}}.$
Let ${\ensuremath{\mathcal{A}}}_{u,d}$ denote the set of all strings $z\in\{0,1\}^n$ such that $f(z)\ge f(y)$ and such that $z$ can be obtained from $y$ by flipping precisely $u$ zeros to ones and $d$ ones to zeros. Then $$p_{keep} := \Pr[{\ensuremath{{\mathcal E}_{keep}}}] = \sum_{u=1}^{n-k}\sum_{d= 0}^k \sum_{z\in {\ensuremath{\mathcal{A}}}_{u,d}} (\frac{c}n)^{u+d}(1-\frac{c}n)^{n-u-d}.$$ To simplify notation we write $$q_{u,d}:={\ensuremath{\left| \mathcal{A}_{ud} \right|}}\left(\frac{c}{n}\right)^{u+d}\left(1-\frac{c}{n}\right)^{n-u-d}$$ and $$a_{u,d}:= \left(\frac{c}{n}\right)^{u+d}\left(1-\frac{c}{n}\right)^{n-u-d}.$$ By the definition of entropy we have $$\begin{aligned}
\mathbb{H}(Y' \mid {\ensuremath{{\mathcal E}_{keep}}})&=-\frac{1}{p_{keep}}\sum_{u=1}^{n-k}\sum_{d= 0}^k \sum_{z\in {\ensuremath{\mathcal{A}}}_{u,d}} a_{ud} \log_2(\frac{a_{u,d}}{p_{keep}})\\
&=\frac{1}{p_{keep}}\sum_{u=1}^{n-k}\sum_{d= 0}^k q_{ud} \log_2(\frac{p_{keep}}{a_{u,d}}).\end{aligned}$$ We want to show that this is at least as large as the expectation in the statement of the proposition. With $$b_{u,d}:={n-k \choose u}{k+u \choose d}$$ we can write this expectation as $${\ensuremath{\mathbbm{E}}}[\log_2\left({n-k\choose U}{k+U \choose D}\right) \mid {\ensuremath{{\mathcal E}_{keep}}}]
=\frac{1}{p_{keep}}\sum_{u=1}^{n-k}\sum_{d=0}^k q_{u,d}\log_2\left(b_{u,d}\right).$$ Hence the proposition follows if we can show that the difference is non-negative. That is, we have to show that $$\frac1{p_{keep}}\sum_{u=1}^{n-k}\sum_{d=0}^k q_{u,d} \log_2\left(\frac{p_{keep}}{a_{u,d}b_{u,d}}\right) \stackrel!\ge 0.$$ Multiplying by $p_{keep} \ge 0$ and partitioning the log amounts to showing that $$\label{eq:Hdiff}
\Delta := p_{keep}\log_2 p_{keep}+\sum_{u=1}^{n-k}\sum_{d=0}^k q_{ud}\log_2\left(\frac{1}{a_{u,d}b_{u,d}}\right) \stackrel!\ge 0.$$ To this end, an elementary calculation shows that the product $a_{u,d}b_{u,d}$ is either maximized by the case $u=1$, $d=0$, or by $u=d=1$. We defer the calculation to the appendix.
Assume first that $a_{1,1}b_{1,1} \leq a_{1,0}b_{1,0}$. Then the left hand side in satisfies $$\Delta \ge p_{keep}\log_2 p_{keep}+p_{keep}\log_2(\frac{1}{a_{1,0}b_{1,0}})$$ which is non-negative, as $p_{keep} \ge q_{1,0}=a_{1,0}b_{1,0}$. For the other case, assume $a_{1,1}b_{1,1} > a_{1,0}b_{1,0}$. Writing $p_{keep} = (1+x)q_{1,0}$ for some $x\ge 0$ and recalling that $q_{1,0} = a_{1,0}b_{1,0}$ we obtain $$\begin{aligned}
\Delta &\ge (1+x)q_{1,0}\cdot(\log_2((1+x) q_{1,0}) +q_{1,0}\log_2(\frac{1}{q_{1,0}}) + xq_{1,0}\log_2(\frac{1}{a_{1,1}b_{1,1}})\\
&=q_{1,0}\cdot\left( (1+x)\log_2(1+x) + x \log_2\frac{a_{1,0}b_{1,0}}{a_{1,1}b_{1,1}}\right).\end{aligned}$$ Now observe that $\frac{a_{1,0}b_{1,0}}{a_{1,1}b_{1,1}} = \frac{1-c/n}c \frac{n}{k+1} > 3/4$, for $n$ sufficiently large. The claim now follows from $(1+x)\log_2(1+x) +x \log_2(3/4) \ge 0$ for all $x\ge 0$.
Entropy of the Markov chain
===========================
The aim of this section is to provide bounds on the entropy of the sequence of updates in the $(1+1)$-EA. Given the sequence $(X_t)_{t=0}^\infty$ of search points, as generated by the algorithm when started in some fixed state $X_0=x\in\{0, 1\}^n$, we define $T$ as the number of updates, that is, the number of times $t=0, 1, \dots$ such that $X_t\neq X_{t+1}$. For each $t=0, 1, \dots T$, we denote by $Y_t$ the state of the algorithm after $t$ update steps. To simplify notation later on, we want $Y_t$ be defined for all $t\geq 0$, so we define $Y_{t}$ to be the all-one string if $t > T$. We note that as $(X_t)_{t=0}^\infty$ is a Markov chain, so is $(Y_t)_{t=0}^\infty$. More precisely, the transition probabilities of $(Y_t)_{t=0}^\infty$ are the ones obtained from transition probabilities of $(X_t)_{t=0}^\infty$ by removing self-transitions from all states besides the all ones state.
We will use ${\# 0}_t$, ${\# 1}_t$ to denote the number of zero-bits and one-bits in $Y_t$, respectively. Furthermore, we use $U_t$ and $D_t$ to denote the number of upflips (zero-to-one) and downflips (one-to-zero) from $Y_t$ and $Y_{t+1}$, respectively.
Recall that the entropy of (the trajectory of) the Markov chain $(Y_t)_{t=0}^{\infty}$, as described above, can be written as $$\mathbb{H}((Y_t)_{t=0}^\infty) =
\sum_{t=0}^\infty \mathbb{H}(Y_{t+1} \mid Y_t),$$ where $$\mathbb{H}(Y_{t+1} \mid Y_t) =-\sum_{y}\Pr[Y_t=y] \mathbb{H}(Y_{t+1} \mid Y_t=y).$$ This entropy is finite almost surely, since the Markov chain will converge to the absorbing state almost surely. The next two propositions bound this entropy from below and above.
\[entropy:lower\] $$\mathbb{H}((Y_t)_{t=0}^{\infty}) \ge {\ensuremath{\mathbbm{E}}}\left[\sum_{t=0}^{T-1}\log_2\left({{\# 0}_t\choose U_{t}}{{\# 1}_t+U_t \choose D_{t}}\right)\right].$$
Observe that, for any $y$, the term $\mathbb{H}(Y_{t+1} \mid Y_t=y)$ does not depend on $t$, as $(Y_t)_{t=0}^{\infty}$ is a time-homogenous Markov chain. If we thus let $g(y):=\mathbb{H}(Y_{t+1}\mid Y_t=y)$, we get $$\mathbb{H}(Y_{t+1} \mid Y_t) ={\ensuremath{\mathbbm{E}}}[g(Y_t)]$$ and thus $$\mathbb{H}((Y_t)_{t=0}^\infty) \geq \sum_{t=0}^\infty \mathbb{E}\left[g(Y_t)\right] = \mathbb{E}\left[\sum_{t=0}^\infty g(Y_t)\right] = \mathbb{E}\left[\sum_{t=0}^{T-1} g(Y_t)\right],$$ where in the last step we have used that the conditional entropy of an update is zero once we have reached the all-ones state. As Proposition \[prop:entropy:update\] implies that ${\ensuremath{\mathbbm{E}}}[g(Y_t)] \ge {\ensuremath{\mathbbm{E}}}[\log_2\left({{\# 0}_t\choose U_{t}}{{\# 1}_t+U_t \choose D_{t}}\right)]$, the proposition follows.
\[entropy:upper\] There exists a constant $C$ such that for all $0<\alpha<1$ and for all $0<c < 2$, $$\mathbb{H}((Y_t)_{t=0}^{\infty}) \le C{\ensuremath{\mathbbm{E}}}[T] -\log_2(1/\alpha) {\ensuremath{\mathbbm{E}}}[T_{bad}] + {\ensuremath{\mathbbm{E}}}\left[\sum_{t=0}^{T-1}\log_2\left({{\# 0}_{t+1}\choose D_{t}}{{\# 1}_{t+1} +D_t\choose U_{t}}\right) \right]\!,$$ where $T_{bad}$ denotes the number of bad updates (as defined in Section \[sec:number:ones\]).
Recall that the entropy represents a lower bound on the expected number of bits needed to represent all information of the process. Thus, the expected length of any encoding strategy for the traces of the chain will form an upper bound on the entropy. We proceed as follows. We encode the process *backwards*, i.e. we start from the all-ones vector that is the unique absorbing state of the process. For each update we encode
[0.8cm0.2cm0.6cm]{}
whether the update is good or bad,
the number of downflips $D_t$ and the number of upflips $U_t$,
the actual choice of which $D_t$ zero-bits in $Y_{t+1}$ were the bits that were flipped from one to zero in the update from $Y_t$ to $Y_{t+1}$, and
the actual choice of which $U_t$ one-bits in $Y_{t+1}$ were the bits that were flipped from zero to one.
To mark the global end of our encoding, we are somewhat wasteful: we start the encoding of each update with a one-bit and conclude the whole encoding with a single zero-bit. For each update we encode $(i)$ with a single bit, and $D_t$ and $U_t$ with $D_t+U_t+2$ bits using a unary encoding. From we deduce that there exists a constant $C > 0$ such the expected length of the encoding of $(i)$ and $(ii)$ for all updates is bounded by $C{\ensuremath{\mathbbm{E}}}[T]$. For the encoding of $(iii)$ observe that the $D_t$ bits that correspond to downflips have to be chosen from the zero-bits in $Y_{t+1}$. We thus can encode the actual choice by $\lceil\log_2{{\# 0}_{t+1}\choose D_t} \rceil$ bits. For the encoding in $(iv)$ we distinguish between good and bad updates. For a good update we proceed similarly as in $(iii)$. As the $D_t$ bits that correspond to upflips have to be chosen from one-bits in $Y_{t+1}$, we can encode the actual choice by $\lceil\log_2{{\# 1}_{t+1}\choose U_t} \rceil$ bits. For a bad update we can be more efficient in $(iv)$. Observe first that a bad update implies $U_t=1$. We thus need to specify only a single bit. Recall also that the definition of bad updates implies that there exist at least $(1-\alpha)n$ one-bits in $Y_t$ that have lower value than the bit that we want to flip. As the number of one-bits in $Y_t$ is bounded by ${\# 1}_{t+1}+D_t$ we thus see that we have at most ${\# 1}_{t+1}+D_t - (1-\alpha)n \le \alpha({\# 1}_{t+1}+D_t )$ bits from which we can choose the bit that corresponds to the (single) upflip. We can thus encode the choice of this bit with $\lceil\log_2(\alpha( {\# 1}_{t+1}+D_t)) \rceil $ bits, which is less than $
\lceil\log_2{{\# 1}_{t+1}+D_t\choose U_t} \rceil -\log_2(1/\alpha)+1$ bits. The claimed bound in the proposition follows by collecting all terms.
Proof of the theorem
====================
We first obtain a bound on the expected number of bad updates by comparing upper and lower bound on the entropy of the Markov chain.
\[bound:bad\] If the algorithm starts in a state with exactly $k$ ones, then $${\ensuremath{\mathbbm{E}}}[T_{bad}] \le {\textstyle\frac{C}{\log_2(1/\alpha)}}{\ensuremath{\mathbbm{E}}}[T] + {\textstyle\frac{1}{\log_2(1/\alpha)}} {\log_2{n\choose k}},$$ where $C$ is the constant from Proposition \[entropy:upper\].
Collecting and rearranging the terms from Propositions \[entropy:lower\] and \[entropy:upper\] we get $$\begin{aligned}
\label{eq:telescoping}
{\ensuremath{\mathbbm{E}}}[T_{bad}] \le {\textstyle\frac{C}{\log_2(1/\alpha)}}{\ensuremath{\mathbbm{E}}}[T] + {\textstyle\frac{1}{\log_2(1/\alpha)}} {\ensuremath{\mathbbm{E}}}\left[\log_2\left(\prod_{t=0}^{T-1}
\frac{{{\# 0}_{t+1}\choose D_{t}}{{\# 1}_{t+1}+D_t \choose U_{t}}}{{{\# 0}_t\choose U_{t}}{{\# 1}_t+U_t \choose D_{t}}}\right)\right].\end{aligned}$$ Using the formulas ${\# 1}_{t+1}={\# 1}_t+U_t-D_t$ and ${\# 0}_{t+1}={\# 0}_t-U_t+D_t$, it is easy to see that, for any $0\leq t\leq T-1$, $$\begin{aligned}
\frac{{{\# 0}_{t+1}\choose D_{t}}{{\# 1}_{t+1}+D_t \choose U_{t}}}{{{\# 0}_t\choose U_{t}}{{\# 1}_t+U_t \choose D_{t}}}&=\frac{{\# 0}_{t+1}!}{{\# 0}_t!}\cdot\frac{{\# 1}_{t+1}!}{{\# 1}_t!}.\end{aligned}$$ Hence, the product in is telescoping, and we get $${\ensuremath{\mathbbm{E}}}[T_{bad}] \le {\textstyle\frac{C}{\log_2(1/\alpha)}}{\ensuremath{\mathbbm{E}}}[T] + {\textstyle\frac{1}{\log_2(1/\alpha)}} {\ensuremath{\mathbbm{E}}}\left[\log_2\left(
\frac{{\# 0}_{T}! \cdot {\# 1}_T!}{{\# 0}_0!\cdot{\# 1}_0!}\right)\right].$$ The claim now follows from $
\frac{{\# 0}_{T}! \cdot {\# 1}_T!}{{\# 0}_0!\cdot{\# 1}_0!} = {{n\choose {\# 1}_0}}/{{n\choose {\# 1}_T}} = {n\choose {\# 1}_0}
$, as ${\# 1}_0=k$ and ${\# 1}_T=n$.
From this and Proposition \[prop:excex\] we obtain an upper bound on the number of updates.
\[bound:total\] There exists an ${\varepsilon}>0 $ and $\beta >0$ such that for any $0<c\le 1+{\varepsilon}$ the following holds. If the algorithm starts in a state with exactly $k$ ones, then $${\ensuremath{\mathbbm{E}}}[T] \le \beta (n-k) + \beta {\log_2{n\choose k}}.$$
Observe that $\sum_{t=0}^\infty({\# 1}_{t+1}-{\# 1}_{t}) = n-{\# 1}_0$. If we thus start in a state with exactly $k$ ones, then whenever $0 \leq \alpha\leq 1/2$ and $0 < c \leq 1/(1-\alpha)$ we get from linearity of expectation and Proposition \[prop:excex\] that $$\begin{aligned}
\label{eq:excex}
n -k &= \sum_{t=0}^\infty\mathbb{E}[{\# 1}_{t+1}-{\# 1}_{t}] \\
&= \sum_{t=0}^\infty\left(\mathbb{E}[{\# 1}_{t+1}-{\# 1}_{t} | {{\ensuremath{{\mathcal E}_{good}}}(t)}] \cdot \Pr[{{\ensuremath{{\mathcal E}_{good}}}(t)}] +\right.\nonumber\\
&\hspace*{2.5cm}\left.\mathbb{E}[{\# 1}_{t+1}-{\# 1}_{t} | {{\ensuremath{{\mathcal E}_{bad}}}(t)}] \cdot \Pr[{{\ensuremath{{\mathcal E}_{bad}}}(t)}] \right)\nonumber\\
&\stackrel{P\ref{prop:excex} }\geq e^{-2\alpha c} (1-(1-\alpha)c)\cdot {\ensuremath{\mathbbm{E}}}[T-T_{bad}] + (1-c)\cdot {\ensuremath{\mathbbm{E}}}[T_{bad}],\nonumber\end{aligned}$$ where we have used $\mathcal{E}_{good/bad}(t)$ to denote the event that $Y_t$ is not the all-ones string and the update from $Y_t$ to $Y_{t+1}$ is bad or good, respectively. (Note that once the Markov chains has reached the all-ones state, neither ${\ensuremath{{\mathcal E}_{bad}}}(t)$ nor ${\ensuremath{{\mathcal E}_{good}}}(t)$ can occur and the contribution of the corresponding term in the last sum is zero, as is desired.)
For ease of notation let $D := e^{-2\alpha c} (1-(1-\alpha)c)$. Then the above can be rewritten as $$n-k\ge D{\ensuremath{\mathbbm{E}}}[T] - (D-1+c) {\ensuremath{\mathbbm{E}}}[T_{bad}].$$ Let $C$ be the constant from Proposition \[bound:bad\]. We may assume $C>1/2$. For a fixed $0<\alpha <1$ (to be determined later) we will choose an ${\varepsilon}>0$ such that $1+{\varepsilon}< 1/(1-\alpha)$. Then holds for all $c \in [0,1+{\varepsilon}]$, and we have $D>0$ for any such $c$. Moreover, since $D$ is a continuous function, it attains a minimum $D_{\min} = D_{\min}(\alpha)>0$ on the compact interval $c \in [0,1+{\varepsilon}]$. We may assume that ${\varepsilon}\leq 1/3$ and ${\varepsilon}< D_{\min}/2$, and by Proposition \[bound:bad\], for all $0<c \leq 1+{\varepsilon}$, $$n-k \ge
\left(D - \frac{C}{\log_2(1/\alpha)}(D-1+c) \right) \cdot {\ensuremath{\mathbbm{E}}}[T] - \frac{D-1+c}{\log_2(1/\alpha)} \log_2{n\choose k}.$$ Set now $\alpha=2^{-2C} <1/2$ and observe that then the term in front of ${\ensuremath{\mathbbm{E}}}[T]$ is equal to $\frac12(D+1-c) \geq \frac14 D_{\min}$; the claim of the proposition follows.
As $\log_2{n\choose k}\le n$, the claim on the number of search points follows immediately from Proposition \[bound:total\]. For the bound on the number of steps we have to be more careful, as we also have to count the number of steps *between* updates. To do so the following observation is useful. Suppose we are in a state with exactly $k$ ones. Then the probability that we flip exactly one zero-bit in the next step is $(n-k)\frac{c}n(1-\frac{c}n)^{n-k} \geq \tfrac12 ce^{-c}\frac{n-k}n$, which holds for $n$ sufficiently large. Note that we may assume $c < 4/3$, in which case we obtain a probability of at least $\tfrac1{10} c\frac{n-k}n$. As we will accept any of these moves, we thus see that the expected number of steps until the next update is at most $\tfrac{10}{c}\cdot n/(n-k)$. This together with Proposition \[bound:total\] implies a bound of $O(n^2/c)$ on the running time of the [$(1 + 1)$-EA]{}. To get a quasilinear bound we partition the trace of the algorithm in phases. For this, let $S_k \subseteq\{0,1\}^n$ denote the set of strings with at most $2^k-1$ zeros, where $0\le k \le \lfloor\log_2 n\rfloor-1 =: k_0$.
Before reaching a state from $S_{k_0}$ the expected time between two updates is just $O(1/c)$ (as then we still have a constant fraction of zeros to choose from). From Proposition \[bound:total\] we thus know that the Markov chain $(X_t)_{t=0}^\infty$ will reach a state from $S_{k_0}$ in $O(n/c)$ steps. Next we consider the phases in which we start in a state from $S_{k}$ and terminate (the phase) when we reach a state from $S_{k-1}$ (for the first time). Denote by $T_k$ the number of update steps in this phase. We can use the bound from Proposition \[bound:total\] (that considers the run of the Markov chain until it reaches the all-ones string) to obtain $${\ensuremath{\mathbbm{E}}}[T_k] \le \beta (n-(n-2^{k}+1)) + \beta \log_2 {n\choose n-2^{k}+1} \le 2 \beta 2^{k}\log n.$$ As we argued above, in this phase the expected number of steps between updates is bounded by $O(n/(2^{k}c))$ (as we always have at least $2^{k-1}$ zeros). The expected number of steps in this phase is thus bounded by $O(n\log n /c)$, where the hidden constant holds uniformly for all phases. Since we assumed $c$ to be a positive constant, the expected number of steps per phase is $O(n\log n)$. As the number of phases is $O(\log n)$, the theorem follows.
Proof that ${\ensuremath{\mathbbm{E}}}[D \mid {\ensuremath{{\mathcal E}_{keep}}}] \leq {\ensuremath{\mathbbm{E}}}[D]$.
======================================================================================================================
In this section we prove that ${\ensuremath{\mathbbm{E}}}[D \mid {\ensuremath{{\mathcal E}_{keep}}}] \leq {\ensuremath{\mathbbm{E}}}[D]$, which is used in Section \[sec:number:ones\]. For a one-bit $j$, let ${\ensuremath{\mathcal{B}}}_j$ denote the event that this bit is flipped into zero. We can divide all potential offspring into pairs $y_0$, $y_1$ which agree in all bits except that $y_0^j = 0$, but $y_1^j=1$. Note that $\Pr[Y' = y_0 \mid {\ensuremath{\mathcal{B}}}_j] = \Pr[Y' = y_1 \mid \neg {\ensuremath{\mathcal{B}}}_j]$, because all bits are flipped independently. Moreover, since $f(y_0) < f(y_1)$, we have the implication “$y_0$ is accepted $\implies$ $y_1$ is accepted”. Hence, $$\begin{aligned}
\label{eq:condition_on_flip}
\Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j] & = \sum_{y_0 \in \{0,1\}^n, y_0^{j}=0} \Pr[Y' = y_0] \cdot {\ensuremath{\mathds 1}}_{{\ensuremath{{\mathcal E}_{keep}}}}(y_0) \nonumber \\
& \leq \sum_{y_1 \in \{0,1\}^n, y_1^{j}=1} \Pr[Y' = y_1] \cdot {\ensuremath{\mathds 1}}_{{\ensuremath{{\mathcal E}_{keep}}}}(y_1) = \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid \neg {\ensuremath{\mathcal{B}}}_j],\end{aligned}$$ and therefore, $$\begin{aligned}
\Pr[{\ensuremath{{\mathcal E}_{keep}}}] & = \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j] \cdot \Pr[{\ensuremath{\mathcal{B}}}_j] + \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid \neg{\ensuremath{\mathcal{B}}}_j] \cdot \Pr[ \neg{\ensuremath{\mathcal{B}}}_j] \\
&\stackrel{\eqref{eq:condition_on_flip}}{\geq} \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j] \cdot \Pr[{\ensuremath{\mathcal{B}}}_j] + \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j] \cdot (1-\Pr[{\ensuremath{\mathcal{B}}}_j]) = \Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j].\end{aligned}$$ In particular, this implies $$\Pr[{\ensuremath{\mathcal{B}}}_j \mid {\ensuremath{{\mathcal E}_{keep}}}] = \frac{\Pr[{\ensuremath{{\mathcal E}_{keep}}}\mid {\ensuremath{\mathcal{B}}}_j] \Pr[{\ensuremath{\mathcal{B}}}_j]}{\Pr[{\ensuremath{{\mathcal E}_{keep}}}]} \leq \Pr[{\ensuremath{\mathcal{B}}}_j].$$ Since this holds for all one-bits $j$, and since $D = \sum_j {\ensuremath{\mathds 1}}_{{\ensuremath{\mathcal{B}}}_j}$, where the sum runs over all one-bits $j$, we obtain ${\ensuremath{\mathbbm{E}}}[D \mid {\ensuremath{{\mathcal E}_{keep}}}] \leq {\ensuremath{\mathbbm{E}}}[D] \leq c$.
Missing details in the proof of Proposition \[prop:entropy:update\]
===================================================================
Here we show that for $c<4/3$ and $n$ large enough, the product $a_{u,d}b_{u,d}$ is maximized either for $u=1$, $d=0$, or for $u=d=1$. As a reminder, we repeat the definitions of $a_{u,d}$ and $b_{u,d}$. $$\begin{aligned}
a_{u,d} & := \left(\frac{c}{n}\right)^{u+d}\left(1-\frac{c}{n}\right)^{n-u-d}.\\
b_{u,d} &:={n-k \choose u}{k+u \choose d}\end{aligned}$$
We first observe $a_{u,d} / a_{u-1,d-1} < \frac{17}9 \frac1{n^2}$ for all $u,d \geq 1$. Thus $$\frac{b_{u,d}}{b_{u-1,d-1}} = \frac{{n-k\choose u}}{{n-k\choose u-1}} \cdot\frac{{k+u\choose d}}{{k+u-1\choose d-1}} = \frac{n-k-u+1}{u}\cdot \frac{k+u-d+1}{d} \le \frac{n^2}{ud}$$ implies that $a_{u,d}b_{u,d} \le a_{u-1,d-1}b_{u-1,d-1}$ for all $u,d\ge 2$. Similarly, since for $u \geq 1$, $$\frac{b_{u,1}}{b_{u-1,1}} =\frac{{n-k\choose u}}{{n-k\choose u-1}} \cdot\frac{{k+u\choose 1}}{{k+u-1\choose 1}} = \frac{n-k-u+1}{u}\cdot \frac{{k+u}}{{k+u-1}} \stackrel{k\ge d=1}\le \frac{3}{2u} \cdot n$$ we deduce that $a_{u,1}b_{u,1} \le a_{1,1}b_{1,1}$ for all $u\ge 2$ and from $$\frac{b_{1,d}}{b_{1,d-1}} = \frac{{k+1\choose d}}{{k+1\choose d-1}} = \frac{k+1-d+1}{d}\le\frac{n}d$$ we get $a_{1,d}b_{1,d} \le a_{1,1}b_{1,1}$ for all $d\ge 2$. Together, this gives $a_{u,d}b_{u,d} \le a_{1,1}b_{1,1}$ for all $u,d \ge 1$. Finally, for $u \geq 2$ and $n$ sufficiently large, $$\frac{a_{u,0}b_{u,0}}{a_{u-1,0}b_{u-1,0}} =\frac{\frac{c}{n}}{(1-\tfrac{c}{n})}\cdot \frac{{n-k\choose u}}{{n-k\choose u-1}} = \frac{c}{n-c}\cdot \frac{n-k-u+1}{u} \le \frac{n}{n-c}\cdot \frac{c}{u} \leq 1,$$ so $a_{u,0}b_{u,0} \le a_{1,0}b_{1,0}$ for all $u\ge 2$. Altogether, we thus have $$a_{u,d}b_{u,d} \le \max\{a_{1,0}b_{1,0}, a_{1,1}b_{1,1}\} \qquad\text{for all $u\ge 1, d\ge 0$},$$ as required.
[^1]: We define “monotone” in a way that otherwise might rather be called “strictly monotone”, for ease of terminology. Note that we can’t expect efficient runtimes for functions which are monotone in a non-strict sense. For example, we could have $f(x) = 0$ for $x\not\equiv 1$ and $f(1)=1$ otherwise and the search for the optimal solution amounts to searching a needle in a hay stack. Therefore, we define monotone functions in this strict sense in our paper.
[^2]: This algorithm is called *Random Local Search*. Despite its good behaviour on monotone functions it has severe limitations in practice. For example, other than the [$(1 + 1)$-EA]{}it can’t escape local optima and is highly susceptible to any form of noise.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the semiclassical propagation of squeezed Gau[ß]{}ian states. We do so by considering the propagation theorem introduced by Combescure and Robert [@CR97] approximating the evolution generated by the Weyl-quantization of symbols $H$. We examine the particular case when the Hessian $H^{\prime\prime}\left(X_{t}\right)$ evaluated at the corresponding solution $X_{t}$ of Hamilton’s equations of motion is periodic in time. Under this assumption, we show that the width of the wave packet can remain small up to the Ehrenfest time. We also determine conditions for “classical revivals” in that case. More generally, we may define recurrences of the initial width. Some of these results include the case of unbounded classical motion. In the classically unstable case we recover an exponential spreading of the wave packet as in [@CR97].'
address: 'Institut für Theoretische Physik, Universität Ulm, Albert-Einstein-Allee 11, D89069-Ulm, Germany'
author:
- Eric Sträng
bibliography:
- '/home/estraeng/Desktop/sp\_prop/bib\_jes.bib'
title: '[Localization of quantum wave packets]{}'
---
Introduction
============
Localization of quantum states in phase space is a prerequisite in some semiclassical treatments of quantum evolution. In the classically chaotic case, the width of an initially localized Gau[ß]{}ian increases exponentially [@CR97] up to the so called Ehrenfest time $T_{E}$, i.e., the time up to which quantum dynamics can be approximated by classical dynamics. In the case of regular classical motion it can be shown that this width grows algebraically in the semiclassical parameter $\hbar$ up to $T_{E}$.
In certain applications, it is necessary to propagate states semiclassically for long times. This demands control of the width of the state. An example is the construction of quasi-modes proposed by Paul and Uribe [@PaUr93]. One refers to a quasi-mode as a state $\psi$ which is a solution of a corresponding spectral problem of an operator $\widehat{H}$ up to some small discrepancy $\delta$, i.e., $$\begin{aligned}
\left\Vert\widehat{H}\psi-E\psi\right\Vert & < & \delta \end{aligned}$$ which insures, in the case of discrete spectrum, that there exist at least one eigenvalue of $\widehat{H}$ in the interval $\left[E-\delta,\,E+\delta\right]$. The approximation proposed by Paul and Uribe uses the semiclassical propagation of coherent states over closed classical trajectories leading to the well known Bohr-Sommerfeld quantization rule in one dimension. The eigenvalues $E_{n}$ of the Hamiltonian $\widehat{H}$ are given by the quantization condition $$\begin{aligned}
\int_{\mathscr{C}_{E_{n}}}p\,\mathrm{d}q & = & 2\pi\hbar\left(n+\frac{1}{2}\right)+\mathcal{O}\left(\hbar^{2}\right),\quad n\in\mathds{N},\end{aligned}$$ where the energy shells $\mathscr{C}_{E}=\left\lbrace \left(p,\,q\right):\,H\left(p,\, q\right)=E\right\rbrace $ are closed curves.
Certain examples suggest that there exist systems (other than the harmonic oscillator) where the propagated width of an initial Gau[ß]{}ian remains small for long times. Such a behavior is exhibited by the propagation of Gau[ß]{}ians generated by some perturbed periodic Schrödinger operators like the Wannier-Stark Hamiltonian $$\begin{aligned}
\widehat{H}_{\mathrm{WS}}\left( \varepsilon \right) & = & -\frac{\hbar^{2}}{2}\Delta_{x}+V_{\Gamma}\left( x\right)+\varepsilon x \end{aligned}$$ in the limit of small perturbations $\varepsilon$. Here $V_{\Gamma}\left(x \right)$ denotes a periodic potential with respect to a lattice $\Gamma\cong \mathds{Z}^{d}$. It is known that the band structure of the spectrum of the unperturbed operator is preserved for small enough perturbations $\varepsilon$. Numerical studies [@Har04; @Witt04] show that an initially localized Gau[ß]{}ian in momentum space defined on such an energy band apparently remains Gau[ß]{}ian for long times. The evolving states carry out oscillations. The center of the Gau[ß]{}ian can oscillate in position space describing so called Bloch oscillations [@Blo28]. Alternatively, the width of the Gaussian can oscillate, whereupon the center remains fixed, describing so called breathing modes. In both cases, the state returns to the initial state after an oscillation period up to a small error.
We study the evolution generated by a class of operators on $L^{2}\left(\mathds{R}^{d}\right)$ defined as Weyl quantizations of classical symbols $H\left(X\right)$ with the property$$\begin{aligned}
H^{\prime\prime}\left(X_{t}\right) & = &
H^{\prime\prime}\left(X_{t+T}\right)
,\quad\forall t\in\mathds{R},\label{eq:percond}\end{aligned}$$ where $X_{t}$ denotes the solutions to Hamilton’s equations of motion and $H^{\prime\prime}\left(X\right)$ is the Hessian of $H$ with respect to $X$. In one dimension, the condition (\[eq:percond\]) is satisfied by bounded classical motion or unbounded motion in a periodic potential. We particularly study the evolution of initial Gau[ß]{}ian (or squeezed) states semi-classically, i.e., asymptotically as $\hbar\searrow0$, when $t\nearrow\infty$. We focus our attention to the spreading of such wave packets.
In sections \[sec:Preliminaries:-semiclassical-propagation\] and \[sec:A semiclassical propagation theorem\] we shortly review the semiclassical propagation of Gau[ß]{}ian coherent states. In section \[EhrT\] we give some known results on the validity of the approximation. We then make, in section \[sec:Floquet-theory\] and section \[sec:A-Uniform-bound\], statements about the approximate Gau[ß]{}ian state given by this semiclassical propagation using Floquet theory. These properties are then brought back to the true evolution in section \[sec:Discussion\].
\[sec:Preliminaries:-semiclassical-propagation\]Preliminaries: semiclassical propagation of wave packets
--------------------------------------------------------------------------------------------------------
We will work in the context of self adjoint operators defined on $L^{2}\left(\mathds{R}^{d}\right)$ that are $\hbar-$Weyl quantizations of symbols. To a smooth $\left(C^{\infty}\right)$ classical symbol $b\left( X\right)$, i.e., a function on the phase space $T^{*}\mathds{R}^{d}\cong\mathds{R}^{2d}$, there corresponds an operator on $L^{2}\left(\mathds{R}^{d}\right)$, $\widehat{b}:=\mathrm{Op}_{\hbar}^{w}\left[b\right]$, defined by$$\begin{aligned}
\mathrm{Op}_{\hbar}^{w}\left[b\right]\psi\left(x\right) & := & \frac{1}{\left(2\pi\hbar\right)^{d}}\int_{\mathds{R}^{2d}}b\left(\frac{x+y}{2},\,\xi\right)\psi\left(y\right)e^{\frac{\mathrm{i}}{\hbar}\left(x-y\right)\xi}\,\mathrm{d}y\mathrm{d}\xi.\end{aligned}$$
The following conditions will be assumed \[conditions\]:
1. The classical Hamiltonian $H:\,\mathds{R}^{2d}\rightarrow\mathds{R}$ is a smooth function.
2. $H\in S\left(m\right)$, i.e., for all multi-indices $\alpha$, $\beta$, there exists $K_{\alpha ,\,\beta}>0$ such that $$\left|\partial_{p}^{\alpha}\partial_{q}^{\beta}H\left(p,\, q\right)\right|\leq K_{\alpha ,\,\beta}\left(1+\left|p\right|^{2}+\left|q\right|^{2}\right)^{\frac{m}{2}}.$$
3. The corresponding classical equation of motion is given by $$\begin{aligned}
\frac{\mathrm{d}X_{t}}{\mathrm{d}t} & = & \mathcal{J}H^{\prime}\left(X_{t}\right),\end{aligned}$$ where $\mathcal{J}$ is the symplectic unity $$\begin{aligned}
\mathcal{J} & := & \left(\begin{array}{cc}
0 & -\mathds{1}_{d\times d}\\
\mathds{1}_{d\times d} & 0\end{array}\right),\end{aligned}$$ and $H^{\prime}$ is the gradient of $H$ with respect to $X$. Furthermore, we denote by $\Phi_{H}^{t}:\,\mathds{R}^{2d}\rightarrow\mathds{R}^{2d},
\,X_{0}\mapsto X_{t}=\Phi_{H}^{t}\left(X_{0}\right)$ the corresponding classical flow.
4. The $\hbar-$Weyl quantization of $H\left(X\right)$, $\widehat{H}:=\mathrm{Op}_{\hbar}^{w}\left[H\right]$, is an essentially self-adjoint operator on $L^{2}\left(\mathds{R}^{d}\right)$ and generates a unitary time evolution $\forall t\in\mathds{R},$$$\begin{aligned}
\widehat{U}\left(t \right) & : & L^{2}\left(\mathds{R}^{d}\right)\rightarrow L^{2}\left(\mathds{R}^{d}\right)\\
& & \psi\left(0\right)\mapsto\psi\left(t\right)\end{aligned}$$ corresponding to the Schrödinger equation$$\begin{aligned}
\mathrm{i}\hbar\frac{\partial\psi}{\partial t} & = & \widehat{H}\psi,\label{eq:Schroedinger}\end{aligned}$$ i.e., we will write $\widehat{U}\left(t\right)=e^{-\frac{\mathrm{i}}{\hbar}\widehat{H}t}.$
The Weyl calculus also allows a representation of quantum mechanical wave functions on phase space. This is given by the Wigner function of the state $u\in L^{2}\left(\mathds{R}^{d}\right),$ $$\begin{aligned}
W\left[u\right]\left(p,\, q\right) & := & \int_{\mathds{R}^{d}}\overline{u}\left(q+\frac{y}{2}\right)u\left(q-\frac{y}{2}\right)e^{\frac{\mathrm{i}}{\hbar} p y}\,\mathrm{d}y\label{WigRep}.\end{aligned}$$ The Wigner function of a Gau[ß]{}ian defines a positive measure on phase space.
\[sec:A semiclassical propagation theorem\]A semiclassical propagation theorem
------------------------------------------------------------------------------
With these assumptions we give a short summary of the method of semiclassical propagation of coherent states introduced by Combescure and Robert [@CR97]. Similar constructions have also been considered in the past. See, e.g., Hagedorn [@Ha81; @Ha85] and references therein.
The idea is to expand the exact Hamiltonian $\widehat{H}$ along the classical flow generated by the symbol $H$ up to second order, whereupon the approximate time dependent Hamiltonian, $\widehat{H}_{2}\left(t\right)$, is the $\hbar-$Weyl quantization of $$\begin{aligned}
H_{2}\left(t,\,Y\right) & := & H\left(X_{t}\right)+\left(Y-X_{t}\right)^{T}
H^{\prime}\left(X_{t}\right)+\frac{1}{2}\left(Y-X_{t}\right)^{T}H^{\prime\prime}
\left(X_{t}\right)\left(Y-X_{t}\right).\end{aligned}$$ The propagation of normalized wave functions (squeezed states)[^1] of the form $$\begin{aligned}
\psi_{Z}^{\left(p,\, q\right)}\left(x\right) & := &
\frac{\det\left(\Im\left(Z\right)\right)^{\frac{1}{4}}}
{\left(\pi\hbar\right)^{\frac{d}{4}}}e^{\frac{\mathrm{i}}{\hbar}
\left(p^{T}\left(x-q\right)+\left(x-q\right)^{T}\frac{Z}{2}\left(x-q\right)\right)},
\,Z\in\Sigma_{d},\,\left(\begin{array}{c}p \\ q\end{array}\right)\in\mathds{R}^{2d}
\label{eq:squstate}\end{aligned}$$ by quadratic Hamiltonians is well known [@Fol89]. By $\Sigma_{d}$, we mean the $d-$dimensional Siegel upper half space, i.e., the set of symmetric $d\times d$ matrices with positive, non-degenerate imaginary part [@Fol89]. The quadratic form $Z$ describes the shape of the wave packet and it should be underlined that it is independent of $\hbar$. In its Wigner representation (see eq.(\[WigRep\])) $\psi_{Z}^{\left(p,\, q\right)}$ is a Gau[ß]{}ian centered around $X=\left(\begin{array}{c}
p\\
q\end{array}\right).$ This Wigner function is given by $$\begin{aligned}
W\left[\psi_{Z}^{X}\right]\left(Y\right) & = & \left(\frac{1}
{\pi\hbar}\right)^{d}\exp\left(-\frac{1}{\hbar}
\left(Y-X\right)^{T}G\left(Y-X\right)\right)\end{aligned}$$ where $$\begin{aligned}
G & := & \left(\begin{array}{ccc}
\Im\left(Z\right)^{-1} & \quad & -\Im\left(Z\right)^{-1}\Re \left(Z\right)\\
\quad & \quad & \quad\\
-\Re\left(Z \right)\Im\left(Z\right)^{-1} & \quad &
\Im\left(Z \right)+\Re\left(Z \right)\Im
\left(Z \right)^{-1}\Re\left(Z \right) \end{array}\right)\end{aligned}$$ is independent of $\hbar$.
The unitary evolution, $\widehat{U}_{2}\left(t\right)$, generated by $\widehat{H}_{2}\left(t\right)$, acts on a squeezed state by translation and metaplectic action, i.e., $$\begin{aligned}
\widehat{U}_{2}\left(t\right) & = & e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}
\widehat{\mathcal{T}}\left( X_{t}\right)
\widehat{\mathcal{M}}\left(S_{t}\right)
\label{eq:metaplect}\end{aligned}$$ where $\widehat{\mathcal{T}}$ is the translation operator on $\mathds{R}^{2d}$, i.e., $\forall Y \in\mathds{R}^{2d},\,
=:\left(
\begin{array}{c}
\xi \\
q
\end{array}
\right)$ and $u\in L^{2}\left(\mathds{R}^{d}\right)$ we have $$\begin{aligned}
\widehat{\mathcal{T}}\left(Y\right)W\left[u\right]\left(Z\right)
& := & e^{\frac{\mathrm{i}}{\hbar}\left(\xi^{T} \widehat{x}-q^{T}\widehat{p}\right)}W\left[u\right]\left(Z\right)\\
& = & W\left[u\right]\left(Z-Y\right),\end{aligned}$$ where we denote by $\widehat{p}$ the momentum operator and by $\widehat{x}$ the position operator. The metaplectic operator $\widehat{\mathcal{M}}\left( F\right)$ is the quantization of a linear symplectomorphism on $\mathds{R}^{2d}$ given by the symplectic matrix $F$. These operators form a double-valued unitary representation[^2] of the linear symplectomorphism of $\mathds{R}^{2d}$. $S_{t}$ denotes the flow differential. The classical flow $\Phi_{H}^{t}$ is a symplectomorphism which ensures that the flow differential is a symplectic matrix. Furthermore, $S_{t}$ satisfies $$\begin{aligned}
\frac{\mathrm{d}S_{t}}{\mathrm{d}t} & = & \mathcal{J}H^{\prime\prime}
\left(X_{t}\right)S_{t},\label{eq:action}\\
\,S_{0} & = & \mathds{1}_{2d\times2d}.\nonumber \end{aligned}$$ In the prefactor of eq. (\[eq:metaplect\]) we have used $$\begin{aligned}
\Theta\left(t\right)&:=&\mathcal{W}\left(t\right)+\hbar\mu, \end{aligned}$$ where $$\begin{aligned}
\mathcal{W}\left(t\right) & := & \int_{0}^{t} \left(p_{\tau}^{T}\dot{q}_{\tau}-H\left(p_{\tau},\,
q_{\tau}\right)\right)\, d\tau,\end{aligned}$$ is the action of the classical trajectory and $\mu$ is the Maslov index of the classical trajectory. We have here expressed the solution of Hamilton’s equations in terms of the canonical variables $$\begin{aligned}
\left(\begin{array}{c}
p_{t}\\
q_{t}\end{array}\right) & := & X_{t}.\end{aligned}$$
Acting with (\[eq:metaplect\]) on (\[eq:squstate\]), one obtains at time $t$ a new Gau[ß]{}ian state $e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}\psi_{Z_{t}}^{X_{t}}$ up to a phase. The Gau[ß]{}ian is centered around $X_{t}$ in phase space and has a quadratic form $Z_{t}\in\Sigma_{d}$, is given explicitly by the group action [@Fol89]
$$\begin{aligned}
Z_{t} & = & S_{t}\left[Z_{0}\right]\\
& = & \left(A_{t}Z_{0}+B_{t}\right)\left(C_{t}Z_{0}+D_{t}\right)^{-1}\quad
,\,Z_{0}\in\Sigma_{d},\end{aligned}$$
i.e., linear fractional transformation on $\Sigma_{d}$. The matrix $S_{t}$ is here written by means of the $d\times d$ blocks $A_{t}$, $B_{t}$, $C_{t}$ and $D_{t}$, i.e., $$\begin{aligned}
S_{t} & = & \left(\begin{array}{cc}
A_{t} & B_{t}\\
C_{t} & D_{t}\end{array}\right).\end{aligned}$$ The quadratic form of the Wigner transform of $\psi_{Z_{t}}^{X_{t}}$, is explicitly given by $$\begin{aligned}
G_{t} & = & \left(S_{t}^{-1}\right)^{T}G_{0}S_{t}^{-1}\label{eq:width}.\end{aligned}$$
The evolution of the approximate wave packet can be perceived as being generated by rotation and scaling of the Gau[ß]{}ian profile in $\mathds{R}^{2d}$.
The difficulty of this scheme resides in the control of errors made by the approximation of $\widehat{U}\left(t\right)$ in terms of $\widehat{U}_{2}\left(t\right)$. We do not dwell on the details, but just state the result (for proof and a thorough description see [@CR97]). The approximation can be checked to all orders in $\hbar$. For this purpose, one defines the following approximating state $$\begin{aligned}
\Psi^{\left(N\right)}\left(t,\, x\right) & := & e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}\sum_{0\leq j\leq N}\hbar^{\frac{j}{2}}\pi_{j}\left(X_{t},\,\frac{x}{\sqrt{\hbar}},\, t\right)\psi_{Z_{t}}^{X_{t}}\left(x\right),\end{aligned}$$ where $\pi_{j}\left(X_{t},\,\frac{x}{\sqrt{\hbar}},\, t\right)$ are polynomials in $x\left/\sqrt{\hbar}\right.$ and $X_{t}$ of degree smaller than or equal to $3j$ with time dependent coefficients.
\[thm:propagation\_control\_CR\](Combescure, Robert) Under the above mentioned assumptions and for an initial Gau[ß]{}ian state $\psi_{Z_{0}}^{X_{0}}$ centered in the phase space representation at $X_{0}\in\mathds{R}^{2d}$, for every $N\in\mathds{N}$ there exists $C<\infty$ such that $\forall\hbar\in\left(0,\,\hbar_{0}\right],\,\hbar_{0}>0$, $$\begin{aligned}
\left\Vert \widehat{U}\left(t\right)\psi_{Z_{0}}^{X_{0}}-\Psi^{\left(N\right)}\left(t\right)\right\Vert &
\leq & C_{N}\hbar^{\frac{N+1}{2}}te^{3\gamma t}\label{eq:normunitaries}\end{aligned}$$ where $0\leq\gamma<\infty$ is the Lyapunov exponent of the classical motion.
We recall that a Lyapunov exponent is a measure of the exponential stability of the solutions of a differential equation upon change of initial conditions. In the case of classical motion, this is given by the Lyapunov exponent defined as $$\begin{aligned}
\gamma & := & \max_{k}\left[ \lim_{t\rightarrow\infty}\sup\left( \frac{\ln\left(s_{k}\left( t\right)\right)}{t}\right) \right] \label{defclasslyapunov}\end{aligned}$$ where $s_{k}\left( t\right)$ are the singular values of $S_{t}$. The Lyapunov exponent $\gamma$ hence satisfies $$\begin{aligned}
\left\Vert S_{t}\right\Vert_{\mathrm{HS}} & < & c_{0}e^{\gamma \left\vert t\right\vert}\label{lyapunovinequality}\end{aligned}$$ where $c_{0}<\infty$ is a positive constant. We denote by $\left\Vert M\right\Vert _{\mathrm{HS}}=
\sqrt{\mathrm{tr}\left(M^{\dagger}M\right)}$ the Hilbert-Schmidt norm of the matrix $M$. Hermitian conjugation is denoted by $M^{\dagger}$, and transposition by $M^{T}$.
Ehrenfest time and spreading of wave packets\[EhrT\]
----------------------------------------------------
It is customary, in this context, to define what is known as the Ehrenfest time. The latter is a time scale up to which the above approximation is valid. We define the Ehrenfest time, denoted $T_{E}\left( \hbar \right)$, as the maximal time up to which
- the error $\left\Vert \widehat{U}\left(t\right)
\psi_{Z_{0}}^{X_{0}}-e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}
\psi_{Z_{t}}^{X_{t}}\right\Vert$ remains small,
- the exact state remains localized.
The latter ensures that the approximation retains a physical meaning, i.e., the above classical approximation makes no sense if the state does not remain localized.
The total width of the semiclassically evolved Gau[ß]{}ian is $$\begin{aligned}
\sigma_{t}\left(X_{0};\,\hbar\right):=\Delta x^{2}\left(\hbar,\, t
\right)+\Delta p^{2}\left(\hbar,\, t\right) & = & \frac{\hbar}{2}
\mathrm{tr}\left(G_{t}\right),\label{eq:semclawidth}\end{aligned}$$ where we have used the variance $$\begin{aligned}
\Delta x^{2} & = & \left\langle e^{\frac{\mathrm{i}}{\hbar}
\Theta\left(t\right)}\psi_{Z_{t}}^{X_{t}},
\, \widehat{x}^{2}e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}
\psi_{Z_{t}}^{X_{t}}\right\rangle _{L^{2}\left(\mathds{R}^{d}\right)}-
\left\langle e^{\frac{\mathrm{i}}{\hbar}\Theta\left(t\right)}
\psi_{Z_{t}}^{X_{t}},\, \widehat{x}e^{\frac{\mathrm{i}}{\hbar}\Theta
\left(t\right)}\psi_{Z_{t}}^{X_{t}}\right\rangle_{L^{2}
\left(\mathds{R}^{d}\right)}^{2},\end{aligned}$$ and similarly for the momentum operator.
From eq. (\[eq:width\]) one obtains that $$\begin{aligned}
\sigma_{t} & \leq & \sigma_{0}\left\Vert S_{t}\right\Vert^{2}_{\mathrm{HS}}.\end{aligned}$$ By Theorem \[thm:propagation\_control\_CR\], there will exist a constant $c_{1}$ such that $$\begin{aligned}
\left\Vert\widehat{U}\left(t\right)\psi_{Z_{0}}^{X_{0}}-e^{\frac{\mathrm{i}}{\hbar}
\Theta\left(t\right)}\psi_{Z_{t}}^{X_{t}}\right\Vert & \leq &
c_{1} \sqrt{\hbar} t e^{3\gamma t}.\end{aligned}$$ The time scale for which the errors are small is thus algebraic in $\hbar$ if $\gamma=0$. In the generic case, the errors remain small for logarithmic times in $\hbar$.
With this result, the errors $$\begin{aligned}
\Delta\left(t\right) & := & \left\langle \widehat{U}\left(t\right)\varphi_{0},\,
\widehat{s}\left(\widehat{U}\left(t\right)
\varphi_{0}\right)\right\rangle_{L^{2}\left(\mathds{R}^{d}\right)} -\left\langle \widehat{U}_{2}\left(t\right)\varphi_{0},\,\widehat{s}\left(\widehat{U}_{2}
\left(t\right)\varphi_{0}\right)\right\rangle_{L^{2}\left(\mathds{R}^{d}\right)} \label{eq:obserror}\end{aligned}$$ for propagating observables $\widehat{s}$ can be approximated explicitly. Of particular interest to us is the width operator $\widehat{s}:=\mathrm{Op}_{\hbar}^{w}\left[\left|Y\right|^{2}\right]$. One can characterize the times for which the error $\Delta\left(t\right)$ is small, e.g., $\Delta\left(t\right)=\mathcal{O}\left(\hbar^{\alpha}\right)$ for some $\alpha>0$. Again, using the Lyapunov inequality (\[lyapunovinequality\]) one finds that the width of the approximate state is bounded by $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & \leq & c_{3} e^{2\gamma\left|t\right|},
\quad c_{3}>0,\end{aligned}$$ and the error remains $\mathcal{O}\left(\hbar^{\alpha}\right)$ up to times of order $\frac{\left|\ln\left(\hbar\right)\right|}{6\gamma}.$ We may thus generically state that the Ehrenfest time is $$\begin{aligned}
T_{E}\left( \hbar \right) & \propto & \frac{\left|\ln\left(\hbar\right)\right|}{6\gamma}.\end{aligned}$$
In the integrable case, implying $\gamma=0$, the width grows like the square of the Hilbert-Schmidt norm of the flow differential, i.e., at most polynomially in time. The error remains small for times up to $\hbar^{-\frac{1}{2}}$. The Ehrenfest time thus is algebraic in $\hbar$.
Our aim is to characterize the spreading of the approximate state for a more specific class of classical motions.
Results
=======
In addition to the conditions imposed on symbols $H$ above (see section \[conditions\]), we assume that the Hessian $H^{\prime\prime}\left(X_{t}\right)$ is $T-$periodic, i.e., $$\begin{aligned}
H^{\prime\prime}\left(X_{t}\right) & = & H^{\prime\prime}\left(X_{t+T}\right),
\quad\forall t\in\mathds{R}.
\label{perhess}\end{aligned}$$
We will also utilize the following definition.
A **classical revival** at a time $t>0$ is the event that the approximate Gau[ß]{}ian given above is the initial one up to a phase factor, i.e., $$\begin{aligned}
\psi_{Z_{t}}^{X_{t}} & = & e^{\mathrm{i}\alpha_{t}}\psi_{Z_{0}}^{X_{0}}.\end{aligned}$$
\[sec:Floquet-theory\]Floquet theory
------------------------------------
According to the Floquet theorem [@F1883], any linear differential equation with continuous $T-$periodic coefficients has a periodic solution of the second type, i.e., a solutions which satisfy $$\begin{aligned}
f\left(t+T\right) & = & \upsilon f\left(t\right),\quad
\upsilon\in\mathds{C},\,\forall t\in\mathds{R}.\end{aligned}$$ In particular, the linear vector differential equation $$\begin{aligned}
\frac{\mathrm{d}}{\mathrm{d}t}f & = & A\left(t\right)f,\end{aligned}$$ where $A\left( t\right) $ is continuous and satisfies $A\left(t+T\right)=A\left(t\right),\,\forall t\in\mathds{R}$, has a fundamental matrix of the form (see e.g. [@CoL55]) $$\begin{aligned}
F_{t} & = & M^{-1}e^{L t}MU\left(t\right),\quad M\in\mathrm{GL}\left(n,\,
\mathds{C}\right),\end{aligned}$$ where $L$ is a diagonal matrix and $U\left(t\right)$ is a $T-$periodic matrix. By definition, a fundamental matrix is a full rank matrix whose columns satisfy the differential equation, i.e., the linear combinations of the columns of the fundamental matrix span the full space of solutions of the differential equation.
We will call the elements of $L$ the *Floquet exponents* of the fundamental system.
The result of this section will be summarized in the following way.
\[pro:periodicwidth\] **If** $H^{\prime\prime}\left(X_{t}\right)$ is periodic **and** the Floquet exponents of $\,S_{t}$ are purely complex **then** the width of a Gau[ß]{}ian propagated semiclassically by $\widehat{U}_{2}\left(t\right)$ will remain unchanged at multiples of the smallest classical period $T$. Furthermore, **if** the classical flow is periodic **then** classical revivals will occur.
Under the condition (\[perhess\]), Floquet theory states the existence of a fundamental Floquet matrix for the linear differential eq. (\[eq:action\]), $$\begin{aligned}
S_{t} & = & M^{-1}e^{\Lambda t}M F_{t},\quad
M\in\mathrm{GL}\left(2d,\,\mathds{C}\right),\end{aligned}$$ where $e^{\Lambda t}$ is the diagonal matrix with entries $e^{2\pi\lambda_{i}\frac{t}{T}},\,\lambda_{i}\in\mathds{C}$, $F_{t}$ has minimal period $T$, and we have chosen $$\begin{aligned}
F_{0} & = & S_{0} = \mathds{1}_{2d\times2d}.\end{aligned}$$
One directly concludes that $$\begin{aligned}
S_{k T} & = & M^{-1}e^{k T\Lambda}M,\quad\forall k\in\mathds{Z}.\end{aligned}$$ for multiples $k$ of the classical period $T$.
Furthermore, the real fundamental matrix defined by $S_{t}\in\mathrm{Sp}\left(2d,\,\mathds{R}\right)$ has a unique polar decomposition, i.e., there exists [@Fol89] an orthogonal matrix $\mathcal{Q}_{t}\in\mathrm{O}\left(2d\right)\cap\mathrm{Sp}\left(2d,\,\mathds{R}\right)$ and a positive definite matrix $\mathcal{P}_{t}\in\mathrm{Sp} \left(2d,\,\mathds{R}\right)$ such that $$\begin{aligned}
S_{t} & = & \mathcal{Q}_{t}\mathcal{P}_{t}.\end{aligned}$$ The width of the approximate squeezed state at time $t$ is (see eq. (\[eq:semclawidth\])) $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & = & \mathrm{tr}\left(
\left(S_{t}^{-1}\right)^{T}G_{0}
S_{t}^{-1}\right),\end{aligned}$$ and since any symplectic matrix $A$ satisfies [@Fol89] $$\begin{aligned}
A^{-1} & = & \mathcal{J}A^{T}\mathcal{J}^{-1},\end{aligned}$$ we can write by eq. (\[eq:width\]), $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & = & \frac{\hbar}{2}\mathrm{tr}
\left(\mathcal{Q}_{t}\mathcal{P}_{t}\mathcal{J}^{T}
G_{0}\mathcal{JP}_{t}^{T}\mathcal{Q}_{t}^{T}\right)=
\frac{\hbar}{2}\mathrm{tr}
\left(\mathcal{P}_{t}\mathcal{J}^{T}G_{0}\mathcal{J}
\mathcal{P}_{t}^{T}\right),\end{aligned}$$ since $\mathcal{Q}_{t}\in\mathrm{O}\left(2d\right)$ .
We are hence confronted with two cases.
- Either $S_{k T}$ is orthogonal, i.e., $\mathcal{P}_{k T}=\mathds{1}_{2d\times2d}$; this corresponds to $$\begin{aligned}
\sigma_{k T}\left(\hbar\right) & = & \frac{\hbar}{2}\mathrm{tr}
\left(G_{0}\right) = \sigma_{0}\left(\hbar\right),\quad\forall k\in\mathds{Z}.\end{aligned}$$ The orthogonality of $S_{k T}$ implies that it has $2d$ singular values $1$. It is furthermore similar to $e^{k T\Lambda}$ $\left( F_{k T}=\mathds{1}_{2d\times2d}\right)$, which hence is unitary so the Floquet exponents $\frac{2\pi\lambda_{i}}{T}$ are purely complex or zero. By definition, the Lyapunov exponent of the classical trajectories is $$\begin{aligned}
\gamma & := & \max_{k}\left[ \lim_{t\rightarrow\infty}\sup\left(\frac{\ln\left(
s_{k}\left(X_{t}\right)\right)}{t}\right)\right] ,\end{aligned}$$ where $s_{k}\left( t\right)$ are the singular values of the flow differential $S_{t}$, i.e., the eigenvalues of $$\begin{aligned}
F_{t}^{\dagger}\left(M^{-1}e^{t\Lambda}M\right)^{\dagger}
M^{-1}e^{t\Lambda}M
F_{t} & = & F_{t}^{\dagger}F_{t}.\end{aligned}$$ Noting that $F_{t}$ is bounded (and periodic), we see directly that if $$\begin{aligned}
\max_{i}\left(\Re\left( \lambda_{i}\right)\right) & = & 0,\end{aligned}$$ the classical motion is linearly stable.
<!-- -->
- $\mathcal{P}_{k T}\neq\mathds{1}_{2d\times2d},$ which corresponds to $$\begin{aligned}
\sigma_{k T}\left(\hbar\right) & = & \frac{\hbar}{2}\mathrm{tr}
\left(\mathcal{P}_{k T}\mathcal{J}^{T}G_{0}\mathcal{JP}_{k T}^{T}
\right)\\
& > & \sigma_{0}\left(\hbar\right).\end{aligned}$$ since $\mathcal{P}_{t}\in\mathrm{Sp}\left(2d,\,\mathds{R}\right)$ is strictly positive definite. This corresponds to the case when $\Re\left(
\lambda_{i}\right)\neq0,$ hence $\max_{i}\left( s_{i}\right)>0$. The classical motion is hence not linearly stable in this case.
Furthermore, if the (purely complex) Floquet exponents are rationally dependent, there exist some multiples $$\begin{aligned}
T_{R} & = & n_{R} T,\quad n_{R}\in\mathds{Z},\end{aligned}$$ of the classical period $T$ such that the orthogonal transformation at times $T_{R}$ reduces to unity. Indeed, if $n_{R}$ is the smallest common multiple of the denominators of the sequence $\left\{ \lambda_{i}\right\} $ defined by the Floquet exponents, i.e., $$\begin{aligned}
\lambda_{i} n_{R} & \in & \mathds{N},\,\forall i,\end{aligned}$$ we find $$\begin{aligned}
M^{-1}e^{\Lambda T_{R}}M & = & M^{-1}\mathrm{diag}\left\{ e^{2\pi\mathrm{i}
n_{i}}\right\} M,\quad n_{i}\in\mathds{Z},\\
& = & \mathds{1}_{2d\times2d}.\end{aligned}$$ The semiclassical approximation then is the initial Gau[ß]{}ian, if it is localized at the initial point $X_{0}$. This is the case if the flow is periodic. In this case we have a classical revival.
Note that these revivals are purely classical in the sense that the conditions only reflect properties of the classical motion and are hence independent of $\hbar$.
If $S_{k T}$ is orthogonal, the approximate Gau[ß]{}ian profile will be the initial one at classical periods. It is just rotated and translated in phase space.
\[sec:A-Uniform-bound\]A uniform bound
--------------------------------------
We can further characterize the approximate state in the case when the Floquet exponents are purely complex.
\[lem:Localization\] **If** $H^{\prime\prime}\left(X_{t}\right)$ is $T-$periodic **and** the Floquet exponents of $S_{t}$ are purely complex **then** the width $\sigma_{t}\left(\hbar\right)$ of a Gau[ß]{}ian state propagated by $\widehat{U}_{2}\left(t\right)$ satisfies the following uniform bound $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & \leq &
K\sigma_{0}\left(\hbar\right),\,\forall t\in\mathds{R}.
\label{eq:regularboundedwidth}\end{aligned}$$ Furthermore, $K:=e^{\kappa}$, where $\kappa$ is defined by $$\begin{aligned}
\kappa & := & 2 T\sup_{t\in\left[0,\, T\right]}\left\Vert \mathcal{J}
H^{\prime\prime}\left(X_{t}\right)\right\Vert _{\mathrm{HS}}.\end{aligned}$$
We consider eq. (\[eq:action\]) $$\left\{
\begin{array}{rll}
\frac{\mathrm{d}S_{t}}{\mathrm{d}t} & = &
\mathcal{J}H^{\prime\prime}\left(X_{t}\right)S_{t}\\
\\S_{0} & = & \mathds{1}.
\end{array}\right.$$ Starting at some initial time $k T,\, k\in\mathds{Z}$, $S_{t+k T}$ satisfies the Grönwall inequality $$\begin{aligned}
\left\Vert S_{t+k T}\right\Vert _{\mathrm{HS}} & \leq &
\left\Vert S_{k T}\right\Vert _{\mathrm{HS}} e^{\kappa},\,\forall t\in\left[0,\, T\right],\end{aligned}$$ during a classical period $T$ where $$\begin{aligned}
\kappa & = & 2 T\sup_{t\in\left[0,\, T\right]}\left\Vert
\mathcal{J}H^{\prime\prime}\left(X_{t}\right)\right\Vert _{\mathrm{HS}}.\end{aligned}$$ Furthermore, $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & = & \frac{\hbar}{2}\mathrm{tr}
\left(S_{t}\mathcal{J}^{T}G_{0}\mathcal{J}
S_{t}^{T}\right)\\
& = & \frac{\hbar}{2}\mathrm{tr}\left(G_{0}
S_{t}^{T}S_{t}\right)\leq\frac{\hbar}{2}\left|
\mathrm{tr}\left(G_{0}\right)\right|\left\Vert
S_{t}\right\Vert _{\mathrm{HS}}^{2}.\end{aligned}$$ Hence $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & \leq & \frac{\hbar}{2}\left|
\mathrm{tr}\left(\mathbf{g}_{0}\right)\right| e^{\kappa},\label{eq:boundbetweenpewriods}\end{aligned}$$ if $S_{t}$ has purely complex Floquet exponents since in this case $\sigma_{0}\left(\hbar\right)=\sigma_{k T}\left(\hbar\right),$ according to proposition \[pro:periodicwidth\].
We conclude with this choice that $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & \leq & K\sigma_{0}\left(\hbar\right)\end{aligned}$$ uniformly in time.
If $S_{t}$ is not orthogonal at the classical periods $T$, the latter bound takes the form $$\begin{aligned}
\sigma_{t}\left(\hbar\right) & \leq & K\sigma_{0}\left(\hbar\right)
e^{2\max_{i}\left(\Re\left(l_{i}\right)\right)\left|t\right|}\end{aligned}$$ where $l_{i}:=\frac{2\pi\lambda_{i}}{T}$ are the corresponding Floquet exponents of $S_{t}$ which is nothing else than a Lyapunov inequality where $\max_{i}\left(\Re\left(l_{i}\right)\right)$ plays the role of the Lyapunov exponent. Using the fact that what we are approximating is the Hilbert-Schmidt norm of the flow differential one can furthermore state the existence of a constant $c_{4}>0$ such that $$\begin{aligned}
\frac{e^{2\max_{i}\left(\Re\left(l_{i}\right)\right)\left|t\right|}}{c_{4}} & \leq\sigma_{t}
\left(\hbar\right)\leq & c_{4} e^{2\max_{i}\left(\Re\left(l_{i}\right)\right)\left|t\right|}.\end{aligned}$$ which determines the asymptotic behavior of the width in that case.
We note that the equality in eq. (\[eq:regularboundedwidth\]) is reached if the classical period is null, i.e., if the Hamiltonian has a constant Hessian $H^{\prime\prime}\left(X_{t}\right)$. This is satisfied by the harmonic oscillator. The flow differential $S_{t}$ is then an orthogonal matrix for all times. It is well known that the propagation is dispersion-less in this case.
\[sec:Discussion\]Discussion
============================
Theorem \[thm:propagation\_control\_CR\] allows us to trace back the properties of the approximation to the exact state $\mathcal{U}\left(t\right)\psi_{Z_{0}}^{X_{0}}$. According to the discussion in [@CR97] (see also section \[EhrT\]), this can be done with an error of order $\mathcal{O}\left(\hbar^{\alpha}\right),\,\alpha>0$ up to times $\hbar^{-\frac{1}{2}}$ if the Floquet exponents have zero real part (stable classical dynamics) and up to times $\frac{\left|\ln\left(
\hbar\right)\right|}{6 \nu},\,\nu=\max_{i}\Re\left(l_{i}\right)$ otherwise (unstable classical dynamics) as $\hbar \searrow 0$ and $t \nearrow \infty$.
In the stable case, the approximate state remains localized ad infinitum if this state remains localized between classical periods since the width of such a state is the same at classical periods. We wish to stress that we have defined the classical period as the minimal period of the Hessian $H^{\prime\prime}\left(X_{t}\right)$ which is not necessarily the period of the classical flow $\Phi_{H}^{t}$. An example is classical motion in a one dimensional periodic potential $V_{\Gamma}$ with energies $E$ such that $$\begin{aligned}
E & > & \sup_{x\in\mathds{R}}
\left(V_{\Gamma}\left(x\right)\right).\end{aligned}$$ In such a case $H^{\prime\prime}\left( X_{t}\right)$ is periodic although the flow $\Phi_{H}^{t}\left( X_{0}\right) $ is not. Furthermore, the approximation remains localized between classical periods if $$\begin{aligned}
\kappa & = & 2T
\sup_{t\in\left[0,\,T\right]}\left\Vert \mathcal{J}H^{\prime\prime}\left(X_{t}\right)
\right\Vert_\mathrm{HS}, \end{aligned}$$ the exponent in eq. (\[eq:boundbetweenpewriods\]), is small enough. We can hence state in that case, up to small errors $\mathcal{O}\left(\hbar^{\alpha}\right)$, localization of the exact propagation up to the Ehrenfest time $T_{E}\propto
\hbar^{-\frac{1}{2}}$. Note that the situation includes unbounded motion.
If the flow differential $S_{t}$ is periodic, it is clear that the shape will be the initial one at classical periods up to a small error. The resonance condition allows to state that this will also occur at some time if the flow differential is merely periodic of the second type and if the Floquet exponents are rationally dependent and purely imaginary. We may state this since the exact state remains localized and that our asymptotic considerations are valid as $\hbar
\searrow 0$ and $t \nearrow \infty$. The property extends also to the general stable case, i.e., in the case of rationally independent Floquet exponents. Recall that the quadratic form $Z_{t}$ is given by linear fractional transformation. In particular, $$\begin{aligned}
Z_{kT} & = & \left(A_{kT}Z_{0}+B_{kT}\right)\left(C_{kT}Z_{0}+D_{kT}\right)^{-1}\end{aligned}$$ where we have used the notation of section \[sec:A semiclassical propagation theorem\]. Since $S_{kT}\sim e^{\Lambda kT}$ which is unitary with eigenvalues $e^{\mathrm{i}\ell_{i}\frac{2\pi}{T}}=e^{\lambda_{i}\frac{2\pi}{T}}$ where $\ell_{i}\in\mathds{R}$, there exist $k-$independent matrices $a_{i}\in\mathds{C}^{d\times d}$ such that $Z_{kT}$ can be viewed as the image of the vector $\ell\in\mathds{T}^{d}$ with entries $\ell_{i}$ under the continuous map $$\begin{aligned}
\mathsf{e} & : & \mathds{T}^{d}\rightarrow\Sigma_{d}\\
& & \omega\mapsto\sum_{i=0}^{d}a_{i}e^{2\pi\mathrm{i}\omega_{i}}.\end{aligned}$$ Furthermore, the endomorphism $$\begin{aligned}
\tau & : & \mathds{T}^{d}\rightarrow\mathds{T}^{d}\\
& & L \mapsto L+\ell,\end{aligned}$$ where $\left( L_{i}+\ell_{i}\right)$ is defined modulo 1, is known to be ergodic with respect to Lebesgue measure on the $d-$torus since the frequencies $\ell_{i}$ are rationally independent. In this notation we have $Z_{0}=\mathsf{e}\left(0\right)=
\mathsf{e}\left(L\left( 0\right) \right)$. From the ergodicity of $\tau$ on $\mathds{T}^{d}$ we may state that for every $\varepsilon>0$ there exist $K\in\mathds{Z}$ such that $\left\Vert \tau^{K}\left(L\left( 0\right) \right)-L\left( 0\right) \right\Vert <\varepsilon$. It follows that for every $\epsilon>0$ there exist some $n\in\mathds{Z}$ such that $\left\Vert \mathsf{e}\circ\tau^{n}\left(0\right)-\mathsf{e}\left(0\right)\right\Vert <\epsilon$ implying $$\begin{aligned}
\forall \epsilon > 0 , \, \exists n\in\mathds{Z},\quad\left\Vert Z_{n T}-Z_{0}\right\Vert < \epsilon, \end{aligned}$$ i.e., $Z_{kT}$ is quasi-periodic. One concludes that the initial shape of the approximated state will up to some small error reoccur at some multiple of the classical period if the Floquet exponents of the flow differential are purely imaginary. We have again no [à]{} priori reasons to exclude unbounded motion.
In the case of a periodic flow, we have localization at the initial point and the initial profile (up to some small error) at classical periods. For bounded classical motion, those recurrences define revivals at periods of the classical flow that have already been described [@APe89] in the past.
We summarize our findings in the following theorem.
Under the following assumptions
1. $H\left(X\right)$ satisfies the conditions of section \[conditions\]
2. $H^{\prime\prime}\left(X_{t}\right)$ is $T-$periodic and $\nu$ is the maximal real part of the Floquet exponents of the flow differential, solution of eq. (\[eq:action\]),
and with $\kappa:=2 T\sup_{t\in\left[0,\, T\right]}\left\Vert \mathcal{J}H^{\prime\prime}\left(X_{t}\right)\right\Vert _{\mathrm{HS}}$ where $K:=e^{\kappa},$
we can make the following statements up to $\mathcal{O}\left(\hbar^{\alpha}\right),\,\alpha>0$ as $\hbar\searrow0$.
**If** $K$ is small enough **then** the approximation described by theorem \[thm:propagation\_control\_CR\] will hold up to times $\frac{\left|\ln\left(\hbar\right)\right|}{6\nu}$ and the approximate width of the state will behave like $e^{2\nu\left|t\right|}$.
**In particular,** **if** the Floquet exponents are purely complex or zero **then** the semiclassical propagation described in theorem \[thm:propagation\_control\_CR\] will hold to times $\hbar^{-\frac{1}{2}}$. The width of the approximated state will be bounded and recurrences will, up to small errors, take place at a multiple of $T$.
We would particularly like to thank Jens Bolte for valuable comments and discussions.
[^1]: We will, with some lack of rigor, call these states Gau[ß]{}ian, coherent or squeezed without distinction.
[^2]: A thorough description of the action of the evolution generated by quadratic operators on Gau[ß]{}ians and the metaplectic representation can be found in [@Fol89].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Dynkin algebras are the hereditary artin algebras of finite representation type. The paper exhibits the number of support-tilting modules for any Dynkin algebra. Since the support-tilting modules for a Dynkin algebra of Dynkin type $\Delta$ correspond bijectively to the generalized non-crossing partitions of type $\Delta$, the calculations presented here may also be considered as a categorification of results concerning the generalized non-crossing partitions. In the Dynkin case $\mathbb A$, we obtain the Catalan triangle, in the cases $\mathbb B$ and $\mathbb C$ the increasing part of the Pascal triangle, and finally in the case $\mathbb D$ an expansion of the increasing part of the Lucas triangle.'
author:
- |
Mustafa A. A. Obaid\
`[email protected]`\
\
S. Khalid Nauman\
`[email protected]`\
\
Wafaa M. Fakieh\
`[email protected]`\
\
Claus Michael Ringel\
`[email protected]`\
\
King Abdulaziz University, P O Box 80200\
Jeddah, Saudi Arabia\
title: |
The Numbers of Support-Tilting Modules\
for a Dynkin Algebra
---
Introduction
============
Let $\Lambda$ be a hereditary artin algebra. Here we consider left $\Lambda$-modules of finite length and call them just modules. The category of all modules will be denoted by $\operatorname{mod}\Lambda$. We let $n = n(\Lambda)$ denote the [*rank*]{} of $\Lambda$; this is by definition the number of simple modules (always, when counting numbers of modules of a certain kind, we actually mean the number of isomorphism classes). Following earlier considerations of Brenner and Butler, tilting modules were defined in [@[HR1]]. In the present setting, a [*tilting module*]{} $M$ is a module without self-extensions with precisely $n$ isomorphism classes of indecomposable direct summands, and we will assume, in addition, that $M$ is multiplicity-free. The endomorphism ring of a tilting module is said to be a tilted algebra. There is a wealth of papers devoted to tilted algebras, and the [*Handbook of Tilting Theory*]{} [@[AHK]] can be consulted for references.
The present paper deals with the Dynkin algebras: these are the connected hereditary artin algebras which are representation-finite, thus their valued quivers are of Dynkin type $\Delta_n = \mathbb A_n, \mathbb B_n, \dots, \mathbb G_2$ (see [@[DR1]]). Its aim is to discuss the number of tilting modules for such an algebra. The corresponding tilted algebras were classified by various authors in the eighties. It seems to be clear that a first step of such a classification result was the determination of all tilting modules, however there are only few traces in the literature (also the Handbook [@[AHK]] is of no help). Apparently, the relevance of the number of tilting modules was seen at that time only in special cases. The tilting modules for a linearly ordered quiver of type $\mathbb A_n$ were exhibited in [@[HR2]] and Gabriel [@[G]] pointed out that here we encounter one of the numerous appearances of the Catalan numbers $\frac1{n+1}\binom{2n}n$. For the cases $\mathbb D_n$, the number of tilting modules was determined by Bretscher-Läser-Riedtmann [@[BLR]] in their study of self-injective representation-finite algebras.
Given a module $M$, we let $\Lambda(M)$ denote its [*support algebra;*]{} this is the factor algebra of $\Lambda$ modulo the ideal which is generated by all idempotents $e$ with $eM = 0$ and is again a hereditary artin algebra (but usually not connected, even if $\Lambda$ is connected). The rank of the support algebra of $M$ will be called the [*support-rank*]{} of $M$. A module $T$ is said to be [*support-tilting*]{} provided $M$ considered as a $\Lambda(M)$-module is a tilting module. It may be well-known that the number of tilting modules of a Dynkin algebra depends only on its Dynkin type; at least for path algebras of quivers we can refer to Ladkani [@[L]]. Section \[algebras\] of the present paper provides a proof in general. It follows that the number of support-tilting modules with support-rank $s$ also depends only on the type $\Delta_n$; we let $a_s(\Delta_n)$ denote the number of support-tilting $\Lambda$-modules with support-rank $s$, where $\Lambda$ is of type $\Delta_n$. Of course, $a_n(\Delta_n)$ is just the number of tilting modules, and we denote by $a(\Delta_n)$ the number of all support-tilting modules; thus $a(\Delta_n) = \sum_{s=0}^n a_s(\Delta_n)$.
The present paper presents the numbers $a(\Delta_n)$ and $a_s(\Delta_n)$ for $0\le s \le n$ in a unified way. Of course, the exceptional cases $\mathbb E_6, \mathbb E_7,\mathbb E_8, \mathbb F_4, \mathbb G_2$ can be treated with a computer (but actually, also by hand); thus our main interest lies in the series $\mathbb A, \mathbb B, \mathbb C, \mathbb D$. In the case $\mathbb A$, we obtain in this way the [**Catalan triangle**]{} , in the case $\mathbb B$ and $\mathbb C$ the increasing part of the [**Pascal triangle,**]{} and finally in the case $\mathbb D$ an [**expansion**]{} of the increasing part of the [**Lucas triangle**]{} (see Section \[triangles\]; an outline will be given later in the introduction).
**The numbers**
---------------
All the numbers which are presented here for the cases $\mathbb A, \mathbb B, \mathbb C, \mathbb D$ are related to the binomial coefficients $\binom st$ and they coincide for $\mathbb B_n$ and $\mathbb C_n$ (as we will show in Section \[algebras\]); thus it is sufficient to deal with the cases $\mathbb A, \mathbb B, \mathbb D$. For $\mathbb B$, the binomial coefficients themselves will play a dominant role. For the cases $\mathbb A$ and $\mathbb D$, suitable multiples are relevant. In case $\mathbb A$, these are the Catalan numbers $C_n = \frac1{n+1}\binom{2n}n$, as well as related numbers. For the case $\mathbb D$, it will be convenient to use the notation $\left[\smallmatrix t\cr s \endsmallmatrix\right] = \frac{s+t}t\binom t s$ as proposed by Bailey [@[B]], since the relevant numbers in case $\mathbb D$ can be written in this way.
Hubery and Krause [@[HK]] have pointed out that the numbers $a(\Delta)$ for the simply laced diagrams $\Delta$ were discussed already in 1987 by Gabriel and de la Peña [@[GP]], but let us quote “although they have the correct number for $\mathbb E_8$, their numbers for $\mathbb E_6$ and $\mathbb E_7$ are slightly wrong”.
The numbers $a(\Delta_n)$ and $a_s(\Delta_n)$ for $0\le s \le n$: $$\hbox{\beginpicture
\setcoordinatesystem units <1.8cm,1.2cm>
\plot -.3 -.5 7.25 -.5 /
\plot 0.4 0.3 0.4 -3.4 /
\put{$\Delta_n$} at 0 0
\put{$\mathbb A_n$} at 1 0
\put{$\mathbb B_n, \mathbb C_n$} at 2 0
\put{$\mathbb D_n$} at 3 0
\put{$\mathbb E_6$} at 4 0
\put{$\mathbb E_7$} at 4.8 0
\put{$\mathbb E_8$} at 5.6 0
\put{$\mathbb F_4$} at 6.4 0
\put{$\mathbb G_2$} at 7 0
\put{$a_n(\Delta_n)$} at 0 -1
\put{$\frac1{n+1}\binom{2n}n$} at 1 -1
\put{$\binom{2n-1}{n-1}$} at 2 -1
\put{$\left[{\smallmatrix 2n-2\cr n-2\endsmallmatrix}\right]$} at 3 -1
\put{$418$} at 4 -1
\put{$2\,431$} at 4.8 -1
\put{$17\,342$} at 5.6 -1
\put{$66$} at 6.4 -1
\put{$5$} at 7 -1
\put{$a_s(\Delta_n)$} at 0 -1.9
\put{$\smallmatrix 0\le s < n \endsmallmatrix$} at 0 -2.25
\put{$\frac{n-s+1}{n+1}\binom{n+s}s$} at 1 -2
\put{$\binom{n+s-1}s$} at 2 -2
\put{$\left[{\smallmatrix n+s-2\cr s\endsmallmatrix}\right]$} at 3 -2
\put{\rm see Section \ref{exceptional}} at 5.5 -2
\setdots <1mm>
\plot 3.8 -2.05 4.8 -2.05 /
\plot 6.2 -2.05 7.1 -2.05 /
\put{$a(\Delta_n)$} at 0 -3
\put{$\frac1{n+2}\binom{2n+2}{n+1}$} at 1 -3
\put{$\binom{2n}n$} at 2 -3
\put{$\left[{\smallmatrix 2n-1\cr n-1\endsmallmatrix}\right]$} at 3 -3
\put{$833$} at 4 -3
\put{$4\,160$} at 4.8 -3
\put{$25\,080$} at 5.6 -3
\put{$105$} at 6.4 -3
\put{$8$} at 7 -3
\endpicture}$$
By analogy with the Bailey notation $\left[\smallmatrix t\cr s\endsmallmatrix\right]$ one may be tempted to introduce the following notation for the Catalan triangle: $\left]\smallmatrix t\cr s\endsmallmatrix\right[ = \frac{t-2s+1}{t-s+1}\binom t s$. Then the numbers for the case $\mathbb A$ are written as follows: $$a_n(\mathbb A_n) = \left]\smallmatrix 2n\cr n\endsmallmatrix\right[,\
a_s(\mathbb A_n) = \left]\smallmatrix n+s\cr s\endsmallmatrix\right[,\
a(\mathbb A) = \left]\smallmatrix 2n+2\cr n+1\endsmallmatrix\right[.$$
The reader should observe that for $\mathbb A_n$ and $\mathbb B_n$, the formula given for $a_s(\Delta_n)$ and $0\le s < n$ works also for $s = n$. This is not the case for $\mathbb D_n$: whereas $\binom{2n-2}{n-2} = \binom{2n-2}n$, the numbers $\left[{\smallmatrix 2n-2\cr n-2\endsmallmatrix}\right]$ and $\left[{\smallmatrix 2n-2\cr n\endsmallmatrix}\right]$ are different (the difference will be highlighted at the end of Section \[triangles\]). The Lucas triangle consists of the numbers $\left[{\smallmatrix t\cr s\endsmallmatrix}\right]$ for all $0 \le s \le t$; it therefore uses the numbers $\left[{\smallmatrix 2n-2\cr n\endsmallmatrix}\right]$ at the positions, whereas the $\mathbb D$-triangle (which we will now consider) uses the numbers $\left[{\smallmatrix 2n-2\cr n-2\endsmallmatrix}\right]$.
The triangles $\mathbb A, \mathbb B, \mathbb D$
-----------------------------------------------
The non-zero numbers $a_s(\Delta_n)$ for $\Delta = \mathbb A, \mathbb B, \mathbb D$ yield three triangles having similar properties. We will exhibit them in Section \[triangles\]; see the triangles \[triangle1\], \[triangle2\], \[triangle3\]. The triangle \[triangle1\] of type $\mathbb A$ is the Catalan triangle itself; this is in Sloane’s OEIS [@[S]]. The triangle \[triangle2\] of type $\mathbb B$ is the triangle , corresponding to the increasing part of the Pascal triangle (thus it consists of the binomial coefficients $\binom ts$ with $2s\le t+1$). The triangle \[triangle3\] of type $\mathbb D$ is an expansion of the increasing part of the Lucas triangle . Taking the increasing part of the rows in the Lucas triangle (thus the numbers $\left[\smallmatrix t\cr s \endsmallmatrix\right]$ with $2s\le t+1$), we obtain numbers which occur in the triangle of type $\mathbb D$, namely the numbers $a_s(\mathbb D_n)$ with $0 \le s < n$; the numbers $a_n(\mathbb D_n)$ on the diagonal however are given by a similar, but deviating formula (they are listed as the sequence ). The Lucas triangle is , but the triangle $\mathbb D$ itself was, at the time of the writing, not yet recorded in OEIS; now it is .
We see that the entries $a_s(n)$ of the triangles $\mathbb A$ and $\mathbb B$, as well as those of the lower triangular part of the triangle $\mathbb D$ can be obtained in a unified way from three triangles with entries $z_s(t)$ which satisfy the following recursion formula $$z_s(t) = z_{s-1}(t-1)+z_s(t-1)$$ (they are exhibited in Section \[triangles\] as triangles S \[triangle1-S\], S \[triangle2-S\], S \[triangle3-S\] using the shearing $a_s(n) = z_s(n+s-1)$). The recursion formula can be rewritten as $z_s(t) = \sum_{i= 0}^{s}z_i(t-s+i+1)$ (sometimes called the hockey stick formula). A consequence of the hockey stick formula is the fact that summing up the rows of any of the three triangles $\mathbb A, \mathbb B, \mathbb D$, we again obtain numbers which appear in the triangle.
Let us provide further details on the triangles to be sheared. Consider first the case $\mathbb B$. Here we start with the Pascal triangle; thus we deal with the triangle with numbers $z_s(t) = \binom ts$ and the initial conditions are $z_0(t) = z_t(t) = 1$ for all $t\ge 0$. In case $\mathbb D$, we start with the Lucas triangle with numbers $z_s(t) = \left[\smallmatrix t\cr s \endsmallmatrix\right]$, and the initial conditions are $z_0(t) = 1,\ z_t(t) = 2$ for all $t\ge 1$ (these initial conditions are the reason for calling the Lucas triangle also the $(1,2)$-triangle). In the case $\mathbb A$ we start with a sheared Catalan triangle, and here the initial conditions are $z_0(t) = 1$ and $z_{t+1}(2t) = 0$ for all $t\ge 0$.
Related results
---------------
Let us repeat that in this paper $a_n(\Delta_n)$ denotes the number of tilting modules, $a(\Delta_n)$ the number of support-tilting modules, for $\Lambda$ of Dynkin type $\Delta_n$. As we have mentioned, the relevance of the numbers $a_n(\Delta_n)$ and $a(\Delta_n)$ was not fully realized in the eighties. It became apparent through the work of Fomin and Zelevinksy when dealing with cluster algebras and the corresponding cluster complexes (see in particular [@[FZ]] and [@[FR]]): the numbers $a_n(\Delta_n)$ and $a(\Delta_n)$ appear in [@[FZ]] as the numbers $N(\Delta_n)$ of clusters and $N^+(\Delta_n)$ of positive clusters, respectively (see Propositions 3.8 and 3.9 of [@[FZ]]). For the numbers $a_s(\Delta_n)$ in general, see Chapoton [@[C]] in case $\mathbb A$ and $\mathbb B$ and Krattentaler [@[Kt]] in case $\mathbb D$. A conceptual proof of the equalities $a(\Delta_n) = N(\Delta_n)$ and $a_n(\Delta_n) = N^+(\Delta_n)$ has been given by Ingalls and Thomas [@[IT]] in case $\Delta_n$ is simply laced (thus of type $\mathbb A, \mathbb D$ of $\mathbb E$). The considerations of Ingalls and Thomas have been extended by the authors [@[ONFR]] to the non-simply laced cases. The papers [@[IT]] and [@[ONFR]] show in which way the representation theory of hereditary artin algebras can be used in order to categorify the cluster complex of Fomin and Zelevinsky: this is the reason for the equalities. Another method to relate clusters and support tilting modules is due to Marsh, Reineke and Zelevinsky [@[MRZ]]. Finally, let us stress that also the Coxeter diagrams $\mathbb H_3$ and $\mathbb H_4$ can be treated in a similar way, using hereditary artinian rings which are not artin algebras; this will be shown in [@[FR]].
The main result of the present paper is the direct calculation of the numbers $a_n(\Delta_n)$ in the case $\Delta = \mathbb B$; see Section \[tilting\]. Of course, using [@[ONFR]], this calculation can be replaced by referring to the determination of the corresponding cluster numbers by Fomin and Zelevinsky in [@[FZ]]. On the other hand, we hope that our proof is of interest in itself.
There is an independent development which has to be mentioned, namely the theory of generalized non-crossing partitions (see for example [@[A]]). It is the Ingalls-Thomas paper [@[IT]] (and [@[ONFR]]; see also the survey [@[R2]]) which provides the basic setting for using the representation theory of a hereditary artin algebra $\Lambda$ in order to deal with non-crossing partitions. It turns out that there is a large number of counting problems for $\operatorname{mod}\Lambda$ which yield the same answer, namely the numbers $a(\Delta_n)$ and $a_s(\Delta_n)$. For example, $a_s(\Delta)$ is also the number of antichains in $\operatorname{mod}\Lambda$ of size $s$: an [*antichain*]{} $A = \{A_1,\dots,A_t\}$ in $\operatorname{mod}\Lambda$ is a set of pairwise orthogonal bricks (a brick is a module whose endomorphism ring is a division ring, and two bricks $A_1, A_2$ are said to be orthogonal provided $\operatorname{Hom}(A_1,A_2) = 0 = \operatorname{Hom}(A_2,A_1)$; antichains are called discrete subsets in [@[GP]] and $\operatorname{Hom}$-free subsets in [@[HK]]).
Since the support-tilting modules for a Dynkin algebra of Dynkin type $\Delta$ correspond bijectively to the non-crossing partitions of type $\Delta$, the calculations presented here may be considered as a categorification of results concerning non-crossing partitions (for a general outline see Hubery-Krause [@[HK]]). Finally, let us mention that there is a corresponding discussion of the number of ad-nilpotent ideals of a Borel subalgebra of a simple Lie algebra; see Panyushev [@[P]].
**Outline of the paper**
------------------------
Let us stress again that there is an inductive procedure using the hook formula (Proposition \[hook-formula\]) and a modified hook formula (Proposition \[modified\]) in order to obtain the numbers $a_s(\mathbb A_n)$ for $0\le s \le n$, as well as the numbers $a_s(\Delta_n)$ for $\Delta = \mathbb B, \mathbb D$ for $0 \le s < n$, provided we know the numbers $a_n(\Delta_n)$. As we have mentioned, for the numbers $a_n(\mathbb D_n)$ we may refer to [@[BLR]]. In Section \[algebras\], we will show that the numbers $a_n(\mathbb B_n)$ and $a_n(\mathbb C_n)$ coincide; thus it remains to determine the numbers $a_n(\mathbb B_n)$. This will be done in Section \[tilting\]. In Section \[summation\], we calculate $a(\Delta_n)$ for $\Delta = \mathbb A, \mathbb B, \mathbb D$.
Section \[triangles\] presents the triangles $\mathbb A, \mathbb B, \mathbb D$ as well as the corresponding Catalan, Pascal, and Lucas triangles, and some observations concerning repetition of numbers in the triangles are recorded. Section \[exceptional\] provides the numbers $a_s(\Delta_n)$ for the exceptional cases $\Delta_n = \mathbb E_6, \mathbb E_7, \mathbb E_8, \mathbb F_4, \mathbb G_2$.
The triangles {#triangles}
=============
\[triangle1\][**The triangle of type $\mathbb A$; this is** ]{} $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,.45cm>
\put{$a_s(\mathbb A_n) = \dfrac{n-s+1}{n+1}\dbinom{n+s}s$} at 7 -1.5
\multiput{1} at 0 0 0 -1 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\put{1} at 1 -1
\put{2} at 1 -2
\put{2} at 2 -2
\put{3} at 1 -3
\put{5} at 2 -3
\put{5} at 3 -3
\put{4} at 1 -4
\put{9} at 2 -4
\put{14} at 3 -4
\put{14} at 4 -4
\put{5} at 1 -5
\put{14} at 2 -5
\put{28} at 3 -5
\put{42} at 4 -5
\put{42} at 5 -5
\put{6} at 1 -6
\put{20} at 2 -6
\put{48} at 3 -6
\put{90} at 4 -6
\put{132} at 5 -6
\put{132} at 6 -6
\put{7} at 1 -7
\put{27} at 2 -7
\put{75} at 3 -7
\put{165} at 4 -7
\put{297} at 5 -7
\put{429} at 6 -7
\put{429} at 7 -7
\put{8} at 1 -8
\put{35} at 2 -8
\put{110} at 3 -8
\put{275} at 4 -8
\put{572} at 5 -8
\put{1001} at 6 -8
\put{1430} at 7 -8
\put{1430} at 8 -8
\put{9} at 1 -9
\put{44} at 2 -9
\put{154} at 3 -9
\put{429} at 4 -9
\put{1001} at 5 -9
\put{2002} at 6 -9
\put{3432} at 7 -9
\put{4862} at 8 -9
\put{4862} at 9 -9
\put{${\smallmatrix n\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\put{} at 0 -10
\put{sum} [r] at 11.5 1.5
\put{1} [r] at 11.5 0
\put{2} [r] at 11.5 -1
\put{5} [r] at 11.5 -2
\put{14} [r] at 11.5 -3
\put{42} [r] at 11.5 -4
\put{132} [r] at 11.5 -5
\put{429} [r] at 11.5 -6
\put{1430} [r] at 11.5 -7
\put{4862} [r] at 11.5 -8
\put{16796} [r] at 11.5 -9
\endpicture}$$
\[triangle2\][**The triangle of type $\mathbb B$; this is** ]{} $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,.45cm>
\put{$a_s(\mathbb B_n) = \dbinom{n+s-1}s$} at 7 -1.5
\multiput{1} at 0 -1 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\put{1} at 0 0
\put{1} at 11.4 0
\put{1} at 1 -1
\put{2} at 1 -2
\put{3} at 2 -2
\put{3} at 1 -3
\put{6} at 2 -3
\put{10} at 3 -3
\put{4} at 1 -4
\put{10} at 2 -4
\put{20} at 3 -4
\put{35} at 4 -4
\put{5} at 1 -5
\put{15} at 2 -5
\put{35} at 3 -5
\put{70} at 4 -5
\put{126} at 5 -5
\put{6} at 1 -6
\put{21} at 2 -6
\put{56} at 3 -6
\put{126} at 4 -6
\put{252} at 5 -6
\put{462} at 6 -6
\put{7} at 1 -7
\put{28} at 2 -7
\put{84} at 3 -7
\put{210} at 4 -7
\put{462} at 5 -7
\put{924} at 6 -7
\put{1716} at 7 -7
\put{8} at 1 -8
\put{36} at 2 -8
\put{120} at 3 -8
\put{330} at 4 -8
\put{792} at 5 -8
\put{1716} at 6 -8
\put{3432} at 7 -8
\put{6435} at 8 -8
\put{9} at 1 -9
\put{45} at 2 -9
\put{165} at 3 -9
\put{495} at 4 -9
\put{1287} at 5 -9
\put{3003} at 6 -9
\put{6435} at 7 -9
\put{12870} at 8 -9
\put{24310} at 9 -9
\put{${\smallmatrix n\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\put{} at 0 -10
\put{sum} [r] at 11.5 1.5
\put{} [r] at 11.5 0
\put{2} [r] at 11.5 -1
\put{6} [r] at 11.5 -2
\put{20} [r] at 11.5 -3
\put{70} [r] at 11.5 -4
\put{252} [r] at 11.5 -5
\put{924} [r] at 11.5 -6
\put{3432} [r] at 11.5 -7
\put{12870} [r] at 11.5 -8
\put{48620} [r] at 11.5 -9
\endpicture}$$
\[triangle3\][**The triangle of type $\mathbb D$; this is now** ]{} $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,.45cm>
\put{$a_s(\mathbb D_n) \,= \left\{\begin{matrix} \cr\cr\cr\end{matrix}\right.$} [l] at 4.2 -1.4
\put{$\left[ \smallmatrix{n+s-2}\cr s \endsmallmatrix\right]$ \ for $0\le s <n;
$} [l] at 6.4 -.3
\put{$ \left[ \smallmatrix{2n-2}\cr n-2 \endsmallmatrix\right]$\qquad for $s=n$.} [l] at 6.4 -2.5
\multiput{1} at 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\multiput{$\cdot$} at 0 0 0 -1 1 -1 /
\multiput{$\cdot$} at 11.4 0 11.4 -1 /
\put{} at 1 -1
\put{2} at 1 -2
\put{1} at 2 -2
\put{3} at 1 -3
\put{5} at 2 -3
\put{5} at 3 -3
\put{4} at 1 -4
\put{9} at 2 -4
\put{16} at 3 -4
\put{20} at 4 -4
\put{5} at 1 -5
\put{14} at 2 -5
\put{30} at 3 -5
\put{55} at 4 -5
\put{77} at 5 -5
\put{6} at 1 -6
\put{20} at 2 -6
\put{50} at 3 -6
\put{105} at 4 -6
\put{196} at 5 -6
\put{294} at 6 -6
\put{7} at 1 -7
\put{27} at 2 -7
\put{77} at 3 -7
\put{182} at 4 -7
\put{378} at 5 -7
\put{714} at 6 -7
\put{1122} at 7 -7
\put{8} at 1 -8
\put{35} at 2 -8
\put{112} at 3 -8
\put{294} at 4 -8
\put{672} at 5 -8
\put{1386} at 6 -8
\put{2640} at 7 -8
\put{4290} at 8 -8
\put{9} at 1 -9
\put{44} at 2 -9
\put{156} at 3 -9
\put{450} at 4 -9
\put{1122} at 5 -9
\put{2508} at 6 -9
\put{5148} at 7 -9
\put{9867} at 8 -9
\put{16445} at 9 -9
\put{${\smallmatrix n\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\put{} at 0 -10
\put{sum} [r] at 11.5 1.5
\put{} [r] at 11.5 0
\put{} [r] at 11.5 -1
\put{4} [r] at 11.5 -2
\put{14} [r] at 11.5 -3
\put{50} [r] at 11.5 -4
\put{182} [r] at 11.5 -5
\put{672} [r] at 11.5 -6
\put{2508} [r] at 11.5 -7
\put{9438} [r] at 11.5 -8
\put{35750} [r] at 11.5 -9
\setdashes <1mm>
\plot 1.5 -1.5 1.5 -2.5
2.5 -2.5 2.5 -3.5
3.5 -3.5 3.5 -4.5
4.5 -4.5 4.5 -5.5
5.5 -5.5 5.5 -6.5
6.5 -6.5 6.5 -7.5
7.5 -7.5 7.5 -8.5
8.5 -8.5 8.5 -9.5
9.5 -9.5 9.5 -10
/
\endpicture}$$
\[triangle1-S\][**The sheared Catalan triangle** ]{} $$\hbox{\beginpicture
\setcoordinatesystem units <.9cm,.45cm>
\put{$\dbinom t s- \dbinom t {s-1} = \dfrac{t-2s+1}{t-s-1}\dbinom ts$} at 7 -2
\multiput{1} at 0 0 0 -1 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\put{1} at 1 -1
\put{2} at 1 -2
\put{} at 2 -2
\put{3} at 1 -3
\put{2} at 2 -3
\put{} at 3 -3
\put{4} at 1 -4
\put{5} at 2 -4
\put{} at 3 -4
\put{} at 4 -4
\put{5} at 1 -5
\put{9} at 2 -5
\put{5} at 3 -5
\put{} at 4 -5
\put{} at 5 -5
\put{6} at 1 -6
\put{14} at 2 -6
\put{14} at 3 -6
\put{} at 4 -6
\put{} at 5 -6
\put{} at 6 -6
\put{7} at 1 -7
\put{20} at 2 -7
\put{28} at 3 -7
\put{14} at 4 -7
\put{} at 5 -7
\put{} at 6 -7
\put{} at 7 -7
\put{8} at 1 -8
\put{27} at 2 -8
\put{48} at 3 -8
\put{42} at 4 -8
\put{} at 5 -8
\put{} at 6 -8
\put{} at 7 -8
\put{} at 8 -8
\put{9} at 1 -9
\put{35} at 2 -9
\put{75} at 3 -9
\put{90} at 4 -9
\put{42} at 5 -9
\put{} at 6 -9
\put{} at 7 -9
\put{} at 8 -9
\put{} at 9 -9
\put{${\smallmatrix t\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\plot 5.5 -10 5.5 -8.5 4.5 -8.5 4.5 -6.5 3.5 -6.5 3.5 -4.5 2.5 -4.5 2.5 -2.5 1.5 -2.5
1.5 -0.5 0.5 -0.5 0.5 .5 /
\setdots <.5mm>
\plot 0 0 1 -1 /
\plot 0 -1 2 -3 /
\plot 0 -2 3 -5 /
\plot 0 -3 4 -7 /
\plot 0 -4 5 -9 /
\plot 0 -5 5 -10 /
\plot 0 -6 4 -10 /
\plot 0 -7 3 -10 /
\plot 0 -8 2 -10 /
\plot 0 -9 1 -10 /
\put{} at 0 -10
\endpicture}$$
\[triangle2-S\][**The Pascal triangle , left of the staircase line is the increasing part**]{} $$\hbox{\beginpicture
\setcoordinatesystem units <.9cm,.45cm>
\put{$\dbinom t s$} at 7 -2
\multiput{1} at 0 0 0 -1 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\put{1} at 1 -1
\put{2} at 1 -2
\put{1} at 2 -2
\put{3} at 1 -3
\put{3} at 2 -3
\put{1} at 3 -3
\put{4} at 1 -4
\put{6} at 2 -4
\put{4} at 3 -4
\put{1} at 4 -4
\put{5} at 1 -5
\put{10} at 2 -5
\put{10} at 3 -5
\put{5} at 4 -5
\put{1} at 5 -5
\put{6} at 1 -6
\put{15} at 2 -6
\put{20} at 3 -6
\put{15} at 4 -6
\put{6} at 5 -6
\put{1} at 6 -6
\put{7} at 1 -7
\put{21} at 2 -7
\put{35} at 3 -7
\put{35} at 4 -7
\put{21} at 5 -7
\put{7} at 6 -7
\put{1} at 7 -7
\put{8} at 1 -8
\put{28} at 2 -8
\put{56} at 3 -8
\put{70} at 4 -8
\put{56} at 5 -8
\put{28} at 6 -8
\put{8} at 7 -8
\put{1} at 8 -8
\put{9} at 1 -9
\put{36} at 2 -9
\put{84} at 3 -9
\put{126} at 4 -9
\put{126} at 5 -9
\put{84} at 6 -9
\put{36} at 7 -9
\put{9} at 8 -9
\put{1} at 9 -9
\put{${\smallmatrix t\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\plot 5.5 -10 5.5 -8.5 4.5 -8.5 4.5 -6.5 3.5 -6.5 3.5 -4.5 2.5 -4.5 2.5 -2.5 1.5 -2.5
1.5 -0.5 0.5 -0.5 0.5 0.5 /
\setdots <.5mm>
\plot 0 0 1 -1 /
\plot 0 -1 2 -3 /
\plot 0 -2 3 -5 /
\plot 0 -3 4 -7 /
\plot 0 -4 5 -9 /
\plot 0 -5 5 -10 /
\plot 0 -6 4 -10 /
\plot 0 -7 3 -10 /
\plot 0 -8 2 -10 /
\plot 0 -9 1 -10 /
\put{} at 0 -10
\endpicture}$$
\[triangle3-S\][**The Lucas triangle , left of the staircase line is the increasing part**]{} $$\hbox{\beginpicture
\setcoordinatesystem units <.9cm,.45cm>
\put{$\left[\begin{matrix} t\cr s \end{matrix}\right]$} at 7 -2
\multiput{1} at 0 -1 0 -2 0 -3 0 -4 0 -5 0 -6 0 -7 0 -8 0 -9 /
\put{$\cdot$} at 0 0
\put{2} at 1 -1
\put{3} at 1 -2
\put{2} at 2 -2
\put{4} at 1 -3
\put{5} at 2 -3
\put{2} at 3 -3
\put{5} at 1 -4
\put{9} at 2 -4
\put{7} at 3 -4
\put{2} at 4 -4
\put{6} at 1 -5
\put{14} at 2 -5
\put{16} at 3 -5
\put{9} at 4 -5
\put{2} at 5 -5
\put{7} at 1 -6
\put{20} at 2 -6
\put{30} at 3 -6
\put{25} at 4 -6
\put{11} at 5 -6
\put{2} at 6 -6
\put{8} at 1 -7
\put{27} at 2 -7
\put{50} at 3 -7
\put{55} at 4 -7
\put{36} at 5 -7
\put{13} at 6 -7
\put{2} at 7 -7
\put{9} at 1 -8
\put{35} at 2 -8
\put{77} at 3 -8
\put{105} at 4 -8
\put{91} at 5 -8
\put{49} at 6 -8
\put{15} at 7 -8
\put{2} at 8 -8
\put{10} at 1 -9
\put{44} at 2 -9
\put{112} at 3 -9
\put{182} at 4 -9
\put{196} at 5 -9
\put{140} at 6 -9
\put{64} at 7 -9
\put{17} at 8 -9
\put{2} at 9 -9
\put{${\smallmatrix t\endsmallmatrix}$} at -2 1
\put{${\smallmatrix 0\endsmallmatrix}$} at -2 0
\put{${\smallmatrix 1\endsmallmatrix}$} at -2 -1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{${\smallmatrix 9\endsmallmatrix}$} at 9 1.5
\plot 5.5 -10 5.5 -8.5 4.5 -8.5 4.5 -6.5 3.5 -6.5 3.5 -4.5 2.5 -4.5 2.5 -2.5 1.5 -2.5
1.5 -0.5 0.5 -0.5 0.5 0 /
\setdots <.5mm>
\plot 0 -1 2 -3 /
\plot 0 -2 3 -5 /
\plot 0 -3 4 -7 /
\plot 0 -4 5 -9 /
\plot 0 -5 5 -10 /
\plot 0 -6 4 -10 /
\plot 0 -7 3 -10 /
\plot 0 -8 2 -10 /
\plot 0 -9 1 -10 /
\put{} at 0 -10
\endpicture}$$
In the triangle \[triangle3\] of type $\mathbb D$ and in the corresponding Lucas triangle S \[triangle3-S\] some values are left open (this is indicated by a dot).
In the Lucas triangle S \[triangle3-S\], this concerns the value at the position $(0,0)$ which could be denoted as $\left[\smallmatrix 0\cr 0 \endsmallmatrix\right]$. The value should be one of the numbers $1$ or $2$ (in OEIS , the number is chosen to be $2$). Note that here we deal with the product $\frac00\binom00$: whereas $\binom 0 0 = 1$ is well-defined, there is the ambiguous fraction $\frac 0 0$.
In the triangle \[triangle3\] of type $\mathbb D$, the positions $(0,0), (0,1), (1,1)$ are left open, since the series of Dynkin diagrams $\mathbb D_n$ starts with $n = 2$ (but see ); by definition $\mathbb D_2 = \mathbb A_1\sqcup \mathbb A_1$ and $\mathbb D_3 = \mathbb A_3$. As a consequence, also the corresponding entries in the summation sequence are missing.
**Some observations concerning the triangles $\mathbb A, \mathbb B, \mathbb D$**
--------------------------------------------------------------------------------
The sum sequence occurs as a diagonal.
In the $\mathbb A$-triangle, the sum sequence is the same sequence as the main diagonal (and these are just the Catalan numbers): $$a(\mathbb A_n) = a_{n+1}(\mathbb A_{n+1}).$$
In the $\mathbb B$-triangle, the sum sequence is the same sequence as the second diagonal $$a(\mathbb B_n) = a_{n}(\mathbb B_{n+1}). \ $$
In the $\mathbb D$-triangle, the sum sequence is the same sequence as the fourth diagonal $$a(\mathbb D_n) = a_{n-1}(\mathbb D_{n+2}).$$
The main diagonal uses the same sequence as one of the other diagonals.
In the $\mathbb A$-triangle, this concerns the main diagonal and the second diagonal: $$a_n(\mathbb A_n) = a_{n-1}(\mathbb A_{n}). \ $$ In the $\mathbb B$-triangle, this concerns the main diagonal and the second diagonal: $$a_n(\mathbb B_n) = a_{n-1}(\mathbb B_{n+1}).$$ In the $\mathbb D$-triangle, this concerns the main diagonal and the fifth diagonal: $$a_n(\mathbb D_n) = a_{n-2}(\mathbb D_{n+2}).$$
It may be of interest to exhibit explicit bijections between the corresponding sets of support-tilting modules. It seems that only in the case $\mathbb A$, this can be done easily (see Remark \[type-a\]).
**Comparison between the Lucas triangle and the $\mathbb D$-triangle**
----------------------------------------------------------------------
The difference between the number $\left[{\begin{matrix} 2n-2\cr n\end{matrix}}\right]$ and $\left[{\begin{matrix} 2n-2\cr n-2\end{matrix}}\right]$ seems to be of interest: $$\left[{\begin{matrix} 2n-2\cr n\end{matrix}}\right]
- \left[{\begin{matrix} 2n-2\cr n-2\end{matrix}}\right]
= \frac{1}{n}\binom{2n-2}{n-1}.$$ This means the following:
\[comparision\] $$\left[\begin{matrix} 2n-2\cr n \end{matrix}\right] - a_n(\mathbb D_n) = a_{n-1}(\mathbb A_{n-1}).$$
We show that $$\frac{3n-4}{n}\binom{2n-2}{n-2}
+\frac{1}{n}\binom{2n-2}{n-1} \ =\
\frac{3n-2}{2n-2}\binom{2n-2}{n}$$ We rewrite $$\begin{aligned}
\binom{2n-2}{n-2} &=& \frac{n}{2n-2} \binom{2n-2}{n}, \\
\binom{2n-2}{n-1} &=& \frac{n}{n-1}\binom{2n-2}{n}. \end{aligned}$$ The assertion now follows from the equality $$\frac{3n-4}{n}\cdot \frac{n}{2n-2}\ +\ \frac{1}{n}\cdot \frac{n}{n-1}\ =\ \frac{3n-2}{2n-2}.$$
Here is a table of these numbers $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,.45cm>
\put{${\smallmatrix n\endsmallmatrix}$} at -2 1
\put{$\left[{\begin{matrix} 2n-2\cr n\end{matrix}}\right]$}
[r] at 0.3 1
\put{$\left[{\begin{matrix} 2n-2\cr n-2\end{matrix}}\right]$} [r] at 2.4 1
\put{$\dfrac{1}{n}\dbinom{2n-2}{n-1}$} [r] at 4.9 1
\put{${\smallmatrix 2\endsmallmatrix}$} at -2 -2
\put{${\smallmatrix 3\endsmallmatrix}$} at -2 -3
\put{${\smallmatrix 4\endsmallmatrix}$} at -2 -4
\put{${\smallmatrix 5\endsmallmatrix}$} at -2 -5
\put{${\smallmatrix 6\endsmallmatrix}$} at -2 -6
\put{${\smallmatrix 7\endsmallmatrix}$} at -2 -7
\put{${\smallmatrix 8\endsmallmatrix}$} at -2 -8
\put{${\smallmatrix 9\endsmallmatrix}$} at -2 -9
\put{$2$} [r] at 0 -2
\put{$7$} [r] at 0 -3
\put{$25$} [r] at 0 -4
\put{$91$} [r] at 0 -5
\put{$336$} [r] at 0 -6
\put{$1254$} [r] at 0 -7
\put{$4719$} [r] at 0 -8
\put{$17875$} [r] at 0 -9
\put{$1$} [r] at 2 -2
\put{$5$} [r] at 2 -3
\put{$20$} [r] at 2 -4
\put{$77$} [r] at 2 -5
\put{$294$} [r] at 2 -6
\put{$1122$} [r] at 2 -7
\put{$4290$} [r] at 2 -8
\put{$16445$} [r] at 2 -9
\put{$1$} [r] at 4 -2
\put{$2$} [r] at 4 -3
\put{$5$} [r] at 4 -4
\put{$14$} [r] at 4 -5
\put{$42$} [r] at 4 -6
\put{$132$} [r] at 4 -7
\put{$429$} [r] at 4 -8
\put{$1430$} [r] at 4 -9
\put{} at 0 -10
\endpicture}$$
Proposition \[comparision\] is essentially the modified hook formula for type $\mathbb D$ which will be presented in Proposition \[modified\]: the Lucas triangle uses the hook formula for the whole triangle, whereas the triangle $\mathbb D$ uses the modified hook formula on the subdiagonal.
The exceptional cases {#exceptional}
=====================
Here are the numbers $a_s(\Delta_n)$ and $a(\Delta_n)$ in the exceptional cases $\mathbb E_6, \mathbb E_7, \mathbb E_8, \mathbb F_4,$ and $\mathbb G_2$ (we add some suitable additional rows in order to stress the induction scheme): $$\hbox{\beginpicture
\setcoordinatesystem units <.97cm,.45cm>
\put{1} at 0 0
\put{3} at 1 0
\put{4} at 2 0
\put{2} at 3 0
\put{1} at 0 -1
\put{4} at 1 -1
\put{9} at 2 -1
\put{14} at 3 -1
\put{14} at 4 -1
\put{1} at 0 -2
\put{5} at 1 -2
\put{14} at 2 -2
\put{30} at 3 -2
\put{55} at 4 -2
\put{77} at 5 -2
\put{1} at 0 -3
\put{6} at 1 -3
\put{20} at 2 -3
\put{50} at 3 -3
\put{110} at 4 -3
\put{228} at 5 -3
\put{418} at 6 -3
\put{1} at 0 -4
\put{7} at 1 -4
\put{27} at 2 -4
\put{77} at 3 -4
\put{187} at 4 -4
\put{429} at 5 -4
\put{1001} at 6 -4
\put{2431} at 7 -4
\put{1} at 0 -5
\put{8} at 1 -5
\put{35} at 2 -5
\put{112} at 3 -5
\put{299} at 4 -5
\put{728} at 5 -5
\put{1771} at 6 -5
\put{4784} at 7 -5
\put{17342} at 8.1 -5
\put{1} at 0 -7
\put{3} at 1 -7
\put{6} at 2 -7
\put{10} at 3 -7
\put{1} at 0 -8
\put{4} at 1 -8
\put{10} at 2 -8
\put{24} at 3 -8
\put{66} at 4 -8
\put{1} at 0 -10
\put{2} at 1 -10
\put{5} at 2 -10
\put{$$} at -2 2
\put{$\mathbb E_3 = \mathbb A_2\sqcup\mathbb A_1$} [l] at -3 -0
\put{$\mathbb E_4 = \mathbb A_4$} [l] at -3 -1
\put{$\mathbb E_5 = \mathbb D_5$} [l] at -3 -2
\put{$\mathbb E_6$} [l] at -3 -3
\put{$\mathbb E_7$} [l] at -3 -4
\put{$\mathbb E_8$} [l] at -3 -5
\put{$$} [l] at -2.2 -6
\put{$\mathbb B_3$} [l] at -3 -7
\put{$\mathbb F_4$} [l] at -3 -8
\put{$$} [l] at -2.2 -9
\put{$\mathbb G_2$} [l] at -3 -10
\plot -2 1.5 -1.5 1 /
\put{${\smallmatrix s\endsmallmatrix}$} at -1.5 1.5
\put{${\smallmatrix 0\endsmallmatrix}$} at 0 1.5
\put{${\smallmatrix 1\endsmallmatrix}$} at 1 1.5
\put{${\smallmatrix 2\endsmallmatrix}$} at 2 1.5
\put{${\smallmatrix 3\endsmallmatrix}$} at 3 1.5
\put{${\smallmatrix 4\endsmallmatrix}$} at 4 1.5
\put{${\smallmatrix 5\endsmallmatrix}$} at 5 1.5
\put{${\smallmatrix 6\endsmallmatrix}$} at 6 1.6
\put{${\smallmatrix 7\endsmallmatrix}$} at 7 1.5
\put{${\smallmatrix 8\endsmallmatrix}$} at 8 1.5
\put{} at 0 -10
\put{sum} [r] at 11 1.5
\put{10} [r] at 11 0
\put{42} [r] at 11 -1
\put{182} [r] at 11 -2
\put{833} [r] at 11 -3
\put{4160} [r] at 11 -4
\put{25080} [r] at 11 -5
\put{} [r] at 11 -6
\put{20} [r] at 11 -7
\put{105} [r] at 11 -8
\put{} [r] at 11 -9
\put{8} [r] at 11 -10
\endpicture}$$
Hereditary artin algebras {#algebras}
=========================
The basic setting
-----------------
Let $\Lambda$ be a hereditary artin algebra. Since by assumption $\operatorname{Ext}_\Lambda^i = 0$ for $i \ge 2$, we write $\operatorname{Ext}(M,M')$ instead of $\operatorname{Ext}_\Lambda^1(M,M')$. The vertices of the quiver $Q(\Lambda)$ are the isomorphism classes $[S]$ of the simple $\Lambda$-modules $S$ and there is an arrow $[S] \to [S']$ provided $\operatorname{Ext}(S,S') \neq 0$. Note that $Q(\Lambda)$ is finite and directed (the latter means that the simple modules can be labeled $S(i)$ such that the existence of an arrow $[S(i)]\to [S(j)]$ implies that $i > j$). We endow $Q(\Lambda)$ with a valuation as follows: given an arrow $[S] \to [S']$, consider $\operatorname{Ext}(S,S')$ as a left $\operatorname{End}(S)^{\text{op}}$-module and also as a left $\operatorname{End}(S')$-module and put $$v([S],[S']) = (\dim {}_{\operatorname{End}(S)}\operatorname{Ext}(S,S'))(\dim {}_{\operatorname{End}(S')^{\text{op}}}\operatorname{Ext}(S,S'))$$ provided $v([S],[S']) > 1$. Given a vertex $i$ of $Q(\Lambda)$, we let $S(i)=S_\Lambda (i),
P(i) = P_\Lambda(i), I(i)=I_\Lambda(i)$, respectively, denote a simple, an indecomposable projective or injective module corresponding to the vertex $i$.
If $M$ is a module, the set of vertices of the quiver $Q(\Lambda(M))$ will be called the [*support*]{} of $M$ and $M$ is said to be [*sincere*]{} provided any vertex of $Q(\Lambda)$ belongs to the support of $M$ (thus provided the only idempotent $e \in \Lambda$ with $eM = 0$ is $e=0$).
We also will be interested in the corresponding valued graph $\overline Q(\Lambda)$ which is obtained from the valued quiver $Q(\Lambda)$ by replacing the arrows by edges: one says that one [*forgets the orientation*]{} of the quiver.
In the special case where $v([S],[S']) = v$ with $v = 2$ or $v = 3$, it is usual to replace the arrow $[S] \longrightarrow [S']$ by a double arrow $[S] \Longrightarrow [S']$ (if $v = 2$) or a similar triple arrow (if $v=3$). Using the bimodule $\operatorname{Ext}(S,S')$ one obtains an embedding either of $\operatorname{End}S$ into $\operatorname{End}S'$, or of $\operatorname{End}S'$ into $\operatorname{End}S$; thus one of the division rings is a subring of the other, with index equal to $v$. One marks the relative size of the endomorphism rings by an additional arrowhead drawn in the middle of the edge, pointing from the larger endomorphism ring to the smaller one (it should be stressed that these inner arrowheads must not be confused with the outer ones). For example, in case there are two simple modules labeled $1$ and $2$ with an arrow $1 \leftarrow 2$ and $v(1,2) = 2$, there are the following two possibilities: $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\put{\beginpicture
\multiput{$\circ$} at 5 0 6 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.6 0.1 5.4 0 5.6 -.1 /
\endpicture} at 0 0
\put{\beginpicture
\multiput{$\circ$} at 5 0 6 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.4 0.1 5.6 0 5.4 -.1 /
\endpicture} at 3 0
\endpicture}$$ On the left we see that $\operatorname{End}S(1)$ is a division subring of $\operatorname{End}S(2)$. On the right, $\operatorname{End}S(2)$ is a division subring of $\operatorname{End}S(1)$. (Let us exhibit corresponding algebras: let $K:k$ be a field extension of degree $2$ and consider the algebras $\Lambda = \left[\smallmatrix k & K \cr
0 & K \endsmallmatrix\right]$ and $\Lambda' = \left[\smallmatrix K & K \cr
0 & k \endsmallmatrix\right]$; the left quiver shown above is $Q(\Lambda)$, and the right quiver is $Q(\Lambda')$.)
Here are the corresponding valued graphs, which are obtained by forgetting the orientation (thus deleting the outer arrowheads, but not the inner ones): $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\put{\beginpicture
\multiput{$\circ$} at 5 0 6 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.6 0.1 5.4 0 5.6 -.1 /
\endpicture} at 0 0
\put{\beginpicture
\multiput{$\circ$} at 5 0 6 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.4 0.1 5.6 0 5.4 -.1 /
\endpicture} at 3 0
\endpicture}$$ They are called $\mathbb B_2$ and $\mathbb C_2$, respectively (observe that there is a difference between $\mathbb B_2$ and $\mathbb C_2$ only if they occur as subgraphs of larger graphs).
We recall the following [@[DR1]]. [*A connected hereditary artin algebra $\Lambda$ is representation-finite if and only if $\overline Q(\Lambda)$ is one of the Dynkin diagrams $$\mathbb A_n, \mathbb B_n, \mathbb C_n, \mathbb D_n, \mathbb E_6, \mathbb E_7,
\mathbb E_8, \mathbb F_4, \mathbb G_2$$ and in this case the indecomposable $\Lambda$-modules correspond bijectively to the positive roots.*]{}
Change of orientation
---------------------
We want to show that the number of basic tilting modules is independent of the orientation. We recall that a module is said to be [*basic*]{} provided it is a direct sum of pairwise non-isomorphic indecomposable modules; an artin algebra $\Lambda$ is [*basic*]{} provided the regular representation ${}_\Lambda\Lambda$ is basic. In case $\Lambda$ is the path algebra of a quiver, we may refer to Ladkani [@[L]]. In the case of the tensor algebra of a species (in particular in the case of the path algebra of a quiver), any change of orientation is obtained by applying a sequence of BGP-reflection functors; see [@[DR2]]. For a general hereditary artin algebra $\Lambda$, we have to deal with APR-tilting functors as defined by Auslander, Platzeck and Reiten [@[APR]]. In order to do so, we may assume that $\Lambda$ is basic. We start with a simple projective module $S$, write ${}_\Lambda\Lambda = S\oplus P$ with a projective module $P$, and consider $W = P\oplus \tau^-S$ (where $\tau = \tau_\Lambda$ is the Auslander-Reiten translation in $\operatorname{mod}\Lambda$) and $\Lambda' = (\operatorname{End}W)^{\text{op}}$. Note that $W$ is a tilting module (called an APR-tilting module) and the quiver $Q(\Lambda')$ is obtained from the quiver $Q(\Lambda)$ by changing the orientation of all the arrows which involve the vertex $\omega = [S]$. We let $\Lambda'' = (\operatorname{End}P)^{\text{op}}$ denote the restriction of $\Lambda$ to the quiver $Q''$ obtained from $Q(\Lambda)$ by deleting the vertex $\omega$ and the arrows ending in $\omega$. Of course, $Q''$ is also a subquiver of $Q(\Lambda')$ and $\Lambda''$ is the restriction of $\Lambda'$ to $Q''$ (thus $\Lambda$ is a one-point coextension of $\Lambda''$, whereas $\Lambda'$ is a one-point extension of $\Lambda''$). We let $S'$ denote the simple $\Lambda'$-module with support $\omega$.
\[APR\] Let $\Lambda$ be a hereditary artin algebra and $S$ a simple projective module. Let $W$ be the APR-tilting module defined by $S$ and $\Lambda' = (\operatorname{End}W)^{\text{op}}$. Then there is a canonical bijection $\eta$ between the basic tilting $\Lambda$-modules and the basic tilting $\Lambda'$-modules.
In order to define $\eta$, we distinguish two cases.
First, if $T$ is a basic tilting module such that $S$ is not a direct summand of $T$, let $\eta(T) = \operatorname{Hom}(W,T)$; this is a basic tilting $\Lambda'$-module and $S'$ is not a direct summand of $\eta(T)$.
Second, consider a basic tilting $\Lambda$-module of the form $S\oplus T$. Let $T'' = T/U$, where $U$ is the sum of the images of all the maps $S \to T$. Obviously, $T''$ is a basic tilting $\Lambda''$-module which we may consider as a $\Lambda'$-module. We form the universal extension $T'$ of $T''$ using copies of $S'$. Then $T'\oplus S'$ is a basic tilting $\Lambda'$-module (and $S'$ is a direct summand).
We may identify the Grothendieck groups $K_0(\Lambda)$ and $K_0(\Lambda')$, using the common factor algebra $\Lambda''$ and identifying the dimension vectors of $S$ and $S'$. Then, in the first case, the dimension vector of $\eta(T)$ is obtained from the dimension vector of $T$ by applying the reflection $\sigma$ defined by $S$. In the second case, the dimension vectors of $T$ and $\eta(T)$ coincide. Actually, here we use twice the internal reflection defined by $S$ in [@[R]], first in the category $\operatorname{mod}\Lambda$, second in the category $\operatorname{mod}\Lambda'$.
The combinatorial backbone
--------------------------
Let $\Lambda$ be a Dynkin algebra and assume that the vertices of $Q(\Lambda)$ are labeled $1\le i \le n$. Let $P(i) = P_\Lambda(i)$ be indecomposable projective. Since we assume that $\Lambda$ is a Dynkin algebra, there is a natural number $q(i) = q(P(i))$ such that $\tau^{-q(i)}P(i)$ is indecomposable injective; the modules $M(i,u) = \tau^{-u}P(i)$ with $0 \le u \le q(i)$ and $1\le i \le n$ furnish a complete list of the indecomposable $\Lambda$-modules.
\[hammock\] Let $\Lambda, \Lambda'$ be Dynkin algebras and assume that the simple modules of both algebras are indexed by $1\le i \le n$. Assume that $q(P_\Lambda(i)) = q(P_{\Lambda'}(i)) = q(i)$ for all $1\le i \le n$. If the support of $M(u,i) = \tau_\Lambda^{-u}P_\Lambda(i)$ and $M'(i,u) = \tau_{\Lambda'}^{-u}P_{\Lambda}(i)$ coincide for all $0 \le u \le q(i)$ and $1\le i \le n$, then $a_s(\Lambda) = a_s(\Lambda')$ for all $s$.
We may interpret the numbers $a_s(\Lambda)$ and $a_s(\Lambda')$ as the number of antichains in $\operatorname{mod}\Lambda$ and $\operatorname{mod}\Lambda'$, respectively, which have support-rank $s$. Note that the support of a module $M$ is the set of numbers $1\le i \le n$ such that $\operatorname{Hom}(P(i),M) \neq 0$.
Note that $\operatorname{Hom}(M(i,u),M(j,v)) = 0$ if and only if $\operatorname{Hom}(M'(i,u),M'(j,v))$ $ = 0$. Namely, if $u\le v$, the Auslander-Reiten translation (see for example [@[ARS]]) furnishes a group isomorphism $$\begin{aligned}
\operatorname{Hom}(M(i,u),M(j,v)) &\simeq& \operatorname{Hom}(M(i,0),M(j,v-u)) \cr
&=&\operatorname{Hom}(P_{\Lambda}(i),M(j,v-u)),\end{aligned}$$ and similarly we have $ \operatorname{Hom}(M'(i,u),M'(j,v)) \simeq \operatorname{Hom}(P_{\Lambda'}(i),M'(j,v-u))$. It follows that $\operatorname{Hom}(M(i,u),M(j,v)) = 0$ if and only if $i$ is not in the support of $M(j,v-u)$ if and only if $i$ is not in the support of $M'(j,v-u)$ if and only if $\operatorname{Hom}(M'(i,u),M'(j,v)) = 0$.
If $u > v$, then $$\operatorname{Hom}(M(i,0),M(j,v)) \simeq \operatorname{Hom}(M(i,u-v),M(j,0)) = 0,$$ since $M(i,u-v)$ is indecomposable and non-projective, whereas $M(j,0)$ is projective. Similarly, we also have $ \operatorname{Hom}(M'(i,0),M'(j,v)) = 0$.
As a consequence we see that given an antichain $A = \{A_1,\dots,A_t\}$ in $\operatorname{mod}\Lambda$, the function $M(i,u)\mapsto M'(i,u)$ yields an antichain $A' = \{A'_1,\dots,A'_t\}$ in $\operatorname{mod}\Lambda'$. Of course, the support-rank of $A$ and $A'$ are the same. This completes the proof.
The numbers $a_s(\Lambda)$ depend only on the Dynkin type of $Q(\Lambda)$, not on $\Lambda$ itself.
According to Proposition \[hammock\], the numbers $a_s(\Lambda)$ depend only on $Q(\Lambda)$. According to Proposition \[APR\], the orientation of $Q(\Lambda)$ does not play a role.
Thus, is $\Lambda$ is of Dynkin type $\Delta$, we write $a_s(\Delta)$ instead of $a_s(\Lambda)$.
For all $0\le s \le n$, we have $a_s(\mathbb B_n) = a_s(\mathbb C_n)$.
Apply the Proposition \[hammock\] to the algebras $\Lambda$ and $\Lambda'$ with valued quivers $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-2\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix n\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.4 0.1 5.6 0 5.4 -.1 /
\multiput{$\circ$} at 6 0 /
\endpicture} at 0 0
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-2\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix n\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.6 0.1 5.4 0 5.6 -.1 /
\multiput{$\circ$} at 6 0 /
\endpicture} at 0 -1
\endpicture}$$ respectively; the first valued quiver is of type $\mathbb B_n$, and the second is of type $\mathbb C_n$. It is well-known (and easy to see) that $q(P_\Lambda(i)) = n\!-\!1
= q(P_{\Lambda'}(i))$ for all $1\le i \le n$ and that the modules $M(i,u)$ and $M'(i,u)$ for $1\le i\le n$ and $0\le u \le n-1$ have the same support.
The tilting modules for $\mathbb B_n$ {#tilting}
=====================================
We are going to determine the number of tilting modules for the Dynkin algebras of type $\mathbb B_n$; namely we will show that $a_n(\mathbb B_n) = \binom{2n-1}{n-1}$. By induction, we assume knowledge about the representation theory of $\mathbb B_i$ with $i < n$, as well as the calculation of $a_s(\mathbb B_n)$ for $s < n$ as shown in Section \[hook\]. We consider a Dynkin algebra $\Lambda$ with quiver $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-2\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-1\endsmallmatrix}$} at 5 -.3
\put{${\smallmatrix n\endsmallmatrix}$} at 6 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.4 0.1 5.6 0 5.4 -.1 /
\multiput{$\circ$} at 6 0 /
\endpicture}$$ We interpret $a_n(\mathbb B_n)$ as the number of sincere antichains (by definition, an antichain $A = \{A_1,\dots, A_t\}$ is [*sincere*]{} provided the module $\bigoplus A_i$ is sincere) and write it as the sum $$a_n(\mathbb B_n) = u(\mathbb B_n) + v(\mathbb B_n)$$ where $u(\mathbb B_n)$ is the number of antichains with a sincere element, whereas $v(\mathbb B_n)$ is the number of sincere antichains without a sincere element. These two numbers will be calculated separately.
The calculation of $u(\mathbb B_n)$
-----------------------------------
We let $w(\mathbb B_n)$ denote the number of antichains which do not contain any injective module.
$$w(\mathbb B_n) = a_n(\mathbb B_n).$$
Let $\mathcal W$ be the set of antichains without injective modules and $\mathcal S$ the set of sincere antichains. We want to construct a bijection $\eta:\mathcal S \to \mathcal W$. Note that an element of $\mathcal S$ contains at most one injective module, since the injective modules are pairwise comparable with respect to $\operatorname{Hom}$. If $A\in \mathcal S$ contains no injective module, then let $\eta(A) = A$. If $A\in \mathcal S$ contains the injective module $I(i)$, let $\eta(A)$ be obtained from $A$ by deleting $I(i)$ and note that $\eta(A)$ is no longer sincere (since all the modules $A_j$ in $\eta(A)$ satisfy $\operatorname{Hom}(A_j,I(i)) = 0$). It follows that $\eta$ is an injective map. In order to show that $\eta$ is surjective, assume that $B$ is an antichain in $\mathcal W$. If $B$ is sincere, then it belongs to $\mathcal S$ and by definition $\eta(B) = B$. If $B$ is not sincere, let $i$ be the smallest number such that $i$ is not in the support of $B$. Let $A$ be obtained from $B$ by adding $I(i)$. Then clearly $A$ is sincere and $\eta(A) = B$.
\[type-a\] A similar proof applies to the linearly oriented quiver of type $\mathbb A_n$. It yields the formula $$a(\mathbb A_{n-1}) = a_n(\mathbb A_n).$$ Also, instead of looking at antichains which do not contain any injective module, we may consider antichains which do not contain any projective module.
Now we are able to determine $u(\mathbb B_n)$.
$$u(\mathbb B_n) = a_{n-1}(\mathbb B_n) = \binom{2n-2}{n-1}.$$
Note that the sincere indecomposable representations of $\Lambda$ are the modules $X(i) = \tau^{-n+i}P(i)$ with $1\le i \le n$. The dimension vector of $X(n)$ is $(1,\dots,1)$, whereas for $1\le i <n$, the length of $X(n)$ is $n+i$ and its dimension vector is of the form $(1,\dots,1)+(0,\dots,0,1,\dots,1)$. It is easy to see that $\operatorname{Hom}(X(i),X(j)) \neq 0$ for $i\ge j$, thus any antichain contains at most one $X(i)$. Let $u_{i}(\mathbb B_n)$ be the antichains which contain $X(i)$, thus $$u(\mathbb B_n) = \sum\nolimits_{i=1}^n u_{i}(\mathbb B_n).$$ Let $\mathcal X_i$ be the set of indecomposable modules $M$ such that $\operatorname{Hom}(X(i),M) = 0 = \operatorname{Hom}(M,X(i))$. Thus, the antichains which contain $X(i)$ correspond bijectively to the antichains in $\mathcal X_i$. In general, the set $\mathcal X_i$ consists of three triangles I, II, III: $$\hbox{\beginpicture
\setcoordinatesystem units <.4cm,.4cm>
\multiput{} at 0 0 27 9.5 /
\plot 0 0 9 9 18 0 27 9 /
\setdots <1mm>
\plot 0 0 18 0 /
\plot 9 9 27 9 /
\setsolid
\plot 4 4 8 0 17 9 /
\plot 2 0 4 2 6 0 /
\plot 11 9 13 7 15 9 /
\plot 10 0 13 3 16 0 /
\setshadegrid span <.5mm>
\vshade -0.3 -.5 .3 <,z,,> 4 3.7 4.3 <z,z,,> 8 -.3 8.3
<z,z,,> 9 0.7 9.2 <z,z,,> 13 4.7 5.3 <z,z,,> 16.8 .7 9.2 <z,z,,> 18 -.3 9.2
<z,,,> 27.3 9 9.2 /
\put{$X(i)$} [l] at 13.5 5
\put{I} at 4 .7
\put{II} at 13 8.3
\put{III} at 13 1.1
\endpicture}$$ The triangle I is the wing at the vertex $\tau^{-1}P(n-i-1)$, the triangle II is the wing at the vertex $\tau^{-n+i-1}P(i+2)$, and the triangle III is the wing at the vertex $\tau^{-n+i+1}P(i-2)$.
We also are interested in a larger triangle II$'$ which contains the triangle II as well as $n-i$ additional modules (all being successors of $X(i)$), namely the wing at the vertex $\tau^{-n+i}P(i+1)$.
The full subcategory $\mathcal X'$ of all direct sums of indecomposable modules in the wings I, II$'$, III is the thick subcategory with simple objects $$S(2), S(3),\dots, S(n-i+1); \quad \tau^{n-i}P(n); \quad S(n-i+3), \dots, S(n-1).$$ The position of these modules is indicated here by bullets: $$\hbox{\beginpicture
\setcoordinatesystem units <.4cm,.4cm>
\multiput{} at 0 0 27 9.5 /
\plot 0 0 9 9 18 0 27 9 /
\setdots <1mm>
\plot 0 0 18 0 /
\plot 9 9 27 9 /
\setsolid
\plot 4 4 8 0 17 9 /
\plot 2 0 4 2 6 0 /
\plot 11 9 14 6 /
\setsolid
\plot 10 0 13 3 16 0 /
\setshadegrid span <.5mm>
\vshade 2 0 0 <,z,,> 4 0 2 <z,,,> 6 0 0 /
\vshade 11 9 9 <,z,,> 14 6 9 <z,,,> 17 9 9 /
\vshade 10 0 0 <,z,,> 13 0 3 <z,,,> 16 0 0 /
\put{$X(i)$} [l] at 13.5 5
\put{I} at 4 .9
\put{II$'$} at 14 8
\put{III} at 13 1.1
\multiput{$\bullet$} at 2 0 6 0 10 0 16 0 17 9 /
\endpicture}$$ (A full subcategory $\mathcal A$ of $\operatorname{mod}\Lambda$ is called a [*thick*]{} subcategory provided it is closed under kernels, cokernels and extensions (see for example [@[K]]). Note that a thick subcategory is an abelian category, and the inclusion functor $\mathcal A \to \operatorname{mod}\Lambda$ is exact.)
The category $\mathcal X'$ is of type $\mathbb B_{n-i} \sqcup \mathbb A_{i-2}$ (the $\mathbb A_{i-2}$-part is given by the triangle III, whereas the $\mathbb B_{n-i}$-part is given by the triangles I and II$'$). Note that the indecomposables in I and II just correspond to the non-injective indecomposables in the $\mathbb B_{n-i}$-part. This shows that $$u_i(\mathbb B_n) = w(\mathbb B_{n-i})a(\mathbb A_{i-2})
= a_{n-i}(\mathbb B_{n-i})a_{i-1}(\mathbb A_{i-1}).$$
In the special cases $i = 1,2,n-1,n$, the same formula holds. Namely, for $i = 1$ and $i=2$, the triangle III is empty, whereas the triangles I and II$'$ together yield a category of type $\mathbb B_{n-i}$. In the cases $i=n-1$ and $i=n$, the triangles I and II are empty, whereas the triangle III yields a category of type $\mathbb A_{i-2}$.
Thus we see $$u(\mathbb B_n) = \sum\nolimits_{i=1}^n u_i(\mathbb B_n) = \sum\nolimits_{i=1}^n
a_{i-1}(\mathbb A_{i-1})a_{n-i}(\mathbb B_{n-i}).$$ But the latter expression is the recursion formula for $a_{n-1}(\mathbb B_n)$, since the number of support-tilting modules $T$ with support $\{1,2,\dots,n\}\setminus\{i\}$ is just $a_{i-1}(\mathbb A_{i-1})a_{n-i}(\mathbb B_{n-i})$.
The calculation of $v(\mathbb B_n)$
-----------------------------------
$$v(\mathbb B_n) = a_{n-2}(\mathbb B_{n+1}) = \binom{2n-2}{n-2}.$$
Let $\mathcal V$ be the set of sincere antichains of $\Lambda$-modules without a sincere element. Let $A = (A_1,\dots, A_r)$ be in $\mathcal V$. Since $A$ is sincere, we may assume that $\operatorname{Hom}(P(1),A_1)\neq 0$. Since $A_1$ is not sincere, we must have $\operatorname{Hom}(P(n),A_1) = 0$, thus $A_1$ is a representation of a Dynkin algebra of type $\mathbb A_{n-1}$ and actually an indecomposable projective representation (also as a $\Lambda$-module), thus $A_1 = P(i)$ for some $i$ with $1\le i < n$. Since an antichain can contain only one indecomposable projective module, we see that $A_1$ is uniquely determined.
We let $\mathcal V_i$ denote the sincere antichains $A$ such that $A_1 = P(i)$. For $2\le j \le r$, we have $\operatorname{Hom}(P(i),A_j) = \operatorname{Hom}(A_1,A_j) = 0$. It follows that $(A_2,\dots, A_r)$ is an antichain with support in $[1,i-1]\cup[i+1,n]$. Altogether, we see that any element of $A$ has support either in $[1,i]$ or in $[i+1,n]$. The elements of $A$ with support in $[1,i]$ but different from $A_1$ form an arbitrary antichain with support in $[2,i-1]$, thus the number of elements is $a(\mathbb A_{i-2})$, at least if $i\ge 2$. Note that $a(\mathbb A_{i-2}) = a_{i-1}(\mathbb A_{i-1})$.
The elements of $A$ with support in $[i+1,n]$ form a sincere antichain for $\mathbb B_{n-i}$. Thus the number of such antichains is $a_{n-i}(\mathbb B_{n-i})$. This shows that for $i\ge 2$, the set $\mathcal V_i$ has cardinality $a_{i-1}(\mathbb A_{i-1})a_{n-i}(\mathbb B_{n-i})$. This formula holds true also for $i=1$, since the number of elements of $\mathcal V_1$ is $a_{n-1}(\mathbb B_{n-1})$ and $a_0(\mathbb A_0) = 1$. Thus we see that
$$\begin{aligned}
v(\mathbb B_n) &=& \sum\nolimits_{i=1}^{n-1} a_{i-1}(\mathbb A_{i-1})a_{n-i}(\mathbb B_{n-i}) \cr
&=& -a_{n-1}(\mathbb A_{n-1})
+\sum\nolimits_{i=1}^{n} a_{i-1}(\mathbb A_{i-1})a_{n-i}(\mathbb B_{n-i}) \cr
&=& -\frac1n\binom{2n-2}{n-1} + \binom{2n-2}{n-1} = \binom{2n-2}{n-2}.\end{aligned}$$
Altogether we see $$u(\mathbb B_n) + v(\mathbb B_n) = \binom{2n-2}{n-1} + \binom{2n-2}{n-2}=
\binom{2n-1}{n-1}.$$
The calculation of $v(\mathbb B_n)$ shows the following relationship between the cases $\mathbb A$ and $\mathbb B$: $$a_{n-1}(\mathbb B_n) = a_{n-2}(\mathbb B_{n+1}) + a_{n-1}(\mathbb A_{n-1}).$$
Support-tilting modules: the hook formula {#hook}
=========================================
The hook formula
----------------
\[hook-formula\] Let $\Delta = \mathbb A, \mathbb B, \mathbb D, \mathbb E$. Then $$a_s(\Delta_n) = a_s(\Delta_{n-1}) + a_{s-1}(\Delta_n)$$ for all $n\ge m$ and $1\le s \le n-c$, where $m = 1,2,3,4$ and $c=0,1,2,3$ for $\Delta = \mathbb A, \mathbb B,
\mathbb D, \mathbb E$, respectively.
Here we use the convention that $\mathbb B_1 = \mathbb A_1,
\mathbb D_2 = \mathbb A_1\sqcup \mathbb A_1,
\mathbb E_3 = \mathbb A_2\sqcup \mathbb A_1, \mathbb E_4 = \mathbb A_4, \mathbb E_5 = \mathbb D_5$. In the triangles \[triangle1\], \[triangle2\], \[triangle3\], as well as in Section \[exceptional\], this equality concerns the following kind of hooks: $$\hbox{\beginpicture
\setcoordinatesystem units <.9cm,.9cm>
\multiput{} at 0 0 3 3 /
\plot 0 0 0 3 3 0 /
\plot 0.7 0.7 1.3 0.7 1.3 1.3 1 1.3 1 1 0.7 1 0.7 0.7 /
\multiput{$\bullet$} at 0.85 0.85 1.15 1.15 /
\put{$\circ$} at 1.15 0.85
\endpicture}$$ The hook formula asserts that the sum of the values at the positions marked by bullets is the value at the position marked by the circle.
The various assertions concern the following general situation: up to the choice of an orientation, we deal with an artin algebra $\Lambda$ with the following valued quiver with $n$ vertices: $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\multiput{} at 0 1 8 -1 /
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-c-1\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-c\endsmallmatrix}$} at 5 -.3
{\arrow <1.5mm> [0.25,0.75] from 5.6 0.3 to 5.2 0.1}
{\arrow <1.5mm> [0.25,0.75] from 5.6 -.3 to 5.2 -.1}
\setdots <.5mm>
\setquadratic
\plot 5.4 0 6 .5 7 .5 7.5 0 7 -.5 6 -.5 5.4 0 /
\endpicture}$$ on the left, we have a quiver of type $\mathbb A_{n-c}$ with arrows $i \leftarrow i\!+\!1$. The remaining $c$ vertices are in the dotted “cloud” to the right. All arrows between the cloud and the $\mathbb A_{n-c}$–quiver end in the vertex $n-c$. We let $Q'$ denote the valued quiver obtained by deleting the vertex $1$ and the arrow ending in $1$; let $\Lambda'$ be the corresponding factor algebra of $\Lambda$. Here are the cases we are interested in.
$$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-1\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n\endsmallmatrix}$} at 5 -.3
\setdots <1mm>
\setquadratic
\plot 5.2 0.1 6 .2 7 .2 7.5 0 7 -.2 6 -.2 5.2 -.1 /
\put{$\mathbb A_n$} at -1 0
\put{$c = 0$} at 9 0
\endpicture} at 0 1
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-2\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-1\endsmallmatrix}$} at 5 -.3
\plot 5.8 0.03 5.2 0.03 /
\plot 5.8 -.03 5.2 -.03 /
\plot 5.3 0.1 5.1 0 5.3 -.1 /
\plot 5.4 0.1 5.6 0 5.4 -.1 /
\multiput{$\circ$} at 6 0 /
\setdots <1mm>
\setquadratic
\plot 5.2 0.1 6 .2 7 .2 7.5 0 7 -.2 6 -.2 5.2 -.1 /
\put{$\mathbb B_n$} at -1 0
\put{$c = 1$} at 9 0
\endpicture} at 0 0
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-3\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-2\endsmallmatrix}$} at 4.9 -.3
{\arrow <1.5mm> [0.25,0.75] from 5.8 0.5 to 5.2 0.1}
{\arrow <1.5mm> [0.25,0.75] from 5.8 -.5 to 5.2 -.1}
\multiput{$\circ$} at 6 .5 6 -.5 /
\setdots <1mm>
\setquadratic
\plot 5.2 0.1 6 .8 7 .8 7.5 0 7 -.8 6 -.8 5.2 -.1 /
\put{$\mathbb D_n$} at -1 0
\put{$c = 2$} at 9 0
\endpicture} at 0 -1.5
\put{\beginpicture
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-4\endsmallmatrix}$} at 4 -.3
\put{${\smallmatrix n-3\endsmallmatrix}$} at 4.9 -.3
{\arrow <1.5mm> [0.25,0.75] from 5.8 0.5 to 5.2 0.1}
{\arrow <1.5mm> [0.25,0.75] from 5.8 -.5 to 5.2 -.1}
{\arrow <1.5mm> [0.25,0.75] from 6.8 -.5 to 6.2 -.5}
\multiput{$\circ$} at 6 .5 6 -.5 7 -.5 /
\setdots <1mm>
\setquadratic
\plot 5.2 0.1 6 .8 7 .8 7.5 0 7 -.8 6 -.8 5.2 -.1 /
\put{} at -1 0
\put{$\mathbb E_n$} at -1 0
\put{$c = 3$} at 9 0
\endpicture} at 0 -3.5
\put{} at 0 -4.7
\endpicture}$$
Let $1\le s \le n-c$. Then $$a_s(\Lambda) = a_s(\Lambda') + a_{s-1}(\Lambda).$$
The support-tilting modules $T$ for $\Lambda$ with $1$ not in the support are just the support-tilting modules for $\Lambda'$. Let $\mathcal S_s(\Lambda;1)$ be the set of the basic support-tilting $\Lambda$-modules $T$ with support-rank $s$ and $\operatorname{Hom}(P(1),T) \neq 0$. Let $\mathcal S_{s-1}(\Lambda)$ be the set of basic support-tilting $\Lambda$-modules $T$ with support-rank $s-1$. We construct a bijection $$\alpha:\mathcal S_s(\Lambda;1) \longrightarrow \mathcal S_{s-1}(\Lambda).$$ This will establish the formula.
Let $X$ be an indecomposable representation with support-rank $s \le n-c$ and $\operatorname{Hom}(P(1),X) \neq 0$. Then the support of $X$ is contained in the $\mathbb A_{n-c}$-subquiver, so $X$ is thin and its support is an interval of the form $[1,v]$ with $1\le v \le n-c$ (a module is said to be [*thin*]{} provided the composition factors are pairwise non-isomorphic; in our setting thin indecomposable modules are uniquely determined by the support, thus we may just write $X = [1,v]$).
Let $T$ be a module in $\mathcal S_s(\Lambda;1)$. At least one of the indecomposable direct summand of $T$, say $X$, satisfies $\operatorname{Hom}(P(1),X)\neq 0$ and we choose $X = [1,v]$ of largest possible length. We claim that $\operatorname{Hom}(P(w),T) = 0$ for any arrow $v \leftarrow w$. Assume, to the contrary, that there is an indecomposable direct summand $Y$ of $T$ with $\operatorname{Hom}(P(w),Y) \neq 0$. The maximality of $X$ shows that $\operatorname{Hom}(P(1),Y) = 0$. But then $\operatorname{Ext}(Y,X) \neq 0$ contradicts the fact that $T$ has no self-extensions. (Namely, if the support of $X$ and $Y$ is disjoint, then the arrow $v\leftarrow w$ yields directly a non-trivial extension of $X$ by $Y$; if the support of $X$ and $Y$ is not disjoint, then there is a proper non-zero factor module of $X$ which is a proper submodule of $Y$, thus there is a non-zero map $X \to Y$ which is neither injective nor surjective — again we obtain a non-trivial extension of $X$ by $Y$.) Thus the support of $T$ is the disjoint union of the set $\{1,2,\dots,v\}$ and a set $S''$ which does not contain a vertex $w$ with an arrow $v\leftarrow w$.
The indecomposable direct summands of $T$ with support in $\{1,2,\dots,v\}$ yield a tilting module for this $\mathbb A_v$-quiver, and $X$ is the indecomposable projective-injective representation of this $\mathbb A_v$-quiver. Deleting $X$ from this tilting module, we obtain a support-tilting representation of $\mathbb A_v$ with support-rank $v-1$.
Thus if we write $T = X\oplus T'$, then $T'$ is a support-tilting $\Lambda$-module with support-rank $s-1$ (namely, it is the direct sum of a support-tilting module with support properly contained in $\{1,2,\dots,v\}$ and a support-tilting module with support $S''$). We define $\alpha(T) = T'$; this yields the map $$\alpha:\mathcal S_s(\Lambda;1) \longrightarrow \mathcal S_{s-1}(\Lambda)$$ we are looking for. It remains to be shown that $\alpha$ is surjective and that we can recover $T$ from $\alpha(T)$.
Thus, let $T'$ be in $\mathcal S_{s-1}(\Lambda)$. Then there are at least $c+1$ vertices outside of the support of $T'$.
Case 1: These are the vertices in the cloud and precisely one additional vertex, say $i$ (with $1\le i \le n-c$). Note that in this case $s = n-c$. Let $T = T'\oplus [1,n-c]$. Since $T'$ is a support-tilting module of $\mathbb A_{n-c}$ with support-rank $n-c-1$ and $[1,n-c]$ is the indecomposable projective-injective representation of $\mathbb A_{n-c}$, we see that $T = T'\oplus[1,n-c]$ is a tilting module for $\mathbb A_{n-c}$.
Case 2: At least two vertices between $1$ and $n-c$ do not belong to $\operatorname{Supp}T'$, say let $i<j$ be the smallest such numbers. Then let $T = T'\oplus [1,j-1]$.
The modified hook formula
-------------------------
\[modified\] $$\begin{aligned}
a_{n-1}(\mathbb D_n) &=& a_{n-1}(\mathbb D_{n-1}) + a_{n-2}(\mathbb D_n) + a_{n-2}(\mathbb A_{n-2}), \\
a_{n-2}(\mathbb E_n) &=& a_{n-2}(\mathbb E_{n-1}) + a_{n-3}(\mathbb E_n) + a_{n-3}(\mathbb A_{n-3}).\end{aligned}$$
Again, we consider a general setting, namely we consider an artin algebra $\Lambda$ with the following valued quiver with $n$ vertices and we assume that $c \ge 2$: $$\hbox{\beginpicture
\setcoordinatesystem units <1cm,1cm>
\multiput{} at 0 1 8 -1 /
\multiput{$\circ$} at 0 0 1 0 2 0 4 0 5 0 6 0.5 6 -0.5 /
\put{$\cdots$} at 3 0
{\arrow <1.5mm> [0.25,0.75] from 0.8 0 to 0.2 0}
{\arrow <1.5mm> [0.25,0.75] from 1.8 0 to 1.2 0}
{\arrow <1.5mm> [0.25,0.75] from 2.5 0 to 2.2 0}
{\arrow <1.5mm> [0.25,0.75] from 4.8 0 to 4.2 0}
\plot 3.8 0 3.5 0 /
\put{${\smallmatrix 1\endsmallmatrix}$} at 0 -.3
\put{${\smallmatrix 2\endsmallmatrix}$} at 1 -.3
\put{${\smallmatrix 3\endsmallmatrix}$} at 2 -.3
\put{${\smallmatrix n-c-1\endsmallmatrix}$} at 3.9 -.3
\put{${\smallmatrix n-c\endsmallmatrix}$} at 4.9 -.3
\put{${\smallmatrix n-c+1\endsmallmatrix}$} at 6 .8
\put{${\smallmatrix n-c+2\endsmallmatrix}$} at 6 -.8
{\arrow <1.5mm> [0.25,0.75] from 5.8 0.4 to 5.2 0.1}
{\arrow <1.5mm> [0.25,0.75] from 5.8 -.4 to 5.2 -.1}
\setdots <.5mm>
\setquadratic
\plot 5.4 0 6 .5 7 .5 7.5 0 7 -.5 6 -.5 5.4 0 /
\endpicture}$$ On the left, we have a quiver of type $\mathbb A_{n-c}$ with arrows $i \leftarrow i+1$, and the remaining $c$ vertices are in the dotted “cloud” to the right. There are precisely two vertices in the cloud, namely $n-c+1$ and $n-c+2$ with arrows $n-c \leftarrow n-c+1$ and $n-c \leftarrow n-c+2$ and there is no other arrows between the cloud and the $\mathbb A_{n-c}$ quiver. Again, we let $Q'$ denote the valued quiver obtained by deleting the vertex $1$ and the arrow ending in $1$ and by $\Lambda'$ the corresponding factor algebra of $\Lambda$ and we show the following:
\[mod-lemma\] $$a_{n-c+1}(\Lambda) = a_{n-c+1}(\Lambda') + a_{n-c}(\Lambda) + a_{n-c}(\mathbb A_{n-c}).$$
The proof follows closely the proof of Proposition \[hook-formula\]. The support-tilting modules $T$ for $\Lambda$ with $1$ not in the support are just the support-tilting modules for $\Lambda'$. We construct a surjection $\alpha$ from the set $\mathcal S_{n-c+1}(\Lambda;1)$ of the support-tilting $\Lambda$-modules $T$ with support-rank $n-c+1$ and $\operatorname{Hom}(P(1),T) \neq 0$ onto the set $\mathcal S_{n-c}(\Lambda)$ of support-tilting $\Lambda$-modules $T$ with support-rank $n-c$. In the present setting, $\alpha$ will not be injective, but will be a double cover: pairs in $\mathcal S(\Lambda;1)$ are identified by $\alpha$; the number of such pairs will be just $a_{n-c}(\mathbb A_{n-c})$.
As above, one shows that any module $T$ in $\mathcal S_{n-c+1}(\Lambda;1)$ is of the form $T = X\oplus T'$ where $X$ is indecomposable, $\operatorname{Hom}(P(1),X)\neq 0$ and $X$ is of maximal possible length. Note that the support of $X$ is contained either in $\{1,2,\dots,n-c+1\}$ or in $\{1,2,\dots,n-c,n-c+2\}$. In particular, $X$ is uniquely determined (since the support of $T$ cannot contain all the vertices $1,2,\dots,n-c+2$). As above, the mapping $\alpha$ will be the deletion of the summand $X$.
Let $Z$ be the indecomposable module with support $\{1,2,\dots,n-c+1\}$ and $Z'$ the indecomposable module with support $\{1,2,\dots,n-c,n-c+2\}$. Starting with a tilting module $T'$ for $\mathbb A_{n-c}$, we may form the direct sums $Z\oplus T'$ and $Z'\oplus T'$. Then these are elements of $\mathcal S_{n-c+1}(\Lambda;1)$, both of which are mapped under $\alpha$ to the same module $T'$. These are the $a_{n-c}(\mathbb A_{n-c})$ pairs of elements of $\mathcal S(\Lambda;1)$ which are identified by $\alpha$.
It follows that $\mathcal S(\Lambda;1)$ has cardinality $a_{n-c}(\Lambda) + a_{n-c}(\mathbb A_{n-c})$.
The two assertions of Proposition \[modified\] are special cases of Lemma \[mod-lemma\]. For the first assertion, $\Lambda$ is of type $\mathbb D_n$, $\Lambda'$ of type $\mathbb D_{n-1}$, and $c = 2$. For the second assertion, $\Lambda$ is of type $\mathbb E_n$, $\Lambda'$ of type $\mathbb E_{n-1}$ and $c = 3$.
For another proof of the modified hook formula, see Hubery [@[H]].
\[hook-cor\] $$a_{n-1}(\mathbb D_n) = \left[
\begin{matrix} 2n-3\cr n-1\end{matrix} \right].$$
We start with the previous observation $$\begin{aligned}
a_{n-1}(\mathbb D_n)
&=&a_{n-1}(\mathbb D_{n-1}) + a_{n-2}(\mathbb D_n) + a_{n-2}(\mathbb A_{n-2}) \\
&=& \frac{3n-7}{2n-4}\binom{2n-4}{n-3}
+\frac{3n-6}{2n-4}\binom{2n-4}{n-2}
+\frac{1}{n-1}\binom{2n-4}{n-2}.\end{aligned}$$ Write $$\begin{aligned}
\binom{2n-4}{n-3} &=& \frac{n-2}{2n-3}\binom{2n-3}{n-1}, \\
\binom{2n-4}{n-2} &=& \frac{n-1}{2n-3}\binom{2n-3}{n-1}.\end{aligned}$$ One easily shows that $$\frac{3n-7}{2n-4}\cdot \frac{n-2}{2n-3} + \frac{3n-6}{2n-4}\cdot\frac{n-1}{2n-3} +
\frac1{n-1}\cdot\frac{n-1}{2n-3}\ =\ \frac{3n-4}{2n-3}.$$ As a consequence, we get $$\begin{aligned}
a_{n-1}(\mathbb D_n)
&=&\frac{3n-4}{2n-3}\binom{2n-3}{n-1} \\
&=& \left[\begin{matrix} 2n-3\cr n-1\end{matrix} \right].\end{aligned}$$
Summation formulas {#summation}
==================
An immediate consequence of the previous section is the following assertion:
Let $\Delta = \mathbb A$, or $\mathbb B$ and $n\ge 0$, or $\Delta =
\mathbb D$ and $n\ge 2$. If $1\le s \le n-1$, then $$\sum\nolimits_{i=0}^s a_i(\Delta_n) = a_s(\Delta_{n+1}).$$
We use induction. For $s = 0$ both sides are equal to $1$. For $s \ge 1$ we have $$\begin{aligned}
\sum\nolimits_{i=0}^s a_i(\Delta_n)
&=& a_s(\Delta_n) + \sum\nolimits_{i=0}^{s-1} a_i(\Delta_n) \\
&=& a_s(\Delta_n) + a_{s-1}(\Delta_{n+1}) \\
&=& a_s(\Delta_{n+1}),\end{aligned}$$ the last equality being the hook formula.
Let $\Delta = \mathbb A, \mathbb B$ or $\mathbb D$. Then $$a(\Delta_n) = a_n(\Delta_n)+a_{n-1}(\Delta_{n+1})$$
[**Case $\mathbb A_n$**]{} $$a(\mathbb A_n) = \frac 1{n+1}\binom{2n}{n}+
\frac 3{n+2}\binom{2n}{n-1}
= \frac 1{n+2}\binom{2n+2}{n+1}.$$
[**Case $\mathbb B_n$**]{} $$a(\mathbb B_n) = \binom{2n-2}{n-1} + \binom{2n-1}n = \binom{2n}n.$$
[**Case $\mathbb D_n$**]{} $$a(\mathbb D_n) = \left[\begin{matrix} 2n-2\cr n-2 \end{matrix}\right] +
\left[\begin{matrix} 2n-2\cr n-1 \end{matrix}\right]
= \left[\begin{matrix} 2n-1\cr n-1 \end{matrix}\right].$$
**Acknowledgment**
==================
The authors are indebted to Henning Krause and Dieter Vossieck for providing the references [@[GP]] and [@[BLR]], and to Lutz Hille for helpful discussions concerning the problem of determining the number of tilting modules. They thank the referee for pointing out mistakes in the proof of Corollary \[hook-cor\]. Andrew Hubery has to be praised for his careful reading of the manuscript. His detailed comments have improved the presentation considerably.
This work is funded by the Deanship of Scientific Research, King Abdulaziz University, under grant No. 2-130/1434/HiCi. The authors, therefore, acknowledge technical and financial support of KAU.
[99]{}
L. Angeleri-Hügel, D. Happel, and H. Krause, [*Handbook of Tilting Theory,*]{} London Math. Soc. Lecture Note Series, Vol. 332, Cambridge University Press, 2007. D. Armstrong, [*Generalized Noncrossing Partitions and Combinatorics of Coxeter Groups.*]{} Mem. Amer. Math. Soc., Vol. 949, 2009. M. Auslander, M. I. Platzeck, and I. Reiten, Coxeter functors without diagrams, [*Trans. Amer. Math. Soc.*]{} [**250**]{} (1979), 1–46. M. Auslander, I. Reiten, and S. Smalø, [*Representation Theory of Artin Algebras,*]{} Cambridge Studies in Advanced Mathematics, Vol. 36, Cambridge University Press, 1997. D. F. Bailey, Counting arrangements of $1$’s and $-1$’s, [*Math. Mag.*]{} [**69**]{} (1996), 128–131. O. Bretscher, Chr. Läser, and Chr. Riedtmann, Selfinjective and simply connected algebras. [*Manuscripta Math.,*]{} [**36**]{} (1981), 253–307. F. Chapoton, Enumerative properties of generalized associahedra, [*Sém. Lothar. Combin.*]{} [**51**]{} (2004). V. Dlab and C. M. Ringel, On algebras of finite representation type, [*J. Algebra*]{} [**33**]{} (1975), 306-394 V. Dlab and C. M. Ringel, [*Indecomposable Representations of Graphs and Algebras,*]{} Mem. Amer. Math. Soc., Vol. 173, 1976. W. Fakieh and C. M. Ringel, The hereditary artinian rings of type $\mathbb H_3$ and $\mathbb H_4$, in preparation. S. Fomin and N. Reading, Root systems and generalized associahedra, in [*Geometric combinatorics.*]{} IAS/Park City Math. Ser., 13, Amer. Math. Soc., 2007, 63–131. S. Fomin and A. Zelevinsky, Y-systems and generalized associahedra. [*Ann. of Math.*]{} (2) [**158**]{} (2003), 977–1018. P. Gabriel, Un jeu? Les nombres de Catalan, [*Uni Zürich, Mitteilungsblatt des Rektorats,*]{} 12. Jahrgang, Heft [**6**]{} (1981), 4–5. P. Gabriel and J. A. de la Peña, Quotients of representation-finite algebras, [*Comm. Algebra*]{} [**15**]{} (1987), 279–307 D. Happel and C. M. Ringel, Tilted algebras. [*Trans. Amer. Math. Soc.*]{} [**274**]{} (1982), 399–443. D. Happel and C. M. Ringel, Construction of tilted algebras, in [*Representations of Algebras: Proceedings ICRA 3.*]{} Lecture Notes in Math., Vol. 903, Springer, 1981, pp. 125–167.
A. Hubery, On the modified hook formula, in preparation. A. Hubery and H. Krause, A categorification of non-crossing partitions, [http://arxiv.org/abs/1310.1907,]{} to appear in [*J. Eur. Math. Soc.*]{}
C. Ingalls and H. Thomas, Noncrossing partitions and representations of quivers. [*Compos. Math.*]{} [**145**]{} (2009), 1533–1562. Chr. Krattenthaler, The F-triangle of the generalised cluster complex. [*Topics in discrete mathematics,*]{} Algorithms Combin. 26, Springer, 2006, pp. 93–126. H. Krause, Thick subcategories of modules over commutative noetherian rings (with an appendix by Srikanth Iyengar), [*Math. Ann.*]{} [**340**]{} (2008), 733–747. S. Ladkani, Universal derived equivalences of posets of tilting modules. [http://arxiv.org/abs/0708.1287.]{} R. Marsh, M. Reineke, and A. Zelevinsky, Generalized associahedra via quiver representations, [*Trans. Amer. Math. Soc.*]{} [**355**]{} (2003), 4171–4186. A. A. Obaid, S. K. Nauman, W. S. Al Shammakh, W. M. Fakieh, and C. M. Ringel, The number of complete exceptional sequences for a Dynkin algebra. [*Colloq. Math.*]{} [**133**]{} (2013), 197–210
A. A. Obaid, S. K. Nauman, W. M. Fakieh, and C. M. Ringel, The Ingalls-Thomas bijections, preprint. D. I. Panyushev, Ad-nilpotent ideals of a Borel subalgebra: generators and duality. [*J. Algebra*]{} [**274**]{} (2004), 822–846
C. M. Ringel, Reflection functors for hereditary algebras. [*J. Lond. Math. Soc.*]{} (2) [**21**]{} (1980), 465–479.
C. M. Ringel, The Catalan combinatorics of the hereditary artin algebras. [http://arxiv.org/abs/1502.06553.]{}
N. J. A. Sloane, Online Encyclopedia of Integer Sequences. [http://oeis.org/]{}
------------------------------------------------------------------------
2010 [*Mathematics Subject Classification*]{}: Primary: 05E10. Secondary: 16G20, 16G60, 05A19, 16D90, 16G70.
*Keywords:* Dynkin algebra, Dynkin diagram, tilting module, support-tilting module, lattice of non-crossing partitions, cluster combinatorics, generalized Catalan number, Catalan triangle, Pascal triangle, Lucas triangle, categorification.
------------------------------------------------------------------------
(Concerned with sequences , , , , , , .)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We prove an index theorem for the existence of Majorana zero modes in a semiconducting thin film with a sizable spin-orbit coupling when it is adjacent to an $s$-wave superconductor. The theorem, which is analogous to the Jackiw-Rebbi index theorem for the zero modes in mass domain walls in one-dimensional Dirac theory, applies to vortices with odd flux quantum in a semiconducting film in which $s$-wave superconductivity and a Zeeman splitting are induced by proximity effect. The momentum-space construction of the zero-mode solution presented here is complementary to the approximate real-space solution of the Bogoliubov-de Gennes equations at a vortex core \[J. D. Sau *et al.*, arXiv:0907.2239\], proving the existence of non-degenerate zero-energy Majorana excitations and the resultant non-Abelian topological order in the semiconductor heterostructure. With increasing magnitude of the proximity-induced pairing potential, the non-Abelian superconducting state makes a topological quantum phase transition to an ordinary $s$-wave superconducting state which no topological order.'
address:
- 'Department of Physics and Astronomy, Clemson University, Clemson, SC 29634'
- 'Condensed Matter Theory Center and Joint Quantum Institute, Department of Physics, University of Maryland, College Park, Maryland 20742-4111'
author:
- Sumanta Tewari
- 'Jay D. Sau'
- 'S. Das Sarma'
title: 'A theorem for the existence of Majorana fermion modes in spin-orbit-coupled semiconductors'
---
Introduction
============
Particle statistics is a genuinely quantum mechanical concept which has no classical analog. In spatial dimension three and above, the wave function of a many-body quantum state of identical particles remains either unchanged (bosons) or undergoes a change of sign (fermions) under a pairwise interchange of the particle coordinates. That there are only two possibilities is a consequence of the fact that there are only two irreducible representations of the permutation group for $N$ particles. However, in $(2+1)$ dimension, where permutation and exchange are not necessarily equivalent, the quantum statistics of particles can be remarkably different from the ordinary statistics of bosons and fermions [@Leinaas; @Wilczek]. In this case, under a pairwise interchange of the particle coordinates, it is possible for the many body wave function to receive an arbitrary phase factor $e^{i\theta}$ where $\theta$ is an angle intermediate between $0$ (boson) and $\pi$ (fermion). The particles obeying statistics given by the angle $\theta$ are called anyons [@Wilczek2]. Even if the anyon statistics is remarkably different from that of bosons and fermions, since the simple phase factor is only a one-dimensional representation of the braid group in 2D, the statistics is still Abelian. A more exotic possibility, one which has remarkable prospect for fault-tolerant topological quantum computation (TQC) [@Nayak-RMP], arises when the quantum ground state of the many-particle system is degenerate. In situations where the many body ground state wave function is a linear combination of states from this degenerate ground state subspace, a pairwise exchange of the particle coordinates can unitarily *rotate* the wave-function in the subspace. In this case, the exchange statistics is given by a multi-dimensional unitary matrix representation of the 2D braid group, and, thus, the statistics is non-Abelian. It has been proposed [@Kitaev] that such a system, where the ground state degeneracy is protected by a gap from local perturbations, can be used as a fault-tolerant platform for TQC.
Recently, some exotic ordered states in condensed matter systems, such as the Pfaffian states in fractional quantum Hall (FQH) systems [@Moore; @Nayak-Wilczek; @Read; @dassarma_prl'05] and chiral $p$-wave superconductors/superfluids [@Ivanov; @Stern; @DasSarma_PRB'06; @tewari_prl'2007], as well as the surface state of a topological insulator (TI) in which $s$-wave superconductivity is induced by proximity effect [@fu_prl'08; @fu_prl'09; @akhmerov_prl'09], have been proposed as systems which support quasiparticles with non-Abelian statistics, and, therefore, can potentially be used as TQC platforms. The common thread between these systems is that they all allow quasiparticle excitations which involve no energy cost (when the mutual separation among the excitations is large). The second quantized operators, $\gamma_i$, corresponding to these zero energy excitations are self-hermitian, $\gamma_i^{\dagger}=\gamma_i$, which is in striking contrast to ordinary fermionic (or bosonic) operators for which $c_i \neq c_i^{\dagger}$. However, since $\gamma_i$ and $\gamma_j$, which are called Majorana fermion operators, anticommute when $i \neq j$, they retain some properties of ordinary fermion operators as well. It is the self-hermitian property of the Majorana operators which lies at the heart of the ground state degeneracy and the resulting non-Abelian statistics [@Nayak-Wilczek; @Ivanov] of quasiparticle excitations in these systems.
It has been shown recently [@Sau] that even a regular semiconducting film with a sizable spin-orbit coupling, such as InGaAs thin films, can host, under suitable conditions, Majorana fermion excitations localized near defects. By an analysis of the real-space Bogoliubov-de Gennes (BdG) equations for a vortex in the semiconductor, in which $s$-wave superconductivity and a Zeeman splitting are proximity induced (Fig. (1a)), it has been shown that the lowest energy quasiparticle excitation is a zero-energy Majorana fermion mode. The Majorana mode is separated by a finite energy gap (so-called mini-gap) from the other conventional fermionic excited states in the vortex core. Thus, for a collection of well-separated vortices, the resulting degenerate ground state subspace is protected from the environment by the mini-gap, enabling the potential use of the semiconductor heterostructure in Fig. (1a) in TQC. Since the basic effects behind the emergence of the Majorana fermion excitations – spin-orbit coupling, $s$-wave superconductivity, and Zeeman splitting – are physically well-understood and experimentally known to occur in many solid state materials, the proposed semiconductor heterostructure [@Sau] is possibly one of the simplest systems to realize, which supports non-Abelian topological order.
In this paper, we show that the existence of the Majorana fermion zero modes at the cores of the vortices in the semiconductor heterostructure in Fig. (1a) is due to an index theorem. The theorem is analogous to the one proven by Jackiw and Rebbi [@Jackiw1; @Jackiw2] for the existence of zero energy eigenstates at mass domain walls in a 1D system of Dirac fermions, which was later applied [@Sumanta] to prove the existence of zero-energy excitations at vortices in a spinless 2D chiral $p$-wave superconductor. In analogy with the chiral $p$-wave superconductor, we find that while an odd flux-quantum vortex in the semiconductor thin film traps a unique zero energy eigenstate, an even flux-quantum vortex does not. Furthermore, since the components of the fermion field in our analog of the 1D Dirac theory are pairwise related by parity reversal and hermitian conjugation, we explicitly show that the second quantized operator corresponding to the zero-energy eigensolution is self-hermitian, $\gamma^{\dagger}=\gamma$. This is in contrast to the corresponding problem solved by Jackiw and Rebbi for the conventional Dirac theory, where the two components of the fermion field are independent, and, therefore, the zero-energy excitations localized in the mass domain walls are conventional fermion excitations. The momentum space construction of the zero-energy solution presented here is complementary to the approximate real space solution [@Sau] of the four-component BdG equations at a vortex in the semiconductor in which superconductivity has been induced by the proximity effect.
Hamiltonian
===========
The single-particle effective Hamiltonian $H_0$ for the conduction band of a spin-orbit-coupled semiconducting thin film in contact with a magnetic insulator (which induces the Zeeman splitting) is given by ($\hbar=1$) $H_0=H_K + H_{SO} + H_Z$, where, $$\begin{aligned}
&H_K&=\sum_{\textbf{k},\beta}(\frac{k^2}{2m^*}-\mu)c^{\dagger}_{\textbf{k},\beta}c_{\textbf{k},\beta}\nonumber\\
&H_{SO}&=-\alpha\sum_{\textbf{k},\beta,\gamma}[(\textbf{k}\times\mathbf{\sigma}).\hat{z}]_{\beta,\gamma}
c^{\dagger}_{\textbf{k},\beta}c_{\textbf{k},\gamma} + {\rm{H. c.}}\nonumber\\
&H_Z&=V_z\sum_{\textbf{k},\beta,\gamma}(\sigma_z)_{\beta,\gamma}
c^{\dagger}_{\textbf{k},\beta}c_{\textbf{k},\gamma}
\label{H_0}\end{aligned}$$ Here, $m^*$, $V_z$ and $\mu$ are the conduction-band effective mass of an electron, effective Zeeman coupling induced by proximity to a magnetic insulator, and chemical potential, respectively. The Zeeman coupling can also be intrinsic if the film is made of a magnetic semiconductor (e.g., GaMnAs). We assume that the Zeeman coupling in the semiconductor is the dominant effect of the proximity to the magnetic insulator and the direct magnetic field induced by the magnetic insulator is neglected.
![(a): The proposed heterostructure of a semiconductor (SM) sandwiched between an $s$-wave superconductor (SC) and a magnetic insulator (MI). In this geometry, the semiconducting film can support non-Abelian topological order. (b): Single-particle band-structure in the semiconducting film with and without the Zeeman splitting. Without the Zeeman splitting, the two spin-orbit shifted bands touch at $k=0$ (red lines). Then, for any value of the chemical potential, the system has two Fermi surfaces. With a finite Zeeman splitting, the bands have an energy gap near $k=0$ (blue lines). If the chemical potential lies in the gap, the system just has one Fermi surface (indicated by the dotted circle), a situation conducive to the emergence of non-Abelian order.[]{data-label="Fig1"}](Fig_C_Index.eps){width="0.65\linewidth"}
The coefficient $\alpha$ describes the strength of the Rashba spin-orbit coupling in the semiconductor, $\beta, \gamma$ are the spin indices, and $\sigma_x, \sigma_y, \sigma_z$ are the Pauli matrices. The energy spectrum of the spin-orbit-coupled semiconductor with $V_z=0$ has two bands crossing the Fermi level. This situation should be contrasted with the metallic surface state of a strong TI where an odd number of bands cross the Fermi level. However, by placing the semiconductor in the proximity to a magnetic insulator, it is possible to open a gap in the spectrum, see Fig. (1b). Then, for $|\mu| < |V_z|$, a single band crosses the Fermi level. This situation is then analogous to the surface of a strong TI in that an odd number of bands cross the Fermi level [@fu_prl'08], and, therefore, is suggestive of supporting non-Abelian topological order if $s$-wave superconductivity can be induced in the film. This way, by replacing the surface of a 3D strong topological insulator by a regular semiconducting thin-film, the scope of solid state systems which can be designed to support non-Abelian topological order can be vastly expanded [@Sau].
In the presence of an adjacent $s$-wave superconductor, which induces an $s$-wave pairing potential in the semiconductor by the proximity effect, the full Hamiltonian in the bulk semiconductor becomes, $H_B=H_0 + H_p$, where, $$H_p = -\Delta_0 \sum_{\textbf{k}}c^{\dagger}_{\textbf{k},\uparrow}c^{\dagger}_{-\textbf{k},\downarrow} + {\rm{H. c.}}
\label{H_p}$$ We show below that the heterostructure in Fig. (1a) has non-Abelian topological order by proving a theorem for the existence of a Majorana zero energy mode at the core of a vortex with an odd number of flux-quantum in the superconductor. Our momentum space construction of the zero mode solution in the form of an index theorem is complementary to the analysis of the real space BdG equations recently carried out [@Sau] for the same system.
In order to describe the spatial dependence of the superconducting order parameter in the presence of a vortex with winding number $1$, we first write [@Sumanta] the pairing part of the Hamiltonian in real space, $$H_{V}=-\int d^2R\int d^2r
e^{i\theta_{\bf{R}}}h(R)g({\bf{r}})c^{\dagger}_{{\bf{R}}+{\bf{r}},\uparrow}c^{\dagger}_{{\bf{R}}-{\bf{r}},\downarrow}+
\rm{H.c.}. \label{Pairingvortex}$$ Here, $\bf{R}$ and $\bf{r}$ are the center-of-mass and the relative coordinates of a Cooper pair, respectively. $h(R)$ and $\theta_{\bf{R}}$ are the amplitude and the phase of the superconducting order parameter, and, in momentum space, $g(\mathbf{k})$ is the $s$-wave order parameter $\Delta_0$. In the vortex core, $h(R)\sim(1-e^{-\frac{R}{\xi}})$ with $\xi$ the coherence length, which describes the suppression of the condensate amplitude at the core of a vortex. In order to prove the existence of a non-degenerate zero energy state at the vortex core, we have to map the 2D Hamiltonian $H_T=H_0+H_V$ on the Jackiw-Rebbi problem of 1D Dirac theory with a mass domain wall. Before we do this, let us briefly discuss the derivation of the zero-mode in the Jackiw-Rebbi problem [@Jackiw1] itself.
Fermion zero mode in 1D Dirac theory
====================================
Let us briefly review the derivation of the Jackiw-Rebbi zero mode for the 1D problem described by the Dirac Hamiltonian, $$H_{D}=\int
dx\Big[-iv_F \psi^{\dagger} \sigma_z \partial_x \psi
+m(x)\psi^{\dagger}\sigma_x \psi\Big],
\label{h2}$$ where $\psi^{\dagger}(x)=(f^{\dagger}_1(x),
f^{\dagger}_2(x))$ with $f_{1,2}(x)$ two independent fermion fields. Here, $v_F$ is the Fermi velocity of the fermions and $m(x)$ is a spatially-varying mass for the fermion fields. To uncover the form of the real space wave functions of the quasiparticle excitations, we first write the second-quantized quasiparticle operator, $$q^{\dagger}=\int
dx~ [\phi_1(x)f^{\dagger}_1(x)+\phi_2(x) f^{\dagger}_2(x)],$$ which is assumed to satisfy the equation, $$[H,q^{\dagger}]=\epsilon q^{\dagger}.$$ This implies the following real-space Dirac equation for the two component wave function $\phi^{\rm{T}}(x)=(\phi_1(x),\phi_2(x))$: $$-iv_F\sigma_z\partial_x\phi(x)+\sigma_x m(x)\phi(x)=\epsilon
\phi(x). \label{wa}$$ As a first step towards solving this equation, we note that, because $\sigma_y$ anticommutes with $\sigma_x$ and $\sigma_z$, if $\phi(x)$ is an eigenfunction of Eq. (\[wa\]) with eigenvalue $\epsilon$, $\sigma_y\phi(x)$ is also an eigenfunction of Eq. (\[wa\]) with eigenvalue $-\epsilon$. As a result, if there is a non-degenerate solution of Eq. (\[wa\]) with the eigenvalue $\epsilon=0$, it can be made a simultaneous eigenstate of $\sigma_y$ since $\sigma_y\phi_(x) \propto \phi_(x)$. Let $\phi_0(x)$ denote such a zero-energy solution and $\sigma_y\phi_0(x)=\lambda\phi_0(x)$. Now, setting $\epsilon=0$ and left-multiplying Eq. (\[wa\]) by $i\sigma_z$ we obtain, $$\partial_x\phi_0(x)={\lambda\over v_F}
m(x)\phi_0(x),$$ which implies $$\phi_0(x)=e^{{\lambda\over v_F}\int_0^x
m(y)dy}\phi_0(0).\label{zero}$$ It can be clearly seen that for $m(x)=\pm {\rm{sign}}(x)|m(x)|$, Eq. (\[zero\]) is normalizable for $\lambda=\mp 1$. Therefore, for each sign change of the mass term $m(x)$ in the 1D system, there is a single normalizable zero energy solution. Such a solution indicates the existence of a zero-energy excitation with a quasiparticle operator (for $\lambda = 1$), $$q^{\dagger}=C\int
dx~ e^{{1\over v_F}\int_0^x
m(y)dy}[f^{\dagger}_1(x)+i f^{\dagger}_2(x)],$$ where $C$ is a normalization constant. Since $f_1(x)$ and $f_2(x)$ are two independent fermion fields, $q^{\dagger}$ above denotes an ordinary fermion operator following fermion anticommutation relations.
Below we map the 2D Hamiltonian $H_T=H_0+H_V$ of the semiconducting thin film with proximity-induced superconductivity on an effective 1D problem which resembles Eq. (\[h2\]), with important differences which render the quasiparticle mode corresponding to the zero-energy solution a Majorana fermion mode.
Map of the Hamiltonian of the 2D semiconductor on an effective 1D theory
========================================================================
In order to map the Hamiltonian of the semiconductor, $H_T=H_0 + H_V$, with proximity-induced superconductivity and a vortex at the center, on an effective one-dimensional problem, we will use the rotational symmetry of $H_T$ to decouple it in the various angular momentum channels [@Sumanta]. We note that, because of the existence of the spin-orbit coupling term $H_{SO}$ in $H_0$, the Hamiltonian $H_T$ is invariant only under the *simultaneous* rotation of the system in the real and spin spaces. Therefore, for angular momentum decoupling of $H_T$, we change to a new representation of the fermion operators where they are expanded in the *total* angular momentum channels (as opposed to the orbital angular momentum channels as employed in the case of a chiral *p*-wave superconductor [@Sumanta]) indexed by the half-odd-integers $m_J$, $$\begin{aligned}
&c_{\mathbf{k}\uparrow}&=\frac{1}{\sqrt{2\pi
k}}\sum_{m_J=-\infty}^{\infty}c_{m_J,k,\uparrow}e^{i(m_J-\frac{1}{2})\theta_{\mathbf{k}}}\nonumber\\
&c_{\mathbf{k}\downarrow}&=\frac{1}{\sqrt{2\pi k}}\sum_{m_J=-\infty}^{\infty}c_{m_J,k,\downarrow}e^{i(m_J+\frac{1}{2})\theta_{\mathbf{k}}}
\label{channel},\end{aligned}$$ where $(m_J \mp 1/2)$ in the exponents on the right indicate the orbital angular momentum quantum numbers for the up and the down spins, respectively. Here, $k=|\mathbf{k}|$ is the magnitude of the vector $\mathbf{k}$, and thus is a one-dimensional variable. The commutation relation $\{c_{\mathbf{k},\beta},c^{\dagger}_{\mathbf{p},\gamma}\}=\delta^2(\mathbf{k}-\mathbf{p})\delta_{\beta,\gamma}$ implies $$\{c_{m_J,k,\beta},c^{\dagger}_{n_J,p,\gamma}\}=\delta_{m_J,n_J}\delta(k-p)\delta_{\beta,\gamma}.$$ Inserting these representations in Eq. (\[H\_0\]), we find for $H_K$, $$H_K=\frac{1}{(2\pi)^2}\sum_{m_J}\sum_{\beta}\int dk (\frac{k^2}{2m^*}-\mu)c^{\dagger}_{m_J,k,\beta}c_{m_J,k,\beta}.
\label{HK}$$ The Zeeman term, $H_Z$, is also diagonal in the total angular momentum as well as the spin quantum numbers, $$H_Z=\frac{V_z}{(2\pi)^2}\sum_{m_J}\int dk(c^{\dagger}_{m_J,k,\uparrow}c_{m_J,k,\uparrow}-c^{\dagger}_{m_J,k,\downarrow}c_{m_J,k,\downarrow}).
\label{HZ}$$ The spin-orbit term, $H_{SO}$, is diagonal in the total angular momentum quantum number, but not so in the spin quantum number: $$H_{SO}=\frac{i\alpha}{(2\pi)^2}\sum_{m_J}\int kdk c^{\dagger}_{m_J,k,\uparrow}c_{m_J,k,\downarrow} + {\rm{H.c.}}
\label{HSO}$$
Let us now focus on the vortex term $H_V$ in Eq. (\[Pairingvortex\]). Substituting $$c^{\dagger}_{\bf{R}\pm\bf{r},\beta}=2\pi\sum_{\mathbf{k}}c^{\dagger}_{\mathbf{k},\beta}e^{i\mathbf{k}.(\bf{R}\pm\bf{r})}$$ in Eq. [\[Pairingvortex\]]{}, we end up with two spatial integrals, $g({\mathbf{k}}-{\mathbf{p}})=\int d^2r
g({\bf{r}})e^{i({\mathbf{k}}-{\mathbf{p}}).{\bf{r}}}=\Delta_0$ and $$I({\mathbf{k}}+{\mathbf{p}})=\int d^2R
e^{i\theta_{\bf{R}}}h(R)e^{i({\mathbf{k}}+{\mathbf{p}}).{\bf{R}}}.\label{I}$$ In order to evaluate $I({\mathbf{k}}+{\mathbf{p}})$, we first note that, if in Eq. (\[I\]) the vector $(\mathbf{k}+\mathbf{p})$ is rotated by an angle $\theta$ in the momentum space, the scalar product in the exponent $\exp[i({\mathbf{k}}+{\mathbf{p}}).{\bf{R}}]$ can be made to remain invariant if simultaneously $\theta_{\bf{R}}$ is rotated by $\theta$. Therefore, $$I(R_{\theta}({\mathbf{k}}+{\mathbf{p}}))=e^{i\theta}I({\mathbf{k}}+{\mathbf{p}}),$$ where $R_{\theta}$ is the operator that rotates $({\mathbf{k}}+{\mathbf{p}})$ by an angle $\theta$ in the momentum space. It follows that, $I({\mathbf{k}}+{\mathbf{p}})=e^{i\theta_{{\mathbf{k}}+{\mathbf{p}}}}I(|{\mathbf{k}}+{\mathbf{p}}|)$. To evaluate $I(|{\mathbf{k}}+{\mathbf{p}}|)$ we choose $({\mathbf{k}}+{\mathbf{p}})$ along the $y$-axis. Performing the $\theta_{{\bf{R}}}$ integral which produces $-2\pi i
J_{-1}(|{\mathbf{k}}+{\mathbf{p}}|R)$, where $J_{-1}$ is the Bessel function of the first kind of order $-1$ [@Table], and then performing the $R$ integral which produces $\frac{(2\pi)^3i}{|{\mathbf{k}}+{\mathbf{p}}|^2}\times\mathcal{O}(1)$, we find, $$H_{V}=-(2\pi)^3 i\Delta_0\sum_{\mathbf{k},\mathbf{p}
}\frac{ke^{i\theta_{\mathbf{k}}}+pe^{i\theta_{\mathbf{p}}}}{ |{\mathbf{k}}+{\mathbf{p}}|^3 }
c^{\dagger}_{{\mathbf{k}},\uparrow}c^{\dagger}_{{\mathbf{p}},\downarrow}+{\rm{H.c.}}$$ Finally, using angular momentum expansion of the fermion operators and noting that a function of $|{\mathbf{k}}+{\mathbf{p}}|$ is periodic in $(\theta_{{\mathbf{k}}}-\theta_{{\mathbf{p}}})$ and hence can be Fourier expanded as $$\frac{1}{|{\mathbf{k}}+{\mathbf{p}}|^3}=\sum_{m}u_m(k,p)e^{im(\theta_{{\mathbf{k}}}-\theta_{{\mathbf{p}}})}$$ with $m$ an integer, we find, after the $\theta_{\mathbf{k}}$ and $\theta_{\mathbf{p}}$ integrals,
$$H_{V}=-i\Delta_0\sum_m\int dkdp \sqrt{kp}u_m(k,p)\Big(k c^{\dagger}_{m+\frac{3}{2},k,\uparrow}c^{\dagger}_{-m-\frac{1}{2},p,\downarrow}
+ p c^{\dagger}_{m+\frac{1}{2},k,\uparrow}c^{\dagger}_{-m+\frac{1}{2},p,\downarrow}\Big).
\label{H_V}$$
From Eqs. (\[HK\],\[HZ\],\[HSO\],\[H\_V\]), it is clear that the $m_J=1/2$ total angular momentum channel separates from the rest (in $H_V$ it is obtained by taking $m=-1$ in the first term and $m=0$ in the second term on the right hand side of Eq. (\[H\_V\])). It is straightforward to check that no other total angular momentum channel is isolated from the rest. One can also check that for a vortex with an even number of flux quanta, for which $e^{i\theta_{\bf{R}}}$ is replaced by $e^{2i\theta_{\bf{R}}}$ in Eq. (\[Pairingvortex\]), no total angular momentum channel can be isolated from the rest of the Hamiltonian. The fact that a single total angular channel can be isolated from the rest is crucial for the following arguments demonstrating the existence of a non-degenerate zero energy solution. The zero energy eigen-solution of the Hamiltonian exists in this channel, and in cases where such an isolated channel does not exist, no non-degenerate zero energy solution exists at the vortex core.
In the $m_J=1/2$ channel, the total Hamiltonian can be written as $H_{T,m_J=\frac{1}{2}}=H_{K,m_J=\frac{1}{2}}+H_{Z,m_J=\frac{1}{2}}+H_{SO,m_J=\frac{1}{2}}+H_{V,m_J=\frac{1}{2}},$ where, $$\begin{aligned}
&&H_{K,m_J=\frac{1}{2}}=\frac{1}{(2\pi)^2}\sum_{\beta}\int dk (\frac{k^2}{2m^*}-\mu)c^{\dagger}_{\frac{1}{2},k,\beta}c_{\frac{1}{2},k,\beta}\nonumber\\
&&H_{SO,m_J=\frac{1}{2}}=\frac{i\alpha}{(2\pi)^2}\int kdk c^{\dagger}_{\frac{1}{2},k,\uparrow}c_{\frac{1}{2},k,\downarrow} + {\rm{H.c.}}\nonumber\\
&&H_{Z,m_J=\frac{1}{2}}=\frac{V_z}{(2\pi)^2}\int dk(c^{\dagger}_{\frac{1}{2},k,\uparrow}c_{\frac{1}{2},k,\uparrow}-c^{\dagger}_{\frac{1}{2},k,\downarrow}c_{\frac{1}{2},k,\downarrow})\nonumber\\
&&H_{V,m_J=\frac{1}{2}}=-i\int dkdp \Delta(k,p)c^{\dagger}_{\frac{1}{2},k,\uparrow}c^{\dagger}_{\frac{1}{2},p,\downarrow} + {\rm{H.c.}},\nonumber\\
\label{H_HALF}\end{aligned}$$ where, in the last line, $\Delta(k,p)=\Delta_0\sqrt{kp}(ku_{m=-1}(k,p)+pu_0(k,p))$. From here onwards we will ignore the subscript $1/2$ in the fermion operators $c, c^{\dagger}$, keeping in mind that all the fermion operators carry the same total angular momentum quantum number $m_J=1/2$.
Demonstration of the zero energy solution
=========================================
To turn the Hamiltonian in Eq. (\[H\_HALF\]) into a form resembling that in Eq. (\[h2\]), we have to first diagonalize the free-electron part, $(H_{K,m_J=\frac{1}{2}}+H_{Z,m_J=\frac{1}{2}}+H_{SO,m_J=\frac{1}{2}})$, using a unitary transformation. Only then it is possible to linearize the band energies around the Fermi surfaces and the combined one-electron part in Eq. (\[H\_HALF\]) may take the Dirac form. To do this, we define the unitary transformation from the $c_{\uparrow/\downarrow,k}^\dagger$ basis to the energy eigenbasis, $$\begin{aligned}
\left(\begin{array}{c}c_{\uparrow,k}^\dagger\\ c_{\downarrow,k}^\dagger\end{array}\right)&=\left(\begin{array}{cc}a_{k}^*&-b_{k}\\b_{k}^*& a_{k}\end{array}\right)\left(\begin{array}{c}f_{+,k}^\dagger\\ f_{-,k}^\dagger\end{array}\right)\\
\left(\begin{array}{c}f_{+,k}^\dagger\\ f_{-,k}^\dagger\end{array}\right)&=\left(\begin{array}{cc}a_{k}&b_{k}\\-b_{k}^*& a_{k}^*\end{array}\right)\left(\begin{array}{c}c_{\uparrow,k}^\dagger\\ c_{\downarrow,k}^\dagger\end{array}\right)\end{aligned}$$ where $|a_k|^2+|b_k|^2=1$. Since the $f_{\pm,k}^\dagger$ operators are related to the original fermion operators by a unitary transformation, they still obey the fermionic anticommutation relations.Calling the corresponding eigenvalues (band energies) $E_{\pm,k}$, the Hamiltonian in this basis in the channel $m_J=\frac{1}{2}$ is written as, $$\begin{aligned}
H_{T}&=&\int \frac{d k}{(2\pi)^2}[(E_{+,k}-\mu)f_{+,k}^\dagger f_{+,k}+(E_{-,k}-\mu)f_{-,k}^\dagger f_{-,k}]\nonumber\\
&-&\imath \int dk dp \Delta(k,p)[a_{k}^*b_{p}^*f^\dagger_{+,k}f^\dagger_{+,p}-b_{k}a_{p}f^\dagger_{-,k}f^\dagger_{-,p}]\nonumber\\&-&\imath \int dk dp \Delta(k,p)[a_{k}^*a_{p}f^\dagger_{+,k}f^\dagger_{-,p}-b_{k}b_{p}^*f^\dagger_{-,k}f^\dagger_{+,p}] + {\rm{H.c.}}\nonumber\\
&=&\int \frac{d k}{(2\pi)^2}[(E_{+,k}-\mu)f_{+,k}^\dagger f_{+,k}+(E_{-,k}-\mu)f_{-,k}^\dagger f_{-,k}]\nonumber\\
&-&\frac{\imath}{2} \int dk dp [\{\Delta(k,p)a_{k}^*b_{p}^*-\Delta(p,k)a_{p}^*b_{k}^*\}f^\dagger_{+,k}f^\dagger_{+,p}]\nonumber\\&-&\frac{\imath}{2} \int dk dp[\{-\Delta(k,p)b_{k}a_{p}+\Delta(p,k)b_{p}a_{k}\}f^\dagger_{-,k}f^\dagger_{-,p}]\nonumber\\
&-&\imath \int dk dp [\{\Delta(k,p)a_{k}^*a_{p}+\Delta(p,k)b_{p}b_{k}^*\}f^\dagger_{+,k}f^\dagger_{-,p}]
+{\rm{H.c.}}
\label{H_T}\end{aligned}$$ Here, in the second expression, we have written the coefficients of the pairing terms in a form so as to highlight the fact that the coefficients of the intra-band pairing terms must be antisymmetric under the interchange of $k$ and $p$, while the coefficients of the inter-band pairing terms have no such symmetry. The antisymmetry of the coefficients of the intra-band pairing terms, which will be crucial for the demonstration of the zero-energy solution below, follow from the fermion anticommutation relations coupled with the fact that the fermion operators involved are from the same band with the same $m_J$ quantum numbers. Using simplified notations for the coefficients of the pairing terms we rewrite Eq. (\[H\_T\]) as, $$\begin{aligned}
H_T&=&\int \frac{d k}{(2\pi)^2}[(E_{+,k}-\mu)f_{+,k}^\dagger f_{+,k}+(E_{-,k}-\mu)f_{-,k}^\dagger f_{-,k}]\nonumber\\
&-&\frac{\imath}{2} \int dk dp [\{\Lambda_{++}(k,p)f^\dagger_{+,k}f^\dagger_{+,p}+\Lambda_{--}(k,p)f^\dagger_{-,k}f^\dagger_{-,p}\nonumber\\
&+&\Lambda_{+-}(k,p)f^\dagger_{+,k}f^\dagger_{-,p}\}] +\rm{H.c.}.\end{aligned}$$
Let us now assume that the Fermi level is in the lower band $f_{-}$, which has a Fermi momentum $k_F$, and that the upper band $f_{+}$ has a set of unoccupied states at energy $E_{+,k}-\mu$ separated from the Fermi level by an energy gap $E_+=|V_z|-\mu$ (without loss of generality we take $\mu$ to be positive). This condition is satisfied only as long as $|\mu|<|V_z|$. In what follows we will implicitly assume that this condition is experimentally satisfied. By limiting the momenta in the lower band to those close to the Fermi momentum and in the upper band to those close to the band minimum at $k=0$, the Hamiltonian can be written as,
$$\begin{aligned}
H_T&=&\int \frac{d q}{(2\pi)^2}[E_{+}f_{+,q}^\dagger f_{+,q}+(E_{-,k_F+q}-\mu)f_{-,k_F+q}^\dagger f_{-,k_F+q}]\nonumber\\
&-&\frac{\imath}{2} \int dq dq^{\prime} [\{\Lambda_{++}(q,q^{\prime})f^\dagger_{+,q}f^\dagger_{+,q^{\prime}}+\Lambda_{--}(k_F+q,k_F+q^{\prime})f^\dagger_{-,k_F+q}f^\dagger_{-,k_F+q^{\prime}}
\nonumber\\
&+&\Lambda_{+-}(q,k_F+q^{\prime})f^\dagger_{+,q}f^\dagger_{-,k_F+q^{\prime}}\}] +\rm{H.c}.\end{aligned}$$
Here, the lower limits on the momenta $q, q^{\prime}$ in the integrals are $0$ or $-k_F$ when the corresponding momentum corresponds to the $+$ or the $-$ bands, respectively (the upper limits on the integrals are taken to be $\infty$). We keep the lowest order dependencies of the functions $\Lambda$ on $q$ and $q^{\prime}$ which are consistent with the symmetries of these functions from the fermion anticommutation relations, namely that $\Lambda_{++}$ and $\Lambda_{--}$ are antisymmetric in $(q-q^{\prime})$. Since $\Lambda_{+-}(q, q^{\prime})$ has no such symmetry requirement, it is generically non-zero for $q, q^{\prime} \rightarrow 0$, and we assume that it is a non-zero constant $\Lambda_{+-}$ in this limit. This yields,
$$\begin{aligned}
H_T&=&\int \frac{d q}{(2\pi)^2}[E_{+}f_{+,q}^\dagger f_{+,q}+(E_{-,k_F+q}-\mu)f_{-,k_F+q}^\dagger f_{-,k_F+q}]\nonumber\\
&-&\frac{\imath}{2} \int dq dq^{\prime} [\{\Lambda_{++}(q-q^{\prime})f^\dagger_{+,q}f^\dagger_{+,q^{\prime}}+\Lambda_{--}(q-q^{\prime})f^\dagger_{-,k_F+q}f^\dagger_{-,k_F+q^{\prime}}\nonumber\\
&+&\Lambda_{+-}f^\dagger_{+,q}f^\dagger_{-,k_F+q^{\prime}}\}] +\rm{H.c}.\end{aligned}$$
Let us now define the Fourier transforms to real space, $$\begin{aligned}
f^\dagger_-(x)&=&\int_{-k_F}^{\infty} dq e^{\imath q x}f^\dagger_{-,k_F+q}\nonumber\\
f^\dagger_+(x)&=&\int_{0}^{\infty} dq e^{\imath q x}f^\dagger_{+,q}.
\label{FT}\end{aligned}$$ Using Eq. (\[FT\]), we fourier transform $H_T$ to real space,
$$\begin{aligned}
H_T&=&\int_{-\infty}^\infty dx dx' [E_{+}\delta(x-x')f_{+}^\dagger(x) f_{+}(x')+E_{-}(x-x')f_{-}^\dagger(x) f_{-}(x')]\nonumber\\
&-&\frac{\imath}{2} \int dx [\{\Lambda_{++}(x)f^\dagger_{+}(x)f^\dagger_{+}(-x)+\Lambda_{--}(x)f^\dagger_{-}(x)f^\dagger_{-}(-x)\nonumber\\
&+&\Lambda_{+-}\delta(x)f^\dagger_{+}(x)f^\dagger_{-}(-x)\}] +\rm{H.c.},
\label{H_Treal}\end{aligned}$$
where the functions $\Lambda_{++}(x)$ and $\Lambda_{--}(x)$ are odd in $x$, and $E_{-}(x-x')=\imath v \partial_x \delta(x-x')$ with $v$ the fermi velocity in the lower band. The Hamiltonian in Eq. (\[H\_Treal\]) is still not in the same form as in Eq. (\[h2\]), so the Jackiw-Rebbi construction does not yet apply. However, as we show below, the BdG equations for a putative zero-energy excitation in the vortex core are of the same form as in Eq. (\[wa\]) (with $\epsilon=0$), guaranteeing the existence of a zero energy solution.
Let us define the quasiparticle operator for the putative zero energy mode as,
$$\gamma^\dagger=\int_{-\infty}^\infty dx (\eta_{+,1}(x)f^\dagger_+(x)+\eta_{-,1}(x)f^\dagger_-(x)+\eta_{+,2}(x)f_+(-x)+\eta_{-,2}(x)f_-(-x)).
\label{QP}$$
The corresponding BdG differential equations can be derived by setting $$[H_T, \gamma^{\dagger}]=0.
\label{BdG1}$$ To derive the commutators involved in Eq. (\[BdG1\]), we need the following anticommutation relations, $$\begin{aligned}
\{f^\dagger_+(x),f_+(x')\}&=&\int_{0}^\infty dk dk' e^{\imath (kx-k'x')}\{f^\dagger_{+,k},f_{+,k'}\}\nonumber\\
&=&\int_{0}^\infty dk e^{\imath k(x-x')}\equiv S_+(x-x')\nonumber\\
&=&\delta(x-x')/2+\imath P(\frac{1}{x-x'})\\
\{f^\dagger_-(x),f_-(x')\}&=&\int_{-k_F}^\infty dk dk' e^{\imath (kx-k'x')}\{f^\dagger_{-,k},f_{-,k'}\}\nonumber\\
&=&\int_{-k_F}^\infty dk e^{\imath k(x-x')}\equiv S_-(x-x'),\end{aligned}$$ where the functions $S_+$ and $S_{-}$ are not exactly $\delta$ functions. The existence of $S_{+}$ and $S_{-}$ necessitates the introduction of a new set of four functions, $$\xi_{\nu,n}(x)=\int_{-\infty}^\infty S_\nu(x_1-x)\eta_{\nu,n}(x_1) d x_1,
\label{new-function}$$ where $\nu = \pm$ and $n$ takes the values $1, 2$.
In terms of the $\xi$ functions, the BdG equations take the simple form, $$\begin{aligned}
\imath v \partial_x \xi_{-,1}(x)-\imath\Lambda_{--}(x)\xi_{-,2}(x)-\imath\Lambda_{+-}\delta(x)\xi_{+,2}(x)/2&=&0\nonumber\\
E_{+}\xi_{+,1}(x)-\imath \Lambda_{++}(x)\xi_{+,2}(x)+\imath \Lambda_{+-}\delta(x)\xi_{-,2}(x)/2&=&0\nonumber\\
-\imath v \partial_x \xi_{-,2}(x)+\imath\Lambda_{--}(x)\xi_{-,1}(x)-\imath\Lambda_{+-}\delta(x)\xi_{+,1}(x)/2&=&0\nonumber\\
-E_{+}\xi_{+,2}(x)+\imath \Lambda_{++}(x)\xi_{+,1}(x)+\imath \Lambda_{+-}\delta(x)\xi_{-,1}(x)/2&=&0
\label{BdG2}\end{aligned}$$ Eliminating the amplitudes in the upper band, $\xi_{+,1}(x)$, $\xi_{+,2}(x)$, using
$$\begin{aligned}
&\left(\begin{array}{c}\xi_{+,1}(x)\\\xi_{+,2}(x)\end{array}\right)=-\frac{\imath \Lambda_{+-}\delta(x)}{2(E_{+}^2+\Lambda_{++}^2(x))}\left(\begin{array}{cc}E_{+}&-\imath \Lambda_{++}(x)\\\imath\Lambda_{++}(x)&-E_{+}\end{array}\right)\left(\begin{array}{c}\xi_{-,2}(x)\\\xi_{-,1}(x)\end{array}\right)
\label{eliminate}\end{aligned}$$
we get the equations for the amplitudes in the lower band,
$$\begin{aligned}
\imath v \partial_x \xi_{-,1}(x)&+&\frac{1}{4}
\frac{\Lambda_{+-}^2\delta^2(x) E_{+}}{E_{+}^2+\Lambda_{++}^2(x)}\xi_{-,1}(x)-\imath\left(\Lambda_{--}(x)+\frac{1}{4}
\frac{\Lambda_{+-}^2\delta^2(x)\Lambda_{++}(x)}{E_{+}^2+\Lambda_{++}^2(x)}\right)\xi_{-,2}(x)=0\nonumber\\
-\imath v \partial_x \xi_{-,2}(x)&-&\frac{1}{4}\frac{\Lambda_{+-}^2\delta^2(x) E_{+}}{E_{+}^2+\Lambda_{++}^2(x)}\xi_{-,2}(x)+\imath\left(\Lambda_{--}(x)+\frac{1}{4}
\frac{\Lambda_{+-}^2\delta^2(x)\Lambda_{++}(x)}{E_{+}^2+\Lambda_{++}^2(x)}\right)\xi_{-,1}(x)=0\nonumber\\\end{aligned}$$
It can now be easily seen that by a redefinition, $$\psi_{n}(x)=e^{-\imath\frac{1}{4 v}\int_0^x dx' \frac{\Lambda_{+-}^2\delta^2(x') E_{+}}{E_{+}^2+\Lambda_{++}^2(x')}}\xi_{-,n}(x),
\label{Psi}$$ the above BdG equations can be recast as, $$\begin{aligned}
& v \partial_x \psi_{1}(x)+\left(\Lambda_{--}(x)+\frac{1}{4}\frac{\Lambda_{+-}^2\delta^2(x)\Lambda_{++}(x)}{E_{+}^2+\Lambda_{++}^2(x)}\right)\psi_{2}(x)=0\nonumber\\
& v \partial_x \psi_{2}(x)+\left(\Lambda_{--}(x)+\frac{1}{4}\frac{\Lambda_{+-}^2\delta^2(x)\Lambda_{++}(x)}{E_{+}^2+\Lambda_{++}^2(x)}\right)\psi_{1}(x)=0
\label{BdG3}\end{aligned}$$ By defining the quantity in the parentheses in Eq. (\[BdG3\]) as $m(x)$, it can be easily checked that $m(x)$ is an odd function, $m(-x)=-m(x)$, since both $\Lambda_{++}(x)$ and $\Lambda_{--}(x)$ are odd under $x \rightarrow -x$. Eq. (\[BdG3\]) can now be written in terms a two-component spinor wave function $\zeta^{\rm{T}}(x)=(\psi_1(x),\psi_2(x))$, $$v\partial_x\left(\begin{array}{c}\psi_1(x)\\\psi_2(x)\end{array}\right)=-m(x)\sigma_x\left(\begin{array}{c}\psi_1(x)\\\psi_2(x)\end{array}\right).
\label{BdG4}$$ which, after multiplication by $\sigma_z$, is of a form similar to Eq. (\[wa\]). Therefore, a unique zero energy eigen-solution of $H_T$ in the $m_J=\frac{1}{2}$ channel is guaranteed.
Notice that, in the above construction for the zero-energy solution of the BdG equations, no special form for the real-space profile of the superconducting gap near the vortex core has been assumed. In particular, we have not assumed the frequently-used step function profile for the gap [@tewari_prl'2007; @Sau] near the core, without which the real space solution of the second-order, coupled, BdG equations for the present system is still an open problem. All we have needed here to prove the existence of the zero-energy solution, irrespective of the real space profile of the gap, is the antisymmetry of the functions $\Lambda_{++}(x)$ and $\Lambda_{--}(x)$ under $x \rightarrow -x$, which is a consequence of the fermion anticommutation relations. Note that, to use the fermion anticommutation relations to determine the symmetry of these functions, it is crucial that a single total angular momentum channel ($m_J=1/2$) could be isolated from the rest. In the cases where this is not possible, *e.g.*, a vortex with an even number of flux quanta, the Jackiw-Rebbi construction does not apply and a non-degenerate zero-energy solution is not expected at the vortex core.
It is worth pointing out that the above construction for the zero-energy solution can be straightforwardly applied to describe the zero modes on the surface of a 3D strong TI, described by a Dirac-like Hamiltonian, in the presence of an $s$-wave superconducting vortex [@fu_prl'08; @Rossi], but only for a non-zero chemical potential. For zero chemical potential, our reduction of the set of $4$ coupled BdG differential equations to a set of $2$ (with the help of Eq. (\[eliminate\])) to bring them to the form in Eq. (\[wa\]) in terms of $2 \times 2$ Pauli spin matrices no longer applies. Therefore, our methods above do not directly apply to the problem of zero modes in a TI in the presence of an $s$-wave superconductor at $\mu=0$. However, for this special case, there exists *exact* real space solution of the zero-mode eigenfunction [@fu_prl'08; @Rossi].
Demonstration of the Majorana condition
=======================================
Now we show that the quasiparticle operator for the zero energy solution found above satisfies the Majorana condition $\gamma^{\dagger}=\gamma$. From Eq. (\[QP\]) it is clear that such a condition is satisfied provided $$\begin{aligned}
&&\eta_{-,1}^{\ast}(-x)=\eta_{-,2}(x)\nonumber\\
&&\eta_{+,1}^{\ast}(-x)=\eta_{+,2}(x)
\label{condition}\end{aligned}$$ Using Eq. (\[new-function\]) and the property for the $S_{\pm}$ functions, $$S_{\pm}^{\ast}(x_1-x)=S_{\pm}(x-x_1)$$ we see that the condition in Eq. (\[condition\]) translates into $$\begin{aligned}
&&\xi_{-,1}^{\ast}(-x)=\xi_{-,2}(x)\nonumber\\
&&\xi_{+,1}^{\ast}(-x)=\xi_{+,2}(x)
\label{condition-xi}\end{aligned}$$
Now, from Eq. (\[BdG4\]) it is clear that $\zeta(x)$ can be taken as an eigen-spinor of $\sigma_x$: $$\sigma_x \zeta(x)=\lambda \zeta(x)$$ Then, solving for Eq. (\[BdG4\]), we find that, $$\begin{aligned}
&&\psi_1(x)=\exp(-\frac{\lambda}{v}\int_0^x m(y)dy)\nonumber\\
&&\psi_2(x)={\rm{sign}}(\lambda)\exp(-\frac{\lambda}{v}\int_0^x m(y)dy),\end{aligned}$$ where, for $m(x)=\pm {\rm{sign}}(x)|m(x)|$, $\lambda=\pm 1$, in which case the zero mode solution is normalizable. Here we take $\lambda=1$ for the purpose of illustration. Using Eq. (\[Psi\]), we get the solutions for the functions $\xi_{-,1}, \xi_{-,2}$,
$$\xi_{-,1}(x)=\exp({\imath\frac{1}{4 v}\int_0^x dx' \frac{\Lambda_{+-}^2\delta^2(x') E_{+}}{E_{+}^2+\Lambda_{++}^2(x')}})\exp(-\frac{\lambda}{v}\int_0^x m(y)dy)=\xi_{-,2}(x)
\label{solution}$$
Now, with the help of the phase factor in Eq. (\[solution\]) being even in $x$ and the mass $m(x)$ being odd in $x$, it can be easily checked that the first line in Eq. (\[condition-xi\]) is satisfied. Subsequently, with the help of Eq. (\[eliminate\]), and recalling that $\Lambda_{++}(x)$ is odd in $x$, the second line of Eq. (\[condition-xi\]) is also satisfied. Therefore, the quasiparticle operator corresponding to the zero mode solution in Eq. (\[solution\]) is a Majorana fermion operator.
Topological protection and topological quantum phase transition
===============================================================
The proof of the index theorem presented above relies on the rotational invariance of the non-interacting part of the Hamiltonian as well as of the vortex profile, together with the assumption that a single band (lower band, $f_{-}$) dominates the wave-function of the Majorana bound state. The latter assumption is valid in the limit $E_+=|V_Z|-\mu\gg\Delta_0$. However, as shown below, the question of the existence of the non-degenerate Majorana mode in this system is robust to deviations from such assumptions, and the Majorana state can only be removed by a bulk phase transition in the system. Thus, a non-degenerate Majorana bound state is topologically protected.
In order to see the topological protection of the Majorana mode, let us consider an isotropic BCS Hamiltonian $H_1$ with a symmetric vortex and a small value of $\Delta_0$ (basically, $H_1$ is the same as $H_T$), for which the derivation presented above can be used to show the existence of a non-degenerate Majorana mode bound to a vortex with an odd number of flux quantum. In addition, let us consider a second BCS Hamiltonian $H_2$ which is possibly anisotropic (*i.e*., breaks rotational symmetry) near the vortex and has a strong pairing potential. Therefore, $H_2$ does not lend itself to a simple analytic treatment as given above. Given these two Hamiltonians $H_1$ and $H_2$, we can construct a family of BCS Hamiltonians parameterized by $0<\lambda<1$ given by $H(\lambda)=\lambda H_1+(1-\lambda)H_2$. Since both $H_1$ and $H_2$ approach the bulk Hamiltonian away from the vortex core, so does $H(\lambda)$. Therefore, as long as the superconductor has a non-zero bulk gap, $E_g$, away from the vortex core, we expect all the eigenstates of $H(\lambda)$ with energy less than $E_g$ to be localized at the vortex core. Furthermore, since only a finite number of states can be localized near the vortex core, there should be a discrete set of eigenstates with energies satisfying $|E_{(s(\lambda)+n)}(\lambda)|<E_g$ bound to the vortex core. Here, in analogy with a $p$-wave superconductor/superfluid [@Kopnin], $(s(\lambda)+n)$ is an index for the states with $E_{(s(\lambda)+n)}\propto s(\lambda)+n $, where $n$ is an integer and $s(\lambda)=0,1/2$. Note that, if for some value of $\lambda$, $s(\lambda)=0$, a single zero energy state is allowed. Conversely, if $s(\lambda)=1/2$, no zero mode is allowed. In general, the shift $s(\lambda)$ can be taken such that it represents an even number (including $0$) zero modes by $s(\lambda)=1/2$ and an odd number of zero modes by $s(\lambda)=0$. The quasiparticle excitations of $H(\lambda)$ are given by the equation, $$[H(\lambda),d_{(s(\lambda)+n)}^\dagger(\lambda)]=E_{(s(\lambda)+n)}(\lambda)d^\dagger_{(s(\lambda)+n)}(\lambda).$$ where $d^{\dagger}, d$ are taken as the quasiparticle creation and annihilation operators. By considering the Hermitian conjugate of the above equation, it can be seen that the solutions of $H(\lambda)$ are paired such that $d_{-(s(\lambda)+n)}^\dagger(\lambda)=d_{(s(\lambda)+n)}(\lambda)$ and $E_{-(s(\lambda)+n)}(\lambda)=-E_{(s(\lambda)+n)}(\lambda)$. In the case where $H(\lambda)$ has a single zero energy solution, $(s(\lambda)=0)$, it is clear that the corresponding zero mode $d_0$ is a Majorana mode $(d_0^\dagger=d_0)$. Since $H_1$ in our problem is assumed to have a non-degenerate Majorana mode (as we have shown by analyzing $H_T$), $s(\lambda=1)=0$. The stability of this non-degenerate Majorana zero mode to small perturbations in the BdG Hamiltonian follows from the fact that as long as the low energy states in the vicinity of the vortex core remain discrete in energy spacing, the eigenstates labeled by the index $s(\lambda)+n$ evolve continuously with $\lambda$, and thus the shift $s(\lambda)$ cannot jump from it’s value $0$ to it’s other allowed value $1/2$. Thus, $s(0)=0$, which implies that $H_2$ has a non-degenerate Majorana mode as well, even though $H_2$ itself does not lend itself to a simple analysis.
From the above argument it is clear that a non-degenerate Majorana bound state at a vortex remains protected as long as the low energy states in the vicinity of the vortex core remain discrete in number for all values of the parameter $0<\lambda<1$. Thus the Majorana bound state is robust to all local deformations of the Hamiltonian in the vicinity of the vortex core since these cannot affect the energy gap away from the vortex. On the other hand, if changing the parameter $\lambda$ changes the bulk parameters of the Hamiltonian, such as $\alpha$, $\mu$ or $V_Z$, then it is possible for the energy gap to close away from the vortex core, causing the states bound to the vortex core to become delocalized. This results in a breakdown of the above argument for the topological stability of the Majorana bound states. Thus, the presence or absence of a non-degenerate Majorana bound state at a vortex core leads to a classification of the phase diagram of two-dimensional superconductors with spin-orbit coupling, such that the phases which support non-degenerate Majorana bound states are separated from the phases that do not by a quantum phase transition (QPT) where the bulk energy gap closes.
The quantum phases on the two sides of the above QPT can be determined by studying a deformation parameter $\lambda$ that controls the proximity-induced pairing potential in the semiconductor, *i.e.*, the pairing potential in $H(\lambda)$ is given by $\lambda \Delta_0$. The pairing potential increases from $0$ ($\lambda=0$) to a maximum value $\Delta_0$ ($\lambda = 1$). The proof presented above shows that for $|\mu|<|V_Z|$, for which there is a single band crossing the Fermi-level, and for $\lambda\Delta_0\ll E_+=(|V_Z|-\mu)$, there is a non-degenerate Majorana mode bound to the core of a vortex. From the bulk Hamiltonian $H(\lambda)$, it is easy to check that the energy gap in the bulk vanishes at a critical value [@Sau; @Sato], $$\Delta_{c}=\lambda_c\Delta_0=\sqrt{V_Z^2-\mu^2}.$$ Thus our argument for the topological protection of the Majorana bound state implies that a Majorana bound state exists at the vortex core in $H(\lambda)$ as long as $\lambda<\lambda_c$ or $\Delta<\Delta_{c}$. Moreover, this phase is separated from the phase with $\Delta>\Delta_c$ by a QPT at which the single particle energy gap vanishes. The spin-dependent terms in the Hamiltonian, such as the Zeeman splitting and the spin-orbit coupling, cease to play a critical role in the phase $\Delta>\Delta_c$, and it is possible to reduce both these couplings to zero without crossing another phase transition (*i.e.*, without $E_g$ going through $0$ again). Thus, the phase $\Delta>\Delta_c$ must be in the same class as a conventional $s$-wave superconductor without spin-orbit coupling or Zeeman splitting, and therefore cannot support a non-degenerate Majorana mode. Furthermore, the symmetries of the superconducting order parameters are identical in these two phases. Therefore, these two phases, distinguished by the presence or absence of a Majorana mode, are separated by a topological QPT at $\Delta=\sqrt{V_z^2-\mu^2}$ where the energy gap closes.
Conclusion
==========
We prove a theorem for the existence of Majorana zero modes in a semiconducting thin film with a sizable spin-orbit coupling, in which $s$-wave superconductivity and a Zeeman splitting are induced by proximity effect (Fig. (1)). Our momentum-space construction of the zero-mode solution in the form of the Jackiw-Rebbi index theorem [@Jackiw1], which is complementary to the approximate real-space solution of the BdG equations at a vortex core [@Sau], proves the existence of non-degenerate Majorana fermion excitations localized at the vortices in the semiconductor heterostructure shown in Fig. (1a).
In our method, no special form (in particular, the frequently-used step-function form [@tewari_prl'2007; @Sau]) for the real-space profile of the superconducting pairing potential near the vortex core is required. We use only the antisymmetry property of the intra-band pairing potentials, $\Lambda_{++}(x)$ and $\Lambda_{--}(x)$ (Eq. (\[H\_Treal\])) under $x \rightarrow -x$, which is a consequence of the fermion anticommutation relations. Therefore, the non-degenerate Majorana mode found here is robust to local spatial deformations of the order parameter. For the construction in this paper to apply, it is crucial that a single total angular momentum channel (for a vortex with a single flux quantum: $m_J=1/2$) can be isolated from the rest of the Hamiltonian. In the cases where this is not possible, *e.g.*, for a vortex with an even number of flux quanta, the theorem does not apply and a non-degenerate zero-energy solution is not expected at the vortex core. The methods of this paper can be straightforwardly applied to the zero modes on the surface of a TI with a non-zero chemical potential in the presence of a superconducting vortex [@fu_prl'08], but not to the case when the chemical potential on the TI surface is zero. However, in the latter special case of zero chemical potential, an exact solution of the BdG equations already exists [@fu_prl'08; @Rossi] for the zero-mode eigenfunction for arbitrary real-space profile of the order parameter.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by DARPA-QuEST, JQI-NSF-PFC, and LPS-NSA. We thank D. H. Lee for previous collaboration on a related work in chiral $p$-wave superconductors [@Sumanta]. ST acknowledges DOE/EPSCoR Grant \# DE-FG02-04ER-46139 and Clemson University start up funds for support.
[99]{}
J. M. Leinaas, and J. Myrheim, Nuovo Cimento **37B**, 1 (1977).
F. Wilczek, Phys. Rev. Lett. 48, 1144 (1982).
F. Wilczek, *Fractional Statistics and Anyon Superconductivity* (World Scientific, Singapore) (1990).
C. Nayak, S. H. Simon, A. Stern, M. Freedman, S. Das Sarma, Rev. Mod. Phys. **80**, 1083 (2008).
A. Kitaev, Ann. Phys. **303**, 2 (2003).
G. Moore, and N. Read, Nucl. Phys. B **360**, 362 (1991).
C. Nayak, and F. Wilczek, Nucl. Phys. B **479**, 529 (1996).
N. Read and D. Green, Phys. Rev. B **61**, 10267 (2000).
S. Das Sarma, M. Freedman, and C. Nayak, Phys. Rev. Lett. **94**, 166802 (2005).
D. A. Ivanov, Phys. Rev. Lett. **86**, 268 (2001).
A. Stern, F. von Oppen, E. Mariani, Phys. Rev. B **70**, 205338 (2004).
S. Das Sarma, C. Nayak, and S. Tewari, Phys. Rev. B **73**, 220502 (R) (2006).
S.Tewari, S. Das Sarma, C. Nayak, C. W. Zhang, and P. Zoller, Phys. Rev. Lett. **98**, 010506 (2007).
L. Fu and C. L. Kane, Phys. Rev. Lett. **100**, 096407 (2008).
L. Fu and C. L. Kane, Phys. Rev. Lett. **102**, 216403 (2009).
A. R.Akhmerov, J. Nilsson, C. W. J. Beenakker, Phys. Rev. Lett. **102**, 216404 (2009).
J. D. Sau, R. M. Lutchyn, S. Tewari, S. Das Sarma, arXiv:0907.2239.
R. Jackiw and C. Rebbi, Phys. Rev. D **13**, 3398 (1976).
R. Jackiw and J. R. Schrieffer, Nuclear Physics B **190**, 253 (1981).
S. Tewari, S. Das Sarma, D. H. Lee, Phys. Rev. Lett. **99**, 037001 (2007). I. S. Gradshteyn and I. M. Ryzhik, [*Table of Integrals, Series, and Products*]{} (Academic, New York, 1980).
R. Jackiw and P. Rossi, Nuclear Physics B **190** , 681 (1981).
N. B. Kopnin, and M. M. Salomaa, Phys. Rev. B **44**, 9667 (1991). M. Sato, Y. Takahashi, S. Fujimoto, Phys. Rev. Lett. **103**, 020401 (2009).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the Voiculescu-Brown entropy of a noncommutative toral automorphism arising from a matrix $S\in GL(d,\mathbb{Z})$ is at least half the value of the topological entropy of the corresponding classical toral automorphism. We also obtain some information concerning the positivity of local Voiculescu-Brown entropy with respect to single unitaries. In particular we show that if $S$ has no roots of unity as eigenvalues then the local Voiculescu-Brown entropy with respect to every product of canonical unitaries is positive, and also that in the presence of completely positive CNT entropy the unital version of local Voiculescu-Brown entropy with respect to every non-scalar unitary is positive.'
address:
- 'Dipartimento di Matematica, Università di Roma “La Sapienza,” P.le Aldo Moro, 2, 00185 Rome, Italy'
- 'Department of Mathematics, University of Toronto, Toronto, Ontario M5S 3G3, Canada'
author:
- David Kerr
- Hanfeng Li
date: 'March 7, 2003'
title: 'Positive Voiculescu-Brown entropy in noncommutative toral automorphisms'
---
Introduction
============
Let $\Theta = (\theta_{jk} )_{1\leq j,k \leq d}$ be a real skew-symmetric $d\times d$ matrix. The noncommutative $d$-torus $A_\Theta$ is defined as the universal $C^*$-algebra generated by unitaries $u_1 , \dots , u_d$ subject to the relations $$u_j u_k = e^{2\pi i \theta_{jk}} u_k u_j$$ for all $1\leq j,k \leq d$ (see [@Rie] for a reference). For any matrix $S= (s_{jk} )_{1\leq
j,k \leq d}$ in $GL(d,\mathbb{Z})$ there is an isomorphism $\alpha
: A_{S^t \Theta S} \to A_\Theta$ determined by $$\alpha_\Theta (u_j ) = u^{s_{1j}}_1 u^{s_{2j}}_2 \cdots u^{s_{dj}}_d$$ for each $j=1, \dots ,d$. Thus when $S^t \Theta S \equiv
\Theta\,\,\, (\text{\rm mod}\,\, M_d (\mathbb{Z}))$ we obtain an automorphism of $A_\Theta$, which we denote by $\alpha_{S,\Theta}$ and refer to as a [*noncommutative toral automorphism*]{}. Note that whenever $\alpha_{S,\Theta}$ exists so does $\alpha_{S,-\Theta}$. These noncommutative analogues of toral automorphisms were initially introduced in [@Wat] and [@Bre] for $d=2$, in which case for any given $S\in
SL(2,\mathbb{Z})$ the automorphism $\alpha_{S,\Theta}$ is defined for all $\Theta$. An indication of their significance from a noncommutative geometry perspective is the fact that, for $d=2$, if $\theta_{12}$ is an irrational number satisfying a generic Diophantine property, then every diffeomorphism of $A_\Theta$ equipped with the canonical differential structure is a composition of an inner automorphism by a smooth unitary in the connected component of the unit, a noncommutative toral automorphism, and an automorphism arising from the canonical action of $\mathbb{T}^2$ [@Ell]. In an arbitrary dimension $d$, if $\Theta$ is rational (i.e., the entries of $\Theta$ are all rational) then the $C^*$-algebra $A_\Theta$ is homogeneous (see [@OPT]) and when $\alpha_{S,\Theta}$ exists we recover the corresponding classical toral automorphism at the level of the pure state space upon restricting $\alpha_{S,\Theta}$ to the center of $A_\Theta$, so that from the noncommutative point of view it is the case of nonrational $\Theta$ that is of primary interest. Unlike classical toral automorphisms, which in the hyperbolic case have served as prototypes for such important phenomena as hyperbolicity and structural stability (see [@DGS; @HK]), noncommutative toral automorphisms as a class have remained somewhat mysterious, although in many cases much has been ascertained from a measure-theoretic viewpoint [@BNS; @Nar; @NT; @Nar2; @Nes].
In [@Voi Sect. 5] it was shown that the Voiculescu-Brown entropy of $\alpha_{S,\Theta}$ is bounded above by the topological entropy of the corresponding classical toral automorphism, i.e., by $\sum_{| \lambda_i | > 1} \log | \lambda_i |$, where $\lambda_1
, \dots , \lambda_d$ are the eigenvalues of $S$ counted with multiplicity (see, e.g., [@DGS Sect. 24] for a calculation of the topological entropy of toral automorphisms). Except in the case that the eigenvalues of $S$ all lie on the unit circle, this does not resolve the basic question of whether the entropy is positive or zero, i.e., of whether the system is “chaotic” or “deterministic.” When $\Theta$ is rational the Voiculescu-Brown entropy can be seen to be $\sum_{| \lambda_i | > 1}\log |
\lambda_i |$ by restricting $\alpha_{S,\Theta}$ to the centre of $A_\Theta$ and applying Propositions 4.4 and 4.8 of [@Voi]. In [@Nes] Neshveyev showed that if $S$ has no roots of unity then the entropic $K$-property (and in particular positive CNT entropy and hence also positive Voiculescu-Brown entropy by [@Voi Prop. 4.6]) follows from a summability condition with respect to a $2$-cocycle $\mathbb{Z}^d
\times\mathbb{Z}^d \to\mathbb{T}$ in terms of which $A_\Theta$ can be described. In the case $d=2$, when we are dealing with a rotation $C^*$-algebra $A_\theta$ ($=A_\Theta$ for $\Theta =
(\theta_{jk})_{j,k=1,2}$ with $\theta_{12} = \theta$), if $S$ has an eigenvalue $\lambda$ with $|\lambda| > 1$ then the set of $\theta$ for which the CNT entropy with respect to the canonical tracial state is positive has zero Lebesgue measure [@NT] and it contains $\mathbb{Z} + 2\mathbb{Z}\lambda^2$ as a consequence of the above-mentioned result of Neshveyev (see [@Nar; @Nes]).
The first aim of this article, which we carry out in Section \[S-pos\], is to show that, in an arbitrary dimension $d$, if the eigenvalues $\lambda_1 , \dots , \lambda_d$ do not all lie on the unit circle, then the Voiculescu-Brown entropy of $\alpha_{S,\Theta}$ is at least $\frac12 \sum_{| \lambda_i | > 1} \log | \lambda_i |$. (This was effectively claimed in [@NT] for $d=2$ but the tensor product argument given there is not correct.) In Section \[S-posloc\] we apply a result from [@EID] to obtain some information concerning the positivity of the local Voiculescu-Brown entropy of $\alpha_{S,\Theta}$ with respect to products of canonical unitaries. We prove in particular that if $S$ has no roots of unity as eigenvalues then the local Voiculescu-Brown entropy of $\alpha_{S,\Theta}$ with respect to any product of canonical unitaries is positive. Finally, in Section \[S-K\] we show in the general unital setting that completely positive CNT entropy of the von Neumann algebraic dynamical system arising from a faithful invariant state implies positivity of the unital version of local Voiculescu entropy with respect to every non-scalar unitary, and also that in the unital case this latter condition is equivalent to a noncommutative extension of the topological-dynamical notion of completely positive entropy which we call “completely positive Voiculescu-Brown entropy.” We then apply these results to the above-mentioned noncommutative toral automorphisms treated by Neshveyev in [@Nes].
[*Acknowledgements.*]{} D. Kerr was supported by the Natural Sciences and Engineering Research Council of Canada. This work was carried out during his stays at the University of Tokyo and the University of Rome “La Sapienza” over the academic years 2001–2002 and 2002–2003, respectively. He thanks Yasuyuki Kawahigashi at the University of Tokyo and Claudia Pinzari at the University of Rome “La Sapienza” for their generous hospitality. H. Li was supported jointly by the Mathematics Department of the University of Toronto and NSERC Grant 8864-02 of George A. Elliott. We thank George Elliott for valuable discussions. H. Li would also like to thank Sergey Neshveyev for helpful discussions.
Positive Voiculescu entropy {#S-pos}
===========================
We begin by recalling the definition of Voiculescu-Brown entropy [@Br], which is based on completely positive approximation (see [@Pau] for a reference on completely positive maps). Let $A$ be an exact $C^*$-algebra and $\alpha$ an automorphism of $A$. Let $\pi : A \to\mathcal{B} (\mathcal{H})$ be a faithful $^*$-representation. For a finite set $\Omega\subseteq A$ and $\delta > 0$ we denote by ${{\rm CPA}}(\pi , \Omega , \delta )$ the collection of triples $(\varphi , \psi , B)$ where $B$ is a finite-dimensional $C^*$-algebra and $\varphi : A\to B$ and $\psi
: B\to\mathcal{B} (\mathcal{H})$ are contractive completely positive maps such that $\| (\psi\circ\varphi )(a) - \pi (a) \| <
\delta$ for all $a\in\Omega$, and we define ${{\rm rcp}}(\Omega , \delta
)$ to be the infimum of ${{\rm rank}}\, B$ over all $(\varphi , \psi ,
B)\in{{\rm CPA}}(\pi , \Omega , \delta )$, with rank referring to the dimension of a maximal Abelian $C^*$-subalgebra. As the notation indicates, ${{\rm rcp}}(\Omega , \delta )$ is independent of the faithful $^*$-representation $\pi$, as shown in the proof of Proposition 1.3 in [@Br]. We then set $$\begin{aligned}
ht(\alpha , \Omega , \delta ) &= \limsup_{n\to\infty}\frac1n \log{{\rm rcp}}(\Omega\cup\alpha\Omega\cup\cdots\cup\alpha^{n-1}\Omega , \delta ) , \\
ht(\alpha , \Omega ) &= \sup_{\delta > 0}ht(\alpha , \Omega , \delta ) , \\
ht(\alpha ) &= \sup_\Omega ht(\alpha , \Omega )\end{aligned}$$ where the last supremum is taken over all finite sets $\Omega\subseteq A$. The quantity $ht(\alpha )$ is a $C^*$-dynamical invariant which we call the [*Voiculescu-Brown entropy*]{} of $\alpha$.
We will also have occasion to use the unital version of $ht(\alpha , \Omega )$ in Section \[S-K\], which we will denote by $ht'(\alpha , \Omega )$. This is defined in the case of unital $A$ by using unital completely positive maps instead of general contractive completely positive maps as above. When $1\not\in\Omega$ we may have $ht(\alpha , \Omega )\neq ht'(\alpha , \Omega )$, but these quantities do agree when $1\in\Omega$ as the proof of Proposition 1.4 in [@Br] shows, so that $ht(\alpha )$ may be alternatively obtained in the unital case by taking the supremum of $ht'(\alpha , \Omega )$ over all finite sets $\Omega\subseteq A$.
\[N-op\] For a $C^*$-algebra $A$ we denote by $A^{{{\rm op}}}$ the opposite algebra (i.e., the $C^*$-algebra obtained from $A$ by reversing the multiplication), and for each $a\in A$ we denote by $\tilde{a}$ the corresponding element in $A^{{{\rm op}}}$.
We would like to thank George Elliott for suggesting the proof we give of the following lemma, which simplifies our original proof.
\[L-op\] Let $A$ and $B$ be $C^*$-algebras and $\varphi : A\to B$ an $n$-positive map. Let $\varphi^{{{\rm op}}} : A^{{{\rm op}}}\to B^{{{\rm op}}}$ be the induced linear map given by $\varphi^{{{\rm op}}} (\tilde{a}) = \widetilde{\varphi (a)}$ for all $a\in A$. Then $\varphi^{{{\rm op}}}$ is also $n$-positive. In particular, if $\varphi$ is completely positive then so is $\varphi^{{{\rm op}}}$.
Suppose that $\varphi$ is $n$-positive, i.e., the map $\varphi\otimes id:A\otimes M_n(\mathbb{C})\to B\otimes
M_n(\mathbb{C})$ is positive. Since for every $C^*$-algebra $D$ the map $D\rightarrow D^{{{\rm op}}}$ given by $a\mapsto \tilde{a}$ for all $a\in D$ is an order isomorphism, it follows that the map $(\varphi\otimes id)^{{{\rm op}}} : (A\otimes M_n(\mathbb{C}))^{{{\rm op}}}\to
(B\otimes M_n(\mathbb{C}))^{{{\rm op}}}$ is positive. But this is just the map $\varphi^{{{\rm op}}}\otimes id^{{{\rm op}}} : A^{{{\rm op}}}\otimes
(M_n(\mathbb{C}))^{{{\rm op}}}\to B^{{{\rm op}}}\otimes
(M_n(\mathbb{C}))^{{{\rm op}}}$. Take an isomorphism $\beta:(M_n(\mathbb{C}))^{op}\to M_n(\mathbb{C})$. Then $(id_{B^{{{\rm op}}}}\otimes \beta)\circ (\varphi^{{{\rm op}}}\otimes
id^{{{\rm op}}})\circ (id_{A^{{{\rm op}}}}\otimes
\beta)^{-1}=\varphi^{{{\rm op}}}\otimes (\beta \circ id^{{{\rm op}}}\circ
\beta^{-1})=\varphi^{{{\rm op}}}\otimes id:A^{{{\rm op}}}\otimes
M_n(\mathbb{C})\rightarrow B^{{{\rm op}}}\otimes M_n(\mathbb{C})$ is positive, i.e., $\varphi^{{{\rm op}}}$ is $n$-positive.
\[P-op\] Let $\alpha$ be an automorphism of an exact $C^*$-algebra $A$ and let $\alpha^{{{\rm op}}}$ be the induced automorphism of $A^{{{\rm op}}}$. Then $ht(\alpha ) = ht(\alpha^{{{\rm op}}})$.
Given a faithful $^*$-representation $\pi : A\to\mathcal{B} (\mathcal{H})$ consider the induced injective $^*$-homomorphism $\pi^{{{\rm op}}} : A^{{{\rm op}}} \to
\mathcal{B} (\mathcal{H} )^{{{\rm op}}}$. Let $\tilde{\mathcal{H}} = \{ \tilde{x} :
x\in\mathcal{H} \}$ be the Hilbert space conjugate to $\mathcal{H}$ with scalar multiplication $\bar{\lambda}
\tilde{x} = \widetilde{\lambda x}$ and inner product $\left<\tilde{x}, \tilde{y}\right> = \left<y,x\right>$. Then we have a natural identification of $\mathcal{B} (\mathcal{H} )^{{{\rm op}}}$ with $\mathcal{B} (\tilde{\mathcal{H}})$ under which we can consider $\pi^{{{\rm op}}}$ as a $^*$-representation. It follows then by Lemma \[L-op\] that for any finite set $\Omega\subseteq A$ and $\delta > 0$ we have ${{\rm rcp}}(\tilde{\Omega} , \delta ) = {{\rm rcp}}(\Omega , \delta )$ where $\tilde{\Omega} = \{ \tilde{a} : a\in\Omega \}$, and so we conclude that $ht(\alpha ) = ht(\alpha^{{{\rm op}}})$.
\[L-circle\] Let $d\geq 2$ and let $\Theta$ be a real skew-symmetric $d\times
d$ matrix. Let $\beta$ be the canonical action of $\mathbb{T}^d$ on the noncommutative torus $A_{\Theta}$, and let $\alpha$ be an automorphism of $A_{\Theta}$ such that $\alpha
\beta(\mathbb{T}^n)\alpha^{-1}=\beta(\mathbb{T}^n)$ in ${{\rm Aut}}(A_{\Theta})$. Then $ht(\beta_x\alpha)=ht(\alpha\beta_x)=ht(\alpha)$ for all $x\in
\mathbb{T}^n$.
It suffices to show $ht(\beta_x\alpha)\leq ht(\alpha)$. Let $\pi : A_\Theta \to \mathcal{B}(\mathcal{H})$ be a faithful $^*$-representation of $A_\Theta$. For each $p\in \mathbb{Z}^d$ there is a unitary $u_p$ in $A_\Theta$ such that ${{\rm span}}\{ u_p :
p\in \mathbb{Z}^d \}$ is dense in $A_\Theta$ and $\beta_x (u_p) =
\left< p, x \right> u_p$, where $\left< \cdot,
\cdot\right>:\mathbb{Z}^d\times \mathbb{T}^d\to \mathbb{T}$ is the canonical pairing. For every $\omega\subseteq \mathbb{Z}^d$ we set $U_\omega = \{ u_p : p\in\omega \}$. Since $U_{\mathbb{Z}^d}$ is total in $A_\Theta$, by Proposition 2.6 of [@Br] $ht(\alpha)$ and $ht(\beta_x\alpha)$ are equal to the supremum of $ht(\alpha,
U_\omega )$ and $ht(\beta_x\alpha, U_\omega)$ over all finite sets $\omega\subseteq\mathbb{Z}^d$ respectively. Thus it suffices to show that $ht(\alpha, U_\omega ) \leq ht(\beta_x \alpha, U_\omega )$ for every finite set $\omega\subseteq\mathbb{Z}^d$, and this will follow once we show that $$\begin{gathered}
{{\rm rcp}}(U_\omega \cup (\beta_x \alpha )(U_\omega )\cup\cdots\cup
(\beta_x \alpha )^{m-1}(U_\omega ), \delta ) \hspace*{25mm} \\
\hspace*{35mm} \leq {{\rm rcp}}(U_\omega \cup\alpha (U_\omega )\cup\cdots\cup\alpha^{m-1} (U_\omega ),
\delta )\end{gathered}$$ for any given finite set $\omega\subseteq\mathbb{Z}^d$, $m\in\mathbb{N}$, and $\delta>0$. Suppose then that $(\varphi, \psi, B)$ is a triple in ${{\rm CPA}}(\pi, U_\omega\cup\alpha (U_\omega )\cup\cdots\cup
\alpha^{m-1} (U_\omega ), \delta )$ such that ${{\rm rank}}(B) = {{\rm rcp}}(U_\omega \cup\alpha (U_\omega )\cup\cdots\cup
\alpha^{m-1} (U_\omega ), \delta)$. For each $j\in \mathbb{Z}_{\geq 0}$ there exists some $x(j)\in\mathbb{T}^d$ such that $(\beta_x \alpha )^j=\alpha^j \beta_{x(j)}$. Then $(\beta_x \alpha )^j(u_p ) = \left< p,x(j) \right> \alpha^j (u_p )$ for every $p\in\mathbb{Z}^d$, and so $$\| (\psi\circ\varphi )((\beta_x \alpha )^j(u_p )) -
\pi ((\beta_x \alpha )^j (u_p )) \| = \| (\psi\circ\varphi )
(\alpha^j (u_p )) - \pi (\alpha^j (u_p )) \| < \delta$$ for all $j=0, \dots , m-1$ and $p\in\omega$. Thus the triple $(\varphi, \psi, B)$ is also contained in ${{\rm CPA}}(\pi, U_\omega\cup (\beta_x \alpha )(U_\omega )
\cup\cdots\cup (\beta_x \alpha )^{m-1} (U_\omega ), \delta )$, finishing the proof.
\[R-tori\] (1) Let $A$ be any exact $C^*$-algebra with a sequence of finite dimensional subspaces $V_1\subseteq V_2\subseteq \cdots$ such that $\bigcup_{j\in\mathbb{N}} V_j$ is dense in $A$, and let $G$ be a subgroup of ${{\rm Aut}}(A)$ preserving every $V_j$. If $\alpha\in{{\rm Aut}}(A)$ satisfies $\alpha G\alpha^{-1}=G$ then $ht(\beta \alpha ) = ht(\alpha \beta ) = ht(\alpha )$ for every $\beta\in G$. The proof of Lemma \[L-circle\] applies with minor modifications.
\(2) It is easy to show that an automorphism $\alpha$ of $A_\Theta$ satisfies the hypothesis of Lemma \[L-circle\] if and only if it is of the form $\alpha_{S,\Theta}\beta_x$ for some noncommutative toral automorphism $\alpha_{S,\Theta}$ and $x\in \mathbb{T}^d$.
\[L-tensor\] Let $\alpha_{S,\Theta}$ be any noncommutative toral automorphism. Then $$ht(\alpha_{S,\Theta}) + ht(\alpha_{S,-\Theta}) \geq
\sum_{| \lambda_i | > 1} \log | \lambda_i |$$ where $\lambda_1 , \dots , \lambda_d$ are the eigenvalues of $S$ counted with multiplicity.
Since the Voiculescu-Brown entropy of an automorphism of a separable commutative $C^*$-algebra agrees with the topological entropy of the induced homeomorphism on the pure state space by [@Voi Prop. 4.8], in the case $\Theta = 0$ we have $ht(\alpha_{S,0} ) =
\sum_{| \lambda_i | > 1} \log | \lambda_i |$. Consider now the tensor product $A_\Theta \otimes A_{-\Theta}$. Denoting by $u_1 , \dots , u_d$ and $v_1 , \dots , v_d$ the canonical unitaries of $A_\Theta$ and $A_{-\Theta}$, respectively, we have that the unitaries $u_j \otimes v_j$ for $j=1, \dots ,d$ form canonical generators for a copy $C$ of $C(\mathbb{T}^d )$. This can been seen from the fact that they operate as shifts in different coordinate directions on the Hilbert subspace $$\overline{{{\rm span}}} \{ \pi_+ (u^{k_1}_1 \cdots u^{k_d}_d )
\xi_+ \otimes\pi_- (v^{k_1}_1 \cdots v^{k_d}_d )\xi_- : (k_1 ,
\dots , k_d ) \in\mathbb{Z}^d \}$$ (identified with $\ell^2
(\mathbb{Z}^d )$) with respect to the tensor product of the canonical tracial state GNS representations $\pi_{\pm}$ of $A_{\pm\Theta}$ with canonical cyclic vectors $\xi_{\pm}$. We furthermore see that this identification of the $\alpha_{S,\Theta}
\otimes \alpha_{S,-\Theta}$-invariant $C^*$-subalgebra $C$ with $C(\mathbb{T}^d )$ establishes a conjugacy between $\alpha_{S,\Theta} \otimes \alpha_{S,-\Theta} \big| {}_C$ and $\alpha_{S,0}$. The monotonicity and tensor product subadditivity of Voiculescu-Brown entropy then yields $$\sum_{| \lambda_i | > 1} \log | \lambda_i | =
ht(\alpha_{S,0} ) = ht(\alpha_{S,\Theta} \otimes \alpha_{S,-\Theta}
\big| {}_C ) \leq ht(\alpha_{S,\Theta}) + ht(\alpha_{S,-\Theta}) .$$
\[T-half\] Let $\alpha_{S,\Theta}$ be any noncommutative toral automorphism. Then $$ht(\alpha_{S,\Theta}) =ht(\alpha_{S,-\Theta})\geq
\frac12 \sum_{| \lambda_i | > 1} \log | \lambda_i |$$ where $\lambda_1 , \dots , \lambda_d$ are the eigenvalues of $S$ counted with multiplicity.
Denoting by $u_1 , \dots , u_d$ and $v_1 , \dots , v_d$ the canonical unitaries of $A_\Theta$ and $A_{-\Theta}$, respectively, we have an isomorphism $A_\Theta \to A^{{{\rm op}}}_{-\Theta}$ given by $u_j \mapsto \tilde{v_j}$. Identify $A_\Theta$ and $A^{{{\rm op}}}_{-\Theta}$ via this isomorphism. By computing $\alpha^{{{\rm op}}}_{S,-\Theta} \in{{\rm Aut}}(A_\Theta )$ on the canonical unitaries we see that it has the form $\alpha_{S,\Theta} \beta_x$ for some $x\in\mathbb{T}^d$, where $\beta$ is the canonical action of $\mathbb{T}^d$ on $A_\Theta$. Proposition \[P-op\] and Lemma \[L-circle\] then yield $ht(\alpha_{S,-\Theta}) = ht(\alpha^{{{\rm op}}}_{S,-\Theta}) =
ht(\alpha_{S,\Theta})$. By Lemma \[L-tensor\] we also have $ht(\alpha_{S,\Theta}) + ht(\alpha_{S,-\Theta}) \geq
\sum_{| \lambda_i | > 1} \log | \lambda_i |$. The assertion of the theorem now follows.
Positive local Voiculescu entropy with respect to products of canonical unitaries {#S-posloc}
=================================================================================
Our goal in this section is to obtain some information concerning positivity of local Voiculescu-Brown entropy with respect to products of canonical unitaries. We will proceed by first relating a noncommutative toral automorphism to the corresponding classical toral automorphism as in the proof of Lemma \[L-tensor\] but at a local level, and then appealing to a result from [@EID] involving local Voiculescu-Brown entropy in the separable unital commutative setting. Throughout this section we will denote the canonical unitaries of the commutative $d$-torus $A_0 \cong C(\mathbb{T}^d )$ by $f_1 , \dots , f_d$. We will continue to denote the canonical unitaries of a general noncommutative $d$-torus by $u_1 , \dots , u_d$.
\[L-loctori\] Let $\alpha_{S,\Theta}$ be a noncommutative toral automorphism, $k_1 , \dots , k_d \in\mathbb{Z}$, and $\lambda\in\mathbb{C}$. Then $$ht(\alpha_{S,\Theta} , \{ \lambda
u^{k_1}_1 \cdots u^{k_d}_d \} ) \geq \frac12 ht(\alpha_{S,0} , \{
\lambda f^{k_1}_1 \cdots f^{k_d}_d \} ) .$$
We may assume that $\lambda = 1$. Let $C$ be the $\alpha_{S,\Theta} \otimes \alpha_{S,-\Theta}$-invariant commutative $C^*$-algebra of $A_\Theta \otimes A_{-\Theta}$ identified in the proof of Lemma \[L-tensor\]. Denoting by $v_1 , \cdots , v_d$ the canonical unitaries of $A_{-\Theta}$, we have $$\begin{aligned}
ht(\alpha_{S,0} , \{ f^{k_1}_1 \cdots f^{k_d}_1 \})
&= ht(\alpha_{S,\Theta} \otimes \alpha_{S,-\Theta} \big| {}_C ,
\{ u^{k_1}_1 \cdots u^{k_d}_d \otimes v^{k_1}_1 \cdots v^{k_d}_d \} ) \\
&\leq ht(\alpha_{S,\Theta} , \{ u^{k_1}_1 \cdots
u^{k_d}_d ) \} ) + ht(\alpha_{S,-\Theta} , \{ v^{k_1}_1 \cdots v^{k_d}_d \} ) ,\end{aligned}$$ where the last inequality follows from an argument similar to that in the proof of Proposition 3.10 in [@Voi]. As in the proof of Theorem \[T-half\] we identify $A_\Theta$ with $A^{{{\rm op}}}_{-\Theta}$ via $u_j \mapsto \tilde{v}_j$ and observe that $\alpha^{{{\rm op}}}_{-\Theta} \in{{\rm Aut}}(A_\Theta )$ has the form $\alpha_{S,\Theta} \beta_x$ for some $x\in\mathbb{T}^d$, where $\beta$ is the canonical action of $\mathbb{T}^d$ on $A_\Theta$. Following Notation \[N-op\], we then have $\widetilde{v^{k_1}_1
\cdots v^{k_d}_d} = \eta \tilde{v}^{k_1}_1 \cdots
\tilde{v}^{k_d}_d$ for some $\eta\in\mathbb{C}$ of unit modulus, and so $$ht (\alpha_{S,-\Theta} , \{ v^{k_1}_1 \cdots v^{k_d}_d \} )
= ht (\alpha^{{{\rm op}}}_{S,-\Theta} , \{ \widetilde{v^{k_1}_1 \cdots
v^{k_d}_d} \} ) = ht(\alpha_{S,\Theta} , \{ u^{k_1}_1 \cdots
u^{k_d}_d \} )$$ in view of the proofs of Proposition \[P-op\] and Lemma \[L-circle\]. The assertion of the lemma now follows from the above two displays.
\[T-localpos\] Let $\alpha_{S,\Theta}$ be a noncommutative toral automorphism and suppose that $S$ has no roots of unity as eigenvalues. Then $$ht(\alpha_{S,\Theta} , \{ \lambda u^{k_1}_1
\cdots u^{k_d}_d \} ) > 0$$ for any non-zero $(k_1 , \dots , k_d )\in\mathbb{Z}^d$ and non-zero $\lambda\in\mathbb{C}$.
Since the measure-theoretic toral automorphism associated to $S$ via Lebesgue measure is ergodic (see [@Wal]) and hence has completely positive (Kolmogorov-Sinai) entropy [@MP], the topological toral automorphism associated to $S$ (i.e., the case $\Theta = 0$ at the level of the pure state space) has completely positive (topological) entropy, i.e., each of its non-trivial factors has positive topological entropy (see [@FPTE]). Thus $ht(\alpha_{S,0} , \{ x \}) > 0$ for every non-scalar $x\in A_0 \cong C(\mathbb{T}^d )$ by Corollary 4.4 of [@EID]. Lemma \[L-loctori\] then yields the result.
For a general noncommutative toral automorphism $\alpha_{S,\Theta}$ it follows from Lemma \[L-loctori\] that to conclude that $ht(\alpha_{S,\Theta} , \{ \lambda u^{k_1}_1 \cdots
u^{k_d}_d \} ) > 0$ we need only show that $ht(\alpha_{S,0} , \{
\lambda f^{k_1}_1 \cdots f^{k_d}_d \})
> 0$. If we are simply dealing with a canonical unitary $u_j$ then this occurs, for example, if the $i$th coordinate axis in $\mathbb{R}^d$ is not orthogonal to some one-dimensional subspace of a eigenspace in $\mathbb{R}^d$ corresponding to a real eigenvalue of $S$ not equal to $\pm 1$. To see this, suppose $L$ is such a one-dimensional subspace and let $\lambda$ be the associated real eigenvalue. We may assume $|\lambda | > 1$ since $ht(\alpha^{-1}_{S,\Theta} , \{ u_j \}) = ht(\alpha_{S,\Theta} ,
\{ u_j \})$ (see the proof of Proposition 2.5 in [@Br]) and $\alpha^{-1}_{S,\Theta} = \alpha_{S^{-1},\Theta} \beta_x$ for some $x\in\mathbb{T}^d$, where $\beta$ is the canonical action of $\mathbb{T}^d$ on $A_\Theta$. Define the pseudo-metric $d_j$ on $\mathbb{T}^d$ by $$d_j (x,y) = | f_j (x) - f_j (y) |$$ for all $x,y\in\mathbb{T}$, with the unitary $f_i$ being considered here in the canonical way as a function on the pure state space $\mathbb{T}^d$. Since the $j$th coordinate subspace of $\mathbb{R}^d$ is not orthogonal to $L$ there exists a $\delta >
0$ such that, for any $x = (x_1 , \dots , x_d )$ and $y = (y_1 , \dots , y_d )$ in $L$, the $j$th coordinate distance $| x_j - y_j |$ is at least $\delta$ times the Euclidean distance between $x$ and $y$. Now since the action of $T$ on $L$ is simply multiplication by $\lambda$ it can be seen via a covering space argument (see the proof of Theorem 24.5 in [@DGS]) that there exists a $C>0$ and an $\varepsilon
> 0$ such that for every $n\in\mathbb{N}$ the image of $L$ under the quotient map onto $\mathbb{T}^d \cong\mathbb{R}^d /
\mathbb{Z}^d$ contains an $(n,\varepsilon )$-separated set of cardinality at least $C |\lambda |^n$, from which it follows that the entropy $h_{d_j} (\bar{S})$ is strictly positive, where $\bar{S}$ is the automorphism of $\mathbb{T}^d$ corresponding to $S$. Here we are using standard notation and terminology from topological dynamics (see [@DGS]) except that we are allowing the metric in the definition of entropy to be merely a pseudo-metric. Now by Theorem 4.3 of [@EID] we conclude that $ht(\alpha_{S,0} , \{ u_j \}) > 0$, as desired.
By a similar argument which allows for the possibility of non-trivial Jordan cells we have the following more general statement.
Let $\alpha_{S,\Theta}$ be a noncommutative toral automorphism. Let $j\in \{ 1 , \dots , d \}$ and suppose that the $j$th coordinate axis in $\mathbb{R}^d$ is not orthogonal to the span of the generalized eigenspaces associated to the set of real eigenvalues of $S$ not equal to $\pm 1$. Then $ht(\alpha_{S,\Theta} , \{ u_j \} ) > 0$.
We could evidently furthermore generalize this theorem to handle products of canonical unitaries, and also formulate a similar result involving pairs of canonical unitaries and complex eigenvalues not on the unit circle.
Completely positive Voiculescu-Brown entropy {#S-K}
============================================
In [@Nes] Neshveyev showed that, for the von Neumann algebraic dynamical system arising from a noncommutative toral automorphism $\alpha_{S,\Theta}$ via the canonical tracial state on $A_\Theta$, the property of being an entropic $K$-system (and in particular of having completely positive CNT entropy), in the case of $S$ having no roots of unity as eigenvalues (which occurs if and only if $S$ is aperiodic (see [@Pet] or [@Wal])), is a consequence of a summability condition which for $d=2$ is satisfied for a certain countable set of rotation parameters. We will show in the general unital setting that completely positive CNT entropy of the von Neumann algebraic dynamical system arising from a faithful invariant state implies “completely positive Voiculescu-Brown entropy” (i.e., every restriction of the automorphism to a non-trivial invariant $C^*$-subalgebra has positive Voiculescu-Brown entropy), and that in the unital case the latter property is equivalent to the positivity of the unital version of local Voiculescu-Brown entropy with respect to every non-scalar unitary.
We will use the standard notation for CNT (Connes-Narnhofer-Thirring) entropy [@CNT]. Let $\alpha$ be an automorphism of a unital $C^*$-algebra $A$ and $\omega$ a faithful $\alpha$-invariant state on $A$. Denoting by $\pi_\omega$ the GNS representation corresponding to $\omega$, we obtain extensions $\bar{\alpha}$ and $\bar{\omega}$ of $\alpha$ and $\omega$, respectively, to $\pi_\omega (A)''$. By definition the automorphism $\bar{\alpha}$ has completely positive CNT entropy if $h_{\bar{\omega} , \bar{\alpha}} (N) > 0$ for all unital finite-dimensional $C^*$-subalgebras of $\pi_\omega (A)''$ which are different from the scalars. Entropic $K$-systems have completely positive CNT entropy, and the two notions coincide in the commutative case. For definitions and discussions see [@QKS; @GN]. Note that in [@GN] what we are referring to for clarity as “completely positive CNT entropy” is simply called “completely positive entropy.”
An automorphism $\alpha$ of a non-trivial exact $C^*$-algebra $A$ is said to have [*completely positive Voiculescu-Brown entropy*]{} if $ht(\alpha\big| {}_B ) > 0$ for every non-zero $\alpha$-invariant $C^*$-subalgebra $B\subseteq A$ which, if $A$ is unital, is not equal to the scalars.
If $A$ is unital then in the above definition we may take the $C^*$-subalgebras $B$ to be unital, for if $B\subseteq A$ is an $\alpha$-invariant $C^*$-subalgebra not containing the unit of $A$ then the Voiculescu-Brown entropies of the restrictions $\alpha\big|
{}_B$ and $\alpha\big| {}_{B+\mathbb{C}1}$ agree by Lemma 1.7 of [@Br]. Thus in the separable unital commutative situation we recover the topological-dynamical notion of completely positive entropy, which refers to the absence of non-trivial factors with zero topological entropy [@FPTE].
\[P-cpe\] Let $\alpha$ be an automorphism of a unital exact $C^*$-algebra $A$ and $\omega$ a faithful $\alpha$-invariant state on $A$, and suppose that the extension $\bar{\alpha}$ of $\alpha$ to $\pi_\omega (A)''$ has completely positive CNT entropy. Then $\alpha$ has completely positive Voiculescu-Brown entropy.
Let $B\subseteq A$ be a unital $\alpha$-invariant $C^*$-subalgebra different from the scalars. By choosing a non-scalar element in $B$ and taking its real part (or imaginary part if its real part is a scalar) we obtain a self-adjoint element $b\in B$ whose spectrum contains more than one point, and we can construct an injective unital $^*$-homomorphism $\gamma : \mathbb{C}^2 \to
\pi_\tau (C({{\rm spec}}(b)))'' \subseteq \pi_\tau (B)''$. Since $\bar{\alpha}$ has completely positive CNT entropy we have $h_{\bar{\omega} , \bar{\alpha}} (\gamma ) > 0$. For any $\delta >
0$ it is easy to construct, using a suitable two-element partition of unity in $C({{\rm spec}}(b))$, a unital positive map $\gamma' :
\mathbb{C}^2 \to C({{\rm spec}}(b))\subseteq B$ (which is necessarily completely positive since $\mathbb{C}^2$ is commutative) such that $\| \gamma - \gamma' \|_\tau < \delta$, where $$\| \gamma - \gamma' \|_\tau =
\sup_{\{ x\in\mathbb{C}^2 \, : \, \| x \| \leq 1 \}} \| \gamma (x)
- \gamma' (x) \|_\tau$$ and $\| a \|_\tau = ( \tau (a^*
a))^{1/2}$ for all $a\in A$. It follows from Theorem VI.3 of [@CNT] that if $\delta$ is chosen small enough then $$h_{\omega|_B , \alpha|_B } (\gamma' ) \geq
h_{\bar{\omega} , \bar{\alpha}} (\gamma' ) > 0$$ with the first inequality being immediate from the definition of CNT entropy. Proposition 9 of [@Dyk] then yields $$ht(\alpha\big| {}_B ) \geq h_{\omega|_B} (\alpha\big| {}_B )
> 0 ,$$ as desired.
Recall from the beginning of Section \[S-K\] that $ht' (\alpha , \Omega )$ denotes the unital version of the local Voiculescu-Brown entropy $ht(\alpha , \Omega )$. We will next show that, in the unital case, completely positive Voiculescu-Brown entropy is equivalent to positivity of the unital version of local Voiculescu-Brown entropy with respect to every non-scalar unitary. For this we will need the following Kolmogorov-Sinai-type property, which is similar to that of Lemma A.2 in [@Oz].
\[L-KS\] Let $\alpha$ be an automorphism of a unital exact $C^*$-algebra $A$. If $\Omega_1 \subseteq \Omega_2 \subseteq \Omega_3
\subseteq\cdots$ is a nested sequence of finite sets of unitaries in $A$ such that $\bigcup_{k\in\mathbb{N},n\in\mathbb{Z}} \alpha^n
\Omega_k$ generates $A$ as a $C^*$-algebra, then $$ht(\alpha ) = \sup_{k\in\mathbb{N}} ht' (\alpha , \Omega_k ) .$$
Given a unital completely positive map $\varphi$ from $A$ into any unital $C^*$-algebra $B$, Lemma 3.1 of [@JOR] yields $$\| \varphi (x^* y) - \varphi (x)^* \varphi (y) \| \leq \| \varphi (x^* x)
- \varphi (x)^* \varphi (x) \|^{1/2}\| \varphi (y^* y) -
\varphi (y)^* \varphi (y) \|^{1/2}$$ for all $x,y\in A$, and so in particular for any unitaries $u,v \in A$ we have $$\begin{aligned}
\| \varphi (uv) - \varphi (u) \varphi (v) \| &\leq \| 1 - \varphi (u)^*
\varphi (u) \|^{1/2} \| 1 - \varphi (v)^* \varphi (v) \|^{1/2} \\
&\leq (\| u^* u - \varphi (u)^* u \| + \| \varphi (u)^* u -
\varphi (u)^* \varphi (u) \| )^{1/2} \\
&\hspace*{8mm}\ \times (\| v^* v - \varphi (v)^* v \| +
\| \varphi (v)^* v - \varphi (v)^* \varphi (v) \|)^{1/2} \\
&\leq 2 \| \varphi (u) - u \|^{1/2} \| \varphi (v) - v \|^{1/2}\end{aligned}$$ whence $$\begin{aligned}
\| \varphi (uv) - uv \| &\leq \| \varphi (uv) - \varphi (u) \varphi (v) \| +
\| \varphi (u) \varphi (v) - u\varphi (v) \| + \| u\varphi (v) - uv \| \\
&\leq (\| \varphi (u) - u \|^{1/2} + \| \varphi (v) - v \|^{1/2} )^2 .\end{aligned}$$ We can now proceed along the lines of the proofs of Propositions 1.4 and 3.4 of [@Voi] to obtain the result.
\[P-cpeunit\] Let $\alpha$ be an automorphism of a unital exact $C^*$-algebra $A$. Then $\alpha$ has completely positive Voiculescu-Brown entropy if and only if $ht' (\alpha , \{ u \} ) > 0$ for every non-scalar unitary $u\in A$.
For the “only if” direction we can consider for any non-scalar unitary $u\in A$ the unital $\alpha$-invariant $C^*$-subalgebra it generates and appeal to Lemma \[L-KS\]. For the “if” direction, we need simply observe that every unital $C^*$-algebra different from the scalars contains a non-scalar unitary, as can be obtained by applying the functional calculus to the real part (or imaginary part if the real part is a scalar) of any non-scalar element.
For a real skew-symmetric $d\times d$ matrix $\Theta$ the noncommutative $d$-torus $A_\Theta$ may alternatively be described as the universal unital $C^*$-algebra generated by unitaries $\{
u_g \}_{g\in \mathbb{Z}^d}$ subject to the relations $$u_g u_h = \beta (g,h) u_{g+h},$$ where $\beta : \mathbb{Z}^d \times\mathbb{Z}^d \to\mathbb{T}$ is a bicharacter satisfying $$\beta (g,h) \beta (h,g)^{-1} = e^{2\pi ig\cdot\Theta h}.$$
\[T-sum\] Let $\alpha_{S,\Theta}$ be a noncommutative toral automorphism, and suppose that $S$ is $\beta$-preserving and has no roots of unity as eigenvalues, and that $$\sum_{n\in\mathbb{Z}} | 1 - \beta (g,S^n h ) | < \infty$$ for all $g,h\in \mathbb{Z}^d$. Then $\alpha_{S,\Theta}$ has completely positive Voiculescu-Brown entropy and $ht' (\alpha_{S,\Theta} , \{ u \} ) > 0$ for every non-scalar unitary $u\in A$.
By Theorem 2 of [@Nes] the hypotheses of the theorem statement imply that the von Neumann algebraic system obtained from $\alpha_{S,\Theta}$ via the canonical tracial state on $A_\Theta$ is an entropic $K$-system, and hence has completely positive CNT entropy. Propositions \[P-cpe\] and \[P-cpeunit\] then yield the desired conclusion.
Theorem \[T-sum\] applies in particular in the case $d=2$ when the matrix $S$ has eigenvalues off the unit circle and the rotation parameter $\theta$ of the rotation $C^*$-algebra $A_\theta$ ($=A_\Theta$ for $\Theta = (\theta_{jk})_{j,k=1,2}$ with $\theta_{12} = \theta$) lies in $\mathbb{Z} + 2\mathbb{Z} \lambda^2$ where $\lambda$ is the (necessarily real) eigenvalue of $S$ of largest absolute value (see [@Nes]).
Finally, we would like to point out that the argument in [@Nes] is not quite complete. Indeed in the proof of the lemma in [@Nes] it is incorrectly taken to be the case that a matrix $T\in GL(n,\mathbb{Z})$ is aperiodic if and only if it has no eigenvalues on the unit circle. (Aperiodicity is defined as the non-existence of non-trivial finite orbits of $T$ acting on $\mathbb{Z}^n$ and is equivalent to $T$ having no roots of unity as eigenvalues and also equivalent to the ergodicity of the measure-theoretic automorphism of $\mathbb{T}^n$ associated to $T$ via Lebesgue measure (see [@Pet] or [@Wal], and also Section 24 of [@DGS]).) However, if we let $\mathbb{R}^n = V_1 \oplus V_2$ be the decomposition of $\mathbb{R}^n$ corresponding to the eigenvalues of $T$ of modulus at least one and strictly less than one, respectively, and denote by $P_1$ and $P_2$ the associated projections, then the proof of the lemma in [@Nes] demonstrates that, for large $n$, if $y_1 + T^n y_2 + \dots + T^{n(k-1)} y_k = 0$ then $P_2 (y_1 ) = \dots = P_2 (y_k ) = 0$. The argument in the first paragraph of p. 191 of [@Katz] then shows that $P_2$ is injective on $\mathbb{Z}^n$, whence $y_1 = \dots = y_k = 0$, as desired.
[999]{}
F. Benatti, H. Narnhofer, and G. L. Sewell. A noncommutative version of the Arnol’d cat map. [*Lett. Math. Phys.*]{} [**21**]{} (1991), 157–172.
F. Blanchard. Fully positive topological entropy and topological mixing. [*Symbolic dynamics and it applications.*]{} New Haven, CT, 1991, pp. 95–105.
B. Brenken. Representations and automorphisms of the irrational rotation algebra. [*Pacific J. Math.*]{} [**111**]{} (1984), 257–282.
N. P. Brown. Topological entropy in exact $C^*$-algebras. [*Math. Ann.*]{} [**314**]{} (1999), 347–367.
A. Connes, H. Narnhofer, and W. Thirring. Dynamical entropy of $C^*$-algebras and von Neumann algebras. [*Commun. Math. Phys.*]{} [**112**]{} (1987), 691–719.
M. Denker, C. Grillenberger, and K. Sigmund. [*Ergodic Theory on Compact Spaces.*]{} Lecture Notes in Math, vol. 527. Springer-Verlag, Berlin, Heidelberg, New York, 1976.
K. Dykema. Topological entropy of some automorphisms of reduced amalgamated free product $C^*$-algebras. [*Ergod. Th. Dynam. Sys.*]{} [**21**]{} (2001), 1683–1693.
G. A. Elliott. The diffeomorphism group of the irrational rotation $C^*$-algebra. [C. R. Math. Rep. Acd. Sci. Canada]{} [**8**]{}, (1986), 329–334.
V. Y. Golodets and S. V. Neshveyev. Non-Bernouillian quantum $K$-systems. [*Commun. Math. Phys.*]{} [**195**]{} (1998), 213–232.
B. Hasselblatt and A. Katok. [*Intoduction to the Modern Theory of Dynamical Systems.*]{} Encyclopedia of Mathematics and its Applications, 54. Cambridge University Press, Cambridge, 1995.
M. Junge, N. Ozawa, and Z.-J. Ruan. On $\mathcal{OL}_\infty$ structure of nuclear $C^*$-algebras. arXiv:math.OA/0206061.
Y. Katznelson. Ergodic automorphisms of $T^n$ are Bernoulli shifts. [*Israel J. Math.*]{} [**10**]{} (1971), 186–195.
D. Kerr. Entropy and induced dynamics on state spaces. arXiv:math.OA/0302118 v3, 2003.
H. Narnhofer. Quantized Arnol’d cat maps can be entropic $K$-systems. [*J. Math. Phys.*]{} [**33**]{} (1992), 1502–1510.
H. Narnhofer. Ergodic properties of automorphisms on the rotation algebra. [*Rep. Math. Phys.*]{} [**39**]{} (1997), 387–406.
H. Narnhofer and W. Thirring. $C^*$-dynamical systems that are asymptotically highly anticommutative. [*Lett. Math. Phys.*]{} [**35**]{} (1995), 145–154.
H. Narnhofer and W. Thirring. Quantum $K$-systems. [*Commun. Math. Phys.*]{} [**125**]{} (1989), 565–577.
S. V. Neshveyev. On the $K$-property of quantized Arnold cat maps. [*J. Math. Phys.*]{} [**41**]{} (2000). 1961–1965.
D. Olesen, G. K. Pedersen, and M. Takesaki. Ergodic actions of compact Abelian groups. [*J. Operator Theory*]{} [**3**]{} (1980), 237–269.
N. Ozawa. Nuclearity of reduced amalgamated free product $C^*$-algebras. Preprint, 2001.
V. I. Paulsen. [*Completely bounded maps and dilations.*]{} Pitman Research Notes in Mathematics [**146**]{}, Longman, 1986.
K. Petersen. [*Ergodic Theory.*]{} Cambridge Studies in Advanced Mathematics, 2. Cambridge University Press, Cambridge, 1983.
M. A. Rieffel. Noncommutative tori—a case study of noncommutative differentiable manifolds. [*Contemporary Math.*]{} [**105**]{} (1990), 191–211.
V. A. Rohlin. Metric properties of endomorphisms of compact commutative groups. [*Amer. Math. Soc. Transl. Ser. 2*]{} [**64**]{} (1967), 244–252.
J. Slawny. On factor representations and the $C^*$-algebra of canonical commutation relations. [*Commun. Math. Phys.*]{} [**24**]{} (1972), 151–170.
E. St[ø]{}rmer. A survey of noncommutative dynamical entropy. In: [*Classification of Nuclear $C^*$-algebras. Entropy in Operator Algebras*]{}, pp. 147–198. Berlin: Springer, 2002.
D. V. Voiculescu. Dynamical approximation entropies and topological entropy in operator algebras. [*Commun. Math. Phys.*]{} [**170**]{} (1995), 249–281.
P. Walters. [*An Introduction to Ergodic Theory.*]{} Graduate Texts in Mathematics, 79. Springer-Verlag, New York, Berlin, 1982.
Y. Watatani. Toral automorphisms on irrational rotation algebras. [*Math. Japon.*]{} [**26**]{} (1981), 479–484.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $W$ be a compact hyperbolic $n$-manifold with totally geodesic boundary. We prove that if $n>3$ then the holonomy representation of $\pi_1 (W)$ into the isometry group of hyperbolic $n$-space is infinitesimally rigid.'
author:
- 'Steven P. Kerckhoff and Peter A. Storm'
bibliography:
- 'bibliography.bib'
title: Local Rigidity of Hyperbolic Manifolds with Geodesic Boundary
---
\[section\] \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Conjecture]{} \[thm\][Question]{}
\[thm\][Definition]{}
[^1]
{#sec:introduction .unnumbered}
Starting with a complete hyperbolic manifold $X$, when is it possible to deform its hyperbolic structure? If $X$ is a finite area surface, its metric always has nontrivial deformations through complete hyperbolic metrics. However, in all higher dimensions, the work of Calabi, Weil, and Garland combine to prove that there are no nontrivial deformations through complete metrics when $X$ has finite volume [@Cal; @We; @Garland]. In search of flexibility, one is led to study infinite volume hyperbolic manifolds.
In dimension $3$ infinite volume hyperbolic manifolds have been studied for many years. A particularly well-understood class of such manifolds are the convex cocompact ones. They are topologically the interior of a compact $3$-manifold with boundary consisting of surfaces of genus at least $2$. Convex cocompact hyperbolic $3$-manifolds have large deformation spaces parametrized by conformal structures on their boundary surfaces inherited from the sphere at infinity. In higher dimensions the situation becomes more mysterious, but one expects to find less flexibility than in dimension $3$.
This article’s research began by looking for a large class of infinite volume hyperbolic manifolds, present in all dimensions, and rigid in dimensions greater than $3$. In high dimensions it is difficult to construct interesting infinite volume hyperbolic manifolds. One method is to begin with a closed hyperbolic manifold (often constructed by arithmetic tools) that contains an embedded totally geodesic hypersurface and cut along the hypersurface to obtain a compact hyperbolic manifold $W^n$ with totally geodesic boundary. There is a canonical extension of such a structure to a complete, infinite volume hyperbolic metric on an open manifold $X^n$ without boundary which is diffeomorphic to the interior of $W^n$. This is the class of manifolds we will study.
Associated to any hyperbolic manifold is a representation, called the *holonomy representation*, of its fundamental group in the group ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ of isometries of $n$-dimensional hyperbolic space. When it is a compact hyperbolic manifold $W$ with totally geodesic boundary, the representation is discrete and faithful. If we denote the image of the representation by $\Gamma$, then $W = K/\Gamma$ where $K$ is a closed subset of ${\ensuremath{\mathbb{H} } }^n$ bounded by a collection of geodesic hyperplanes that map onto the boundary of $W$. The quotient ${\ensuremath{\mathbb{H} } }^n/\Gamma$ of all of hyperbolic space by $\Gamma$ is a complete, infinite volume hyperbolic $n$-manifold $X$ that we call the extension of $W$. It is diffeomorphic to the interior of $W$ and contains $W$ as a compact, convex subset. In particular, $\pi_1 X = \pi_1 W$ and the holonomy representations of $X$ and $W$ are the same.
The holonomy representation $\rho$ of $\pi_1 W$ is uniquely determined, up to conjugation in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$, by the hyperbolic structure on $W$. The space ${\ensuremath{\operatorname{Hom} } }(\pi_1 W, {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n))$ is topologized by the compact-open topology. The holonomy representation $\rho$ is said to be *locally rigid* if a neighborhood of it in ${\ensuremath{\operatorname{Hom} } }(\pi_1 W, {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n))$ consists entirely of conjugate representations. In this case, we will refer to the hyperbolic manifold $W$ (or $X$) as locally rigid. Note that nearby representations of $\pi_1 W$ need not correspond to hyperbolic structures with geodesic boundary.
The main result of this paper is:
\[thm:intromain\] Let $W$ be a compact hyperbolic $n$-manifold with totally geodesic boundary. If $n>3$ then the holonomy representation of $W$ is locally rigid.
We will actually prove (Theorem \[thm:main\]) the slightly stronger result that such hyperbolic manifolds $W$ are *infinitesimally rigid*, a concept that will be discussed in Section \[sec:algebraic preliminaries\].
Theorem \[thm:intromain\] was known only in a couple special cases prior to this paper. Initially, M. Kapovich observed that Mostow rigidity could be applied to prove an orbifold version of Theorem \[thm:intromain\] for the group generated by reflections in $119$ of the $120$ walls ($3$-cells) of the hyperbolic right-angled $120$-cell in $\mathbb{H}^4$. This group produces an orbifold quotient of $\mathbb{H}^4$ with totally geodesic boundary in a suitable orbifold sense. Again starting with the hyperbolic $120$-cell, Aougab and the second author verified Theorem \[thm:intromain\] for a specific group generated by reflections in $96$ walls of the hyperbolic right-angled $120$-cell [@St13]. These $96$ walls were chosen so the complementary $24$ walls form a maximal collection of pairwise disjoint walls.
We remark that, at least when $n = 4$, it is essential that the manifolds $W$ with totally geodesic boundary be compact. Specifically, the hypotheses of Theorem \[thm:intromain\] cannot be loosened to permit $W$ to have finite volume. With such a hypothesis Theorem \[thm:intromain\] is false. This is due to the $4$-dimensional example constructed in [@KS]. There, the authors study in detail deformations of reflection groups derived from the $4$-dimensional hyperbolic $24$-cell. These deformations involved a $4$-dimensional and infinite volume analog of $3$-dimensional hyperbolic Dehn filling. The proof of Theorem \[thm:intromain\] shows that, in a suitable sense, the deformations of [@KS] are the only type possible for $4$-manifolds with totally geodesic boundary. In particular, we conjecture that this is a purely $4$-dimensional phenomenon and that Theorem \[thm:intromain\] remains true in dimensions $\geq 5$ even if $W$ is allowed to be non-compact with finite volume. We will comment on this further in Section \[sec:conjectures\].
One of the central ideas in the proof of Theorem \[thm:intromain\] is the interaction between the deformations of the holonomy representation of the $n$-dimensional manifold $W$ and those of its $(n-1)$-dimensional geodesic boundary. Starting with a hyperbolic $n$-manifold $W^n$ with totally geodesic boundary $M^{n-1}$, we can restrict the holonomy representation $\rho$ of $\pi_1 W$ to the fundamental group of any component of $M$. For simplicity of discussion, we will assume that $M$ is connected. We denote the restricted holonomy representation by $\rho_M$. Since $M$ is an $(n-1)$-dimensional hyperbolic manifold it is the quotient of a geodesic hyperplane which is preserved by the isomorphic image of $\pi_1 M$ under $\rho_M$. Now glue two copies of $W$ along the boundary to obtain a hyperbolic structure on the double of $W$. The double is a closed manifold which can’t have any nontrivial deformations when $n\geq 3$. This implies, in these dimensions, that there are no nontrivial deformations of the holonomy representation of $W$ through hyperbolic structures with totally geodesic boundary, ie. deformations where the restriction $\rho_M$ continues to preserve a geodesic hyperplane.
Let $\Gamma$ denote the image $\rho_M(\pi_1 M)$ in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ and consider the $n$-dimensional hyperbolic manifold $X = {\ensuremath{\mathbb{H} } }^n/\Gamma$. Topologically, $X$ is a product $M \times {\ensuremath{\mathbb{R} } }$. Geometrically, it has infinite volume and contains a totally geodesic submanifold isometric to $M$ that completely determines its geometry. We will call such an $n$-dimensional hyperbolic manifold *Fuchsian*.
In all dimensions, Johnson and Millson [@JMbending] found Fuchsian hyperbolic manifolds $X = M \times {\ensuremath{\mathbb{R} } }$ with nontrivial deformations. These deformations depend on the existence of a codimension $1$ totally geodesic submanifold in $M.$ This geodesic submanifold allows one to deform the holonomy representation of $X$ in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ using the process of bending. (See Example 8.7.3 of [@Th] or [@JMbending] for an explanation of bending.) On the other hand, specific examples of $4$-dimensional Fuchsian manifolds without deformations were discovered by Kapovich [@Kap2] and Scannell [@Scannell].
The flexibility or rigidity of an $n$-dimensional Fuchsian manifold $X = M \times {\ensuremath{\mathbb{R} } }$ is really a question about the $(n-1)$-dimensional hyperbolic manifold $M$. We will say that $M^{n-1}$ is either *deformable* or *locally rigid* in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ depending on whether or not its associated Fuchsian manifold $X^n$ has nontrivial deformations. Although both Kapovich and Scannell construct infinite families of rigid examples, it remains unclear what geometric or topological condition on $M^{n-1}$ would imply local rigidity in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$, even when $n=4$. This is a very interesting question about which little is known. We do not address it here (though see Theorem \[harmonicbends\]).
The discussion above implies that, for the holonomy representation $\rho$ of a hyperbolic $n$-manifold $W$ with geodesic boundary to have nontrivial deformations, its $(n-1)$-dimensional boundary $M$ must be deformable in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$. This suggests the possibility that Theorem \[thm:intromain\] is true simply because any closed $(n-1)$-dimensional hyperbolic manifold $M$ which is the totally geodesic boundary of an $n$-dimensional manifold $W$ is, in fact, locally rigid in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$. However, examples of Gromov and Thurston show this is definitely not the case [@GromThur].
For each dimension $n>2$ they construct an infinite number of closed hyperbolic $n$-manifolds $V$ with the following properties: $V$ has a codimension $1$ embedded, totally geodesic submanifold $M$ which itself has a codimension $1$ embedded, totally geodesic submanifold $P$ (so $P$ is codimension $2$ in $V$). For each $V$, cutting along $M$ provides an $n$-dimensional hyperbolic manifold $W$ with geodesic boundary $M$. The existence of the geodesic submanifold $P$ in $M$ implies that $\rho_M$ is deformable in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ by bending.
This point of view provides another interpretation of Theorem \[thm:intromain\]. It says that, for $n\ge 4$, when a closed $(n-1)$-dimensional hyperbolic manifold $M$ is the boundary of a compact $n$-dimensional manifold $W$, even when $M$ is deformable in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$, this deformation does not extend to $W$.
In dimension $2$ it is easy to construct compact hyperbolic surfaces of any genus with totally geodesic boundary. Thurston’s hyperbolization theorem for Haken manifolds implies that any compact, irreducible $3$-manifold with genus $g \geq 2$ boundary has a hyperbolic structure with totally geodesic boundary as long as it contains no nontrivial annuli or tori. In higher dimensions it is much more difficult to find explicit examples. It follows from a theorem of Wang [@Wa] that in any fixed dimension $n > 3$ the number $N(V)$ of (isometry classes of) compact hyperbolic $n$-manifolds with nonempty totally geodesic boundary and volume less than $V$ is finite. (It becomes infinite in dimension $3$.) Nonetheless, $N(V)$ grows faster than $V^{aV}$ for some $a>0$. In particular, in [@Rat2] Ratcliffe and Tschantz construct an explicit infinite family of compact hyperbolic $4$-manifolds with nonempty totally geodesic boundary that provides such a lower bound. For general $n\geq 4$ it is shown in [@BGLM] that the number of *closed* hyperbolic $n$-manifolds of volume at most $V$ grows at least this fast. But their examples (which are covering spaces of a single example) all contain embedded totally geodesic hypersurfaces along which one can cut to obtain a sufficient number of manifolds with totally geodesic boundary. (One needs also to bound the number of pieces that become isometric after cutting. We will not provide that argument here.) Thus, while the structures covered by Theorem \[thm:intromain\] are not as prevelant as those in lower dimensions, they are, nonetheless, numerous.
A compact hyperbolic $n$-manifold $W$ with totally geodesic boundary has an extension $X$ which is an infinite volume, complete hyperbolic manifold without boundary. Since $\pi_1 X = \pi_1 W$ and the images of the holonomy representations of $X$ and $W$ are the same, Theorem \[thm:intromain\] can be rephrased in terms of the local rigidity of the holonomy representation of $X$. This, in turn, can be described directly in Riemannian geometric terms.
The hyperbolic structure on $X$ is convex cocompact, which means that it contains an $n$-dimensional convex, compact submanifold $C$ that is homotopy equivalent to $X$. (This concept is discussed in more detail in Section \[sec:hyperbolic preliminaries\].) When $X$ is the extension of a hyperbolic manifold $W$ with geodesic boundary, $C$ can be taken to equal $W$. Conversely, if $X$ has such an $n$-dimensional convex, compact submanifold $C$ with totally geodesic boundary, it is the extension of $C$. We say that $X$ has *Fuchsian ends* in this case. Each component of the complement $X-C$ is diffeomorphic to the product $\partial C \times (0,\infty)$. The term “Fuchsian end” reflects the fact that each component of $\partial C$ is an $(n-1)$-dimensional hyperbolic manifold and hence is determined by a discrete group of isometries of ${\ensuremath{\mathbb{H} } }^{n-1}.$ (This term also applies to the case when $X$ is actually Fuchsian, but we are ruling out this case by the assumption that $C$ is $n$-dimensional.) Thus we have a $1{-}1$ correspondence between (non-Fuchsian) hyperbolic manifolds with Fuchsian ends and compact hyperbolic manifolds with totally geodesic boundary.
If $X$ is deformable, any nearby structure will still be convex cocompact. When $n \geq 3$ any nontrival deformation cannot still have Fuchsian ends (by local rigidity of the double of $C$). If $X$ is $3$-dimensional, it has a large deformation space, parametrized by conformal structures on surfaces at infinity (which are diffeomorphic to $\partial C$). However, for $n >3$, quite the opposite is true and there are no nearby non-isometric hyperbolic structures on $X$. This is the Riemannian version of Theorem \[thm:intromain\] which we state as a separate theorem below. In Section \[sec:hyperbolic preliminaries\] the relationship between the two theorems will be discussed in more detail.
\[thm:intro\] Let $X$ be an infinite volume convex cocompact complete hyperbolic $n$-manifold without boundary. Assume that $X$ has Fuchsian ends but is not Fuchsian. If $n>3$ then $X$ has no nontrivial deformations. Specifically, if $g_n$ is a sequence of complete hyperbolic metrics on $X$ converging to the given metric in the compact-$\mathcal{C}^\infty$ topology, then for sufficiently large $n$ the Riemannian manifolds $(X,g_n)$ are all isometric.
Theorem \[thm:intro\] can be viewed as a generalization of the local rigidity theorems of Calabi and Weil to the infinite volume setting. It is natural to ask whether a generalization of the global rigidity theorem of Mostow holds in this case; in other words, does the smooth manifold underlying $X$ support a unique complete hyperbolic structure? (See Question \[quest:Mostow\].) Our methods cannot answer this global question.
Preliminaries {#sec:preliminaries}
=============
Hyperbolic preliminaries {#sec:hyperbolic preliminaries}
------------------------
In this subsection we will define some useful terms from hyperbolic geometry, including some already mentioned in the introduction. We will use the standard convention that a *manifold* has empty boundary, unless explicitly stated otherwise. However, we will occasionally refer to a *manifold without boundary* for the sake of clarity.
A hyperbolic structure on an $n$-dimensional manifold $X$ determines a developing map from the universal cover $\widetilde X$ of $X$ to $n$-dimensional hyperbolic space ${\ensuremath{\mathbb{H} } }^n$. It is equivariant with respect to the action of $\pi_1 X$ on $\widetilde X$ and ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ on ${\ensuremath{\mathbb{H} } }^n$, inducing a representation $\rho: \pi_1 X \to {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$, called the *holonomy representation* of $X$. The developing map is well-defined up to post-multiplication by an element in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$, and the holonomy representation is uniquely determined up to conjugation in ${\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ by the hyperbolic structure. If the hyperbolic structure on $X$ is complete then the developing map is a diffeomorphism and the holonomy representation is faithful. In this case we can identify $\widetilde X$ with ${\ensuremath{\mathbb{H} } }^n$ by choosing a particular developing map. This choice determines $\rho$, identifies $\pi_1 X$ with $\Gamma = \rho(\pi_1 X)$, and induces an isometry between $X$ and ${\ensuremath{\mathbb{H} } }^n/\Gamma$.
Associated to any complete hyperbolic structure on $X$ is its convex core, which we now define. For more information see [@EM].
\[defn:convex core\] Let $X$ be a complete hyperbolic $n$-manifold, inducing an action of $\pi_1 (X)$ on ${\ensuremath{\mathbb{H} } }^n$ by isometries. Let $\mathcal{S}$ be the set of subsets $Y$ of ${\ensuremath{\mathbb{H} } }^n$ satisfying:
- $Y$ is closed and nonempty.
- $Y$ is convex.
- $\gamma \cdot Y = Y$ for all $\gamma \in \pi_1(X)$.
The *convex hull* $\operatorname{Hull}(X) \subseteq {\ensuremath{\mathbb{H} } }^n$ is defined as the intersection $\cap_{Y \in \mathcal{S}} Y$. The quotient $$C(X) := \operatorname{Hull}(X) / \pi_1(X) \subseteq X$$ is the convex core of $X$.
The convex hull of $X$ is convex, and thus contractible. Therefore $C(X)$ is homotopy equivalent to $X$. If $X$ has finite volume then $\operatorname{Hull}(X) = {\ensuremath{\mathbb{H} } }^n$ and $C(X)=X$. When $X$ has infinite volume, the convex core is typically an $n$-dimensional submanifold with boundary, but, at one extreme, it can be all of $X$, or, at the other extreme, a lower dimensional submanifold. In all cases $X$ is homeomorphic to the gluing $$C_{\varepsilon}(X) \cup_{\partial C_{\varepsilon}(X)} \left( \partial C_{\varepsilon} (X) \times [0,\infty) \right),$$ where $C_{\varepsilon}(X)$ is a closed $\varepsilon$-neighborhood of $C(X)$ in $X$.
\[defn:convex cocompact\] If the convex core of $X$ is compact then $X$ is *convex cocompact*.
Although we will not need it in this paper, for cultural context we mention that $X$ is called geometrically finite if $C_{\varepsilon}(X)$ has finite volume for some $\varepsilon>0$. This allows the image of the holonomy representation to contain parabolic isometries, which is not possible when $X$ is convex cocompact.
\[defn:Fuchsian\] A complete hyperbolic $n$-manifold $X$ is *Fuchsian* if it has a holonomy representation $\pi_1(X) \to {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ whose image $\Gamma$ lies in a subgroup ${\ensuremath{\operatorname{Isom} } }(\mathbb{H}^m)$ for $m<n$. We assume that the quotient ${\ensuremath{\mathbb{H} } }^m/\Gamma$ has finite volume.
A Fuchsian manifold $X$ comes from an $m$-dimensional hyperbolic manifold $M = {\ensuremath{\mathbb{H} } }^m/\Gamma_M$ whose holonomy image $\rho_M(\pi_1(M)) = \Gamma_M$ is equal to $\Gamma$. The convex core $C(X) \subset X$ is a totally geodesic submanifold of dimension $m$, isometric to $M$, which uniquely determines the geometry of $X$. We will only be considering the case when $m = n-1$. Then $C(X)$ has codimension $1$ and $X$ is isometric to the warped product metric on $C(X) \times \mathbb{R}$ given by $$\label{eqn:warped}
(\cosh^2 t) g \oplus dt^2,$$ where $g$ is the metric on $C(X)$.
The next definition is used to describe hyperbolic structures whose ends are isometric to a Fuchsian structure with codimension $1$ convex core.
\[defn:Fuchsian ends\] Let $X$ be a complete hyperbolic $n$-manifold without boundary. Then $X$ *has Fuchsian ends* and is not Fuchsian if its convex core $C(X)$ is an $n$-manifold with totally geodesic boundary, and some $\varepsilon$-tubular neighborhood of $C(X)$ has finite volume.
In general, this terminology allows $C(X)$ to have finite volume ends. We will avoid this possibility by assuming that $X$ is convex compact. Each end of $X$ is geometrically determined by a component $S$ of $\partial C(X)$. The end beginning at $S$ is isometric to $S \times [0,\infty)$ equipped with the Fuchsian warped product metric (\[eqn:warped\]).
Given a hyperbolic manifold $W$ with totally geodesic boundary, we can form a manifold $X$ with Fuchsian ends and empty boundary by simply isometrically gluing ends $\partial W \times [0,\infty)$ onto the boundary of $W$ equipped with the Fuchsian warped product metric (\[eqn:warped\]). The convex core of $X$ will be $W$. This relationship is canonical: a hyperbolic manifold with nonempty totally geodesic boundary determines a unique manifold with Fuchsian ends and vice versa.
Given this tight relationship, it is natural to ask whether it’s possible to phrase Theorem \[thm:intro\] in terms of deformations of compact hyperbolic manifolds with totally geodesic boundary. This is possible, but requires some care. Let $X$ be a convex cocompact hyperbolic manifold with Fuchsian ends, dimension $n \ge 4$, and $n$-dimensional convex core $C$. Then simply replacing $X$ with $C$ everywhere in Theorem \[thm:intro\] produces a false statement. There are many hyperbolic metrics on the underlying manifold with boundary which is diffeomorphic to $C$. One could, for example, take a closed $\varepsilon$-neighborhood $C_\varepsilon$ of $C$ inside $X$. The manifold $C_\varepsilon$ would be diffeomorphic but not isometric to $C$ because the boundary of $C_\varepsilon$ would be strictly convex rather than totally geodesic. However, it is clear that such a deformation is, in some sense, trivial. One can rule out such trivial deformations, yielding the following restatement of Theorem \[thm:intro\].
\[thm:intro with boundary\] Let $C$ be a compact hyperbolic $n$-manifold with nonempty totally geodesic boundary. Let $X$ be the manifold with Fuchsian ends determined by $C$. If $n>3$ then $C$ has no nontrivial deformations in the following sense: if $g_n$ is a sequence of hyperbolic metrics on $C$ converging to the given metric in the $\mathcal{C}^\infty$ topology, then for sufficiently large $n$ the Riemannian manifolds $(C,g_n)$ admit isometric embeddings into $X$, and the boundaries of the images converge in the $\mathcal{C}^{\infty}$ topology to $\partial C$.
The proof of Theorem \[thm:intro with boundary\] uses Theorem \[thm:intro\], the fact that the manifolds $(C,g_n)$ admit analytic thickenings to open manifolds [@CEG], and that $\mathcal{C}^\infty$ convergence guarantees the boundaries are nearly totally geodesic. As it is not central to this paper, the details are omitted.
Algebraic preliminaries {#sec:algebraic preliminaries}
-----------------------
The holonomy representation leads to an algebraic description of the deformation theory of hyperbolic structures. In this subsection we briefly outline this algebraic structure. For more detailed presentations of this material see [@We; @We2; @Rag; @Gold1; @HK1; @HK3].
Let $X$ be a complete hyperbolic $n$-manifold with universal cover $\widetilde X$. Let $G$ denote the group of isometries of ${\ensuremath{\mathbb{H} } }^n$. Choose an isometric identification of $\widetilde X$ with ${\ensuremath{\mathbb{H} } }^n$ and let $\rho: \pi_1(X) \to G$ be the associated holonomy representation. Denote by $\Gamma$ the image group $\rho(\pi_1(X))$ so that we can identify $X$ with the quotient ${\ensuremath{\mathbb{H} } }^n/\Gamma$. The representation $\rho$ is a point in the space of representations ${\ensuremath{\operatorname{Hom} } }(\pi_1(X),G)$, which is equipped with the compact-open topology. A sequence of hyperbolic metrics $g_n$ on $X$ determines a sequence of holonomy representations, each of which is only determined up to conjugacy. However, it is a standard fact that if the metrics converge smoothly then it is possible to choose a convergent sequence of holonomy representations.
\[lem:Riem2holonomy\] Suppose $g_n$ is a sequence of hyperbolic metrics on $X$ converging in the compact-$\mathcal{C}^\infty$ topology to the initial metric $g$. Then it is possible to choose holonomy representations $\rho_n$ of $(X,g_n)$ such that $\rho_n \to \rho$ in ${\ensuremath{\operatorname{Hom} } }(\pi_1(X),G)$.
A hyperbolic manifold $X$ is *locally rigid* if a holonomy representation $\rho$ for $X$ has a neighborhood $U \subseteq {\ensuremath{\operatorname{Hom} } }(\pi_1(X),G)$ such that any representation in $U$ is conjugate to $\rho$.
By Lemma \[lem:Riem2holonomy\], if $X$ is locally rigid then $X$ is locally rigid in the sense of Theorem \[thm:intro\]. The converse is not generally true. However, it is true when $X$ is compact or, more generally, when it is convex cocompact. This converse direction will not be used here, and is not pursued further.
Suppose $\rho_t$ is a smooth path in ${\ensuremath{\operatorname{Hom} } }(\pi_1(X),G)$ such that $\rho_0 = \rho$. Then for each $\gamma \in \pi_1(X)$, $\rho_t(\gamma)$ is a smooth path in $G$ and the derivative $$\frac{d}{dt} \left( \rho_t (\gamma) \cdot \rho(\gamma)^{-1}\right)$$ at $t=0$ determines an element in the Lie algebra $\mathfrak{g}$, where the latter is identified with the tangent space to $G$ at the identity. This determines a set map $z: \pi_1(X) \to \mathfrak{g}$. The fact that $\rho_t$ is a homomorphism implies that $z(\gamma_1 \gamma_2) = z(\gamma_1) + \rho(\gamma_1) z(\gamma_2) \rho(\gamma_1)^{-1}.$ Define a $\pi_1(X)$-action on $\mathfrak{g}$ by $\gamma \cdot V := {\ensuremath{\operatorname{Ad} } }(\rho(\gamma))\cdot V$, where ${\ensuremath{\operatorname{Ad} } }$ denotes the adjoint action. The previous equation becomes $z(\gamma_1 \gamma_2) = z(\gamma_1) + \gamma_1 \cdot z(\gamma_2).$ A map satisfying this equation is defined to be a *$1$-cocycle* for the group cohomology of $\pi_1(X)$ with coefficients in $\mathfrak{g}$ twisted by the above $\pi_1(X)$-action. The group of such cocycles is denoted by $ Z^1(\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho} ).$
Given a smooth path $g_t \in G$ so that $g_0 = e$ one can define a path of representations $\rho_t = g_t \rho g_t^{-1}$. The associated cocycle $z$ in this case equals $z(\gamma) = V - \gamma \cdot V$ where $V \in \mathfrak{g}$ is the vector tangent to $g_t$ at $t=0$. The right hand side of this equation defines a cocycle for any $V \in \mathfrak{g}$. The set of such cocycles, called coboundaries, is denoted by $ B^1(\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho} ).$
The *cohomology group* $H^1 (\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho})$ is defined to be the quotient $$\frac{Z^1 (\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho})}{B^1(\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho} )}.$$ This quotient group can be interpreted as an algebraic description of the tangent space at $\rho$ to the representation space ${\ensuremath{\operatorname{Hom} } }(\pi_1(X),G)$ modulo conjugation. Under certain smoothness hypotheses it is isomorphic to the actual tangent space of this quotient space. In any case, it is referred to as the space of *infinitesimal deformations* of the hyperbolic structure on $X$. This leads us naturally to the next definition.
\[defn:infrig\] A hyperbolic manifold $X$ is *infinitesimally rigid* if $$H^1 (\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho}) = 0.$$
Infinitesimal rigidity states that any path $\rho_t$ through the holonomy representation $\rho$ agrees to first order at $\rho$ with a geometrically trivial path consisting entirely of conjugate representations. An application of the implicit function theorem proves the following lemma, known as Weil’s lemma.
\[lem:Weil\][@We2] If $X$ is infinitesimally rigid then it is locally rigid.
With this we are ready to state the main theorem of this paper.
\[thm:mainFuchs\] Let $X$ be a convex cocompact hyperbolic $n$-manifold without boundary which has Fuchsian ends but is not Fuchsian. If $n>3$ then $X$ is infinitesimally rigid.
Theorem \[thm:mainFuchs\] implies Theorem \[thm:intro\] of the introduction, but the converse implication need not hold. There do exist locally rigid manifolds that are not infinitesimally rigid [@GoldMill].
Despite its simple description the cohomology group $H^1 (\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }h})$ is extremely difficult to compute. Given an explicit presentation of $\pi_1(X)$ and a formula for $\rho$, computations are occasionally possible, at least with a computer. However, a direct algebraic computation for any large class of examples is generally infeasible. We will need to use analytic and geometric tools to describe this group.
Let $X$ be a complete hyperbolic manifold which has been identified with a quotient ${\ensuremath{\mathbb{H} } }^n/\Gamma$ where $\Gamma = \rho(\pi_1(X))$ is the image in $G$ of a particular holonomy representation $\rho$. Consider the product $\widetilde {\ensuremath{\mathcal{E} } }:= \mathbb{H}^n \times \mathfrak{g}$ equipped with the diagonal $\Gamma$-action. The quotient is a $\mathfrak{g}$-bundle ${\ensuremath{\mathcal{E} } }$ over $X$. Note that the $\Gamma$-action on $\widetilde {\ensuremath{\mathcal{E} } }$ takes a constant section to a constant section. Thus constant sections define a flat bundle structure on $\widetilde {\ensuremath{\mathcal{E} } }$ that descends to ${\ensuremath{\mathcal{E} } }$.
The group cohomology $H^1 (\pi_1(X); \mathfrak{g}_{{\ensuremath{\operatorname{Ad} } }\rho})$ is isomorphic to the singular cohomology group $H^1(X;{\ensuremath{\mathcal{E} } })$. (This is true for any connected CW complex with fundamental group equal to $\pi_1(X)$.) In turn, an extension of the usual deRham theorem shows that singular cohomology equals deRham cohomology with values in ${\ensuremath{\mathcal{E} } }$. The latter can be described as follows. Consider the space $\Omega^{*}({\ensuremath{\mathbb{H} } }^n; \widetilde {\ensuremath{\mathcal{E} } })$ of $\widetilde {\ensuremath{\mathcal{E} } }$-valued differential forms on ${\ensuremath{\mathbb{H} } }^n$. Given a basis $\{B_i \}$ for $\mathfrak{g}$, such a $k$-form is given concretely as a finite linear combination $\sum B_i \otimes \theta_i$, where each $\theta_i$ is a real-valued differential $k$-form. Then $\Gamma$ acts on $\Omega^{*}({\ensuremath{\mathbb{H} } }^n; \widetilde {\ensuremath{\mathcal{E} } })$ via the adjoint action on $\widetilde {\ensuremath{\mathcal{E} } }$ and the usual action on differential forms. The $\widetilde {\ensuremath{\mathcal{E} } }$-valued forms fixed by this action form the space $\Omega^{*}(X;{\ensuremath{\mathcal{E} } })$ of ${\ensuremath{\mathcal{E} } }$-valued differential forms on $X$. Using the constant basis, the differential $d_{\ensuremath{\mathcal{E} } }$ can be defined on $\Omega^{*}({\ensuremath{\mathbb{H} } }^n;\widetilde {\ensuremath{\mathcal{E} } })$ as simply $$d_{\ensuremath{\mathcal{E} } }\left( \sum B_i \otimes \theta_i \right) := \sum B_i \otimes d\theta_i .$$ On $\Omega^{*}(X;{\ensuremath{\mathcal{E} } })$ the differential $d_{\ensuremath{\mathcal{E} } }$ is defined by applying the above definition on small open neighborhoods of $X$. With this differential on the space of forms $\Omega^{*}(X;{\ensuremath{\mathcal{E} } })$, we can define the deRham cohomology groups $H_{dR}^{*} (X;{\ensuremath{\mathcal{E} } })$ as the closed ${\ensuremath{\mathcal{E} } }$-valued forms modulo the exact ones.
Of course, deRham cohomology is also difficult to compute directly. In order to understand these groups, it will be necessary to use Hodge theory. We need a metric on the fibers of ${\ensuremath{\mathcal{E} } }$ to define the notion of a harmonic form in $\Omega^{*}(X;{\ensuremath{\mathcal{E} } })$. For this a little terminology is required. Represent $G= {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$ explicitly as the index $2$ subgroup of $\operatorname{O}(1,n)$ preserving the upper hyperboloid and let $K$ denote the maximal compact subgroup $\operatorname{O}(n)$ of $G$ fixing $(1,0,\ldots,0)$, so that ${\ensuremath{\mathbb{H} } }^n = G/K$. Then consider the $1$-parameter family $g_t$ in $G$ given by the matrices of the form $$\begin{pmatrix}
A & 0 \\
0 & I
\end{pmatrix}, \quad \text{where} \quad A =
\begin{pmatrix}
\cosh t & \sinh t \\
\sinh t & \cosh t
\end{pmatrix}$$ and $I$ is the $(n-1)$-dimensional identity matrix. Taking the derivative at $t = 0$ determines a tangent vector at the identity element $e \in G$, hence an element of the Lie algebra $V \in \mathfrak{g}.$ The family $g_t$ acts on ${\ensuremath{\mathbb{H} } }^n$ by isometries, translating along a geodesic through the basepoint $x_0 \in {\ensuremath{\mathbb{H} } }^n$ which is the image of $e \in G$ under the projection to ${\ensuremath{\mathbb{H} } }^n/K$. (Viewing ${\ensuremath{\mathbb{H} } }^n$ as the upper hyperboloid in ${\ensuremath{\mathbb{R} } }^{1,n}$, $x_0$ is the point $(1,0,\ldots,0)$.) Taking the derivative at $t=0$ determines a tangent vector based at $x_0.$ Conjugation by an element $\gamma \in G$ determines a family $\gamma g_t \gamma^{-1}$ that acts on ${\ensuremath{\mathbb{H} } }^n$ by translation along a geodesic through $\gamma x_0$. The Lie algebra element obtained by taking the derivative of the conjugated family at $t=0$ equals ${\ensuremath{\operatorname{Ad} } }(\gamma)\cdot V$. Taking the derivative of the conjugated action on ${\ensuremath{\mathbb{H} } }^n$ at $t=0$ determines a unit tangent vector in $T_{\gamma x_0} {\ensuremath{\mathbb{H} } }^n$. In this way a unit vector at a point $p \in {\ensuremath{\mathbb{H} } }^n$ uniquely determines a Lie algebra element, and we refer to that element as the *infinitesimal translation* at $p$ in that direction. We view it as an element of the $\mathfrak{g}$-bundle $\widetilde {\ensuremath{\mathcal{E} } }$ over ${\ensuremath{\mathbb{H} } }^n.$
Given an orthonormal frame $\{ e_i \}$ on an open patch of ${\ensuremath{\mathbb{H} } }^n$, we define the corresponding sections $\{ E_i \}$ of the $\mathfrak{g}$-bundle $\widetilde {\ensuremath{\mathcal{E} } }$, where $E_i (p)$ is the infinitesimal translation at $p$ in the direction $e_i (p)$. Define $R_{ij}(p) := \left[ E_i (p) , E_j (p) \right]$ (for $i \neq j$) to be a unit infinitesimal rotation at $p$. It represents the skew-symmetric mapping of the tangent space at $p$ that takes $e_j$ to $e_i$, reflecting the fact that ${\ensuremath{\mathbb{H} } }^n$ has negative curvature. It annihilates all vectors orthogonal to the plane spanned by $e_i$ and $e_j.$ Define a positive metric $\langle \cdot, \cdot \rangle$ on the bundle $\widetilde {\ensuremath{\mathcal{E} } }= {\ensuremath{\mathbb{H} } }^n \times \mathfrak{g}$ by taking $\{E_i(p), R_{ij}(p)\}_{i \neq j}$ as an orthonormal basis of the fiber over $p$. With this metric, the action of $G$ on $\widetilde {\ensuremath{\mathcal{E} } }$ is isometric, and descends to ${\ensuremath{\mathcal{E} } }$ over $X$. Equipped with this metric on ${\ensuremath{\mathcal{E} } }$ we can define a pairing on $\Omega^{*}(X;{\ensuremath{\mathcal{E} } })$ by $$\langle \alpha, \beta \rangle := \int_X \alpha \wedge {*} \beta,$$ where the coefficients in ${\ensuremath{\mathcal{E} } }$ of $\alpha$ and $\beta$ are paired to produce a real number. Use this pairing to define an adjoint $\delta_{\ensuremath{\mathcal{E} } }: \Omega^{k+1}(X;{\ensuremath{\mathcal{E} } }) \to \Omega^k (X;{\ensuremath{\mathcal{E} } })$ to $d_{\ensuremath{\mathcal{E} } }$ on forms with compact support.
It is important to note that none of the sections $p \mapsto E_i (p)$ and $p \mapsto R_{ij}(p)$ are flat. They never represent a locally constant element in $\mathfrak{g}$. This can be seen geometrically from the fact that if the $1$-parameter subgroup determined by an infinitesimal rotation at a point $p$ does not fix $q$, then it cannot correspond to the same Lie algebra element as an infinitesimal rotation at $q$. Similarly, since an infinitesimal translation at $p$ exponentiates to translation along a geodesic it can only correspond to infinitesimal translations at other points along that geodesic; the corresponding Lie algebra elements cannot be constant on an open set.
However, it is possible to relate the flat derivative $d_{\ensuremath{\mathcal{E} } }$ and its adjoint to the covariant derivative (and its adjoint) determined by the hyperbolic metric on ${\ensuremath{\mathbb{H} } }^n.$ The covariant derivative has the advantange that it can be understood in local differential geometric terms. The difference between the flat and the geometric derivative is a purely algebraic operator.
It is convenient to describe this relationship in terms of a local orthonormal frame $\{ e_i \}$ of $X$ and the resulting orthonormal basis $\{ E_i, R_{ij} \}$ of ${\ensuremath{\mathcal{E} } }$. Let $\{ \omega_i \}$ be the dual coframe of $X$. Then [@Wu Ch.6][@MatMur; @HK1] $$\begin{aligned}
d_{\ensuremath{\mathcal{E} } }(V \otimes \alpha) & = \sum_j \left( \omega_j \wedge \nabla_{e_j} (V \otimes \alpha ) +
\omega_j \wedge ( [E_j,V] \otimes \alpha ) \right)\\
\delta_{\ensuremath{\mathcal{E} } }(V \otimes \alpha) & = - \sum_j i(e_j) \left( \nabla_{e_j} ( V \otimes \alpha) - [E_j,V] \otimes \alpha \right).\end{aligned}$$
To make sense of these equations we need to explain the term $\nabla_{e_j} (V \otimes \alpha)$, where $V$ is a local section of ${\ensuremath{\mathcal{E} } }$ and $\alpha$ is a differential form on $X$. It equals $(\nabla_{e_j} V) \otimes \alpha + V \otimes \nabla_{e_j} \alpha,$ in other words the connection acts on both $V$ and $\alpha$. Here $\nabla_{e_j} \alpha$ denotes the covariant derivative with respect to the Levi-Civita connection associated to the hyperbolic metric. Similarly, the connection $\nabla$ is defined on ${\ensuremath{\mathcal{E} } }$ using the Levi-Civita connection as follows. First, $\nabla_{e_j} E_k$ is the Lie algebra element corresponding to $\nabla_{e_j} e_k$. (Compute $\nabla_{e_j} e_k$ and then capitalize all the $e$’s.) Next, $\nabla_{e_k} R_{ij}$ is defined by considering $R_{ij}$ as a section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TX)$, the bundle of skew symmetric endomorphisms of $TX$, and using the Levi-Civita connection to differentiate this section. The result will again be skew symmetric, and one converts it back into a linear combination of infinitesimal rotations. In short, if one writes a local section $V$ as a linear combination of the $\{E_i\}$ and of the $\{R_{ij}\}$, it can be viewed as a pair $(u,\widetilde u)$ consisting of a section $u$ of $TX$ and $\widetilde u$ of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TX)$, and $\nabla V = (\nabla u, \nabla \widetilde u)$ where the Levi-Civita connection is used on these bundles. Finally, $i(e_j)$ indicates left contraction of the resulting form along the vector $e_j$ (eg. $i(v) \alpha \wedge \beta = \alpha(v) \cdot \beta - \beta(v) \cdot \alpha$).
It is useful to express this decomposition as $d_{\ensuremath{\mathcal{E} } }= D + T$ and $\delta_{\ensuremath{\mathcal{E} } }= D^{*} + T^{*}$, where $$\label{DTeqns} \begin{split} D & = \sum_j \omega_j \wedge \nabla_{e_j} \\
T & = \sum_j \omega_j \wedge {\ensuremath{\operatorname{ad} } }(E_j) \\
D^{*} & = - \sum_j i(e_j) \nabla_{e_j} \\
T^{*} & = \sum_j i(e_j) {\ensuremath{\operatorname{ad} } }(E_j),
\end{split}$$ and ${\ensuremath{\operatorname{ad} } }(E_j) (V \otimes \alpha) := [E_j,V]\otimes \alpha$. Note that $D$ and $D^{*}$ take infinitesimal translations (resp. rotations) to infinitesimal translations (resp. rotations), while $T$ and $T^{*}$ take infinitesimal translations (resp. rotations) to infinitesimal rotations (resp. translations).
In particular, if we write a local section $s$ of ${\ensuremath{\mathcal{E} } }$ as a pair $(u,\widetilde u)$ consisting of a local section of $TX$ (i.e., a vector field) and of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TX)$, then both $Du$ and $T \tilde u$ are $1$-forms with values in the infinitesimal translations of $X$ and so can be thought of as local sections of $TX \otimes T^{*} X$. Using the identification of $TX \otimes T^{*} X$ with ${\ensuremath{\operatorname{Hom} } }(TX,TX)$, $T \widetilde u$ is skew-symmetric and corresponds to $-\widetilde u$. Of particular importance are sections of ${\ensuremath{\mathcal{E} } }$ of where $T \widetilde u$ cancels out the skew-symmetric part of $Du$.
\[defn:canon\_lift\] A local section $s = (u, \widetilde u)$ is a *canonical lift* if the skew symmetric part of $D u$ equals $- T \widetilde u$. A canonical lift is clearly determined by its vector field part $u$, and will be called the canonical lift of $u$.
To help the reader digest this concept, we give two equivalent definitions. First, the local section $s = (u, \widetilde u)$ is a canonical lift if and only if the skew symmetric part of $D u$ equals $\widetilde u$ viewed as a local section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TX)$. Second, if we write $$\label{eqn:canon_lift}
d_{\ensuremath{\mathcal{E} } }s = \left( \sum b_{ij} E_i \otimes \omega_j \right) +
\left( \sum c_{k \ell m} R_{k \ell} \otimes \omega_m \right).$$ then $s$ is a canonical lift if and only if $b_{ij} = b_{ji}$. This implies a locally constant section is a canonical lift, because its differential is zero. We leave the proof of these equivalences to the reader.
On a contractible neighborhood a closed $1$-form $\omega$ (with coefficients in ${\ensuremath{\mathcal{E} } }$) has a section $s$ satisfying $d_{\ensuremath{\mathcal{E} } }s = \omega$. This section is unique up to adding a locally constant section, implying that being a canonical lift is a property of the $1$-form $\omega$. If, on any contractible neighborhood, there is always a canonical lift $s$ satisfying $d_{\ensuremath{\mathcal{E} } }s = \omega$, then we will say that $\omega$ locally admits a canonical lift. Using equation (\[eqn:canon\_lift\]) it follows that $\omega$ admits a canonical lift if and only if $b_{ij} = b_{ji}$.
\[lem:canon\_lift\][@HK1 Prop.2.3(c)] Given (globally defined) closed $1$-form $\omega \in \Omega^1 (X,{\ensuremath{\mathcal{E} } })$, there is a global section $s$ such that $\omega + d_{\ensuremath{\mathcal{E} } }s$ locally admits a canonical lift.
Define the (flat) Laplacian on ${\ensuremath{\mathcal{E} } }$ as $\lap_{\ensuremath{\mathcal{E} } }= d_{\ensuremath{\mathcal{E} } }\delta_{\ensuremath{\mathcal{E} } }+ \delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }.$ Then $$\lap_{\ensuremath{\mathcal{E} } }= (D D^{*} + D^{*} D + T T^{*} + T^{*} T) + (D T^{*} + T D^{*} + D^{*} T + T^{*} D).$$ Conveniently, the second term $(D T^{*} + T D^{*} + D^{*} T + T^{*} D)$ is always zero [@MatMur]. In particular, the Laplacian preserves the decomposition into infinitesimal rotations and translations. Thus we can write $$\lap_{\ensuremath{\mathcal{E} } }= \lap_D + H,$$ where $\lap_D = D D^{*} + D^{*} D$ and $H= T T^{*} + T^{*} T$.
The idea of the proof of Theorem \[thm:mainFuchs\] is to represent every cohomology class in $H^1(X;{\ensuremath{\mathcal{E} } })$ by an ${\ensuremath{\mathcal{E} } }$-valued harmonic $1$-form $\omega$, one where $d_{\ensuremath{\mathcal{E} } }\omega = 0 = \delta_{\ensuremath{\mathcal{E} } }\omega$, hence $\lap_{\ensuremath{\mathcal{E} } }= 0$, and then show that such a form must equal $0$. We will, in fact, not do this for the complete manifold $X$ with Fuchsian ends, but rather for its convex core which is a hyperbolic manifold $W$ with totally geodesic boundary. On $W$ we will represent each class in $H^1(W; {\ensuremath{\mathcal{E} } })$ by a harmonic form satisfying certain boundary conditions and show that such a form must equal $0$. Since $H^1(X;{\ensuremath{\mathcal{E} } }) = H^1(W;{\ensuremath{\mathcal{E} } })$, this will prove Theorem \[thm:mainFuchs\].
We now state a key formula, called the formula, which will motivate the boundary conditions on $W$ that we will choose. It is obtained via integration by parts.
\[prop:byparts\][@HK1 Prop.1.3] Let $W$ be a compact oriented hyperbolic manifold with boundary and let $\omega \in \Omega^1 (W;{\ensuremath{\mathcal{E} } })$ be a smooth ${\ensuremath{\mathcal{E} } }$-valued $1$-form. Then $$\label{Weitzformula}
(d_{\ensuremath{\mathcal{E} } }\omega, d_{\ensuremath{\mathcal{E} } }\omega) + (\delta_{\ensuremath{\mathcal{E} } }\omega, \delta_{\ensuremath{\mathcal{E} } }\omega) = (D \omega, D \omega) + (D^{*} \omega, D^{*} \omega) + (H \omega, \omega) + B,$$ where $B$ denotes the boundary term $$\label{Weitzbdry}
B = - \int_{\partial W} \left( {*}T\omega \wedge \omega + T^{*} \omega \wedge {*} \omega \right) .$$
If $\omega$ is harmonic, which we always take to mean that it is closed and co-closed ($d_{\ensuremath{\mathcal{E} } }\omega = 0 = \delta_{\ensuremath{\mathcal{E} } }\omega$), the left hand side of the formula (\[Weitzformula\]) equals $0.$ When $W$ is at least $3$ dimensional, Weil proved the existence of a $c>0$ such that $$( H \omega, \omega) \ge c \, (\omega,\omega)$$ for all $1$-forms $\omega \in \Omega^1 (W;{\ensuremath{\mathcal{E} } })$ [@We]. In the case when $W$ is closed, the boundary term $B$ is trivial and this formula implies that any harmonic ${\ensuremath{\mathcal{E} } }$-valued $1$-form must be $0.$ By the Hodge theorem, this means that $H^1(W;{\ensuremath{\mathcal{E} } }) = 0$ which proves infinitesimal (hence local) rigidity in the closed case for these dimensions.
Similarly, when $W$ has boundary, if one can show that any class in $H^1(W;{\ensuremath{\mathcal{E} } })$ has a harmonic representative where the the boundary term $B$ of (\[Weitzbdry\]) is non-negative, this will again prove that $H^1(W;{\ensuremath{\mathcal{E} } }) = 0.$ This is the basis of the proof of our main theorem, which we now state in the form it will be proved.
\[thm:main\] Let $W$ be a compact hyperbolic $n$-manifold with totally geodesic boundary. If $n>3$ then $H^1(W;{\ensuremath{\mathcal{E} } }) = 0$ so that $W$ is infinitesimally rigid. Hence, $W$ is locally rigid.
Harmonic forms on the boundary {#sec:bdryforms}
==============================
In this section we will begin the proof of Theorem \[thm:main\].
Let $W$ be a compact hyperbolic manifold with totally geodesic boundary and dimension $n \geq 4$. Equivalently, let $X$ be a complete hyperbolic $n$-manifold which has Fuchsian ends and is not Fuchsian, and let $W$ be its convex core. Define $M := \partial W$. The goal is to prove $H^1(W;{\ensuremath{\mathcal{E} } }) \cong H^1(X;{\ensuremath{\mathcal{E} } }) = 0$. The idea is to consider an infinitesimal deformation of the hyperbolic structure of $W$, described as a deRham cohomology class in $H^1(W;{\ensuremath{\mathcal{E} } })$. We find a harmonic representative $\omega$ of this cohomology class which is an ${\ensuremath{\mathcal{E} } }$-valued $1$-form on $W$ satisfying certain boundary conditions. We show that these boundary conditions can be chosen so that the boundary term $B$ in the formula from Proposition \[prop:byparts\] will equal $0$. This implies that $\omega = 0$ and hence that $H^1(W;{\ensuremath{\mathcal{E} } })$ is trivial.
There are three main steps in this proof. Each has its own section. First, we analyze the structure of ${\ensuremath{\mathcal{E} } }$-valued harmonic forms on the boundary $M$ and show how to extend them to model harmonic forms in a neighborhood of the boundary. Any class in $H^1(W;{\ensuremath{\mathcal{E} } })$ is represented by an ${\ensuremath{\mathcal{E} } }$-valued $1$-form ${\hat \omega}$ which is closed and has this model harmonic structure near the boundary. Thus, $d_{\ensuremath{\mathcal{E} } }{\hat \omega} = 0$ on all of $W$, and $\delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ is $0$ near the boundary, but $\delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ is not necessarily zero on all of $W$. To find a globally harmonic representative we must solve the problem of finding a section $s$ of ${\ensuremath{\mathcal{E} } }$ satisfying $\delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }s = - \delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ on $W$. Then $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s$ is a harmonic form in the same cohomology class. In the second part we describe a boundary value problem for finding such a section $s$, and show that it is uniquely solvable. In the final section we compute the boundary term $B$ in the formula (\[Weitzformula\]) and show that, for the harmonic form $\omega$ constructed from our boundary value problem, this term is trivial. This implies that $\omega$ itself must be trivial.
Because much of the analysis takes place on the boundary $M$ of the manifold $W$, we will denote the dimension of $M$ by its own letter, $m$. The dimension of $W$ will be then be denoted by $n = m+1$. Abusing notation, it will be convenient to let the letter $n$ also denote a vector normal to $M$.
As sketched above, in this section we will describe harmonic ${\ensuremath{\mathcal{E} } }$-valued $1$-forms on, and in a neighborhood of the boundary $M$ of $W$. For this we choose local frames and coframes that reflect the geometry of the situation. At any point in $M$, choose a local orthonormal frame $\{n,e_1,\ldots,e_m\}$, where $n$ is an outward normal vector, and the $e_i$ are tangent to $M$. Let $\{N,E_1,\ldots,E_m\}$ denote the corresponding infinitesimal translations, which are local sections of ${\ensuremath{\mathcal{E} } }$. There are $m$ rotational generators $R_{ni}$, $i =1, \ldots, m$ corresponding to infinitesimal rotations in planes perpendicular to $M$. They will play a special role in what follows.
As $W^{n}$ has totally geodesic boundary $M^m$, the restriction to a connected component of $M$ induces a representation of its fundamental group into the orientation preserving isometries of ${\ensuremath{\mathbb{H} } }^m$ viewed as a subgroup of the isometries of ${\ensuremath{\mathbb{H} } }^{n}$. For simplicity, we will discuss the case when the boundary is connected. In general, the following argument goes through component by component, identifying a corresponding invariant hyperplane in the universal cover with ${\ensuremath{\mathbb{H} } }^m$ in each case.
The adjoint action of the subgroup preserving ${\ensuremath{\mathbb{H} } }^m$ preserves a direct sum decomposition $\mathfrak{g} = \mathfrak{h} \oplus \mathfrak{s}$ of the Lie algebra $\mathfrak{g}$ of $G = {\ensuremath{\operatorname{Isom} } }({\ensuremath{\mathbb{H} } }^n)$. Here $\mathfrak{h} \cong {\ensuremath{\operatorname{\mathfrak{so} } } }(1,m)$ is the Lie algebra of the subgroup preserving ${\ensuremath{\mathbb{H} } }^m.$ The other factor $\mathfrak{s}$ does not come from a subalgebra of $\mathfrak{g}$ but is preserved under the adjoint action of the isometries of ${\ensuremath{\mathbb{H} } }^m$. The infinitesimal version of this statement is that $[\mathfrak{h}, \mathfrak{s}] \subset \mathfrak{s}$. We will describe this factor in more detail below.
This decomposition of $\mathfrak{g}$ induces a decomposition of the bundle $\widetilde {\ensuremath{\mathcal{E} } }$ restricted to ${\ensuremath{\mathbb{H} } }^m.$ It is an orthogonal decomposition with respect to the metric on this bundle. Because it is invariant under the adjoint action, it descends to a decomposition of ${\ensuremath{\mathcal{E} } }$ over $M$; we denote the corresponding sub-bundles by $\mathcal{H}$ and $\mathcal{S}.$ Furthermore, it induces a direct sum decomposition of the cohomology group $H^1(M;{\ensuremath{\mathcal{E} } }).$ We write these decompositions as ${\ensuremath{\mathcal{E} } }\cong \mathcal{H} \oplus \mathcal{S}$ and $H^1(M;{\ensuremath{\mathcal{E} } }) \cong H^1(M;\mathcal{H}) \oplus H^1(M;\mathcal{S})$ .
The cohomology group $H^1(M;\mathcal{H})$ parametrizes the infinitesimal deformations of the $m$-dimensional hyperbolic structure on $M$. In terms of the basis above, $\mathcal{H}$ is generated by $\{E_i\}$ and $\{R_{jk}\}$ for $i, j, k \in \{ 1,2,\ldots, m\}$. In other words, there is no normal component.
As a vector space, the factor $\mathcal{S}$ is generated by $N$ and $\{R_{ni}\}$. This factor measures how the hyperplane ${\ensuremath{\mathbb{H} } }^m$ is being moved away from itself. At any point in ${\ensuremath{\mathbb{H} } }^m$, the element $N$ corresponds to infinitesimal translation in the direction normal to the hyperplane. An element $R_{ni}$ rotates in the plane spanned by $n$ and $e_i$, rotating the tangent vector $e_i$ in the hyperplane to the normal vector $n$. The subspace in ${\ensuremath{\mathbb{H} } }^m$ orthogonal to $e_i$ is fixed under this rotation. More intrinsically, we can consider the space of all hyperplanes in ${\ensuremath{\mathbb{H} } }^{m+1}$, called $(m+1)$-dimensional deSitter space. Then the fibers $\mathfrak{s}$ of $\mathcal{S}$ can be identified with the tangent space at our particular hyperplane of deSitter space.
With respect to the positive definite metric on $\mathcal{S}$ induced as a subspace of ${\ensuremath{\mathcal{E} } }$, $N$ is orthogonal to the $R_{ni}$. Thus, $\mathcal{S}$ splits (as a metric bundle) into a direct sum $\mathcal{S} \cong N \oplus \mathcal{B}$ (“$\mathcal{B}$" for “bending"). This is a reflection of the following useful observation. We can identify each infinitesimal rotation $R_{ni}$ with the corresponding infinitesimal translation $E_i$. Extending this identification linearly, we can identify the subspace generated by these rotations with the tangent space of ${\ensuremath{\mathbb{H} } }^m$. The subspace $N$ corresponds to the normal bundle of ${\ensuremath{\mathbb{H} } }^m$ inside ${\ensuremath{\mathbb{H} } }^{n}$. These identifications descend to $M$ and the bundle $\mathcal{S}$ can be identified with the tangent bundle of $M$ plus a trivial line bundle. Note that this decomposition of the bundle is not compatible with the local flat structure; in particular, it does not induce a decomposition of $H^1(M;\mathcal{S})$.
Now given an equivalence class $[\hat \omega] \in H^1(W;{\ensuremath{\mathcal{E} } })$ of infinitesimal deformations of the hyperbolic structure on $W$, we can restrict it to the boundary $M$ and consider the resulting class $[{\hat \omega}_M] \in H^1(M;\mathcal{H}) \oplus H^1(M;\mathcal{S}).$ When $M$ is compact and has dimension at least $3$, the first factor is always trivial. We assume that we are in this case. (The case when $M$ has dimension $2$ will be discussed briefly in Section \[sec:conjectures\].)
Consider $[{\hat \omega}_M] \in H^1(M;\mathcal{S}).$ Even when $M$ is compact and $m\geq 3$, it can be nontrivial. Our first step is to describe the harmonic elements in $ H^1(M;\mathcal{S})$.
\[harmonicbends\] Let $M$ be a closed hyperbolic $m$-manifold, $m\geq 2.$ Then $ H^1(M;\mathcal{S})$ is isomorphic to the space of $TM$-valued $1$-forms $B$ on $M$ satisfying the equations $D B = 0 = D^{*} B$ and so that, when viewed as elements of ${\ensuremath{\operatorname{Hom} } }(TM,TM)$, they are symmetric and traceless.
In this theorem the operators $D$, $T$, $D^*$, $T^*$, $d_{\ensuremath{\mathcal{E} } }$, and $\delta_{\ensuremath{\mathcal{E} } }$ are understood to live on $M$, so, in particular, the covariant derivatives and bracket operations in the formulae (\[DTeqns\]) for these operators are only in directions tangent to ${\ensuremath{\mathbb{H} } }^m$. Note that this theorem holds even in the case $m=2$, where the $1$-form $B$ corresponds to a holomorphic quadratic differential, and we can view the general case as being some kind of higher dimensional infinitesimal Schwarzian derivative.
The Hodge Theorem says that we can represent the cohomology class $[{\hat \omega}_M]$ by a unique harmonic element $B$. Harmonic here means closed and co-closed with respect to the flat structure on $\mathcal{S}$ so that $d_{\ensuremath{\mathcal{E} } }B = 0 = \delta_{\ensuremath{\mathcal{E} } }B$. We can decompose $d_{\ensuremath{\mathcal{E} } }$ and $\delta_{\ensuremath{\mathcal{E} } }$ into $D + T$ and $D^{*} + T^{*}$.
The laplacian $\lap_{\ensuremath{\mathcal{E} } }= d_{\ensuremath{\mathcal{E} } }\delta_{\ensuremath{\mathcal{E} } }+ \delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }$ decomposes into $\lap_D + H$ where $\lap_D = D^{*} D + D D^{*}$ and $H = T^{*} T + T T^{*}$. Via integration by parts, both $\lap_D$ and $H$ are non-negative operators. We will now identify the kernel of $H$. To do this, note that $T$ and $T^{*}$ switch the two factors $N$ and $\mathcal{B}$ in the orthogonal decomposition of $\mathcal{S}$ and hence that $T^{*} T$ and $T T^{*}$ preserve this decomposition. It is straightforward to compute that on $N$ the operator $T^{*} T$ is multiplication by $m-1$ and $T T^{*}$ is multiplication by $1$. So the sum is multiplication by $m$. Viewing an element of $\Omega^1(M;\mathcal{B}) \cong \Omega^1(M;TM)$ as an element of ${\ensuremath{\operatorname{Hom} } }(TM,TM)$, one computes that $T T^{*}$ annihilates the traceless part and is multiplication by $m$ on multiples of the identity transformation. Also, $T^{*} T$ is multiplication by $2$ on the skew symmetric part and $0$ on the symmetric part. Thus, the sum of the two operators is multiplication by $m$ on multiples of the identity, multiplication by $2$ on the skew symmetric part and $0$ on the traceless, symmetric part. (This can also be computed directly using formula (4) in [@HK1].) We conclude that $H$ is positive semi-definite with kernel equal to the symmetric, traceless elements of $\Omega^1(M;\mathcal{B})$. Note that such elements are, in fact, in the kernels of both $T$ and $T^{*}$ individually. It follows that the harmonic elements $B$ of $H^1(M;\mathcal{S})$ satisfy $T B = 0 = T^{*} B$ and $DB = 0 = D^{*}B$.
Now, given a harmonic element $B \in \Omega^1(M;\mathcal{S})$, we extend it to a harmonic element in $\Omega^1(M\times {\ensuremath{\mathbb{R} } };{\ensuremath{\mathcal{E} } })$ where $M \times {\ensuremath{\mathbb{R} } }$ is equipped with the Fuchsian hyperbolic structure coming from the inclusion ${\ensuremath{\mathbb{H} } }^m \subset {\ensuremath{\mathbb{H} } }^n$. This extension can be defined by simply pulling back $B$ via the orthogonal projection map $\pi: M \times {\ensuremath{\mathbb{R} } }\to M$ onto the totally geodesic copy of $M = M \times \{0 \}$. A pull-back of a closed form (using the flat coefficients) is always closed. Whether or not a pull-back is co-closed depends on the underlying metrics and the map.
To see that the pull-back is co-closed in our case, we express everything in terms of an orthonormal frame. Given an orthonomal frame and co-frame $\{e_i\}$ and $\{\omega_i\}$ on $M$, the product structure determines tangent vectors and $1$-forms, $\{\pi^{*} e_i\}$ and $\{\pi^{*} \omega_i\}$ along the slices $M \times \{r\}$. The corresponding orthonormal frames and coframes with respect to the Fuchsian hyperbolic metric on $M \times {\ensuremath{\mathbb{R} } }$, denoted again by $\{e_i\}$ and $\{\omega_i\}$, equal ${\ensuremath{\operatorname{sech} } }r \{\pi^{*} e_i\}$ and $\cosh r \{\pi^{*} \omega_i\}$. We denote by $n$ the unit vector orthogonal to the slices $M \times \{r\}$, pointing in the positive direction along ${\ensuremath{\mathbb{R} } }$.
For a fixed value of $r$ the points on the hypersurface $M \times \{r\}$ have constant distance $|r|$ from the totally geodesic $M \times \{0\}$. This implies that it is totally umbillic with constant normal curvature $\tanh r.$ From this we conclude that ${\nabla }_{e_i} n = \tanh r \, e_i$ and that $({\nabla }_{e_i} e_j) \cdot n = - \tanh r \, \delta_{ij}$ at any point of $M \times \{r\}.$ Furthermore, ${\nabla }_n e_i = 0 = {\nabla }_n n.$ The values $c_{ijk} = \langle \nabla_{e_j} e_i , e_k \rangle$ depend on our choice of frame. However, when the frame on $M \times \{r\}$ is determined as above by the frame on $M \times \{0\}$, the value of any $c_{ijk}$ at a point $(x,r)$ equals ${\ensuremath{\operatorname{sech} } }r$ times its value at $(x,0)$. In particular, if we choose a frame near $(x,0)$ using geodesic coordinates, the $c_{ijk}$ will equal $0$ at $(x,r)$ for all values of $r$.
With these observations we can now describe our extended form and show that it has the properties we want.
\[pullback\] Let $B \in \Omega^1(M;\mathcal{B})$ be closed and co-closed where $M$ is an $m$-dimensional hyperbolic manifold, $m\geq 2$. Suppose that we can write $B = \sum b_{ij} R_{ni} \otimes \omega_j $, where the $b_{ij}$ are functions on $M$ which determine a matrix that is symmetric and traceless at each point. With respect to the extended orthonormal frame and coframe on $M \times {\ensuremath{\mathbb{R} } }$, let $\widetilde B = \sum b_{ij} R_{ni} \otimes \omega_j$ and $\widetilde A = -\tanh r \, \sum b_{ij} E_i \otimes \omega_j$. Then the ${\ensuremath{\mathcal{E} } }$-valued $1$-form $\omega_0 = \widetilde B + \widetilde A$ is closed, co-closed, equal to $\pi^{*} B$, and satisfies $D^{*} \omega_0 = 0$.
In other words, $\widetilde B$ trivially extends $B$ to each slice $M \times \{r\}$ using the extended frame and coframe and, under the identification of $\mathcal{B}$ with $TM$, $\widetilde A$ simply equals $-\tanh r \, \widetilde B$.
First, we claim that the $1$-form $\omega_0 = \widetilde B + \widetilde A$ equals the pull-back $\pi^{*} B$. To see this let $x\in M$ and $r \in {\ensuremath{\mathbb{R} } }$. Observe that the element $R_{ni}$ of ${\ensuremath{\mathcal{E} } }$ at $(x,0)$ equals $\cosh r \, R_{ni} - \sinh r \, E_i$ at $(x,r)$. On the other hand, the $1$-form $\omega_j$ at $(x,0)$ pulls back to ${\ensuremath{\operatorname{sech} } }r \, \omega_j$ at $(x,r)$. The formula $\pi^{*} B = \widetilde B + \widetilde A$ follows, implying that $\omega_0$ is closed.
To see that it is co-closed refer to the formula (\[DTeqns\]) for $\delta_{\ensuremath{\mathcal{E} } }= T^* + D^*$ on ${\ensuremath{\mathcal{E} } }$-valued $1$-forms and see how the value of $\delta_{\ensuremath{\mathcal{E} } }\omega_0$ differs at $(x,r)$ from that at $(x,0)$. First, consider the algebraic operator $T^{*}$. When applied to $\widetilde B$ on $M \times {\ensuremath{\mathbb{R} } }$, the only difference from being applied to $B$ on $M$ is the addition of the term corresponding to the normal direction. Since there is no $dr$ term in $\widetilde B$ this adds nothing and $T^{*} \widetilde B = 0$ since $T^{*} B = 0$. Also, $T^{*} (E_i \otimes \omega_j) = [E_j,E_i] = R_{ji}$. Since $b_{ij} = b_{ji}$ and $R_{ij} = - R_{ji}$ we obtain $T^{*} \widetilde A = 0$.
Now consider the differential operator $D^{*}$. At any point $(x,0)$ we can choose a frame and coframe based on geodesic coordinates. As noted above this implies that $\nabla_{e_i} e_j = -\tanh r \, \delta_{ij} \, n$ for all $i,j $ with respect to the extended frame at $(x,r)$ for any $r$. One computes that $$D^{*}(R_{ni} \otimes \omega_j) =
-\nabla_{e_j} R_{ni} + (\sum_k \omega_j(\nabla_{e_k} e_k)) \, R_{ni}.$$ Because of our choice of frame the second term vanishes. Furthermore, $$\nabla_{e_j} R_{ni} = (\tanh r \, R_{ji} + \sum_k c_{ijk} R_{nk})$$ where $c_{ijk} = \langle \nabla_{e_j} e_i , e_k \rangle.$ In our choice of frame only the first term is nontrivial. Thus, using geodesic coordinates, we obtain $$D^{*}(b_{ij} (R_{ni} \otimes \omega_j)) = - ( b_{ij} \tanh r \, R_{ji} +
(\nabla_{e_j} b_{ij}) R_{ni}).$$ Summing over $i,j \in {1,2,\cdots, m}$, the contribution from the terms involving $R_{ji}$ equals $0$ because $b_{ij} = b_{ji}$ and $R_{ij} = -R_{ji}$. Since the functions $b_{ij}$ are independent of $r$ and the orthonormal frame at $(x,r)$ equals ${\ensuremath{\operatorname{sech} } }r$ times the frame pulled-back from $(x,0)$, the terms involving derivatives of $b_{ij}$ at $(x,r)$ equal ${\ensuremath{\operatorname{sech} } }r$ times those at $(x,0)$. Since $D^{*} B = 0$ on $M$ we conclude that $D^{*} \widetilde B = 0$ on $M \times {\ensuremath{\mathbb{R} } }$.
Similarly, to compute $D^{*} \widetilde A$ we observe that $$D^{*}(E_i \otimes \omega_j) = -\nabla_{e_j} E_i
+ (\sum_k \omega_j(\nabla_{e_k} e_k)) \, E_i.$$ In geodesic coordinates the second term is trivial and the first term is nontrivial only when $i=j$ in which case it equals $\tanh r \, N.$ We obtain $$D^* (\tanh r \, b_{ij} \, (E_i \otimes \omega_j)) =
(\delta_{ij} b_{ij}) (\tanh r)^2 \, N - (\nabla_{e_j} b_{ij})\, \tanh r \, E_i$$ Summing over $i,j$, the terms involving $N$ only occur when $i = j$ so the contribution from them is zero because $B$ has trace zero. The contribution from the terms involving derivatives of the $b_{ij}$ equals zero as before.
We will refer to $\omega_0$ as the model harmonic form near the boundary $M$ of $W$.
Boundary value problem {#sec:bdryvalue}
======================
In this section we will describe and solve a boundary value problem that will provide us with a harmonic representative $\omega$ for any class in $H^1(W;{\ensuremath{\mathcal{E} } }).$ The boundary conditions have been chosen so that the boundary term (\[Weitzbdry\]) in the formula is reasonably computable. In the next section we will show that the boundary term of (\[Weitzbdry\]) is $0$, which will imply that $\omega$ is trivial.
Given an element $[\hat \omega] \in H^1(W;{\ensuremath{\mathcal{E} } })$ where $W$ is a hyperbolic $(m+1)$-dimensional manifold with totally geodesic boundary $M^m$, we know that, when $m \ge 3$, the restriction $[{\hat \omega}_M] \in H^1(M;{\ensuremath{\mathcal{E} } })$ lies completely in $H^1(M; \mathcal{S})$. By the results of the previous section, there is a unique harmonic $B \in \Omega^1(M;\mathcal{B})$ representing this class. Under the identification of $\mathcal{B}$ with $TM$, and $\Omega^1(M;TM)$ with ${\ensuremath{\operatorname{Hom} } }(TM,TM)$, the linear transformation corresponding to $B$ is symmetric and traceless. By Proposition \[pullback\], $B$ determines a model harmonic form $\omega_0$ defined in a neighborhood of $M$. The element $[{\hat \omega}] \in H^1(W;{\ensuremath{\mathcal{E} } })$ will be cohomologous to $\omega_0$ in a neighborhood of the boundary of $W$. We can therefore assume that $[{\hat \omega}]$ is represented by a form ${\hat \omega}$ which agrees with $\omega_0$ in a neighborhood of the boundary $M$. In particular, it is closed and co-closed near $M$.
However, ${\hat \omega}$ will not necessarily be harmonic on all of $W$. It is closed, but $\delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ will generally not be trivial away from the boundary. To find a harmonic representative we need to find a section $s$ of the bundle ${\ensuremath{\mathcal{E} } }$ over $W$ satisfying $$\label{correctionprob}
\delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }s = -\delta_{\ensuremath{\mathcal{E} } }{\hat \omega}.$$ Then $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s$ will be a harmonic representative of our given cohomology class.
There are many solutions to (\[correctionprob\]). We will require $s$ to satisfy certain boundary conditions that will make the operator $\lap_{\ensuremath{\mathcal{E} } }= \delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }$ self-adjoint and elliptic with trivial kernel. Since we want to apply the formula (\[Weitzformula\]) to this harmonic representative, our choice of boundary conditions will also be motivated by the resulting boundary term in that formula. Recall that the boundary term (\[Weitzbdry\]) is an integral over $M$ with integrand $-({*}T \omega \wedge \omega + T^{*} \omega \wedge {*}\omega).$ In this section we will show that our chosen boundary conditions imply that $T^{*} \omega = 0$, greatly simplifying our computation of the boundary integral in the next section.
By Lemma \[lem:canon\_lift\] we can assume ${\hat \omega}$ admits a canonical lift, meaning it is locally the image under $d_{\ensuremath{\mathcal{E} } }$ of a canonical lift.
Equivalently, we can assume the translational part of ${\hat \omega}$, viewed as a element of ${\ensuremath{\operatorname{Hom} } }(TW,TW),$ is symmetric.
The operator $\lap_{\ensuremath{\mathcal{E} } }= \delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }$ acting on sections preserves the decomposition of ${\ensuremath{\mathcal{E} } }$ into translational and rotational parts. It follows that solving equation (\[correctionprob\]) is equivalent to solving it separately for the translational and rotational parts; in other words, to solving the corresponding problem for vector fields and for sections of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW).$ Using canonical lifts, we will see below that it suffices to solve the vector field problem. In particular, when solving (\[correctionprob\]) we will be able to assume that the global section $s$ is a canonical lift. This will imply that the translational part of $d_{\ensuremath{\mathcal{E} } }s$ will be symmetric. Thus, the harmonic representative $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s$ will also have this property.
Using the metric on $TW$ we can identify vector fields with $1$-forms and linear transformations with $(0,2)$-tensors. We will use the convention that a linear transformation $\phi$ is identified with the $(0,2)$-tensor $\theta$ defined by $$\theta (u_1, u_2) := \langle \phi u_1, u_2 \rangle.$$ Notice that a skew symmetric transformation is identified with a $2$-form. For example, with this convention $R_{ij}$ is identified with $\omega_j \wedge \omega_i$. A section of ${\ensuremath{\mathcal{E} } }$ can then be identified with a pair $(\tau, \widetilde \tau)$ consisting of a $1$-form and a $2$-form. Via this identification a canonical lift is identified with the pair $(\tau, \frac {1}{2} d \tau)$, where $\tau$ is dual to $u$ and $d$ is exterior differentiation on $W$. (In low dimensions the skew symmetric part of $Du$ is identified with $\frac {1}{2} {\ensuremath{\operatorname{curl} } }(u).$) On the translational part of ${\ensuremath{\mathcal{E} } }$, the operator $T^{*}T$ is multiplication by $m$, so $\lap_{\ensuremath{\mathcal{E} } }u = D^{*}D u + T^{*}T u = D^{*} D u + m u$. Similarly, on the rotational part of ${\ensuremath{\mathcal{E} } }$, $T^{*}T$ is multiplication by $2$ so that $\lap_{\ensuremath{\mathcal{E} } }\widetilde u = D^{*}D \widetilde u + 2 \widetilde u.$ If we view a section of ${\ensuremath{\mathcal{E} } }$ as a pair consisting of a vector field and a section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$, the operator $D^{*}D$ equals $\nabla^{*} \nabla$ on each component of the pair. Under the identification of vector fields with $1$-forms and sections of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$ with $2$-forms, $\nabla^{*} \nabla$ becomes $\nabla^{*} \nabla$ on these forms. The fact that the Ricci curvature of ${\ensuremath{\mathbb{H} } }^{n}$ equals $m$ implies that $\nabla^{*} \nabla$ on $1$-forms equals $ \lap + m$ where $ \lap = \delta d + d \delta$ is the usual (exterior) laplacian on differential forms. More generally, for an $n$-dimensional space of constant curvature $-1$, $\nabla^{*}\nabla$ on $p$-forms equals $ \lap + p (n-p)$ [@Pet Ch.7.4]. When $p=2$ and $n = m+1$, $p (n-p) = 2 (m-1) = 2m-2.$ Thus, under the identification of $u$ and $\widetilde u$ with the forms $\tau$ and $\widetilde \tau$, the operator $\lap_{\ensuremath{\mathcal{E} } }$ corresponds to $ \lap + 2m$ in both cases.
Being a canonical lift means that the $2$-form $\widetilde \tau = \frac {1}{2} d \tau.$ Since $d$ commutes with $\lap + 2m$, it follows that $\lap_{\ensuremath{\mathcal{E} } }$ preserves the property of being a canonical lift. We are assuming that ${\hat \omega}$ is the local image under $d_{\ensuremath{\mathcal{E} } }$ of a canonical lift. It follows that $\delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ is a canonical lift. If $\zeta$ denotes the $1$-form dual to the translational part of $ - \delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$, then the $2$-form corresponding to its rotational part equals $\frac{1}{2} d \zeta$. We conclude that in order to solve equation (\[correctionprob\]), it suffices to solve $$\label{formprob}
(\lap + 2m) (\tau) = \zeta$$ for a $1$-form $\tau$ on $W$. The solution to (\[correctionprob\]) will then be the section $s$ which is the canonical lift of the vector field $u$ dual to $\tau.$
We will now describe our boundary conditions both in terms of the translational part $u$ of the section $s$ and the $1$-form $\tau$ dual to $u$. At any point on the boundary $M$ of $W$ we can decompose vector fields and $1$-forms into their normal and tangential components. Using surfaces equidistant from the boundary and orthogonal projection, this decomposition extends to a neighborhood of the boundary. Choose the unit normal $n$ to be pointing outward and the normal coordinate to be denoted by $r$, so that $dr$ is dual to $n$. Near the boundary write $u = (h,v)$, where $h$ is a function and $v$ is a vector field on $M$. The dual $1$-form is $\tau = h dr + \sigma$, where $\sigma$ is dual to $v$ on $M$.
Our boundary conditions are $$\label{vfbdcond1}
h = 0$$ $$\label{vfbdcond2}
{\nabla }_n v = 0.$$ In other words, the normal component of $u$ is zero and the normal derivative of the tangential part of $u$ is zero.
We must show that solving (\[correctionprob\]) subject to boundary conditions (\[vfbdcond1\]) and (\[vfbdcond2\]) is an elliptic problem with a unique solution. Furthermore, we need to show that the resulting harmonic ${\ensuremath{\mathcal{E} } }$-valued $1$-form $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s$ on $W$ satisfies $T^{*} \omega = 0$.
Let $s$ be a global section of ${\ensuremath{\mathcal{E} } }$ and write its decomposition into translational and rotational parts as $s = (u,\widetilde u)$. Decompose $u = (h,v)$ into its normal and tangential components. Let $\tau$ be the $1$-form dual to $u$ and $\widetilde \tau$ the $2$-form associated to $\widetilde u$ (which is a section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$). Assume that $s$ is a canonical lift so that $2 \widetilde \tau = d \tau$. Decompose $\tau$ and $\widetilde \tau$ into their normal and tangential parts as $\tau = h dr + \sigma$ and $\widetilde \tau = dr \wedge \widetilde \sigma + \widetilde \phi$. Here $\sigma$ and $\widetilde \sigma$ are tangential $1$-forms and $\widetilde \phi$ is a tangential $2$-form. Since we are assuming that $2 \widetilde \tau = d \tau$, we obtain $2 \widetilde \sigma = {\nabla }_n \sigma - d_M h$, where $d_M$ is the exterior derivative in the tangential directions.
Using canonical lifts, our boundary problem (\[correctionprob\]) reduces to the analogous problem for the vector field part of the section $s$, subject to boundary conditions (\[vfbdcond1\]) and (\[vfbdcond2\]). With the above notation, the boundary conditions for the dual problem (\[formprob\]) on $1$-forms are:
$$\label{formbdcond1}
h = 0$$
$$\label{formbdcond2}
{\nabla }_n \sigma = 0.$$
Using equation (\[formbdcond1\]) we have that $d_M h \equiv 0$ on the boundary, so we could equivalently replace the second condition with
$$\label{formbdcond3}
\widetilde \sigma = 0.$$
In the formalism of differential operators on the bundle of $\mathcal{C}^{\infty}$ $1$-forms we write the boundary value problem as solving $P(\tau) = -\zeta$ subject to the conditions $p_1(\tau) = 0 = p_2(\tau)$ where $(P;\{p_1,p_2\})$ is the operator $$\begin{aligned}
P(\tau) &= (\lap +2m) (\tau)\\
p_1(\tau) &= h\\
p_2(\tau) &= {\nabla }_n \sigma. \end{aligned}$$ To show that this boundary problem has a unique solution, it suffices to show that this differential operator is elliptic and that, on the subspace where $p_1(\tau) = 0 = p_2(\tau)$, it is self-adjoint and positive.
These are standard facts because the main operator $P$ has the same symbol as the laplacian on $1$-forms, and because, near the boundary, the operator decomposes into the normal part, where the boundary condition is the Dirichlet condition, and the tangential part, where the boundary condition is the Neumann condition. However, for completeness, we include the argument here.
We will first show that this operator is self-adjoint and positive. Recall that $\lap = d \delta + \delta d$. Then for real-valued $1$-forms $\tau$ and $\psi$ on $W$, integration by parts yields: $$\langle \lap \tau, \psi\rangle = \langle d \tau, d \psi \rangle ~+~
\langle \delta \tau, \delta \psi \rangle + \beta(\tau,\psi),$$ where $\langle \cdot, \cdot \rangle$ is the $L^2$ inner product on $W$ and $\beta(\tau,\psi)$ is a term given by an integral over the boundary.
The operator $\lap$ (hence $\lap + 2m$) will be self-adjoint as long as $$\langle \lap \tau, \psi\rangle ~-~ \langle \lap \psi, \tau\rangle
~=~ \beta(\tau,\psi)\,-\, \beta(\psi, \tau) ~=~ 0.$$ The operator $\lap + 2m$ will have trivial kernel as long as $\langle (\lap +2m) \tau, \tau \rangle > 0$ for any nonzero $\tau$. Letting $\tau = \psi$ above, this will be guaranteed as long as we have $\beta(\tau,\tau) \geq 0$.
Using Green’s identity, we obtain the following formula for the boundary term, where the boundary is oriented with respect to the outward normal: $$\beta(\tau, \psi) ~=~ \int_M ~ {*} d\tau \wedge \psi ~+~ \delta \tau \wedge {*} \psi.$$ As before we decompose $\tau$ as $\tau = h \, dr + \sigma$ and $\psi$ as $\psi = k \,dr + \phi$. Boundary condition (\[formbdcond1\]) implies that $h=0=k$, so that the tangential parts of $ {*} \tau$ and $ {*} \psi$ are trivial. Similarly, boundary condition (\[formbdcond3\]) implies that the normal parts of $d \tau$ and $d \phi$, hence the tangential parts of $ {*} d \tau$ and $ {*} d \phi$, are trivial. It is then clear that $\beta (\tau, \psi) = 0$ for $\tau$ and $\psi$ satisfying the boundary conditions and the operator is thus self-adjoint and positive.
To see that the boundary conditions lead to an elliptic boundary value problem, we consider the operators, $P$, $p_1$, and $p_2$, where the ranges of the two boundary operators are the sub-bundles of normal and tangential parts, respectively, of $1$-forms on the boundary. We then take the top order terms of each of these operators. It is a subtlety of differential operators on bundles that ellipticity may depend on the choice of decomposition of the target bundle, since this affects what the top order terms are. In our case, the boundary operators themselves decompose (ie., you can decompose the domain in the same way).
To show that the system is elliptic, one considers the symbols of the operators. This amounts to looking at the system in local coordinates, fixing the coefficients of the operators by evaluating at a boundary point, and taking only the top order terms in each operator. One then considers the homogeneous, constant coefficient problem in the upper half space of ${\ensuremath{\mathbb{R} } }^n$ given by these simplified operators.
The symbol of $\lap + 2m$ is simply the symbol of the standard laplacian on ${\ensuremath{\mathbb{R} } }^n$, which is well-known to be elliptic. Clearly $p_1$ is of order $0$ and $p_2$ is of order $1$. Taking the top order terms, the operators become the same operators viewed in the upper half-space ${\ensuremath{\mathbb{R} } }^n_+ = \{(x_1, \ldots, x_m,t) \, | \, t \geq 0\}$ with the standard Euclidean metric. We are now left with the simplified system of solving $\lap \tau = 0$ in the upper half-space with homogeneous boundary conditions determined by these Euclidean boundary operators. By definition, the original system is elliptic if and only if this simplified system has no nontrivial bounded solutions. It suffices to show that there are no nontrivial bounded solutions $\tau(x,t)$ of the form $f (t) e^{i(\zeta \cdot x)},$ where $x = (x_1,\ldots, x_m)$ and $\zeta$ is any nonzero vector in the boundary hyperplane. The solutions $f (t)$ to $\lap \tau = 0$ for a given choice of $\zeta$ are linear combinations of $e^{|\zeta| \, t}$ and $e^{-|\zeta| \,t}$. Since we are interested in bounded solutions, only scalar multiples of the latter function will appear. In particular, we have that $f' (t) = -|\zeta| f (t)$.
We decompose $f$ into its normal and tangential components, which we denote by $h \, dr$ and $\sigma$, respectively. The boundary conditions are then $h(0) = 0$ and $\sigma^{\prime}(0) = 0$. But since $\sigma ^{\prime}(t) = -|\zeta| \sigma(t)$, the latter condition implies that $\sigma(0) = 0$ and we conclude that any solution must be trivial.
Thus we have shown that our boundary value problem is uniquely solvable. Standard elliptic theory implies that the solution will be smooth, even on the boundary. Having solved the equation $\delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }s = - \delta_{\ensuremath{\mathcal{E} } }{\hat \omega}$ we define $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s.$ Then $\omega$ is a closed and co-closed representative in the same comology class as ${\hat \omega}$. We will now show that $T^{*} \omega = 0.$
Define the trace of $\omega$ as the trace of its translational part, viewed as an element of ${\ensuremath{\operatorname{Hom} } }(TW,TW)$. We denote this trace function on $W$ by ${\ensuremath{\operatorname{tr} } }$. To prove $T^{*} \omega = 0$ it will suffice to show that $\omega$ has trace zero. Since $\omega$ is closed, it is the image under $d_{\ensuremath{\mathcal{E} } }$ of a locally defined section of ${\ensuremath{\mathcal{E} } }$. The trace is just the divergence of the locally defined vector field which is the translational part of this locally defined section. It also equals $-\delta$ of the local $1$-form which is dual to this local vector field. Note that, although this vector field and its dual $1$-form are only locally defined, the trace is, nonetheless, globally defined.
Our proof that $\omega$ has trace zero is similar to an argument of [@HK2]. We know that $\omega$ is harmonic. Since $\omega$ is closed, it is locally equal to $d_{\ensuremath{\mathcal{E} } }$ of a local section, and being co-closed as well means that $\lap_{\ensuremath{\mathcal{E} } }$ of this local section is zero. $\lap_{\ensuremath{\mathcal{E} } }$ preserves the decomposition into translational and rotational parts, so $\lap_{\ensuremath{\mathcal{E} } }$ of the corresponding local vector field is zero. As we showed above, this means that $ d \delta + \delta d + 2m$ applied to the dual local $1$-form is zero. Applying $ \delta$ and using that ${\ensuremath{\operatorname{tr} } }$ equals $-\delta$ of this local $1$-form, we conclude that $(\delta d + 2m){\ensuremath{\operatorname{tr} } }= 0$. This equation holds on all of $W$. Taking the $L^2$ dot product on $W$ of ${\ensuremath{\operatorname{tr} } }$ with this equation, we conclude that $\langle ( \delta d + 2m){\ensuremath{\operatorname{tr} } }, {\ensuremath{\operatorname{tr} } }\rangle = 0.$ Integrating by parts gives $$\label{eqn:tr}
\langle d {\ensuremath{\operatorname{tr} } }, d {\ensuremath{\operatorname{tr} } }\rangle + 2m \langle {\ensuremath{\operatorname{tr} } }, {\ensuremath{\operatorname{tr} } }\rangle -
\int_M~ {\ensuremath{\operatorname{tr} } }\wedge {*} d {\ensuremath{\operatorname{tr} } }~=~ 0$$ where the boundary is oriented using the outward normal. Below we will show that the integral term of equation (\[eqn:tr\]) equals $0$. It will follow that ${\ensuremath{\operatorname{tr} } }= 0$ on all of $W$.
Since ${\hat \omega}$ equals the model harmonic ${\ensuremath{\mathcal{E} } }$-valued $1$-form $\omega_0$ of Proposition \[pullback\] in a neighborhood of the boundary, and since model forms all have trace zero, the trace of $\omega = {\hat \omega} + d_{\ensuremath{\mathcal{E} } }s$ is just the trace of $d_{\ensuremath{\mathcal{E} } }s$ in a neighborhood of the boundary. In turn, this is the divergence of the vector field $u$ which is the translational part of $s$; this equals $-\delta \tau$, where $\tau$ is the $1$-form dual to $u$.
Since ${\ensuremath{\operatorname{tr} } }= -\delta \tau$ in a neighborhood of the boundary, we can use the boundary conditions on $\tau$ when computing the boundary integral. Also, since ${\hat \omega}$ is harmonic in a neighborhood of the boundary, $\delta_{\ensuremath{\mathcal{E} } }{\hat \omega} = 0$ in a neighborhood of the boundary which in turn implies that $\delta_{\ensuremath{\mathcal{E} } }d_{\ensuremath{\mathcal{E} } }s$ is zero near the boundary. This implies that $(\lap + 2m) \tau = 0$ near the boundary.
The term $ {*} d {\ensuremath{\operatorname{tr} } }$ in the integrand becomes $- {*} d \delta \tau$. Because $\tau$ satisfies $( d \delta + \delta d + 2m) \tau = 0,$ this equals $ {*}( \delta d + 2m) \tau$. Since the integral is over the boundary, only the tangential part of the integrand appears. The tangential part of ${*} \delta d \tau = d {*} d \tau$ equals $d_M$ of the tangential part of ${*} d \tau$, where $d_M$ is the exterior derivative on the boundary. But, the tangential part of $ {*} d \tau$ equals ${*} (dr \wedge \widetilde \sigma)$, where the normal part of $ d \tau$ equals $dr \wedge \widetilde \sigma$. However, boundary condition (\[formbdcond3\]) implies $\widetilde \sigma = 0$ on $M$. Similarly, the tangential part of $ {*} \tau = {*} (h dr + \sigma)$ equals $h$ times the volume form of $M$. We know $h=0$ on $M$ by (\[formbdcond1\]). Thus, the boundary integral is zero and we conclude that the infinitesimal deformation $\omega$ has trace zero, as desired.
Finally, we need to show that, if $\omega$ is harmonic, has trace zero, and comes from a canonical lift, then $T^{*} \omega = 0$ (hence $D^{*} \omega = 0$ since $\delta_{\ensuremath{\mathcal{E} } }\omega = 0$). This is really an algebraic fact arising from the basic identities coming from being harmonic and a canonical lift.
Because $\omega$ is a canonical lift, its translational part is symmetric when viewed as a section of ${\ensuremath{\operatorname{Hom} } }(TW,TW)$. But $T^{*}(\sum_{i,j} a_{ij} E_i \otimes \omega_j) = \sum_{i,j} (a_{ji} -a_{ij}) R_{ij}$ so it is trivial on symmetric elements of ${\ensuremath{\operatorname{Hom} } }(TW,TW)$. Next we need to compute $T^{*}$ on the rotational part of $\omega$. Since $d_{\ensuremath{\mathcal{E} } }\omega = 0$ it is locally the image under $d_{\ensuremath{\mathcal{E} } }$ of a section of ${\ensuremath{\mathcal{E} } }$. Expressing this section as a pair $(u,\widetilde u)$ consisting of its translational and rotational parts, we see that the rotational part of $\omega$ equals $Tu + D\widetilde u$. Thus, we want to compute $T^{*}T u + T^{*}D \widetilde u$.
One computes that $T^{*}T(E_i) = \sum_k [E_k,[E_k,E_i]] = m E_i$, so that $T^{*}T(u) = m u$. To compute $T^{*} D \widetilde u$, recall that being a canonical lift means that the skew symmetric transformation $\widetilde u$ corresponds to the $2$-form $\frac {1}{2} d \tau$, where $\tau$ is the $1$-form dual to $u$. Below we will see that the vector field obtained by applying $T^{*}D$ to a section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$ is dual to the $1$-form obtained by applying $\delta$ to the corresponding $2$-form. Assuming this, we have that $T^{*} D \widetilde u$ is dual to $\frac {1}{2} \delta d \tau$. We conclude that $T^{*}$ of the rotational part of $\omega$ is dual to $\frac {1}{2} (\delta d + 2 m) \tau$. Since $\omega$ is harmonic this equals $- \frac {1}{2} d \delta \tau$. But $-\delta \tau $ equals the divergence of $u$ which, by definition, equals the trace of $\omega.$ Therefore $-\delta \tau = 0$ on all of $W$ so that $- \frac {1}{2} d \delta \tau = 0$, as desired.
To see the relation between $\delta$ and $T^{*} D$, we recall that, under the identification between sections of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$ and $2$-forms, the element $R_{ij}$ corresponds to $- \omega_i \wedge \omega_j$. We first compute $T^{*}D(a R_{ij})$ where $a$ is a function. From (\[DTeqns\]) we conclude that $D (a R_{ij}) = \sum_k {\nabla }_{e_k}(a R_{ij}) \otimes \omega_k.$ and that $T^{*}({\nabla }_{e_k}(a R_{ij}) \otimes \omega_k) = [E_k, {\nabla }_{e_k}(a R_{ij})].$ Thus we obtain $$T^{*} D (a R_{ij}) = \sum_k [E_k, {\nabla }_{e_k}(a R_{ij})] =
({\nabla }_{e_i} a) E_j - ({\nabla }_{e_j} a) E_i + a \sum_k [E_k, {\nabla }_{e_k}(R_{ij})].$$ Similarly, using a standard expression for $\delta$ on forms, we have $$- \delta (a \omega_i \wedge \omega_j) =
\sum_k i(e_k) {\nabla }_{e_k}(a \omega_i \wedge \omega_j) =
({\nabla }_{e_i} a) \omega_j - ({\nabla }_{e_j} a) \omega_i
+ a \sum_k i(e_k) ({\nabla }_{e_k}(\omega_i \wedge \omega_j)).$$ The first two terms of these two expressions are clearly dual. At any point we can choose a frame and coframe using geodesic coordinates and see that all the terms ${\nabla }_{e_k}(R_{ij})$ and ${\nabla }_{e_k}(\omega_i \wedge \omega_j)$ are zero. Since $\delta$ and $T^*D$ are independent of this choice, we see that they are always dual.
Computing the boundary term {#sec:bdryterm}
===========================
We now assume that we have a closed and co-closed form $\omega$ satisfying $T^{*} \omega = 0$. The boundary term (\[Weitzbdry\]) in the formula is then the integral of $- {*}T\omega \wedge \omega$ over the boundary, oriented by the outward normal. If this term is non-negative, $\omega$ must be zero. We have found it easier to keep track of signs by writing the boundary integrand as $(-1)^m \omega \wedge {*}T\omega$. We will give a general formula for this boundary integrand. This requires some preliminary notation. We will see that when $s$ is a solution to (\[correctionprob\]) satisfying the boundary conditions of the previous section, and we let $\omega = \hat \omega + d_{\ensuremath{\mathcal{E} } }s$, the boundary integral (\[Weitzbdry\]) will equal $0$.
We first describe the matrix corresponding to $\omega \in \Omega^1(W,{\ensuremath{\mathcal{E} } })$ in terms of $8$ sub-matrices or blocks that are determined by dividing the columns into $2$ groups, the rows into $4$ groups, and taking all possible combinations. The columns are indexed by the $1$-forms $dr, \omega_1,\omega_2, \ldots, \omega_m$ corresponding to our coframe. They are grouped according to normal and tangential directions. (So the first “subset" consists of the single element $dr$.). The first $m+1$ rows correspond to the infinitesimal translations $N,E_1,E_2, \ldots, E_m$. Again, they are grouped according to normal (single element) and tangential ($m$ elements) directions. The next subset of rows correspond to the infinitesimal rotations $R_{ab}$ that do not involve the normal direction. There are $k$ such elements, where $k$ is the dimension of ${\ensuremath{\operatorname{\mathfrak{so} } } }(m)$. These can be put in any order; they will not play much of a role. The final group of $m$ rows correspond to the infinitesimal rotations $R_{ni}$ that do involve the normal direction.
The resulting $8$ blocks are labeled by letters $\bA - \bH$ in the following way, where the labeling has been chosen so that the most important terms come first. (The case when $m=3$ is pictured in Table \[1forms\].) $\bA$ is the $m \times m$ block of elements of the form $E_i \otimes \omega_j$ and $\bB$ is $m\times m$ block whose entries are the coefficients of $R_{ni} \otimes \omega_j$. They have already made an appearance and are our main players. Then there is an $m \times 1$ block with coordinates $R_{ni} \otimes dr$ and a $k \times 1$ block with coordinates $R_{ab} \otimes dr$. These are labeled $\bC$ and $\bD$ respectively. These are followed by a block whose entries are the coefficients of $E_i \otimes dr$ ($m \times 1$, labeled $\bE$) and the coefficient of $N \otimes dr$ ($1 \times 1$ block, labeled $\bF$). The remaining two blocks with tangential forms have entries that are the coefficients of $N \otimes \omega_j$ ($1 \times m$) and of $R_{ab} \otimes \omega_j$ ($k \times m$), respectively and are labeled $\bG$ and $\bH$.
$$\kbordermatrix{ & dr & & \omega_1 & \omega_2 & \omega_3 \\
N & \bF &\vrule & & \bG & \\ \cline{2-6}
E_1 & &\vrule & & & \\
E_2 & \bE & \vrule & & \bA & \\
E_3 & & \vrule & & & \\ \cline{2-6}
R_{12} & & \vrule & & & \\
R_{13} & \bD & \vrule & & \bH & \\
R_{23} & & \vrule & & & \\ \cline{2-6}
R_{n1} & & \vrule & & & \\
R_{n2} & \bC & \vrule & & \bB & \\
R_{n3} & & \vrule & & &
}$$
We want to compute $\omega \wedge {*}T \omega$. Since we will be integrating this form over the boundary, we are only interested in the tangential terms of the forms. So the formula will only involve the tangential part $\omega_M$ of $\omega$ and the normal part of $T \omega$. (The ${*}$ operator interchanges the forms with a $dr$ term and those without.) The tangential $1$-forms have a basis given by $\{\omega_i\}$, and the normal $2$-forms have a natural basis given by $\{dr \wedge \omega_j\}$. The decomposition of ${\ensuremath{\mathcal{E} } }$ into $4$ groups decomposes the matrix of normal $2$-forms into $4$ blocks. These blocks have the same shapes as $\bG$, $\bA$, $\bH$ and $\bB$, respectively.
Note that $\omega_i \wedge {*}(dr \wedge \omega_j)$ equals $0$ when $i \neq j$ and is a positive orientation of $M$ when $i = j$. It follows that the integral of $\omega \wedge {*}T \omega$ over $M$ equals the $L^2$ dot product on $M$ of the $1$-forms $\omega_M$ and $i(n) T\omega$, where $i(n)$ denotes the interior product with the outward normal. The form $i(n) T\omega$ is simply the normal part of $T \omega$ without the $dr$ term, so we can continue to speak of the $4$ blocks of $i(n) T \omega$.
The $L^2$ product of $\omega_M$ and $i(n) T\omega$ on $M$ equals the integral of the function obtained by taking the dot products of each of $\bG$, $\bA$, $\bH$, and $\bB$ with its corresponding block in $i(n) T \omega$ and then adding them together.
This leaves the task of computing the blocks of $i(n) T\omega$. For $m=3$ the answer is shown in Table \[2forms\]. The next few paragraphs explain this computation. There will be terms of two different types: those coming from the tangential part $\omega_M$ of $\omega$ and those coming from the normal part. Looking at formula (\[DTeqns\]) for $T$, we see that this can be expressed as $$\label{eqn:inTw}
i(n) T\omega = [N, \omega_M] - T_M (i(n) \omega).$$ Here $T_M$ is the operator $T$ restricted to $M$; the sum in formula (\[DTeqns\]) is taken only over the vectors $\{ e_1 , \ldots, e_m \}$ tangent to $M$.
$$\kbordermatrix{ & dr & & \omega_1 & \omega_2 & \omega_3\\
N & 0 &\vrule & & \bC^T & \\ \cline{2-6}
E_1 & &\vrule & & & \\
E_2 & 0 &\vrule & & \bB+ \text{skew}(\bD) & \\
E_3 & &\vrule & & & \\ \cline{2-6}
R_{12} & &\vrule & & & \\
R_{13} & 0 &\vrule & & \text{rot}(\bE) &\\
R_{23} & &\vrule & & & \\ \cline{2-6}
R_{n1} & &\vrule & & & \\
R_{n2} & 0 &\vrule & & \bA + \bF (\text{Id}) & \\
R_{n3} & &\vrule & & &
}$$
Taking the bracket of $N$ with $\omega_M$ simply interchanges the translational Lie algebra elements $\{E_i\}$ with the normal rotational ones $\{R_{ni}\}$ in the obvious way and annihilates the others. This has the effect of interchanging the blocks labelled $\bA$ and $\bB$. In other words, the first summand of (\[eqn:inTw\]) consists of a copy of $\bB$ in the second block, a copy of $\bA$ in the fourth block, and zeros elsewhere. This is visible in Table \[2forms\].
To understand the contribution from the second summand of (\[eqn:inTw\]), we simply drop the normal $1$-form $dr$ from the first column $c$ of $\omega$ and think of $c$ as a section of ${\ensuremath{\mathcal{E} } }$. We then compute $- T_M c$. We can describe the contribution to the first block ($1 \times m$) as $\bC^T$, to the second block ($m \times m$) as ${\ensuremath{\operatorname{skew} } }(\bD)$, the third block ($k \times m$) as ${\ensuremath{\operatorname{rot} } }(\bE)$, and the fourth block ($m \times m$) as $\bF(Id)$. In these descriptions $\bC^T$ is just the transpose of the block $\bC$ in $\omega$, and $\bF (Id)$ is a scalar matrix. The block $\bD$ belongs to ${\ensuremath{\operatorname{\mathfrak{so} } } }(m) \otimes dr$ so that, dropping the $dr$ term, it represents an element of ${\ensuremath{\operatorname{\mathfrak{so} } } }(m).$ Then ${\ensuremath{\operatorname{skew} } }(\bD)$ denotes this skew symmetric transformation, where we are identifying $TM$-valued $1$-forms with sections of ${\ensuremath{\operatorname{Hom} } }(TM,TM)$ in the usual way. Dropping the $dr$ part of the block $\bE$, it becomes simply a vector in $TM$. Its image under $T$ is a $1$-form with values in the bundle of tangential infinitesimal rotations, defined by sending $E_i$ to $\sum_j R_{ij} \otimes \omega_j$ and extending linearly. We define ${\ensuremath{\operatorname{rot} } }(\bE)$ to be the image of the vector $\bE$ under this map. In low dimensions this operator is related to the cross-product [@HK3], which motivates our notation. However, we don’t have a good geometric description in general and use it simply as a definition here.
Adding these two summands of (\[eqn:inTw\]) we obtain the result shown in Table \[2forms\] (for $m=3$).
With this computation and notation we can now write down the integrand $\omega \wedge {*}T \omega$ on $M$. We record the answer in the following proposition.
\[integrand\] Let $\omega$ be an ${\ensuremath{\mathcal{E} } }$-valued $1$-form on an oriented hyperbolic manifold $W$ of dimension $n = m+1$ with totally geodesic boundary $M$. Then, with respect to the notation above, the real-valued $1$-form $\omega \wedge {*}T\omega$ restricted to the boundary equals the volume form on $M$ times the function $$\label{bdterm}
\bG \cdot \bC^T + \bA \cdot \bB + \bA \cdot {\ensuremath{\operatorname{skew} } }(\bD) + \bH \cdot {\ensuremath{\operatorname{rot} } }(\bE) +\bB \cdot \bA + \bB \cdot \bF(Id)$$
In the situation of interest to us $\omega = \omega_0 + d_{\ensuremath{\mathcal{E} } }s$ where $\omega_0$ is a model harmonic form near the boundary and $d_{\ensuremath{\mathcal{E} } }s$ is a correction term. We will use the letters $\hat \bA - \hat \bH$ to denote the blocks of $\omega_0$ and $\bA - \bH$ to denote those of $d_{\ensuremath{\mathcal{E} } }s$. The form $\omega_0$ is particularly simple with all the blocks equal to $0$ except for $\hat \bB$ which is symmetric and traceless. In the correction term, $s$ is a global section of ${\ensuremath{\mathcal{E} } }$. In particular, it is a global section on $M$. Furthermore, the section $s$ is a canonical lift so that $\bA$ will be the symmetric part of $D v$ where $v$ is a global vector field on $M$.
Below we will prove the following lemma.
\[0blocks\] Let $d_{\ensuremath{\mathcal{E} } }s$ be an ${\ensuremath{\mathcal{E} } }$-valued $1$-form, where $s$ is a global section of ${\ensuremath{\mathcal{E} } }$ which is a canonical lift satisfying boundary conditions (\[vfbdcond1\]) and (\[vfbdcond2\]). Denote the blocks of $d_{\ensuremath{\mathcal{E} } }$ by the letters $\bA - \bH$ as in Table \[1forms\] above. Then $\bB$, $\bD$, $\bE$, and $\bG$ are all zero.
Assuming Lemma \[0blocks\], we can show that, for a harmonic form $\omega$ obtained from our boundary value problem, the boundary term in the formula (\[Weitzformula\]) is always trivial. This implies infinitesimal, hence local, rigidity.
\[boundaryzero\] Let $W$ be a compact hyperbolic $n$-manifold with totally geodesic boundary $M$, where $n \geq 4$. Let $\omega \in H^1(W;{\ensuremath{\mathcal{E} } })$ be a harmonic representative constructed using the boundary value problem in Section \[sec:bdryvalue\]. Then the boundary term $$- \int_M \left( {*}T\omega \wedge \omega + T^{*} \omega \wedge {*} \omega \right) .$$ in the formula (\[Weitzformula\]) equals $0.$ Hence $\omega = 0$.
Since $T^{*} \omega = 0$, it suffices to show that the integral of $\omega \wedge {*}T \omega$ over $M$ is trivial. We use Proposition \[integrand\] and the notation above to compute the integrand of this integral.
Using the decomposition $\omega = \omega_0 + d_{\ensuremath{\mathcal{E} } }s$, we expand to obtain $$\omega \wedge {*}T \omega = \omega_0 \wedge {*}T \omega_0 + d_{\ensuremath{\mathcal{E} } }s \wedge {*}T d_{\ensuremath{\mathcal{E} } }s + \omega_0 \wedge {*}T d_{\ensuremath{\mathcal{E} } }s + d_{\ensuremath{\mathcal{E} } }s \wedge {*}T \omega_0$$ Each of the four terms is a sum of dot products with the same terms as those in equation (\[bdterm\]) except that some of the letters have hats. In the first term all letters have hats, in the second none do, in the third the first letter in each dot product has a hat, and in the fourth the second letter in each dot product has a hat.
From Lemma \[0blocks\] we know that blocks $\bB$, $\bD$, $\bE$ and $\bG$ are all zero. In $\omega_0$ every block except $\hat \bB$ is zero. One quickly sees that the first two terms are trivial because at least one matrix in each dot product equals $0$. The third term equals $\hat \bB \cdot (\bA + \bF (Id)) = \hat \bB \cdot \bA + \bF ({\ensuremath{\operatorname{tr} } }\hat \bB) = \hat \bB \cdot \bA$ since $\hat \bB$ is traceless. The final term equals $\bA \cdot \hat \bB$. Thus the integrand equals $ 2 \bA \cdot \hat \bB$. Integrating this over $M$ is just $2$ times the $L^2$ dot product $\langle \bA, \hat \bB \rangle$ of the forms $\bA$ and $\hat \bB$ (where $\hat \bB$ is interpreted as a $TM$-valued $1$-form on $M$). Since $\hat \bB$ is symmetric, its dot product with $Dv$, for $v$ any global vector field on $M$, equals its dot product with the symmetric part of $Dv$ (this is true pointwise, just a fact about dot products of matrices). Since the section $s$ is a canonical lift, $\bA$ is the symmetric part of $Dv$ for a global vector field $v$ on $M$. So, taking the $L^2$ dot product, $\langle\cdot,\cdot\rangle_M$, on $M$, we obtain $$\langle \bA, \hat \bB \rangle_M = \langle Dv, \hat \bB \rangle_M = \langle v, D^{*} \hat \bB \rangle_M.$$ By Proposition \[pullback\], $0 = D^{*} \omega_0 = D^{*} \hat \bB$. We conclude that integrating $\bA \cdot \hat \bB$ over $M$ produces $0$.
We conclude this section with the proof of Lemma \[0blocks\].
Recall that ${\nabla }_n e_i = 0 = {\nabla }_n n$ in a neighborhood of $M.$ Since $M$ is totally geodesic, ${\nabla }_{e_i} n = 0$ and ${\nabla }_{e_i} e_j$ has no normal component. The tangential components of ${\nabla }_{e_i} e_j$ depend on the choice of moving frame, but are not relevant to our computations.
Near the boundary the bundle ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$ decomposes into infinitesimal rotations $\{R_{ab}\}_{n \notin \{a,b\}}$ fixing $n$, called tangential rotations, and those of the form $R_{ni}$, which we will call normal rotations. The above discussion
about covariant derivatives implies that, on $M$, ${\nabla }_{e_i} R_{ab}$ involves only tangential rotations and ${\nabla }_{e_i} R_{nj}$ involves only normal rotations. The normal derivatives of all rotations are zero.
These observations show that the exterior differentiation operator $D: \Omega^0(M;{\ensuremath{\mathcal{E} } }) \to \Omega^1(M;{\ensuremath{\mathcal{E} } })$ preserves a decomposition of the domain and range into blocks in the following way. Near the boundary, an element of the bundle ${\ensuremath{\mathcal{E} } }$ is decomposed into $4$ blocks corresponding to the normal translational part, tangential translational part, tangential rotational part, and normal rotational part, respectively. Similarly, a matrix representing an ${\ensuremath{\mathcal{E} } }$-valued $1$-form is decomposed into the $8$ blocks of Table \[1forms\]. $D$ preserves these decompositions in the sense that the only contribution to a normal or tangential block in $\Omega^1(M;{\ensuremath{\mathcal{E} } })$ comes from normal or tangential derivatives respectively of the corresponding block in ${\ensuremath{\mathcal{E} } }$. For example, the contribution from $Ds$ to the block $ \bB$ whose coordinates are of the form $R_{ni} \otimes \omega_j$ come from tangential derivatives of the normal rotational block of $s$.
With these preliminary observations, it is easy to see the effects of our boundary conditions on the blocks of $Ds$. As before, let $s = (u, \widetilde u)$ for $u$ a vector field on $W$ and $\widetilde u$ a section of ${\ensuremath{\operatorname{\mathfrak{so} } } }(TW)$. Applying boundary condition (\[vfbdcond1\]) we see that, on $M$, $({\nabla }_{e_i} u)\cdot n = {\nabla }_{e_i} (u \cdot n) = {\nabla }_{e_i} h = 0$ because $h \equiv 0$ and where we have used that ${\nabla }_{e_i} n = 0$ on $M$. This is the contribution from $Ds$ to block $ \bG$, whose coordinates are $N \otimes \omega_i$. Similarly, because the $e_i$ are parallel in the normal direction, the boundary condition (\[vfbdcond2\]) implies that $({\nabla }_n u)\cdot e_i = {\nabla }_n (u \cdot e_i) = 0$. This is the $ \bE$ block of $Ds$.
Since $s = (u,\widetilde u)$ is a canonical lift, its rotational part, $\widetilde u$, is the skew symmetric part of $Du$ and corresponds to the $2$-form $ \frac {1}{2} d \tau$, where $\tau$ is dual to $u$. The normal rotational part of $\widetilde u$, which is a linear combination of the $R_{ni}$, is just the part of the $2$-form $ \frac {1}{2} d \tau$ involving $dr$. It equals $\frac {1}{2} \sum_i (({\nabla }_n u)\cdot e_i - ({\nabla }_{e_i} u)\cdot n)\, dr \wedge \omega_i.$ We have just shown that this is zero, so we conclude that the normal rotational part of $s$ is trivial on $M$. Since this is true on all of $M$, its tangential derivative is zero. We saw above that this is precisely the image in $ \bB$ of $Ds$.
Finally, the image in $ \bD$ of $Ds$ comes from the normal derivative of the tangential rotational part of $s$. The tangential rotational part of $s$ is described by the coefficients of $R_{ab}$. Since $s$ is a canonical lift the coefficient of $R_{ab}$ equals $- \frac{1}{2} (({\nabla }_{e_a} u)\cdot e_b - ({\nabla }_{e_b} u)\cdot e_a).$ We don’t know anything about these values except that, since ${\nabla }_{e_a} n = 0 =
{\nabla }_{e_b} n$, they equal $- \frac{1}{2} (({\nabla }_{e_a} v)\cdot e_b - ({\nabla }_{e_b} v)\cdot e_a),$ where $v$ is the tangential part of the vector field $u$. However, since it is totally geodesic, ${\nabla }_{e_i}$ and ${\nabla }_n$ commute on $M$ and the normal derivative of these tangential derivatives of $v$ can be expressed as tangential derivatives of the normal derivative of $v$. But, by boundary condition (\[vfbdcond2\]), these are all zero. Thus, we have shown that, under our boundary conditions, $Ds$ has trivial image in the blocks $ \bB$, $ \bD$, $ \bE$, and $ \bG$.
Since $d_{\ensuremath{\mathcal{E} } }= D + T$ we also need to compute the value of $T(s)$ in these blocks. We can quickly compute that $T(N) = - \sum_i R_{ni} \otimes \omega_i$, so its image is in $ \bB$, and $T(E_i) = R_{ni} \otimes dr + \sum_j R_{ji} \otimes \omega_j$ with image in $ \bH$ and $ \bC$. Similarly, $T(R_{ab}) = E_b \otimes \omega_a - E_a \otimes \omega_b$ is contained in $ \bA$ and $T(R_{ni}) = E_i \otimes dr - N \otimes \omega_i$ contributes to $ \bE$ and $ \bG$. However, for the section $s$ the coefficient of $N$ and all the coefficients of the $R_{ni}$ are zero (by boundary conditions (\[formbdcond1\]) and (\[formbdcond3\]), respectively) so the value of $T s$ in each of $ \bB$, $ \bE$, and $ \bG$ is trivial. The block $ \bD$ is not in the image of $T$. Thus, we have shown that, under our boundary conditions, the $4$ blocks $ \bB$, $ \bD$, $ \bE$, and $ \bG$ of $ds$ are all trivial.
This completes the proof of Theorem \[thm:main\].
Extensions and Conjectures {#sec:conjectures}
==========================
In this section we will discuss some fairly immediate extensions of the results in this paper and some conjectures suggested by our methods.
A first obvious question to ask is whether these results generalize to the case of manifolds with parabolic elements. The situation turns out to be more subtle than one might expect at first glance. As we will discuss below, the answer depends on whether or not the cusps are on the boundary. Furthermore, we conjecture that the answer is different in dimension $4$ than in higher dimensions.
Let $W$ be a finite volume hyperbolic $n$-manifold with totally geodesic boundary. In Theorem \[thm:intromain\] $W$ is assumed to be compact, but in general it might have cusps and hence be noncompact. If the boundary of $W$ is nonetheless compact, then the methods used in this paper extend easily to prove local rigidity. The main thing to check is that, when the dimension $n$ of $W$ is at least $4$, then the cusped ends remain complete under any small deformation. This follows from the fact that all of the cusps will have rank $n-1$ and a simple analysis of the algebraic deformations of rank $n-1$ parabolic subgroups in ${\ensuremath{\mathbb{H} } }^n.$ Once it is established that the cusps remain complete, standard $L^2$ techniques allow one to deal with the boundary terms of a harmonic representative in $H^1(X;{\ensuremath{\mathcal{E} } })$ on these ends. The rest of our analysis goes through without change. This proves the following theorem.
\[thm:rank3\] Let $W$ be a finite volume hyperbolic $n$-manifold with compact, totally geodesic boundary. Assume $n>3$. Then the holonomy representation of $W$ is infinitesimally rigid.
On the other hand, if the boundary of $W$ is allowed to have cusps, hence to be noncompact, then in dimension $4$ it is no longer true that $W$ is always locally rigid. The first example of a such a flexible, finite volume hyperbolic $4$-manifold with non-compact geodesic boundary is constructed and studied in detail in [@KS]. (Actually, the example is an orbifold, but it has manifold covers with similar properties.)
It is worth looking at how the arguments of the current paper break down in that case, as it provides some insight into the situation in other dimensions. If we denote by $W$ the finite volume hyperbolic manifold and let $M$ denote its geodesic boundary, then we consider the restriction map $H^1(W;{\ensuremath{\mathcal{E} } }) \to H^1(M;{\ensuremath{\mathcal{E} } }).$ The image of this map corresponds to those infinitesimal deformations in ${\ensuremath{\mathbb{H} } }^n$ of the $n-1$ dimensional hyperbolic manifold $M$ that extend over $W$. We have seen that there is a decomposition $$\label{cohdecomp}
H^1(M;{\ensuremath{\mathcal{E} } }) \cong H^1(M;\mathcal{H}) \oplus H^1(M;\mathcal{S}).$$ When $M$ is compact and has dimension at least $3$, the first factor is trivial. This was crucial to our argument. When $M$ is complete and finite volume but noncompact, the factor $H^1(M;\mathcal{H})$ is *nontrivial* when $M$ has dimension $3.$ This corresponds to the fact that such an $M$ has local deformations where the new structure is no longer complete; in particular, at least one cupsed end will no longer be complete. This phenomenon is the basis for hyperbolic Dehn surgery which is an important tool in the study of hyperbolic $3$-manifolds.
The analysis in this paper strongly suggests that, for any nontrivial element of $H^1(W;{\ensuremath{\mathcal{E} } })$, the image in $H^1(M;\mathcal{H})$ under the restriction map must be nontrivial. Indeed, this is the case for the example in [@KS]. In that example some of the cusps on the (orbifold) boundary of the original structure do not remain complete and undergo an orbifold version of hyperbolic Dehn surgery. Furthermore, nearby representations of the boundary groups no longer preserve a $3$-dimensional totally geodesic hyperplane; in other words the boundary of convex hull is no longer totally geodesic. This means that the image in the other factor, $H^1(M;\mathcal{S})$ is also nonzero. We conjecture that this must always be true, that a nontrivial deformation of $W$ must always have nontrivial image in *both* factors in (\[cohdecomp\]) under the restriction map.
Similar reasoning leads us to suspect that this failure of local rigidity only occurs in dimension $4$. For, when the dimension $n$ of $W$ is at least $5$, the dimension of the boundary $M$ is at least $4$. In these dimensions, the results of Garland-Raghunathan [@GR] imply that $H^1(M;\mathcal{H}) = 0$ whenever $M$ is complete and finite volume, even in the noncompact case.
\[conj:dimn5\] Let $W$ be a finite volume hyperbolic $n$-manifold with totally geodesic boundary. If $n>4$ then $W$ is infinitesimally rigid.
We have necessarily restricted ourselves to hyperbolic manifolds of dimension at least $4$ in proving local rigidity results. Our results are simply false in dimension $3$, where infinite volume convex cocompact manifolds have large deformation spaces corresponding to the Teichm[ü]{}ller spaces of their conformal boundaries. However, our analysis does provide some information about infinitesimal deformations even in dimension $3$.
Suppose $W$ is $3$-dimensional with totally geodesic boundary $M$ equal to a finite collection of closed surfaces of genus $g\geq 2.$ In this dimension both factors of the decomposition (\[cohdecomp\]) of $H^1(M;{\ensuremath{\mathcal{E} } })$ are nontrivial. In fact they are actually isomorphic to each other and the harmonic representatives can be identified with holomorphic quadratic differentials on $M$. The boundary value problem discussed in this paper is solvable in the same way as before. Analysis of the boundary term shows that any nontrivial deformation must have nontrivial image under the restriction map in both of the factors of $H^1(M;{\ensuremath{\mathcal{E} } }) \cong H^1(M;\mathcal{H}) \oplus H^1(M;\mathcal{S}).$ This means that, not only can’t the boundary remain totally geodesic (which is clear by doubling and applying local rigidity in the closed case), but its hyperbolic metric must also change infinitesimally. Viewed as quadratic differentials the two images of the restriction map must have a nontrivial $L^2$ pairing on $M.$ This leads to the following theorem:
\[thm:3dim\] Let $W$ be a compact hyperbolic $3$-manifold with totally geodesic boundary $M$ equal to a collection of hyperbolic surfaces of genus $g \geq2.$ If $\omega \in H^1(W;{\ensuremath{\mathcal{E} } })$ is a nontrivial infinitesimal deformation of $W$ then its image under the restriction map in $H^1(M;{\ensuremath{\mathcal{E} } }) \cong H^1(M;\mathcal{H}) \oplus H^1(M;\mathcal{S})$ is nontrivial in *both* factors. Furthermore, when the two factors are suitably identified with holomorphic quadratic differentials, their $L^2$ inner product on $M$ is positive.
With this collection of theorems and conjectures we have begun to develop a picture in all dimensions of the deformation theory of an $n$-dimensional hyperbolic manifold with totally geodesic boundary. The increasing rigidity of the boundary as the dimension increases is reflected in the rigidity of the manifold itself. Before this, the authors were unable to discern any significant structure in the deformation theory of high dimensional hyperbolic manifolds, and it seemed possible that the rigidity or flexibility in the infinite volume case was just a chaotic phenomenon. At one extreme there are flexible (convex cocompact) examples with free fundamental group; they are always very flexible due to the lack of relations. At the other extreme lie the closed manifolds; they are always rigid both locally and globally by Mostow rigidity [@Mos]. However, little was known about cases in between. At least now we have a new class of examples that are consistently locally rigid.
However, it remains unclear what the implications of these results are for the deformation theory of general higher dimensional convex cocompact groups. If one considers an infinite volume complete manifold $X$ with compact convex core, the condition that the convex core have totally geodesic boundary implies certain topological properties of $X.$ For example, it implies that the fundamental group of the boundary of the convex core injects into the fundamental group of $X$ and that $X$ has the homotopy type of an $(n-1)$-complex. This is in contrast to the case when the fundamental group is free, where the holonomy representation is very flexible. In this case $X$ is homotopy equivalent to a $1$-dimensional complex. Whether or not these topological properties are central remains to be seen.
In dimensions $n \geq 4$ it is very difficult to construct examples from which one can formulate a conjectural picture. In particular, if we consider the smooth manifold underlying an $n$-dimensional hyperbolic manifold $X$ with Fuchsian ends, we can ask whether it has other locally rigid hyperbolic structures or whether there are deformable ones. We are unable to answer this question because we do not know if there even are any other hyperbolic structures. It is quite possible that the hyperbolic manifold $X$ satifies an infinite volume version of Mostow rigidity. While the results in this paper suggest that this is plausible, other methods than those used here would be needed to make progress on this issue. We conclude by raising it as a question.
\[quest:Mostow\] Let $f: X_1 \to X_2$ be a homotopy equivalence between complete hyperbolic $n$-manifolds without boundary, where $n>3$. Assume that $X_1$ has Fuchsian ends, is convex cocompact, and is not Fuchsian. Is $f$ necessarily homotopic to an isometry?
[^1]: Kerckhoff and Storm were partially supported by NSF grants DMS-0605151 and DMS-0741604, respectively. Storm also received support from the Roberta and Stanley Bogen Visiting Professorship at Hebrew University.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We consider stable solutions of semilinear equations in a very general setting. The equation is set on a Polish topological space endowed with a measure and the linear operator is induced by a carré du champs (equivalently, the equation is set in a diffusion Markov triple).
Under suitable curvature dimension conditions, we establish that stable solutions with integrable carré du champs are necessarily constant (weaker conditions characterize the structure of the carré du champs and carré du champ itéré).
The proofs are based on a geometric Poincaré formula in this setting. [F]{}rom the general theorems established, several previous results are obtained as particular cases and new ones are provided as well.
address:
- 'Dipartimento di Matematica “Federigo Enriques”, Università degli studi di Milano, Via Saldini 50, 20133 Milano (Italy).'
- 'Dipartimento di Matematica, Università di Trento, Via Sommarive 14, 38123 Povo, Trento (Italy).'
- 'Dipartimento di Matematica “Federigo Enriques”, Università degli studi di Milano, Via Saldini 50, 20133 Milano (Italy), and School of Mathematics and Statistics, University of Melbourne, Grattan Street, Parkville, VIC-3010 Melbourne (Australia), and Istituto di Matematica Applicata e Tecnologie Informatiche, Via Ferrata 1, 27100 Pavia (Italy). '
author:
- 'Serena Dipierro, Andrea Pinamonti, Enrico Valdinoci'
title: Rigidity results in Diffusion Markov Triples
---
Introduction
============
A quasilinear equation is an expression of the form $$\label{EQUAFP}
Lu+F(u)=0.$$ Here, $L$ is a linear operator and identity often turns out to provide a significant constraint for the solution $u$: namely, at any point $x$, the operator $Lu$ at $x$ has to be perfectly balanced by the nonlinear source $F(u(x))$ and, as a consequence, the operator is constant along the level sets of the solution.
It is conceivable that this rigid constraint implies suitable classification results: for this, one has typically to consider problems in which arises from a variational structure and focus on solutions with “sufficiently small energy”, since high energy solutions may develop some “wild behavior”. To this end, one often considers solutions which are local minimizers of the energy functional. Nevertheless, this minimal property may be uneasy to check in practice and it is therefore customary to look at a more general class of solutions, the so called “stable” solutions, for which the second derivative of the energy functional is nonnegative (in this setting, local minimizers become a special subclass of the stable solutions). We refer to the monograph [@DUPAI] for a throughout discussion on stable solutions and on several classification results.
After [@FarHab; @FSV], a very useful tool towards the classification of stable solutions has been provided by a series of geometric Poincaré inequalities, originally introduced by Sternberg and Zumbrun in the celebrated articles [@SZ1] and [@SZ2]. Roughly speaking, in this approach a weighted $L^2$-norm of any test function is controlled by a weighted $L^2$ norm of its gradient. The advantage of this method is that the weights are nonnegative and possess a geometric interpretation, hence the possible vanishing of the integral in the Poincaré-type inequalities implies the vanishing of the corresponding geometric weight, which in turn provides a series of useful geometric rigidities. Rigidity results via Poincaré-type inequalities have been recently obtained in different settings, including, among the others, systems of equations [@DI; @fazly; @FAZSI1], manifolds [@fsv1; @fsv2; @FMV; @punzo1; @FAZ17; @DPV-VAR], stratified groups [@FV1; @PV; @BIRI10; @FP], equations with drift [@FNP], stratified media [@SAVV; @CHE; @DP2] and fractional equations [@FRAC1; @FRAC2; @DP1], and there are also applications for equations in infinite dimensional spaces [@CNV]. The method can be also applied to deduce new weighted Poincaré inequalities from the explicit knowledge of a stable solution, see [@INDI], and it is also flexible enough to deal with Neumann boundary conditions [@punzo2; @DPV].
In this framework, a special role is often played by the geometry of the ambient space. To understand this phenomenon, one can think about the case of one-dimensional solutions in the Euclidean flat space for a bistable nonlinearity, for instance heteroclinic solutions of the mechanical pendulum. If one wants to “bend” these objects to construct solutions on a sphere, a geometric difficulty arises from the fact that the asymptotics “at the point infinity” is not well-determined, thus making it difficult to perform such a bending operation. This very heuristic example suggests that for “curved” manifolds the number and the structure of the stable solutions could be very different from the flat case.
The main objective of this paper is to deal with stable solutions of semilinear equations in the very general setting provided by the Markov triples, see the monograph [@BLG]. Though we have not attempted to reach the widest generality, the setting of Markov triples is an excellent setting that comprises several particular cases at once, by developing an appropriate form of calculus and often providing a general, elegant and unified treatment. Moreover, the setting of Markov triples finds applications in probability and mathematical physics, e.g. to describe quantum ensembles, see e.g. [@CHENG] and the references therein.
The environment provided by the Markov triple (described here in details in Section \[89237ef2S\]) is that of a Polish space endowed by a measure and a carré du champ operator, which provide a variational framework for a general form of equation . See e.g. [@HIR; @BOU; @BH] for a classical introduction to these kinds of Dirichlet forms and also [@AmbTre] and the references therein for recent developments.
The main result of this paper (stated in details in Theorem \[main1\]) is that stable solutions with integral carré du champ in environments with suitable geometric assumptions satisfy additional rigidity properties. Roughly speaking, under strong curvature assumptions, these solutions are necessarily constant and under weak curvature assumptions their level sets need to satisfy suitable geometric identities.
These types of results will be obtained here as a consequence of a very general Poincaré-type inequality (given explicitly in Theorem \[eqnlin\]).
The rest of this paper is organized as follows. In Section \[89237ef2S\], we introduce in details the setting of Markov triples in which we work and we state our main results, the proofs of which are given in Section \[0pqwdlfjvsw3hb\].
Then, in Section \[oqidwuwiefyetytyyt\], we deduce several particular cases from our main results (some of these results were already known in the literature, but follow here as a byproduct of our general and unified approach; some other results seem to be new to the best of our knowledge).
Functional setting and main results {#89237ef2S}
===================================
A triple $(X,\mu,\Gamma)$ is called a *diffusion Markov triple* if it is composed by a Polish topological space $(X,\tau)$ (i.e. a separable completely metrizable topological space $X$ with topology $\tau$) endowed with a $\sigma-$finite Borel measure $\mu$ with full support, a class $\mathcal{A}_0$ of real-valued measurable functions on $X$ and a symmetric bilinear map (the *carré du champ*) such that $\Gamma:\mathcal{A}_0\times\mathcal{A}_0\to \mathcal{A}_0$ satisfying the following conditions:
1. $\mathcal{A}_0\subset L^{1}(\mu)$ is a vector space, dense in every $L^p(\mu)$, $1\leq p<\infty$, such that $$\forall f,g\in\mathcal{A}_0\Longrightarrow fg\in \mathcal{A}_0$$ and $$\forall f_1,\ldots , f_k\in \mathcal{A}_0,\ \Psi\in C^{\infty}(\mathbb{R}^k),\ \Psi(0)=0\Longrightarrow \Psi(f_1,\ldots, f_k)\in \mathcal{A}_0,$$
2. The map $\Gamma$ is bilinear, symmetric and such that, for any $f\in\mathcal{A}_0$, it holds that $$\Gamma(f):=\Gamma(f,f)\geq 0.$$ Moreover, for any $f_1,\ldots , f_k,g\in \mathcal{A}_0$, and any smooth $C^{\infty}$ function $\Psi : \mathbb{R}^k\to \mathbb{R}$ such that $\Psi(0)=0$, we have that $
\Psi(f_1,\ldots, f_k)\in\mathcal{A}_0$ and $$\Gamma(\Psi(f_1,\ldots, f_k), g)=\sum_{i=1}^k \partial_i\Psi(f_1,\ldots, f_k)\Gamma(f_i,g)\qquad \mu-\mbox{a.e.}$$
3. For every $f\in \mathcal{A}_0$, there is $C=C(f)>0$ such that for every $g\in\mathcal{A}_0$ $$\left| \int_X \Gamma(f,g)\, d\mu\right| \leq C\|g\|_{L^2(\mu)}.$$ The Dirichlet form $\mathcal{E}$ is defined for every $(f,g)\in\mathcal{A}_0\times\mathcal{A}_0$ by $$\mathcal{E}(f,g):=\int_X \Gamma(f,g)\, d\mu.$$ The domain $\mathcal{D}(\mathcal{E})$ of $\mathcal{E}$ is the completion of $\mathcal{A}_0$ with respect to the norm $$\|f\|_{\mathcal{E}}:=(\|f\|_{L^2(\mu)}^2+\mathcal{E}(f,f))^{1/2}.$$ The Dirichlet form $\mathcal{E}$ is extended to $\mathcal{D}(\mathcal{E})$ by continuity together with the map $\Gamma$.
4. $L$ is a linear operator on $\mathcal{A}_0$ defined as $$\int_X g Lf\, d\mu=-\int_X \Gamma(f,g)\, d\mu$$ for all $f,g\in\mathcal{A}_0$. The domain of the operator $L$, $\mathcal{D}(L)$, is defined as the set of $f\in \mathcal{D}(\mathcal{E})$ for which there exists a constant $C=C(L)>0$ such that for any $g\in \mathcal{D}(\mathcal{E})$ $$|\mathcal{E}(f,g)|\leq C\|g\|_{L^2(\mu)}.$$ On $\mathcal{D}(L)$, the operator $L$ is extended via integration by parts formula for every $g\in\mathcal{D}(\mathcal{E})$. The operator $L$ defined on $\mathcal{D}(L)$ is always self-adjoint.
5. There exists an increasing sequence $(\xi_k)_{k\geq 1}\subset \mathcal{A}_0$ of functions such that $\xi_k(x)\in[0,1]$ for any $x\in X$, and $$\lim_{k\to+\infty} \xi_k=\mathrm{1}\qquad \mu-\mbox{a.e. in $X$}$$ and $$\label{xi2}
\Gamma(\xi_k)\leq \frac{1}{k}\qquad k\geq 1.$$
We also assume the existence of an algebra $\mathcal{A}_0\subset \mathcal{A}$ of measurable functions on $E$, containing the constant functions and satisfying the following requirements:
1. Whenever $f\in\mathcal{A}$ and $h\in\mathcal{A}_0$, $hf\in\mathcal{A}_0$;
2. For any $f\in\mathcal{A}$, if $$\int_E hf d\mu \geq 0\qquad \mbox{for every positive}\qquad h\in \mathcal{A}_0,$$ then $f\geq 0,$\
3. $\mathcal{A}$ is stable under composition with smooth $C^{\infty}$-functions $\Psi: \mathbb{R}^k\to \mathbb{R}$ with $\Psi(0)=0$, namely if $f_1,\dots,f_k\in\mathcal{A}$ then $\Psi(f_1,\dots,f_k)\in
\mathcal{A}$. Furthermore, for all $f_1,\dots,f_k,g\in\mathcal{A}$, it holds that $$\label{9090}
\Gamma(\Psi(f_1,\ldots, f_k), g)=\sum_{i=1}^k \partial_i\Psi(f_1,\ldots, f_k)\Gamma(f_i,g)\qquad \mu-\mbox{a.e.}$$\
4. The operator $L : \mathcal{A}\to \mathcal{A}$ is an extension of $L$ on $\mathcal{A}_0$. The carré du champ operator $\Gamma$ is also defined on $\mathcal{A}\times\mathcal{A}$ by the formula, for every $(f, g) \in \mathcal{A}\times\mathcal{A}$, $$\label{f per g}
\Gamma(f,g)=\frac{1}{2}\left[L(fg)-fLg-gLf\right]\in\mathcal{A},$$ and for any $f\in\mathcal{A}$ we set $\Gamma(f):=\Gamma(f,f)$,
5. For every $f\in\mathcal{A}$, $\Gamma(f,f)\geq 0$,\
6. For every $f\in\mathcal{A}$ and $g\in\mathcal{A}_0$, the integration by part formula $$\int_X \Gamma(f,g)\, d\mu=-\int_X g Lf d\mu=-\int_X f Lg\, d\mu$$ holds true.
7. If $f\in\mathcal{A}$ is such that $\Gamma(f)=0$ then $f$ is constant.
The assumption on the sign of $\Gamma(f)$ in $(5')$ and the nondegeneracy condition in $(7')$ are important and nontrivial structural assumptions. Roughly speaking, they reflect the “ellipticity’’ of the operator $L$. For instance, if, for ${{\mathbb{R}}}^2\ni(x,y)\mapsto f(x,y)$, one considers the d’Alembert operator $
Lf=f_{xx}-f_{yy}$, one obtains $$\Gamma(f) = \frac12\,L(f^2) - f Lf
= ( (f_x)^2 + f f_{xx} ) - ( (f_y)^2 + f f_{yy} ) - f ( f_{xx}-f_{yy} )
= (f_x)^2 - (f_y)^2 ,$$ which has indefinite sign.
In addition, if, for ${{\mathbb{R}}}\ni x\mapsto f(x)$, one considers the derivative operator $Lf = f_x$, then it follows that $$\Gamma(f) = \frac12\,L(f^2) - f Lf
=\frac12\,(f^2)_x - ff_x = 0,$$ therefore there are nontrivial operators producing carré du champs which vanish identically.
It is interesting to point out that operators in “divergence” and “nondivergence” form share the same carré du champs. For instance, if $a_{ij}$ is a smooth symmetric matrix, and $$\label{DIVandNONDIV} L_{D}f:= \sum_{i,j=1}^n (a_{ij} f_i)_j
\quad{\mbox{ and }}\quad
L_{ND}f:= \sum_{i,j=1}^n a_{ij} f_{ij},$$ a direct computation shows that both $L_D$ and $L_{ND}$ satisfy $$\Gamma(f,g)=\sum_{i,j=1} a_{ij} \,f_i\,g_j.$$ Remarkably, the difference between the operators $L_D$ and $L_{ND}$ is read, in our setting, by condition (6’), which is satisfied by $L_D$ and not by $L_{ND}$. That is, in a sense, while conditions (5’) and (7’) reflect an elliptic condition into a positive definiteness of an associated quadratic form, condition (6’) detects the variational structure of the associated operator.
Taking $\Psi(x,y)=xy$ in we find that $$\label{9191}
\Gamma(f_1f_2, g)= f_1 \Gamma(f_2, g) + f_2\Gamma(f_1, g)\qquad \mu-\mbox{a.e.}$$ for any $f_1,f_2, g\in \mathcal{A}$.
In addition, from , we infer that, for any $f,g\in
\mathcal{A}$, $$\label{9292}
\Gamma(f^2, g^2)=
2f\Gamma(f, g^2)=
2f\Gamma(g^2,f)=4fg
\Gamma(g,f)=4fg
\Gamma(f,g)
\qquad \mu-\mbox{a.e.}$$
Moreover, exploiting with $f_1=f_2=g=1$, we see that $$\Gamma(1)= \Gamma(1) + \Gamma(1)\qquad \mu-\mbox{a.e.}$$ and therefore $$\Gamma(1)=0\qquad \mu-\mbox{a.e.}$$ Using this identity and formula with $f=g=1$, we also infer that $$0=\Gamma(1)=\frac{1}{2}\left[L(1)-L(1)-L(1)\right]
=-\frac{L(1)}{2},$$ and so $$\label{and sp}
L(1)=0\qquad \mu-\mbox{a.e.}$$
To detect the behavior of second derivative operators, it is also classical to introduce the following notation.
\[defGamma2\] The *carré du champ itéré* is the bilinear form $\Gamma_2:\mathcal{A}\times\mathcal{A}\to \mathcal{A}$ defined as $$\label{GammaD2}
\Gamma_2(f,g):=\frac{1}{2}\left[L\Gamma(f,g)-\Gamma(f, Lg)-\Gamma(g,Lf)\right].$$ We define $\Gamma_2(f):=\Gamma_2(f,f)$.
As an example, we point out that when $L$ is the Laplace operator in ${{\mathbb{R}}}^n$, then the carré du champ itéré reduces to the square of the norm of the Hessian matrix (see also Appendix C.5 in [@BLG] for more general formulas for Riemannian manifolds).
Now we recall a classical notion of curvature dimension condition in our setting:
\[DEF2.3\] We say that $(X,\mu,\Gamma)$ satisfies the $CD(K,\infty)$ condition, for some $K\in\mathbb{R}$, if for any $f\in\mathcal{A}$ $$\Gamma_2(f)\geq K\Gamma(f)
.$$
The following result will be crucial in our setting (see page ). For the proof of it, see [@BLG formula (3.3.6)].
\[QUI0\] Assume that $(X,\mu,\Gamma)$ satisfies the $CD(K,\infty)$ condition in Definition \[DEF2.3\], for some $K\in\mathbb{R}$. Then $$\begin{aligned}
\label{Dis}
4\Gamma(f)\left(\Gamma_2(f)-K\Gamma(f)\right)\geq \Gamma(\Gamma(f))\end{aligned}$$ for every $f\in \mathcal{A}$.
In this paper we study the solutions to the following boundary value problem: $$\begin{aligned}
\label{eqncomp}
L u+F(u)=0\quad \mbox{in}\ X,\end{aligned}$$ where $F\in C^{\infty}({{\mathbb{R}}})$. As customary, we say that $u$ is a weak solution to if $u\in \mathcal{A}$ and $$\begin{aligned}
\label{weak}
\int_{X}\Gamma(u,\varphi)\, d\mu =\int_{X} F(u)\varphi \, d\mu,
\quad {\mbox{ for any }} \varphi\in \mathcal{A}_0.\end{aligned}$$ Moreover, we say that a weak solution $u$ is stable if $$\begin{aligned}
\label{hstab}
\int_{X}\Gamma(\varphi) \, d\mu- \int_{X} F'(u)\varphi^2\, d\mu\ge 0, \qquad {\mbox{ for any }}
\varphi\in \mathcal{A}_0.\end{aligned}$$
\[main1\] Assume that $(X,\mu,\Gamma)$ satisfies the curvature dimension condition $CD(K,\infty)$ for some $K\geq 0$ and $\Gamma:\mathcal{A}\times \mathcal{A}\to \mathcal{A}$. For any $(x,y)\in X\times X$, let $$\label{DIS}
d(x,y):={{\mathrm{esssup}}}\{f(x)-f(y)\}$$ be the distance function in $(X,\mu,\Gamma)$, where the essential supremum is computed on bounded functions $f\in\mathcal{A}$ with $\Gamma(f)\leq 1$. Let $x_0\in X$ and $\rho(x):=d(x,x_0)$, for any $x\in X$, and suppose that there exists a sequence of functions $\rho_k\in{\mathcal{A}}$ such that $$\label{DIS2}
\rho_k\to\rho \qquad\mu-\mbox{a.e. in }X\quad{\mbox{ and }}\quad
\|\Gamma(\rho_k)\|_{L^{\infty}}\leq C_0,$$ for some $C_0>0$.
Let $u\in \mathcal{A}$ be a stable solution to with $$\label{mai us}
\int_{X}\Gamma(u)\, d\mu <\infty
.$$ Then: $$\begin{aligned}
&K>0\Longrightarrow \Gamma(u)=0\qquad\mu-\mbox{a.e. in }X;\\
\label{KPOSI0}
&K=0 \Longrightarrow \Gamma_2(u)-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)=0\qquad\mu-\mbox{a.e. in }X.\end{aligned}$$ In addition, if $K>0$ and $\Gamma(u)\in \mathcal{A}\cap C^0(X)$ then $$\begin{aligned}
\label{KPOSI}
{\mbox{$u$ is constant in~$X$}}.\end{aligned}$$
The distance function in , often called intrinsic distance, has been considered also in [@BLG] and it coincides with the Riemannian distance if $X$ is a Riemannian manifold and with the Carnot-Carathéodory distance if $X$ is a Carnot-Carathéodory space (see [@BLU] for the definition). In this setting, Assumption is related to the fact that $\rho$ is Lipschitz as defined in [@BLG Definition 3.3.24]. We recall that the same assumption appears in [@Sturm] and it is the analogous to $|\nabla\rho|\leq 1$ which is satisfied by geodesic distances on any manifold.
The proof of Theorem \[main1\] is based on a geometric Poincaré-type inequality, which we state in this setting as follows:
\[eqnlin\] Let $u\in\mathcal{A}$ be stable weak solution to . Then, $$\label{GF}
\int_{X}\left(\Gamma_2(u)-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\varphi^2\, d\mu\leq \int_{X} \Gamma(u)\Gamma(\varphi)$$ for any $\varphi\in \mathcal{A}_0$.
Proof of Theorems \[main1\] and \[eqnlin\] {#0pqwdlfjvsw3hb}
==========================================
We fix $\varepsilon>0$ and take $\varphi\in \mathcal{A}_0$. By $(4')$, we know that $\Gamma(u)\in\mathcal{A}$. Since $\mathcal{A}$ is an algebra (and therefore a vector space) it follows that $$\label{02}
\Gamma(u)+\varepsilon\in\mathcal{A}.$$ Now, we consider a function $\Psi\in C^\infty({{\mathbb{R}}})$ such that $\Psi(r)=0$ for any $r\le\varepsilon/4$ and $\Psi(r)=\sqrt{r}$ for any $r\ge\varepsilon/2$. In view of $(3')$ and , we have that $$\sqrt{\Gamma(u)+\varepsilon}=\Psi\big(\Gamma(u)+\varepsilon\big)
\in\mathcal{A}.$$ Consequently, by $(1')$ we conclude that $$\psi_{\varepsilon}:=\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\in \mathcal{A}_0.$$ Hence, applying with $\varphi$ replaced by $\psi_{\varepsilon}$, we get $$\label{9393}
0\leq \int_{X} \Gamma\left(\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)\, d\mu-\int_{X}F'(u)\left(\Gamma(u)+\varepsilon\right)\varphi^2\, d\mu.$$ Furthermore, by , we have that $$\begin{split}&
\Gamma\left(\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)\\
=\;&
\Gamma\left(\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi,\,
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)\\
=\;&
\sqrt{\Gamma(u)+\varepsilon}\,
\Gamma\left(\varphi,\,
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)
+\varphi
\Gamma\left( \sqrt{\Gamma(u)+\varepsilon},\,
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)
\\ =\;&
\sqrt{\Gamma(u)+\varepsilon}\,
\Gamma\left(
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi, \varphi\right)
+\varphi
\Gamma\left(
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi,\,
\sqrt{\Gamma(u)+\varepsilon}\right)
\\ =\;&
\big(\Gamma(u)+\varepsilon\big)\,
\Gamma\left(\varphi, \varphi\right)
+
\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\,
\Gamma\left(
\sqrt{\Gamma(u)+\varepsilon}, \varphi\right)
\\&\quad+\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\,
\Gamma\left(
\varphi,\,
\sqrt{\Gamma(u)+\varepsilon}\right)
+\varphi^2
\Gamma\left( \sqrt{\Gamma(u)+\varepsilon},\,
\sqrt{\Gamma(u)+\varepsilon}\right)\\
=\;&
\big(\Gamma(u)+\varepsilon\big)\,
\Gamma(\varphi)
+
2\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\,
\Gamma\left(
\sqrt{\Gamma(u)+\varepsilon}, \varphi\right)
+\varphi^2
\Gamma\left( \sqrt{\Gamma(u)+\varepsilon}\right).
\end{split}$$ This and imply that $$\Gamma\left(\left(\sqrt{\Gamma(u)+\varepsilon}\right)\varphi\right)=
\big(\Gamma(u)+\varepsilon\big)\,
\Gamma(\varphi)
+\frac12\,
\Gamma\left(
\Gamma(u)+\varepsilon,\, \varphi^2\right)
+\varphi^2
\Gamma\left( \sqrt{\Gamma(u)+\varepsilon}\right).$$ Plugging this information into , we obtain that $$\label{9494}\begin{split}&
\int_{X}F'(u)(\Gamma(u)+\varepsilon)\varphi^2\, d\mu\\
\leq\;& \int_{X} \left(\Gamma(u)+\varepsilon\right)\Gamma(\varphi)+ \frac{1}{2}\,\Gamma(\varphi^2, \Gamma(u)+\varepsilon)+\varphi^2 \Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right)\, d\mu
.\end{split}$$ Now, we remark that, by Fatou’s Lemma, $$\label{9595}
\liminf_{\varepsilon\to0}
\int_{X}F'(u)(\Gamma(u)+\varepsilon)\varphi^2\, d\mu\ge
\int_{X}F'(u) \Gamma(u)\varphi^2\, d\mu.$$ Moreover, from $(1)$ we know that $
\Gamma(\varphi)$ is a bounded function and therefore $$\label{9696}
\lim_{\varepsilon\to0}
\int_{X} \left(\Gamma(u)+\varepsilon\right)\Gamma(\varphi)\, d\mu
=
\int_{X} \Gamma(u)\Gamma(\varphi)\, d\mu
+
\lim_{\varepsilon\to0}
\varepsilon
\int_{X} \Gamma(\varphi)\, d\mu
=\int_{X} \Gamma(u)\Gamma(\varphi)\, d\mu.$$ We also remark that, for any $f,g\in{\mathcal{A}}$, $$\label{100}\begin{split}
\Gamma(f,g+\varepsilon)\;&=\;
\frac{1}{2}\left[L(f(g+\varepsilon))-fL(g+\varepsilon)-(g+\varepsilon)
Lf\right]\\
&=\;
\frac{1}{2}\left[L(fg)+\varepsilon Lf-
fLg-\varepsilon fL(1)-gLf-\varepsilon Lf\right]
\\ &=\;
\frac{1}{2}\left[L(fg)-
fLg-gLf\right]\\
&=\;\Gamma(f,g),
\end{split}$$ thanks to and . As a consequence, we have that $$\label{9797}
\int_{X}\Gamma(\varphi^2, \Gamma(u)+\varepsilon)\, d\mu
=
\int_{X}\Gamma(\varphi^2, \Gamma(u))\, d\mu.$$ Furthermore, we claim that $$\label{9898}
\limsup_{\varepsilon\to0}
\int_{X}
\varphi^2 \Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right)\, d\mu
\le
\int_{X}
\varphi^2 \Gamma\left(\sqrt{\Gamma(u)}\right)\, d\mu
.$$ To prove this, we can assume that $$\label{9999}
\int_{X}
\varphi^2 \Gamma\left(\sqrt{\Gamma(u)}\right)\, d\mu
<+\infty,$$ otherwise is true by default. Also, in view of (used here with $f=g=\sqrt{\Gamma(u)+\varepsilon}$), we see that $$\Gamma(\Gamma(u)+\varepsilon)=4(\Gamma(u)+\varepsilon)
\,\Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right).$$ [F]{}rom this and , we obtain that $$\label{101}
\Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right)=\frac{
\Gamma(\Gamma(u)+\varepsilon)}{4(\Gamma(u)+\varepsilon)}
=
\frac{
\Gamma(\Gamma(u))}{4(\Gamma(u)+\varepsilon)}
.$$ Similarly, using with $f=g=\sqrt{\Gamma(u)}$, we see that $$\label{sei pr}
\Gamma(\Gamma(u))=4(\Gamma(u))
\,\Gamma\left(\sqrt{\Gamma(u)}\right).$$ Inserting this into , we conclude that $$\varphi^2\Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right)=
\varphi^2\,\frac{\Gamma(u)}{\Gamma(u)+\varepsilon}
\,\Gamma\left(\sqrt{\Gamma(u)}\right)\le
\varphi^2\,\Gamma\left(\sqrt{\Gamma(u)}\right),$$ and the latter is a summable function, thanks to . Therefore, $$\begin{aligned}
\int_{X}
\varphi^2 \Gamma\left(\sqrt{\Gamma(u)+\varepsilon}\right)\, d\mu
&=&
\int_{X}
\varphi^2\,\frac{\Gamma(u)}{\Gamma(u)+\varepsilon}
\,\Gamma\left(\sqrt{\Gamma(u)}\right)
\, d\mu
\\ &=&
\int_{X}
\varphi^2\,\left(1-\frac{\varepsilon}{\Gamma(u)+\varepsilon}
\right)\,\Gamma\left(\sqrt{\Gamma(u)}\right)
\, d\mu
\\
&\le&
\int_{X}
\varphi^2\,\Gamma\left(\sqrt{\Gamma(u)}\right)
\, d\mu,\end{aligned}$$ which in turn implies .
Therefore, letting $\varepsilon \to 0$ in , and exploiting , , and , we conclude that $$\int_{X}F'(u)\Gamma(u)\varphi^2\, d\mu \leq \int_{X}\Gamma(u)\Gamma(\varphi)+ \frac{1}{2}\Gamma(\varphi^2, \Gamma(u))+\varphi^2 \Gamma\left(\sqrt{\Gamma(u)}\right)\, d\mu.$$ As a consequence, by $(6')$, Definition \[defGamma2\] (used here with $f=g=u$) and , we obtain that $$\label{102}
\begin{split}
\int_{X}F'(u)\Gamma(u)\varphi^2\, d\mu &\leq \int_{X} \Gamma(u)\Gamma(\varphi)- \frac{1}{2}\,\varphi^2 \,L\Gamma(u)+ \varphi^2\,\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\, d\mu\\
&=\int_{X} \Gamma(u)\Gamma(\varphi)-\varphi^2\,
\Big(\Gamma_2(u)+\Gamma(u, Lu)\Big) +\varphi^2 \Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\, d\mu\\
&=\int_{X} \Gamma(u)\Gamma(\varphi)- \varphi^2 \,\Big(\Gamma_2(u)+\Gamma(u, -F(u))\Big)+\varphi^2 \Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\, d\mu.
\end{split}$$ Besides, using with $\Psi=-F$, we see that $$\Gamma(u,-F(u))
\Gamma(-F(u),u)=
-F'(u)\,\Gamma(u)\qquad \mu-\mbox{a.e.}$$ and thus becomes $$\int_{X}F'(u)\Gamma(u)\varphi^2\, d\mu \le
\int_{X} \Gamma(u)\Gamma(\varphi)- \Gamma_2(u)\varphi^2+F'(u)\Gamma(u)\varphi^2 +\varphi^2 \Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\, d\mu.$$ Then, canceling one term, we obtain , as desired.
With this, we are able to prove Theorem \[main1\]:
Using the identity in and Theorem \[QUI0\] \[QUI\] we have that $$\label{11-039}
\begin{split}&
4\Gamma(u)\,\Big(
\Gamma_2(u)-\Gamma\big(\Gamma(u)^{\frac{1}{2}}\big)\Big)
=
4\Gamma(u)\,
\Gamma_2(u)-
4\Gamma(u)\,
\Gamma\big(\Gamma(u)^{\frac{1}{2}}\big)
\\ &\qquad= 4\Gamma(u)\,
\Gamma_2(u)-\Gamma(\Gamma(u))
\ge 4K\,(\Gamma(u))^2,
\end{split}$$ which is always nonnegative if $K\ge0$. Also, we know that $\Gamma(0)=\Gamma_2(0)=0$, due to and , and therefore $$\Gamma_2(u)-\Gamma\big(\Gamma(u)^{\frac{1}{2}}\big)=0
\qquad \mbox{in}\ \{\Gamma(u)=0\}.$$ Using this and Theorem \[eqnlin\] we can write $$\begin{split}\label{stimaGamma2}
\int_{\{ \Gamma(u)\neq 0\}}\left(\Gamma_2(u)-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\varphi^2\, d\mu
=\;&
\int_{X}\left(\Gamma_2(u)-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\varphi^2\, d\mu
\\ \leq \;&\int_{X} \Gamma(u)\Gamma(\varphi)\, d\mu
\end{split}$$ for every $\varphi\in\mathcal{A}_0$.
Now, we fix $R>1$ and define $\Phi=\Phi_R\in C^{\infty}({{\mathbb{R}}})$, with $$\label{test2}
|\Phi'(t)|\leq 3$$ for any $|t|\in [R, R+1]$ and $$\label{test}
\Phi(t):=\left\{\begin{array}{lll}
1 & \mbox{if}& |t|\leq R,\\
0 & \mbox{if}& |t|\geq R+1.
\end{array}
\right.$$ We also define $$\label{WI}\widetilde{\Phi}(t):=\Phi(t)-1,$$ for any $t\in{{\mathbb{R}}}$. In this way, we have that $\widetilde{\Phi}(0)=0$.
Moreover, we consider $\xi_k$ as given by (5). We remark that, in light of and , $$\Gamma(\xi_k^2)=4\xi_k^2\Gamma(\xi_k)\le\frac{4\xi_k^2}{k}.$$ Iterating this, we have that $$\label{xi2-tris}
\Gamma(\xi_k^4)=4\xi_k^4\Gamma(\xi_k^2)\le\frac{16\xi_k^4}{k}.$$ Therefore, by possibly renaming $\xi_k$ into $\xi_k^4$, we can suppose in view of (5) that $\xi_k(x)\in[0,1]$ for any $x\in X$, and $$\label{inse}
\lim_{k\to+\infty} \xi_k=\mathrm{1}\qquad \mu-\mbox{a.e. in $X$}$$ and $$\label{xi2-bis}
\Gamma(\xi_k)\le\frac{16\xi_k^4}{k}.$$ Then, using (1’) and the setting in , we obtain that for every $k\in\mathbb{N}$ the function $\rho_k\xi_k$ belongs to $\mathcal{A}_0$. Since $\mathcal{A}_0$ is a vector space we conclude that $$\label{WI2}
\varphi_k:=\widetilde{\Phi}(\rho_k\xi_k)+\xi_k\in \mathcal{A}_0.$$ Also, exploiting the bilinearity of $\Gamma$ we get $$\label{zero99}
\Gamma(\varphi_k)=\Gamma\big(\widetilde{\Phi}(\rho_k\xi_k)+\xi_k\big)
=\Gamma\big(\widetilde{\Phi}(\rho_k\xi_k)\big)
+2\Gamma\big(\widetilde{\Phi}(\rho_k\xi_k),\xi_k\big)+\Gamma(\xi_k).$$ Moreover, using formula with $\Psi:=\widetilde{\Phi}$, we obtain that $$\begin{split}\label{1}
\Gamma(\widetilde{\Phi}(\rho_k\xi_k))&=(\widetilde{\Phi}'(\rho_k\xi_k))^2\Gamma(\rho_k\xi_k)\\
&=(\Phi'(\rho_k\xi_k))^2\Big(\xi_k^2\Gamma(\rho_k)+
2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+\rho_k^2\Gamma(\xi_k)\Big)
\end{split}$$ and $$\begin{split}\label{2}
\Gamma\big(\widetilde{\Phi}(\rho_k\xi_k),\xi_k\big)&=\Phi'(\rho_k\xi_k)\Gamma(\rho_k\xi_k, \xi_k)\\
&=\Phi'(\rho_k\xi_k)\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\,\Gamma(\xi_k)\Big).
\end{split}$$ Plugging and into , we have that $$\begin{split}\label{tre}
\Gamma(\varphi_k)
=\;&(\Phi'(\rho_k\xi_k))^2\Big(\xi_k^2\Gamma(\rho_k)+
2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+\rho_k^2\Gamma(\xi_k)\Big)\\&\qquad
+ 2\Phi'(\rho_k\xi_k)\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\Gamma(\xi_k)\Big)
+\Gamma(\xi_k).
\end{split}$$ Now, taking $\varphi:=\varphi_k$ into and making use of , we get that $$\begin{split}\label{ineqk:PRE}
& \int_{\{ \Gamma(u)\neq 0\}}\left(\Gamma_2(u)
-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\varphi_k^2\, d\mu\\
\leq\;& \int_{X} \Gamma(u)\Big[(\Phi'(\rho_k\xi_k))^2\Big(
\xi_k^2\Gamma(\rho_k)+2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+
\rho_k^2\Gamma(\xi_k)\Big)\\ &\qquad\qquad+ \Phi'(\rho_k\xi_k)
\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\,\Gamma(\xi_k)\Big)+
\Gamma(\xi_k)\Big]\, d\mu.
\end{split}$$ We also point out that, in view of and , and recalling and , $$\begin{split}\label{0q3481-1-P}
& \int_{X} \Gamma(u)\Big[(\Phi'(\rho_k\xi_k))^2\Big(
\xi_k^2\Gamma(\rho_k)+2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+
\rho_k^2\Gamma(\xi_k)\Big)\\ &\qquad\qquad+ \Phi'(\rho_k\xi_k)
\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\,\Gamma(\xi_k)\Big)+
\Gamma(\xi_k)\Big]\, d\mu\\
\le\;&
\int_{X_{R,k}} \Gamma(u)\left[9\left(
C_0+2\rho_k\xi_k\,|\Gamma(\rho_k,\xi_k)|+
\frac{16\rho_k^2\xi_k^2}{k}\right)+ 3
\left(|\Gamma(\rho_k,\xi_k)|+\frac{16\rho_k\xi_k}{k}\right)\right]\,d\mu\\
&\qquad+\frac1k\int_X \Gamma(u)\,d\mu,
\end{split}$$ where $$\label{0138400}
X_{R,k} := \{ x\in X{\mbox{ s.t. }} \rho_k(x)\,\xi_k(x)\in[R,R+1]\}.$$ [F]{}rom and , we have $$\begin{split}\label{0q3481-1}
& \int_{X} \Gamma(u)\Big[(\Phi'(\rho_k\xi_k))^2\Big(
\xi_k^2\Gamma(\rho_k)+2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+
\rho_k^2\Gamma(\xi_k)\Big)\\ &\qquad\qquad+ \Phi'(\rho_k\xi_k)
\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\Gamma(\xi_k)\Big)+
\Gamma(\xi_k)\Big]\, d\mu\\
\le\;&
\int_{X_{R,k}} \Gamma(u)\left[9\left(
C_0+2(R+1)\,|\Gamma(\rho_k,\xi_k)|+
\frac{16(R+1)^2}{k}\right)+ 3
\left(|\Gamma(\rho_k,\xi_k)|+\frac{16(R+1)}{k}\right)\right]\,d\mu\\
&\qquad+\frac1k\int_X \Gamma(u)\,d\mu.
\end{split}$$ Now, we observe that the following Cauchy-Schwarz inequality for $\Gamma$ holds true: given $f$, $g\in{\mathcal{A}}$ and $\alpha>0$, the fact that $\Gamma(\cdot,\cdot)$ is a symmetric bilinear map implies that $$0\le \Gamma\left(\alpha f\pm\frac1\alpha g\right)=
\Gamma\left(\alpha f\right)+\Gamma\left(\frac1\alpha g\right)
\pm2\Gamma\left(\alpha f,\frac1\alpha g\right)
=\alpha^2
\Gamma(f)+\frac1{\alpha^2}\Gamma(g)
\pm2\Gamma\left(f, g\right)$$ and therefore $$2\left|\Gamma\left( f, g\right)\right|\le \alpha^2
\Gamma(f)+\frac1{\alpha^2}\Gamma(g).$$ Then, by choosing $\alpha$ appropriately, we obtain that $$\left|\Gamma\left( f, g\right)\right|\le \big(\Gamma(f)\big)^{\frac12}\big(
\Gamma(g)\big)^{\frac12}.$$ As a consequence, we infer from that $$\begin{aligned}
&& \int_{X} \Gamma(u)\Big[(\Phi'(\rho_k\xi_k))^2\Big(
\xi_k^2\Gamma(\rho_k)+2\rho_k\xi_k\Gamma(\rho_k,\xi_k)+
\rho_k^2\Gamma(\xi_k)\Big)\\ &&\qquad\qquad+ \Phi'(\rho_k\xi_k)
\Big(\xi_k\Gamma(\rho_k,\xi_k)+\rho_k\Gamma(\xi_k)\Big)+
\Gamma(\xi_k)\Big]\, d\mu\\
&\le&
\int_{X_{R,k}} \Gamma(u)\left[9\left(
C_0+\frac{32(R+1)\sqrt{C_0}}{\sqrt{k}}+
\frac{16(R+1)^2}{k}\right)+ 3
\left(\sqrt{ \frac{C_0}{{k}} }+\frac{16(R+1)}{k}\right)\right]\,d\mu\\
&&\qquad+\frac1k\int_X \Gamma(u)\,d\mu.\end{aligned}$$ We insert this information into and we take the limit as $k\to+\infty$: in this way, we obtain that $$\label{S-12:01}
\begin{split}
&
\lim_{k\to+\infty} \int_{\{ \Gamma(u)\neq 0\}}\left(\Gamma_2(u)
-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\varphi_k^2\, d\mu
\\ \le\;&\lim_{k\to+\infty}
\int_{X_{R,k}} \Gamma(u)\left[9\left(
C_0+\frac{32(R+1)\sqrt{C_0}}{\sqrt{k}}+
\frac{16(R+1)^2}{k}\right)+ 3
\left(\sqrt{\frac{C_0}{{k}}}+\frac{16(R+1)}{k}\right)\right]\,d\mu\\
&\qquad+\frac1k\int_X \Gamma(u)\,d\mu
\\ =\;&
9C_0\int_{X_{R}} \Gamma(u),
\end{split}$$ where we have used , and in the last step, and $$X_{R} := \{ x\in X{\mbox{ s.t. }} \rho(x)\in[R,R+1]\}.$$ We also notice that, by , and , $$\lim_{k\to+\infty}
\varphi_k=\widetilde{\Phi}(\rho)+1=\Phi(\rho),$$ and therefore we deduce from that $$\label{S-12:02}
\int_{\{ \Gamma(u)\neq 0\}}\left(\Gamma_2(u)
-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\big(\Phi(\rho)\big)^2\, d\mu
\le
9C_0\int_{X_{R}} \Gamma(u).$$ Noticing that $\Phi(\rho(x))=1$ for any $x\in X$ for which $\rho(x)\le R$, thanks to , and recalling , we can take the limit as $R\to+\infty$ in , obtaining that $$\label{S-12:03}
\int_{\{ \Gamma(u)\neq 0\}}\left(\Gamma_2(u)
-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)\right)\, d\mu
\le0.$$ Since the integrand in the left hand side of is nonnegative (recall ), this gives .
Now we suppose that $K>0$. Then, by Definition \[DEF2.3\], we have that $(X,\mu,\Gamma)$ also satisfies the $CD(K,\infty)$ condition with $K=0$. Then, from and , we infer that $
\Gamma(u)=0$ $\mu-$a.e. in $ X$. If in addition $\Gamma(u)\in C^0(X)$, we get $\Gamma(u)=0$ in $ X$, which in light of (7’) gives .
Applications to vector fields satisfying the Hörmander condition {#oqidwuwiefyetytyyt}
================================================================
Here we take $\mathcal{A}_0:=\mathrm{Lip}_c(\mathbb{R}^n)$ and $\mathcal{A}:=\mathrm{Lip}(\mathbb{R}^n)$. Moreover, we let $\eta\in\mathcal{A}$ and define $$d\mu:=e^{\eta} dx.$$ Let $Z_1,\ldots, Z_m$ be smooth vector fields in $\mathbb{R}^n$ with $$Z_j=\sum_{i=1}^n Z_i^j\partial_i.$$ We define $$\label{ewiegrbgv567u6576u}
Z_0 f:= \sum_{j=1}^m Z_j\eta\,Z_jf,
\qquad
\mathrm{div}{Z_j}:=\sum_{i=1}^n \partial_i Z_i^j\quad{\mbox{ and }}\quad
\Delta_{Z}:=\sum_{j=1}^m Z_jZ_j.$$ We assume that the family $(Z_1,\ldots, Z_m)$ satisfies the Hörmander condition: at any point $x\in\mathbb{R}^n$, consider the vector spaces $V_p$ generated by the vector fields $Z_j$ at $x$, namely $$\begin{aligned}
&V_1:=\mathrm{span}\{Z_j\ |\ 1\leq j\leq m\},\\
&V_2:=\mathrm{span}\{Z_j, [Z_j,Z_k]\ |\ 0\leq j,k\leq m\},\\
&\cdots\\
&V_d:=\mathrm{span}\{V_{d-1}\cup \{[Z_j,V]\ |\ V\in V_{d-1},\
0\leq j\leq m\},\end{aligned}$$ then there exists $d\in\mathbb{N}$ such that, for any $x\in\mathbb{R}^n$, $V_d=\mathbb{R}^n$.
We recall the following result from [@GN]:
\[Lip\] Let $Z=(Z_1,\ldots, Z_m)$ be a family of smooth vector fields in $\mathbb{R}^n$ satisfying the Hörmander condition and let $d$ be the associated Carnot-Carathéodory distance [@gromov], which we assume to be continuous with respect to the Euclidean topology. If $f: \mathbb{R}^n\to \mathbb{R}$ is a function such that, for some $\Lambda\geq 0$, $$\label{sss3}
|f(x)-f(y)|\leq \Lambda d(x,y)\qquad \mbox{for all}\ x,y\in\mathbb{R}^n,$$ then the derivatives $Z_j f$, with $j=1,\ldots, m$, exist in distributional sense, are measurable functions and $|(Z f)(x)|\leq\Lambda$ for a.e. $x\in\mathbb{R}^n$.
With the notation introduced in , we consider the linear operator $$\label{LHO}
Lg=\sum_{j=1}^m Z_j g\, \mathrm{div}{Z_j} +\Delta_{Z} g
+ Z_0 g.$$ We remark that $$\label{HOR1}
Z_j(fg)=\sum_{i=1}^m Z_i^j \partial_i(fg)=
\sum_{i=1}^m Z_i^j f \partial_i g+\sum_{i=1}^m Z_i^j g\partial_i f=
fZ_j g+gZ_j f.$$ Therefore $$Z_jZ_j(fg) = Z_j (fZ_j g+gZ_j)=f Z_jZ_j g+gZ_jZ_j f+2Z_j fZ_jg,$$ that is, recalling , $$\label{HOR2}
\Delta_Z(fg) = f \Delta_Z g + g\Delta_Z f+2\sum_{j=1}^m Z_j fZ_jg.$$ Moreover, implies that $$\label{ngrhnrhth54865u8}
Z_0 (fg)= fZ_0 g + gZ_0f.$$ In view of , , and , and recalling (4’), we define $$\begin{aligned}
\label{Gam}
\Gamma(f,g):=\sum_{j=1}^m Z_j f Z_j g.\end{aligned}$$ Notice that $$\Gamma(f)=\sum_{j=1}^m (Z_j f)^2= |Zf|^2.$$
We can now consider the diffusion Markov triple $(\mathbb{R}^n,\mu,\Gamma)$. We observe that, with this choice, properties (1),(2) and (3) are satisfied.
We now prove that condition (4) is also satisfied. For this, using an integration by parts, we point out that, for any $f\in C^\infty({{\mathbb{R}}}^n)$ and any $\varphi\in C^\infty_c({{\mathbb{R}}}^n)$, $$\begin{aligned}
&& \int_{{{\mathbb{R}}}^n} Z_j f\,\varphi\,d\mu =\sum_{i=1}^n \int_{{{\mathbb{R}}}^n} Z_i^j \,\partial_i f\,\varphi\,e^\eta\,dx
=-\sum_{i=1}^n \int_{{{\mathbb{R}}}^n} f\, \partial_i(Z_i^j \,\varphi\,e^\eta)\,dx \\
&&\qquad =-\sum_{i=1}^n \int_{{{\mathbb{R}}}^n} f\, (\partial_iZ_i^j \,\varphi\,e^\eta
+Z_i^j \,\partial_i\varphi\,e^\eta+Z_i^j \,\varphi\,\partial_i\eta\,e^\eta)\,dx \\
&&\qquad=
-\int_{{{\mathbb{R}}}^n} f\, (\varphi\,{\rm div}Z_j
+Z_j \varphi+Z_j\eta \,\varphi)\,d\mu . \end{aligned}$$ Taking $\varphi:=Z_j g$, we thereby conclude that, for every $f\in C^{\infty}(\mathbb{R}^n)$ and $g\in C_c^{\infty}(\mathbb{R}^n)$ (and, more generally, for every $f\in\mathrm{Lip}(\mathbb{R}^n)$ and $g\in\mathrm{Lip}_c(\mathbb{R}^n)$ by a density argument), it holds that $$\begin{aligned}
\int_{\mathbb{R}^n} \Gamma(f,g)\ d\mu &=&\sum_{j=1}^m
\int_{\mathbb{R}^n} Z_j f \,Z_j g\,d\mu\\&=&
-\int_{\mathbb{R}^n} f\left(\sum_{j=1}^m Z_j g\,
\mathrm{div}{Z_j} +\Delta_{Z} g+Z_0g\right) \,d\mu\\
&=&-\int_{\mathbb{R}^n} f\,Lg\,d\mu ,\end{aligned}$$ which is (4) in this setting.
We now prove condition (5). We denote by $d(x)=d(x,0)$, and we consider a function $\Phi\in C^{\infty}({{\mathbb{R}}}, [0,1])$, with $|\Phi'(t)|\leq 1$ for any $|t|\in [1/8, 1/4]$, and $$\Phi(t):=\left\{\begin{array}{lll}
1 & \mbox{if}& |t|\leq 1/8,\\
0 & \mbox{if}& |t|\geq 1/4.
\end{array}
\right.$$ For every $k\in\mathbb{N}$, we define $
\xi_k(x):=\Phi(d(x)^2/k^2)\in \mathrm{Lip}_c(\mathbb{R}^n)$. Then $(\xi_k)_{k\in\mathbb{N}}$ is an increasing sequence with $
\xi_k(x)\in [0,1]$ and $\xi_k(x)\to 1$ as $k\to+\infty$ for every $x\in\mathbb{R}^n$. Moreover, we observe that in Theorem \[Lip\] is satisfied taking $f:=d$ with $\Lambda=1$, and therefore we have that $|Z d|\le 1$. As a consequence $$|\Gamma(\xi_k)|= |Z\xi_k| = \frac{2|\Phi'(d^2/k^2)|}{k^2}d
|Zd|\leq \frac{1}{k},$$ which completes the proof of (5).
Notice that also (1’), (2’), (3’), (4’), (5’) and (6’) easily follow from the very definition of $\mathcal{A}$, $L$ and $\Gamma$.
We claim that also (7’) holds. Indeed, if $f\in{\mathcal{A}}$ with $\Gamma(f)=0$, then, by definition, $$\sum_{j=1}^m (Z_j f)^2 =0,$$ and so we have that $Z_j f=0$ for any $ j=1,\ldots, m$. Thus, also all iterated derivatives vanish. Then, the conclusion follows, since, by the Hörmander condition, every $\partial_{x_i}f$ can be written as a linear combination of iterated derivatives. Therefore, $f$ is a constant function, and so condition (7’) is satisfied.
We now describe some interesting applications of the setting introduced above.\
Riemannian Manifolds
--------------------
Let $(M,g)$ be a connected Riemannian manifold of dimension $n$ equipped with the standard Levi-Civita connection $\nabla$ and let $G\in C^2(M)$. As customary in Riemannian geometry, we define the gradient vector of $f\in C^{\infty}(M)$ as the vector field whose coordinates are $$\nabla^i f=g^{ij}\partial_{x_j} f, \quad {\mbox{ for any }}\,1\leq i\leq n,$$ where $g^{ij}$ are the coefficients of the inverse matrix $(g_{ij})_{1\leq i,j\leq n}$, and the repeated indices notation has been used. We consider the Markov triple $(M, \mu,\Gamma)$, where $$\Gamma(f,g):=\nabla^i f\partial_{x_i} g.$$ and $$\mu:=e^{-G} dV.$$ Here $dV$ denotes the Riemannian volume element, namely, in local coordinates, $$dV = \sqrt{|g|}\,dx^1\wedge \dots \wedge dx^m,$$ where $\{ dx^1,\dots,dx^n\}$ is the basis of $1$-forms dual to the vector basis $\{\partial_1,\dots,\partial_n\}$. The Laplace-Beltrami operator $\Delta_g$ is defined on $
f\in C^{\infty}(M)$ as $$\Delta_g f=\frac{1}{\sqrt{|g|}}\partial_{x_i}\left(\sqrt{|g|}g^{ij}\partial_{x_j}
f\right),$$ whereas the Hessian matrix $\nabla^2 f$ of a smooth function $f$ is defined as the symmetric $2$-tensor given in a local patch by $$(\nabla^2 f)_{ij}=\partial^2_{ij}f-\Gamma^k_{ij}\partial_k f,$$ where $\Gamma^k_{ij}$ are the Christoffel symbols, namely $$\Gamma_{ij}^k=\frac12 g^{hk} \left( \partial_i g_{hj} +\partial_j g_{ih} -\partial_h g_{ij} \right) .$$ Given a tensor $A$, we also define its norm by $|A|=\sqrt{A A^*}$, being $A^*$ the adjoint of $A$.
As proved in [@BLG 1.11.10] and [@BLG 1.16.4], we see that $$\label{0swo:1}
Lf=\Delta_gf-\left\langle \nabla G,\nabla f\right\rangle_g$$ and $$\label{G GA2}
\Gamma_2(f)=|\nabla^2 f|+\mathrm{Ric}(L)(\nabla f,\nabla f),$$ where $\mathrm{Ric}(L)$ is a symmetric tensor defined from the Ricci tensor $\mathrm{Ric}_g$ by $$\mathrm{Ric}(L)=\mathrm{Ric}_g+\nabla^2 G.$$ Observing that $$\Gamma\left(\Gamma(f)^{\frac{1}{2}}\right)=|\nabla|\nabla f||^2,$$ we use and conclude that $$\label{G GA3}
\Gamma_2(u)-\Gamma\left(\Gamma(u)^{\frac{1}{2}}\right)=
|\nabla^2 u|+\mathrm{Ric}(L)(\nabla u,\nabla u)-|\nabla|\nabla u||^2.$$ Consequently, for any stable weak solution $u\in C^{\infty}(M)$ to $$\label{EQM}
Lu+F(u)=0 \quad{\mbox{ in }}\; M,$$ the Poincaré inequality in of Theorem \[eqnlin\] reads as follows: $$\begin{aligned}
\label{GFRiem}
\int_M\Big(|\nabla^2 u|+\mathrm{Ric}(L)(\nabla u,\nabla u)
-|\nabla|\nabla u||^2\Big)\varphi^2 d\mu\leq \int_{M} |\nabla u|^2
|\nabla \varphi|^2 d\mu,\end{aligned}$$ for any $\varphi\in C^{\infty}_c(M)$.
In particular, on ${{\mathbb{R}}}^n$ equipped with the flat Euclidean metric, and for the usual Laplacian $\Delta$, inequality reads as $$\int_{{{\mathbb{R}}}^n}\Big(|\nabla^2 u|^2+\nabla^2 G(\nabla u,\nabla u)
-|\nabla|\nabla u||^2\Big)\varphi^2 d\mu\leq \int_{{{\mathbb{R}}}^n}
|\nabla u|^2|\nabla \varphi|^2 d\mu,$$ for any $\varphi\in C^{\infty}_c({{\mathbb{R}}}^n)$, which is precisely the inequality already proved in [@FarHab; @CNP; @CNV; @FNP].
Furthermore, if $G:= 0$ then was proved in [@fsv1; @fsv2; @FMV], whereas the general case seems to be new in the literature.
In our setting, we take $\mathcal{A}_0:=\mathrm{Lip}_c(M)$ and $\mathcal{A}:=\mathrm{Lip}(M)$, where $f\in\mathrm{Lip}(M)$ if $$\label{LIL0}
\sup_{x\ne y\in M}\frac{f(x)-f(y)}{d(x,y)}<+\infty,$$ being $d$ the distance defined in . Notice that, if we fix $x_0\in M$ and we set $
\rho(x):=d(x,x_0)$, it holds that $$\label{LIL}
\rho(x)-\rho(y)\le d(x,y).$$ Indeed, by , for any ${\epsilon}>0$ there exists $f_{\epsilon}$ with $\Gamma(f_{\epsilon})\le1$ such that $\rho(x)\le {\epsilon}+ f_{\epsilon}(x)
-f_{\epsilon}(x_0)$, and therefore $$\rho(x)-\rho(y)\le
{\epsilon}+ f_{\epsilon}(x)-f_{\epsilon}(x_0) - \big( f_{\epsilon}(y)-f_{\epsilon}(x_0)\big)
={\epsilon}+f_{\epsilon}(x)-f_{\epsilon}(y)\le{\epsilon}+d(x,y).$$ [F]{}rom this, taking ${\epsilon}$ as small as we wish, we obtain .
Then, comparing and , we see that $\rho\in
\mathrm{Lip}(M)=\mathcal{A}$.
We also assume that $$\lambda\delta^{ij}\le g^{ij}\le \frac{ \delta^{ij} }\lambda,$$ for some $\lambda\in(0,1]$. In this way, if $|\cdot|_E$ is the Euclidean norm of a vector, it holds that $$\lambda \,|v|_E^2 =\lambda \delta^{ij}v_i v_j\le g^{ij} v_i v_j.$$ Then, by , for any ${\epsilon}>0$ there exists $\tilde f_{\epsilon}$ with $\Gamma(\tilde f_{\epsilon})\le1$ such that $$\begin{aligned}
d(x,y) &\le&
{\epsilon}+ \tilde f_{\epsilon}(x)-\tilde f_{\epsilon}(y)\\
&=&
{\epsilon}+ \int_0^1 \partial_{x_i}\tilde f_{\epsilon}(x+ty)\cdot(x_i-y_i)\,dt\\
&\le&
{\epsilon}+ \frac1\lambda\,
\int_0^1 \sqrt{g^{ij}\partial_{x_i}\tilde f_{\epsilon}(x+ty)\,
\partial_{x_j}\tilde f_{\epsilon}(x+ty)}\,|x-y|_E\,dt
\\&=&
{\epsilon}+ \frac1\lambda\,
\int_0^1 \sqrt{\Gamma(\tilde f_{\epsilon})(x+ty)}\,|x-y|_E\,dt
\\&=&
{\epsilon}+ \frac{|x-y|_E}\lambda.\end{aligned}$$ Hence, taking ${\epsilon}$ arbitrary small, it follows that $d(x,y)\le\frac{|x-y|_E}\lambda$. Recalling , we thereby conclude that $|\nabla \rho|$, and therefore $\Gamma(\rho)$, is bounded by a universal constant depending on $\lambda$. Consequently, the assumptions of Theorem \[main1\] are satisfied. Then, using Theorem \[main1\] and , we obtain that, in this setting, if the curvature dimension condition $CD(K,\infty)$ holds true for some $K\geq 0$, and $u$ is a stable solution with $\int_M |\nabla u|^2\,d\mu<+\infty$, then: $$\begin{aligned}
&K>0\Longrightarrow {\mbox{$u$ is constant in~$M$}};\\
\label{KPOSI0-B}
&K=0 \Longrightarrow
|\nabla^2 u|+\mathrm{Ric}(L)(\nabla u,\nabla u)-|\nabla|\nabla u||^2
=0\qquad\mu-\mbox{a.e. in }M.\end{aligned}$$ To grasp a geometric flavor of , one can fix a point $p
\in M$ with $\nabla u(p)\ne0$ and consider normal coordinates at $p$ for which $$\label{NORMALC}
g^{ij}(p)=\delta^{ij},\qquad \partial_{x_k}g^{ij}(p)=0\qquad{\mbox{
and }}\qquad\Gamma_{ij}^k(p)=0,$$ see e.g. page 55 in [@2009]. Then, the level set $S$ of $ u$ passing through $p$ is locally a submanifold of $M$ of codimension $1$, endowed with a Riemannian structure induced by that of $M$ (namely if $v,w\in T_p S\subseteq T_p M$ one can consider $g(v,w)$ as defining a metric on $S$). Consequently, in view of , we can reduce the Riemannian term $
|\nabla^2 u|-|\nabla|\nabla u||^2$ to its Euclidean counterpart, which, due to the classical Sternberg-Zumbrun identity (see formula (2.1) of [@SZ2]) is larger than $K^2\,|\nabla u|^2$, being $K^2$ the sum of the square of the eigenvalues of the second fundamental form of $S$, according to the induced Riemannian structure, see Proposition 18 in [@FMV]. Therefore, by , if the Ricci tensor is nonnegative, it follows that the second fundamental form of $S$ at $p$ vanishes, and the Ricci tensor must vanish at $p$ as well.
Submanifolds with vanishing second fundamental form are called totally geodesic (see e.g. page 104 in [@ONE] or Proposition 1.2 in [@RADE]) and are characterized by the property that any geodesic on the submanifold is also a geodesic on the ambient manifold.
Carnot groups
-------------
We recall that a Carnot group $\mathbb{G}$ is a connected Lie group whose Lie algebra $\mathcal{G}$ is finite dimensional and stratified of step $s\in\mathbb{N}$. Precisely, there exist linear subspaces $V_1,\dots,V_s$ of $\mathcal{G}$ such that $$\mathcal{G}=V_1\oplus \cdots \oplus V_s$$ with $$[V_1,V_{i-1}]=V_{i}\;\; \mbox{ if }2\leq i\leq s \quad
\mbox{ and }\quad [V_1,V_s]=\{0\}.$$ Here $[V_1,V_i]:=\mathrm{span}\{[a,b]: a\in V_1,\ b\in V_i\}.$ Since $\mathcal{G}$ is stratified, then every element of $\mathcal{G}$ is the linear combination of commutators of elements of $V_1$. We refer to [@BLU] for a complete introduction to the subject.
Let $\mathrm{dim}(V_1)=m$ and $Z=(Z_1,\ldots, Z_m)$ be a basis of $V_1$. The family $Z$ satisfies the Hörmander condition. Moreover, in this setting, $$\label{feejiyh5}
\Gamma(f,g)=\sum_{i=1}^m Z_i f Z_i g\quad {\mbox{ and }}
\quad L f= \Delta_Z f.$$ In order to compute $\Gamma_2$ we will use the following Bochner-type formula proved in [@G Proposition 3.3] coupled with [@G Lemma 3.1]:
\[bochner\] Let $u$ be a smooth function. Then, $$\begin{aligned}
&&\frac12 \Delta_Z |Zu|^2\\
&=&\|Z^2 u \|^2+\sum_{j=1}^m Z_j u\,
Z_j(\Delta_Zu)+ 2\sum_{i,j=1}^m
Z_j u [Z_i,Z_j]Z_iu +\sum_{i,j=1}^m Z_j u [Z_i,[Z_i,Z_j]]u ,\end{aligned}$$ where $Z^2 u$ denotes the horizontal Hessian matrix associated to the family $Z$, namely the $m\times m$ matrix whose elements are given by $
u_{ij}:=Z_iZ_j u$, with $i,j=1,\dots, m$.
Now, let $u\in C^{\infty}(\mathbb{G})$ be a stable solution to $$\label{8768bnfkle}
\Delta_Z u+F(u)=0\quad {\mbox{ in }}\mathbb{G}.$$ Therefore, recalling and using Theorem \[bochner\], $$\begin{aligned}
&\Gamma_2(u) =\frac12L\Gamma (u) -\Gamma(u,Lu)=
\frac{1}{2}\Delta_Z |Zu|^2- \Gamma(u, \Delta_Z u)\\
&\quad
= \|Z^2 u \|^2+\sum_{j=1}^m Z_j u\,
Z_j(\Delta_Zu)+ 2\sum_{i,j=1}^m
Z_j u [Z_i,Z_j]Z_iu +\sum_{i,j=1}^m Z_j u [Z_i,[Z_i,Z_j]]u
- \Gamma(u, \Delta_Z u)\\
&\quad=
\|Z^2 u \|^2+ 2\sum_{i,j=1}^m
Z_j u [Z_i,Z_j]Z_iu +\sum_{i,j=1}^m Z_j u [Z_i,[Z_i,Z_j]]u
\\&\quad
=\|Z^2 u \|^2+\mathcal{R}(u),\end{aligned}$$ where $$\mathcal{R}(u):=2\sum_{i,j=1}^m Z_j u [Z_i,Z_j]Z_iu +\sum_{i,j=1}^m Z_j u [Z_i,[Z_i,Z_j]]u.$$ Therefore, if $u\in C^{\infty}(\mathbb{G})$ is a stable solution to , inequality reads as $$\begin{aligned}
\label{Carnot}
\int_{\mathbb{G}}\left(\|Z^2 u \|^2-|Z|Z u||^2
+\mathcal{R}(u)\right) \varphi^2\ dx\leq
\int_{\mathbb{G}} |Z u|^2 |Z \varphi|^2\ dx,\end{aligned}$$ for any $\varphi\in C^{\infty}_c(\mathbb{G})$.
Formula generalizes to general Carnot groups the Poincaré inequality obtained in [@FV1] in the Heisenberg group and in [@PV] in the Engel group (we refer the reader to [@BLU] for the definitions, and we remark that the divergence of the Heisenberg and Engel vector fields vanish in the setting of ). In particular, in the case of the Heisenberg group, formula here reduces to formula (7) in [@FV1], and in the case of the Engel group, formula here reduces to the formula in Proposition 3.7 of [@PV] and ${\mathcal{R}}$ here coincides with ${\mathcal{J}}$ in Theorem 1.1 of [@PV]. In its full generality, our formula seems to be new in the literature.
In addition, the distance in coincides with that of Carnot-Carathéodory in this setting, see [@BLG], and so holds true in this case. For completeness, we state and we apply Theorem \[main1\] to obtain this original result:
Let $\mathbb{G}$ be a Carnot group whose Lie algebra $\mathcal{G}=
V_1\oplus \cdots \oplus V_s$ is stratified of step $s$, with $V_1$ generated by the basis of vector fields $(Z_1,\dots,Z_m)$ that satisfy the Hörmander condition. Let $u\in C^\infty(\mathbb{G})$ be a stable weak solution to $\Delta_Zu+F(u)=0$ in $\mathbb{G}$. Then, $$\int_{\mathbb{G}}\left(\|Z^2 u \|^2-|Z|Z u||^2
+\mathcal{R}(u)\right) \varphi^2\ dx\leq
\int_{\mathbb{G}} |Z u|^2 |Z \varphi|^2\ dx,$$for any $\varphi\in C^{\infty}_c(\mathbb{G})$.
Assume also that $$\begin{aligned}
&& \mathcal{R}(u)\ge 0\\{\mbox{and }}&&
\int_{\mathbb{G}} |Zu|^2\, dx <\infty.\end{aligned}$$ Then $$\begin{aligned}
&& \mathcal{R}(u)= 0\\{\mbox{and }}&& \|Z^2 u\|^2=|Z|Z u||^2
\mbox{ a.e. in }\mathbb{G}.\end{aligned}$$
Now, we prove that, for a particular family of Carnot groups, formula provides a geometric inequality for every stable solution to . Model filiform groups are the Carnot groups with the simplest Lie brackets possible while still having arbitrarily large step, see [@Mon02]. They have previously been investigated in connection with non-rigidity of Carnot groups [@O08], quasiconformal mappings between Carnot groups [@Xia05], geometric control theory [@Mon02] and geometric measure theory [@PS].
The formal definition is as follows:
\[filiform\] Let $n\geq 2$. The *model filiform group of step $n-1$* is the Carnot group $\mathbb{E}_{n}$ whose Lie algebra $
\mathcal{E}_{n}$ admits a basis $Z_{1}, \ldots, Z_{n}$ satisfying $[Z_{i},Z_{1}]=Z_{i+1}$ for $1<i<n$, with all other Lie brackets among the $Z_{i}$ equal to zero.
The stratification of $\mathcal{E}_{n}$ is $$\mathcal{E}_{n}=V_{1}\oplus \cdots \oplus V_{n-1}$$ with $V_{1}=\mathrm{Span}\{Z_{1}, Z_{2}\}$ and $V_{i}=\mathrm{Span}\{Z_{i-1}\}$ for $1<i<n$.
Proceeding exactly as in [@FV1 formula (19)], we get $$|Z|Z u||^2=\frac{1}{|Z u|^2}
\left\langle H_Zu \,Z u, Zu\right\rangle\quad
\mbox{in}\ \{Z u\neq 0\},$$ where $$H_Z u:=Z^2 u(Z^2 u)^T.$$ Whenever $P\in \{u=k\}\cap \{Z u\neq 0\}$, we can consider the smooth surface $\{u=k\}$ and define the intrinsic normal to $\{u=k\}$ and the intrinsic unit tangent direction to $\{u=k\}$ as $$\nu:=\frac{Z u(P)}{|Z u(P)|}\quad \mbox{and}\quad v:=\frac{(Z_2 u(P), -Z_1u(P))}{|Z u(P)|},$$ respectively. We observe that [@FV1 Lemma 2.1] only depends on the fact that $\mathrm{dim} V_1=2$ and $\mathrm{dim} V_2=1$, which still hold in every model filiform Carnot group. Therefore, the following result holds:
On $\{u=k\}\cap \{Z u \neq 0\}$, it holds that $$\begin{aligned}
\label{ddw}
\|Z^2 u\|^2-\left\langle (H_Z u) \nu, \nu\right\rangle=|Z u|^2\left[h^2+ \left(p+\frac{\left\langle (Hu) v, \nu\right\rangle}{|Z u|}\right)^2\right]\end{aligned}$$ where $$h=\mathrm{div}_Z \nu=Z_1 \nu_1+Z_2 \nu_2$$ and $$p=-\frac{Z_3u}{|\nabla_Z u|}.$$
Plugging into we get the following geometric Poincaré inequality: $$\begin{aligned}
\int_{\{Zu\neq 0\}}\left(|Z u|^2\left[h^2+ \left(p+\frac{\left\langle (Hu) v, \nu\right\rangle}{|Z u|}\right)^2\right]
+\mathcal{R}(u)\right) \varphi^2\ dx\leq
\int_{\mathbb{E}_{n}} |Z u|^2 |Z \varphi|^2\ dx,\end{aligned}$$ for any $\varphi\in C^{\infty}_c(\mathbb{E}_{n})$.
We summarize this statement and that of Theorem \[main1\] in the following original result:
\[saq\] Let $\mathbb{E}_{n}$ be a model filiform group of step $n-1$, as in Definition . Let $u\in C^\infty(\mathbb{E}_{n})$ be a stable weak solution to $\Delta_Zu+F(u)=0$ in $\mathbb{E}_{n}$. Then, $$\begin{aligned}
\int_{\{Zu\neq 0\}}\left(
|Z u|^2\left[h^2+ \left(p+\frac{\left\langle (Hu) v, \nu\right\rangle}{|Z u|}\right)^2\right]
+\mathcal{R}(u)\right) \varphi^2\ dx\leq
\int_{\mathbb{E}_{n}} |Z u|^2 |Z \varphi|^2\ dx,\end{aligned}$$ for any $\varphi\in C^{\infty}_c(\mathbb{E}_{n})$.
Assume also that $$\begin{aligned}
&&
\mathcal{R}(u)
\ge 0\\{\mbox{and }}&&
\int_{\mathbb{E}_{n}} |Zu|^2\, dx <\infty.\end{aligned}$$ Then: $$\begin{aligned}
&& \nonumber
\mathcal{R}(u)
= 0\quad{\mbox{and }}\\&&
\left.\begin{matrix}
h=0\\
\\
p+\displaystyle\frac{\left\langle (Hu) v, \nu\right\rangle}{|Z u|}
=0\quad\end{matrix}\right\}\quad
\mbox{ a.e. in } {\mathbb{E}_{n}}\cap\{ Zu\ne0\}.\label{09222133}\end{aligned}$$
The two equations in on $ \{ Zu\ne0\}$ can be seen as “intrinsic geodesic equations” on the noncritical level sets of the solution $u$.
Grushin plane
-------------
For a given $\alpha\in \mathbb{N}$, the vector fields $Z_1=\partial_x$ and $Z_2=|x|^{\alpha}\partial_y$ satisfy the Hörmander condition in $\mathbb{R}^2$. We call Grushin plane the metric space $\mathbb{G}_{\alpha}
=(\mathbb{R}^2, d)$, where $d$ is the Carnot-Carathéodory distance induced by $Z_1$ and $Z_2$. Background on the Grushin plane may be found in [@Bel; @Mon02].
Since $\mathrm{div}Z_1=\mathrm{div}Z_2=0$ and $Z_0=0$, then, from what we proved for vector fields satisfying the Hörmander condition, we get $$Lu=\Delta_Z u,\quad \Gamma(f,g)=Z_1 f Z_1 g+Z_2 fZ_2 g\quad
\mbox{and}\quad d\mu= dx.$$ For every solution $u\in C^{\infty}(\mathbb{G}_{\alpha})$ to $\Delta_Z u+ F(u)=0$, we see that, for any $i\in\{1,2\}$, $$\label{9yhn83tugrgbe}
Z_i\Delta_Z u= -F'(u)\,Z_i u.$$ Let also $Z_3:=[Z_1,Z_2]$. We observe that, by [@FV3 Lemma 2.1] and , we have $$\begin{aligned}
\label{ddsa}
&\Delta_Z Z_1 u=Z_1\Delta_Z u- 2Z_3Z_2 u
= -F'(u)\,Z_1 u- 2Z_3Z_2 u
\\
\label{ddsa1} {\mbox{and }}\quad
&\Delta_Z Z_2 u=Z_2\Delta_Z u+ 2Z_3Z_1 u=-F'(u)\,Z_2 u+ 2Z_3Z_1 u.\end{aligned}$$ As a further consequence of , we obtain that $$\begin{split}\label{rekhtbfndb}
& \Gamma(u,\Delta_Z u)=
Z_1 u Z_1\Delta_Z u + Z_2 u Z_2\Delta_Z u\\&\qquad=
-F'(u)\left( |Z_1u|^2 +|Z_2u|^2\right)
=-F'(u)|Zu|^2 .
\end{split}$$ Moreover, direct calculations give $$\begin{aligned}
Z_1Z_1(Z_1u)^2&=2(Z_1Z_1u)^2+2Z_1 uZ_1Z_1Z_1 u\\
&=2(Z_1Z_1u)^2+2Z_1 u \Delta_Z Z_1 u-2Z_1u Z_2Z_2Z_1u,\\
Z_2Z_2(Z_1u)^2&=2(Z_2Z_1u)^2+2Z_1u Z_2Z_2Z_1 u\\
Z_1Z_1(Z_2u)^2&=2(Z_1Z_2u)^2+2Z_2u Z_1Z_1Z_2 u,\\
{\mbox{and }}\quad
Z_2Z_2(Z_2u)^2&=2(Z_2Z_2u)^2+2Z_2 uZ_2Z_2Z_2 u
\\&=2(Z_2Z_2u)^2+2Z_2 u \Delta_Z Z_2u -2Z_2u Z_1Z_1Z_2 u.\end{aligned}$$ Summing up these equalities and recalling and , we get $$\begin{aligned}
\Delta_Z|Zu|^2&=2\|Z^2 u\|+2Z_1 u \Delta_Z Z_1 u+2Z_2 u \Delta_Z Z_2u\\
&=2\|Z^2 u\|+2Z_1u (-F'(u)Z_1 u -2Z_3Z_2 u)+2Z_2 u(-F'(u)Z_2 u+ 2Z_3Z_1 u)\\
&=2\|Z^2 u\| -2F'(u)|Zu|^2 -4Z_1uZ_3Z_2 u+ 4Z_2uZ_3Z_1 u.\end{aligned}$$ Using this and , we are now in position to compute $\Gamma_2(u)$ as follows: $$\begin{aligned}
\Gamma_2(u)&=\frac{1}{2}L\Gamma(u)-\Gamma(u, Lu)
=\frac{1}{2} \Delta_Z|Z u|^2-\Gamma(u,\Delta_Z u)\\
&=\|Z^2 u\| -2Z_1uZ_3Z_2 u+ 2Z_2uZ_3Z_1 u.\end{aligned}$$ Therefore, if $u\in C^{\infty}(\mathbb{G}_{\alpha})$ is a stable solution to $\Delta_Z u+ F(u)$, inequality reads as $$\begin{aligned}
\int_{\mathbb{G}_{\alpha}}\Big(\|Z^2 u \|^2-|Z|Z u||^2
-2Z_1uZ_3Z_2 u+ 2Z_2uZ_3Z_1 u\Big) \varphi^2\ dx\leq
\int_{\mathbb{G}_{\alpha}} |Z u|^2 |Z \varphi|^2\ dx,\end{aligned}$$ for any $\varphi\in C^{\infty}_c(\mathbb{G}_{\alpha})$, which coincides with formula (1.10) in Theorem 1.1 of [@FV3].
We conclude pointing out that, since the proof of Theorem \[saq\] is based on the commutator relations of the vectors $Z_1$ and $Z_2$, the same result also holds in the Grushin plane (compare with (1.15) and (1.16) of [@FV3]).
[100]{}
Ambrosio, L., Trevisan, D.: *Well posedness of Lagrangian flows and continuity equations in metric measure spaces*, Anal. PDE **7**, no. 5, 1179–1234 (2014).
Bakry, D. Gentil, I., Ledoux, M.: *Analysis and Geometry of Markov Diffusion Operators*, Grundlehren der mathematischen Wissenschaften, Springer, 2014.
Bandle, C., Mastrolia, P., Monticelli, D. D., Punzo, F.: *On the stability of solutions of semilinear elliptic equations with Robin boundary conditions on Riemannian manifolds*, SIAM J. Math. Anal. **48**, no. 1, 122–151 (2016).
Bandle, C., Punzo, F., Tesei, A.: *Existence and nonexistence of patterns on Riemannian manifolds*, J. Math. Anal. Appl., **387**, 33–47 (2012).
Bellaïche, A.: *The tangent space in sub-Riemannian geometry. In Sub-Riemannian geometry*, volume 144 of Progr. Math., pages 1–78. Birkhäuser, Basel, 1996.
Birindelli, I., Ferrari, F., Valdinoci, E.: *Semilinear PDEs in the Heisenberg group: the role of the right invariant vector fields*, Nonlinear Anal. **72**, no. 2, 987–997 (2010).
Bonfiglioli, A., Lanconelli, E., Uguzzoni, F.: *Stratified Lie groups and Potential theory for their Sub-Laplacians*, Springer Monographs in Mathematics, 26. New York, NY: Springer-Verlag, 2007.
Bouleau, N.: *Opérateur carré du champ, subordinateurs et processus de Dirichlet*, Seminar on potential theory, Paris, no. 7, 82–91, Lecture Notes in Math., 1061, Springer, Berlin, 1984.
Bouleau, N., Hirsch, F.: *Dirichlet forms and analysis on Wiener spaces*, De Gruyter studies in Mathematics **14**, De Gruyter, 1991.
Cesaroni, A., Novaga, M., Valdinoci, E.: *A simmetry result for the Ornstein-Uhlenbeck operator*, Discrete Contin. Dyn. Syst. **34**, no. 6, 2451–2467 (2014).
Cesaroni, A., Novaga, M., Pinamonti, A.: *One-dimensional symmetry for semilinear equations with unbounded drift*, Commun. Pure Appl. Anal. **12**, no. 5, 2203–2211 (2013).
Cheng, H.-C., Hsieh, M.-H., Tomamichel, M.: *Exponential decay of matrix $\Phi$-entropies on Markov semigroups with applications to dynamical evolutions of quantum ensembles*, J. Math. Phys. **58**, no. 9, 092202, 24 pp. (2017).
Chermisi, M., Valdinoci, E.: *A symmetry result for a general class of divergence form PDEs in fibered media*, Nonlinear Anal. **73**, no. 3, 695–703 (2010)
Dipierro, S.: *Geometric inequalities and symmetry results for elliptic systems*, Discrete Contin. Dyn. Syst. **33**, no. 8, 3473–3496 (2013).
Dipierro, S., Pinamonti, A.: *A geometric inequality and a symmetry result for elliptic systems involving the fractional Laplacian*, J. Differential Equations **255**, no. 1, 85–119. (2013).
Dipierro, S., Pinamonti, A.: *Symmetry results for stable and monotone solutions to fibered systems of PDEs*, Commun. Contemp. Math. **17**, no. 4, 1450035, 22 pp. (2015).
Dipierro, S., Pinamonti, A., Valdinoci, E.: *Rigidity results for elliptic boundary value problems*, Submitted. ArXiv:1709.07934.
Dipierro, S., Pinamonti, A., Valdinoci, E.: *Classification of stable solutions for boundary value problems with nonlinear boundary conditions on Riemannian manifolds with nonnegative Ricci curvature*, Submitted. ArXiv:1710.07329.
Dupaigne, L.: *Stable solutions of elliptic partial differential equations*. Chapman & Hall/CRC, Boca Raton, FL, 2011.
Farina, A.: *Propriétés qualitatives de solutions d’équations et systèmes d’équations non-linéaires*, Habilitation à diriger des recherches, Paris VI, (2002).
Farina, A., Mari, L., Valdinoci, E.: *Splitting theorems, symmetry results and overdetermined problems for Riemannian manifolds*, Comm. Partial Differential Equations **38**, no. 10, 1818–1862 (2013).
Farina, A., Novaga, M., Pinamonti, A.: *Symmetry results for nonlinear elliptic operators with unbounded drift*, NoDEA Nonlinear Differential Equations Appl. **21**, no. 6, 869–883 (2014).
Farina, A., Sire, Y., Valdinoci, E.: *Stable solutions of elliptic equations on Riemannian manifolds*, J. Geom. Anal. **23**, no. 3, 1158–1172 (2013).
Farina, A., Sire, Y., Valdinoci, E.: *Stable solutions of elliptic equations on Riemannian manifolds with Euclidean coverings*, Proc. Amer. Math. Soc. **140**, no. 3, 927–930 (2012).
Farina, A., Sciunzi, B., Valdinoci, E.: *Bernstein and De Giorgi type problems: new results via a geometric approach*, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) **7**, no. 4, 741–791 (2008).
Fazly, M.: *Stable solutions of symmetric systems on Riemannian manifolds*, Proc. Amer. Math. Soc. **145**, no. 12, 5435–5449 (2017).
Fazly, M., Ghoussoub, N.: *De Giorgi type results for elliptic systems*, Calc. Var. Partial Differential Equations **47**, no. 3–4, 809–823 (2013).
Fazly, M., Sire, Y.: *Symmetry results for fractional elliptic systems and related problems*, Comm. Partial Differential Equations **40**, no. 6, 1070–1095 (2015).
Ferrari, F., Pinamonti, A.: *Nonexistence results for semilinear equations in Carnot groups*, Anal. Geom. Metr. Spaces **1**, 130–146 (2013).
Ferrari, F., Valdinoci, E.: *A geometric inequality in the Heisenberg group and its applications to stable solutions of semilinear problems*, Math. Ann. **343**, no. 2, 351–370 (2009).
Ferrari, F., Valdinoci, E.: *Some weighted Poincaré inequalities*, Indiana Univ. Math. J. **58**, no. 4, 1619–1637 (2009).
Ferrari, F., Valdinoci, E.: *Geometric PDEs in the Grushin plane: weighted inequalities and flatness of level sets*, Int. Math. Res. Not. IMRN, no. 22, 4232–4270 (2009).
Garofalo, N.: *Geometric second derivative estimates in Carnot groups and convexity*, Manuscripta Math. **126**, no. 3, 353–373 (2008).
Garofalo, N., Nhieu, D. M.: *Lipschitz continuity, global smooth approximations and extension theorems for Sobolev functions in Carnot Carathéodory spaces*, J. Anal. Math. **74**, 67–97 (1998).
Gromov, M.: *Carnot-Carathéodory spaces seen from within.* Sub-Riemannian geometry, 79–323, Progr. Math., 144, Birkhäuser, Basel, 1996.
Hirsch, F.: *Opérateurs carré du champ (d’après J. P. Roth)*. Séminaire Bourbaki, 29e année (1976/77), Exp. no. 501, pp. 167–182, Lecture Notes in Math., 677, Springer, Berlin, 1978.
Moore, J. D.: *Lectures on Differential Geometry*, [http://web.math.ucsb.edu/$\sim$moore/riemanniangeometry.pdf]{} (2009).
Montgomery, R.: *A tour of subriemannian geometries, their geodesics and applications*, American Mathematical Society, Mathematical Surveys and Monographs 91 (2006).
O’Neill, B.: *Semi-Riemannian geometry. With applications to relativity*, Pure and Applied Mathematics, Academic Press, 1983.
Ottazzi, A.: *A sufficient condition for nonrigidity of Carnot groups*, Math. Z. **259**, no. 3, 617–629 (2008).
Pinamonti, A., Speight, G.: *Universal Differentiability Sets in Carnot Groups of Arbitrarily High Step*, Submitted. ArXiv:1711.11433.
Pinamonti, A., Valdinoci, E.: *A geometric inequality for stable solutions of semilinear elliptic problems in the Engel group*, Ann. Acad. Sci. Fenn. Math. **37**, no. 2, 357–373 (2012).
Rademacher, H.-B.: *Totally geodesic submanifold – definition*, Bulletin of the Manifold Atlas, [http://www.boma.mpim-bonn.mpg.de/data/49print.pdf]{} (2013).
Savin, O., Valdinoci, E.: *Elliptic PDEs with fibered nonlinearities*, J. Geom. Anal. **19**, no. 2, 420–432 (2009).
Sire, Y., Valdinoci, E.: *Fractional Laplacian phase transitions and boundary reactions: a geometric inequality and a symmetry result*, J. Funct. Anal. **256**, no. 6, 1842–1864 (2009).
Sire, Y., Valdinoci, E.: *Rigidity results for some boundary quasilinear phase transitions*, Comm. Partial Differential Equations **34**, no. 7-9, 765–784 (2009).
Sternberg, P., Zumbrun, K.: *A Poincaré inequality with applications to volume-constrained area-minimizing surfaces*, J. Reine Angew. Math. **503**, 63–85 (1998).
Sternberg, P., Zumbrun, K.: *Connectivity of phase boundaries in strictly convex domains*, Arch. Ration. Mech. Anal. **141**, no. 4, 375–400 (1998).
Sturm, K. T.: *Analysis on local Dirichlet spaces. I. Recurrence, conservativeness and $L^p$-Liouville properties*, J. Reine Angew. Math. **456**, 173–196 (1994).
Xie, X.: *Quasi-conformal maps on model filiform groups*, Michigan Math. J. **64**, no. 1, 169–202 (2015).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Closed-form evaluations of certain integrals of $J_{0}(\xi)$, the Bessel function of the first kind, have been crucial in the studies on the electromagnetic field of alternating current in a circuit with two groundings, as can be seen from the works of Fock and Bursian, Schermann etc. Koshliakov’s generalization of one such integral, which contains $J_s(\xi)$ in the integrand, encompasses several important integrals in the literature including Sonine’s integral. Here we derive an analogous integral identity where $J_{s}(\xi)$ is replaced by a kernel consisting of a combination of $J_{s}(\xi)$, $K_{s}(\xi)$ and $Y_{s}(\xi)$ that is of utmost importance in number theory. Using this identity and the Vorono" summation formula, we derive a general transformation relating infinite series of products of Bessel functions $I_{{\lambda}}(\xi)$ and $K_{{\lambda}}(\xi)$ with those involving the Gaussian hypergeometric function. As applications of this transformation, several important results are derived, including what we believe to be a corrected version of the first identity found on page $336$ of Ramanujan’s Lost Notebook.'
address:
- 'Discipline of Mathematics, Indian Institute of Technology Gandhinagar, Palaj, Gandhinagar 382355, Gujarat, India'
- 'Department of Mathematics, University of North Carolina at Charlotte, 9201 University City Blvd., Charlotte, NC 28223, USA'
author:
- Atul Dixit and Arindam Roy
title: 'Analogue of a Fock-type integral arising from electromagnetism and its applications in number theory'
---
[^1]
Introduction {#intro}
============
In his famous memoir on the propagation of waves in wireless telegraphy, Sommerfeld [@sommerfeld] developed a method of integral representation for calculating the electromagnetic field on a flat boundary where the solution sought is expressed in terms of a Fourier integral consisting of Bessel functions. The Sommerfeld integral, given by [@praus p. 366] $$\label{popovintfock}
\int_{0}^{\infty}tJ_{0}(\rho t)\frac{e^{-a\sqrt{t^2+\xi^2}}}{\sqrt{t^2+\xi^2}}\, dt=\frac{e^{-\xi\sqrt{a^2+\rho^2}}}{\sqrt{a^2+\rho^2}},$$ is valid for Re$(a)>|$Im$(\rho)|$ and Re$(\xi)>0$, and is actually a special case of [@watson-1966a p. 416, Equation (2)] (see also [@grn p. 693, formula 6.596.7]) $$\label{grn693}
\int_{0}^{\infty}t^{s+1}J_{s}(\rho t)\frac{K_{\nu}\left(a\sqrt{t^2+\xi^2}\right)}{(t^2+\xi^2)^{\nu/2}}\, dt=\frac{\rho^{s}}{a^{\nu}}\left(\frac{\sqrt{a^2+\rho^2}}{\xi}\right)^{\nu-s-1}K_{\nu-s-1}\left(\xi\sqrt{a^2+\rho^2}\right),$$ where[^2] Re$(a)>|$$(\rho)|$, $|\arg\rho|<\pi$, Re$(s)>-1$, and Re$(\xi)>0$. Here $J_{s}(\xi)$, the Bessel function of the first kind of order $s$, is defined by [@watson-1966a p. 40] $$\begin{aligned}
\label{sumbesselj}
J_{s}(\xi):=\sum_{m=0}^{\infty}\frac{(-1)^m(\xi/2)^{2m+s}}{m!\Gamma(m+1+s)}, \quad |\xi|<\infty,
\end{aligned}$$ and $K_{s}(\xi)$ is the modified Bessel function of the second kind of order $s$ defined by [@watson-1966a p. 78, eq. (6)], $$K_{s}(\xi):=\frac{\pi}{2}\frac{\left(I_{-s}(\xi)-I_{s}(\xi)\right)}{\sin\pi s},$$ where $I_{s}(\xi)$ is the modified Bessel function of the first kind of order $s$ given by [@watson-1966a p. 77] $$\label{besseli}
I_{s}(\xi):=
\begin{cases}
e^{-\frac{1}{2}\pi s i}J_{s}(e^{\frac{1}{2}\pi i}\xi), & \text{if $-\pi<$ arg $\xi\leq\frac{\pi}{2}$,}\\
e^{\frac{3}{2}\pi s i}J_{s}(e^{-\frac{3}{2}\pi i}\xi), & \text{if $\frac{\pi}{2}<$ arg $\xi\leq \pi$}.
\end{cases}$$ Seventeen years after Sommerfeld’s aforementioned work, Fock and Bursian [@fockbursian p. 361–363] encountered integrals of the type of in their study on the electromagnetic field of alternating current in a circuit with two groundings. Their study rested on the following important result [@fockbursian Equation (20)] valid for Re$(z)>$ Re$(w)\geq 0$ and Re$(\nu)>-3/4$: $$\begin{aligned}
\label{fock}
\int_{0}^{\infty}tJ_{0}(\rho t)\left(\frac{\sqrt{t^2+z^2}-\sqrt{t^2+w^2}}{\sqrt{t^2+z^2}+\sqrt{t^2+w^2}}\right)^{\nu}\frac{\, dt}{\sqrt{t^2+z^2}\sqrt{t^2+w^2}}=I_{\nu}\left(\frac{\rho(z-w)}{2}\right)K_{\nu}\left(\frac{\rho(z+w)}{2}\right).\end{aligned}$$ Identity was given without proof in [@fockbursian] and it was claimed there that its proof will be published later. Seven years later, Fock proved it in his paper [@fock]. The integrals of Fock have been successfully used to explore the electrical properties of earth’s subsurface, for example, see [@praus]. See also Schermann’s work on electromagnetism [@schermann Equation (11)] which makes use of .
Popov [@popovfock] obtained a short proof of and Koshliakov [@kosh34] obtained its remarkable generalization, namely, for Re$(s)>-1$, Re$\left(s+2\nu+\tfrac{3}{2}\right)>0$ and Re$(z)>$ Re$(w)\geq0$, $$\begin{aligned}
\label{koshfock}
&\int_{0}^{\infty}t^{s+1}J_{s}(\rho t)\left(\frac{\sqrt{t^2+z^2}-\sqrt{t^2+w^2}}{\sqrt{t^2+z^2}+\sqrt{t^2+w^2}}\right)^{\nu}\frac{1}{\sqrt{t^2+z^2}
\sqrt{t^2+w^2}}\left(\frac{1}{\sqrt{t^2+z^2}}+\frac{1}{\sqrt{t^2+w^2}}\right)^{2s}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{\nu-s, -s}
{\nu+1} \bigg| {\left(\frac{\sqrt{t^2+z^2}-\sqrt{t^2+w^2}}{\sqrt{t^2+z^2}+\sqrt{t^2+w^2}}\right)^{2}} \right)}}\, dt\nonumber\\
&=\frac{{\Gamma}(\nu+1)}{{\Gamma}(\nu+s+1)}(2\rho)^{s}I_{\nu}\left(\frac{\rho(z-w)}{2}\right)K_{\nu}\left(\frac{\rho(z+w)}{2}\right),\end{aligned}$$ where the hypergeometric function $_2F_1$ is defined by $$_2F_1\left(\begin{matrix} a,b\\c\end{matrix}\,\Bigr|\, \xi\right):=\sum_{n=0}^{\infty}{\dfrac}{(a)_n(b)_n}{(c)_nn!}\xi^n,\qquad |\xi|<1,$$ with $(a)_n:=a(a+1)\cdots(a+n-1)$.
Indeed, if we let $s=0$ in , we recover . As Koshliakov remarks [@kosh34 p. 146], the above integral contains many important integral evaluations apart from . For example, if we let $s=\nu$, divide both sides of by $(z-w)^{\nu}$ and then let $z\to w$, he records that one obtains Sonine’s integral, namely, for Re$(\rho)>0$, Re$(w)>0$ and Re$(\nu)>-1/2$, $$\label{sonine}
\int_{0}^{\infty}\frac{x^{\nu+1}J_{\nu}(\rho x)}{(x^2+w^2)^{2\nu+1}}\, dx=\left(\frac{\rho}{2\sqrt{w}}\right)^{2\nu}\frac{K_{\nu}(\rho w)}{{\Gamma}(2\nu+1)}.$$
In a recent paper [@bdkz Theorem 1.6], Berndt, Kim, Zaharescu and one of the present authors successfully employed in the Vorono" summation formula for $r_k(n)$, the number of representations of $n$ as sum of $k$ squares, where $k\geq2$, to obtain an important transformation which gives, as corollaries, several important results in analytic number theory, for example, those of Dixon and Ferrar [@dixfer2 Equation (3.12)], Hardy [@hardyqjpam1915 Equation (2.12)] and Popov [@popov1935 Equation (6)].
Loosely speaking, the Vorono" summation formula provides an avenue to investigate the order of magnitude of the error term in the asymptotic estimates of the summatory function of the product of an arithmetic function $a(n)$ and a function $f$ by linking it with certain infinite series whose summand is an integral transform of $f$ with respect to a certain kernel. For different arithmetic functions $a(n)$, the function $f$ needs to satisfy corresponding hypotheses which validate these summation formulas.
The Vorono" summation formula for $a(n)=r_2(n)$ is given in [@landau p. 274] (or [@dixfer2 Thm. A]), whereas for $k\geq 2$, it is given by Popov [@popov Equation (3)], and in an equivalent form by Guinand [@guinandconcord p. 264]. This equivalent form is $$\begin{aligned}
\label{guirknsuminfallk}
&\sum_{n=1}^{\infty}r_k(n)n^{\frac{1}{2}-\frac{k}{4}}f(n)-\frac{\pi^{\frac{k}{2}}}{{\Gamma}(\frac{k}{2})}\int_{0}^{\infty}
x^{\frac{k}{4}-\frac{1}{2}}f(x)\, dx\nonumber\\
&=\sum_{n=1}^{\infty}r_k(n)n^{\frac{1}{2}-\frac{k}{4}}g(n)-\frac{\pi^{\frac{k}{2}}}{{\Gamma}(\frac{k}{2})}\int_{0}^{\infty}
x^{\frac{k}{4}-\frac{1}{2}}g(x)\, dx,\end{aligned}$$ where $$g(y)=\pi\int_{0}^{\infty}f(t)J_{\frac{k}{2}-1}(2\pi\sqrt{yt})\, dt.$$ Unfortunately, neither Popov nor Guinand gives any conditions on $f$ and $g$. One can refer to [@bdkz Theorem 1.5] for the same and also for a detailed discussion on this topic.
While the integral transform of $f$ in involves the Bessel function $J_{s}(t)$ in the kernel, the corresponding summation formula for the generalized sum-of-divisors function $\sigma_{-s}(n):=\sum_{d|n}d^{-s}$ involves the Koshliakov kernel $$\label{kk}
F_{s}(t):=M_{s}(t)\cos\left(\frac{\pi s}{2}\right)-J_{s}(t)\sin\left(\frac{\pi s}{2}\right),$$ where $$\label{msxi}
M_{s}(\xi):=\frac{2}{\pi}K_{s}(\xi)-Y_{s}(\xi),$$ with $Y_{s}(\xi)$ being the Bessel function of the second kind defined by [@watson-1966a p. 64] $$\begin{aligned}
Y_{s}(\xi)=\frac{J_{s}(\xi)\cos(\pi s)-J_{-s}(\xi)}{\sin{\pi s}}.\end{aligned}$$ For a comprehensive discussion and recent new results on the Vorono" summation formula for $\sigma_{-s}(n)$, the reader is referred to a paper by Berndt, Zaharescu and the present authors [@bdrz Section 6].
Guinand’s version [@guinandsumself2 Theorem 6], [@guinand Equation (1)] of the Vorono" summation formula for $\sigma_{-s}(n)$ reads, if $f(x)$ and $f'(x)$ are integrals, $f$ tends to zero as $x\to\infty$, $f(x), xf'(x)$, and $x^2f''(x)$ belong to $L^{2}(0,\infty)$, $-\frac{1}{2}<$ Re$(s)<\frac{1}{2}$, and $$\label{recpairgui}
g(x) = 2\pi \int_{0}^{\infty} f(t) F_s(4\pi\sqrt{xt}) \,dt,$$ where $F_s(t)$ is defined in , then $$\begin{aligned}
\label{guitra}
&\sum_{n=1}^{\infty} \sigma_{-s} (n) n^{\frac{s}{2}} f(n) - \zeta(1+s) \int_{0}^{\infty} y^{\frac{s}{2}} f(y) \,dy - \zeta(1-s) \int_{0}^{\infty} y^{-\frac{s}{2}} f(y) \,dy\nonumber\\
&= \sum_{n=1}^{\infty} \sigma_{-s} (n) n^{\frac{s}{2}} g(n) - \zeta(1+s) \int_{0}^{\infty} y^{\frac{s}{2}} g(y) \,dy - \zeta(1-s) \int_{0}^{\infty} y^{-\frac{s}{2}} g(y) \,dy.\end{aligned}$$ The Vorono" summation formula for $\sigma_0(n)=d(n)$, the number of divisors of $n$, has been found to be enormously useful in physics; for example, S. Egger and F. Steiner [@es1; @es2] have shown that it plays the role of an exact trace formula for a Schrödinger operator on a certain non-compact quantum graph.
The two notoriously difficult, and still unsolved problems, in analytic number theory are the Gauss circle and the Dirichlet divisor problems. They ask for the correct order of magnitude of the error terms $P(x)$ and $\Delta(x)$ involved in the summatory functions of $r_2(n)$ and $d(n)$ respectively. A natural tool to tackle these problems is to use the Vorono" summation formula corresponding to each of these arithmetic functions. While they are inherently distinct in that one counts the number of lattice points inside a circle and the other the number of lattice points under a hyperbola, they have lots of similarities. For example, if an upper bound on $P(x)$ is obtained, then a similar one can be found for $\Delta(x)$. The reader is referred to a recent survey article by Berndt, Kim and Zaharescu [@bkz-unsolved] on these two problems. For the generalized Gauss circle problem, which counts the number of lattice points on multidimensional spheres and where the associated arithmetic function is naturally $r_k(n)$, the reader is referred to the two papers of Lursmana$\check{s}$vili [@lurs1], [@lurs2].
Considering the fact that Fock’s integral evaluation is one of the key steps in his and Bursian’s work [@fockbursian] on electromagnetism, the first goal of this paper is to derive an analogue of Koshliakov’s formula (of which is a special case) where the kernel $J_{s}(\rho t)$ is replaced by the Koshliakov kernel $F_s(\rho t)$. The second goal is to apply this evaluation in to obtain a general summation formula from which many interesting and important identities result as special cases.
However, before we proceed to establish the former, it is important to mention the thoughts of Dixon and Ferrar [@dixfer3 p. 168] on obtaining analogues of known integral identities where the $J$-Bessel function is replaced by the corresponding kernel $M_0(\xi)$, where $M_s(\xi)$ is defined in . They say[^3]:\
*Our recent work on summation-formulae suggests that it would be of interest to find a formula corresponding to (1) \[Equation here\] when $J_{s}(\rho t)t^{s+1}$ is replaced by $tM_{0}(\rho t)$. The problem of finding such a formula has proved to be a difficult one: we have found that any single integral of the type which occurs in (1) leads to a complicated and ill-balanced formula; the only way we have discovered of obtaining a reasonably well-balanced formula is to combine two integrals.*
To give an example of their principle, consider the following identity due to Koshliakov [@koshliakov Equation (5)] for Re$({\beta})>0$: $$\label{koshvor}
2 \sum_{n=1}^{\infty} d(n) \left( K_{0} \left( 4 \pi e^{\frac{i \pi}{4}} \sqrt{n{\beta}} \right) + K_{0} \left( 4 \pi e^{-\frac{i \pi}{4}} \sqrt{n{\beta}} \right)\right)=- \gamma - \frac{1}{2} \log{\beta}- \frac{1}{4 \pi{\beta}} +
\frac{{\beta}}{\pi} \sum_{n=1}^{\infty} \frac{d(n)}{{\beta}^{2} + n^{2}},$$ which is instrumental in his clever proof of the Vorono" summation formula for the divisor function $d(n)$. The latter is, in fact, shown in [@soni] to be equivalent to the above formula. Here ${\gamma}$ denotes Euler’s constant. Koshliakov’s identity can be proved by replacing $a$, first by ${\beta}e^{i\frac{\pi}{2}}$, and then by ${\beta}e^{-i\frac{\pi}{2}}$, in the identity [@voronoi Equations (5), (6)], [@lnb p. 254] $$\label{vorram}
2\sum_{n=1}^{\infty}d(n)K_{0}(4\pi\sqrt{an})=\frac{a}{\pi^2}\sum_{n=1}^{\infty}\frac{d(n)\log(a/n)}{a^2-n^2}-\frac{\gamma}{2}-\left(\frac{1}{4}+\frac{1}{4\pi^2a}\right)\log a-\frac{\log 2\pi}{2\pi^2a},$$ and then by adding the two resulting identities. Note that even though each of the individual series $\sum_{n=1}^{\infty} d(n) K_{0} \left( 4 \pi e^{\pm\frac{i \pi}{4}} \sqrt{n{\beta}} \right)$ admits a representation through , adding the two gives a simpler and a well-balanced formula . A generalization of Koshliakov’s above identity, which is also a key ingredient in obtaining a short proof of Vorono" summation formula for $\sigma_{{\lambda}}(n)$ [@bdrz Section 6.1], was first studied by Moll and one of the present authors in [@dixitmoll Section 6]. For Re$({\beta})>0$ and Re$({\lambda})>-1$, it is given by $$\begin{aligned}
\label{dixmol}
&2 \sum_{n=1}^{\infty} \sigma_{-{\lambda}}(n) n^{\frac{{\lambda}}{2}}
\left( e^{\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi e^{\frac{\pi i}{4}} \sqrt{n{\beta}} ) +
e^{-\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi e^{-\frac{\pi i}{4}} \sqrt{n{\beta}} ) \right)\nonumber\\
&=- \frac{\Gamma({\lambda}) \zeta({\lambda})}{(2 \pi \sqrt{{\beta}})^{{\lambda}}} +
\frac{{\beta}^{\frac{{\lambda}}{2}-1}}{2 \pi} \zeta({\lambda}) - \frac{{\beta}^{\frac{{\lambda}}{2}}}{2} \zeta({\lambda}+1) +\frac{{\beta}^{\frac{{\lambda}}{2}+1}}{\pi} \sum_{n=1}^{\infty} \frac{\sigma_{-{\lambda}}(n)}{n^{2}+{\beta}^{2}},\end{aligned}$$ Here one can observe the same aforementioned principle. To see this, we first note the following generalization of , due to Cohen [@cohen Theorem 3.4], valid for Re$(a)>0$ and ${\lambda}\notin\mathbb{Z}$: $$\begin{aligned}
&8\pi a^{{\lambda}/2}\sum_{n=1}^{\infty}\sigma_{-{\lambda}}(n)n^{{\lambda}/2}K_{{\lambda}}(4\pi\sqrt{n a})=
A({\lambda}, a)\zeta({\lambda})+B({\lambda}, a)\zeta({\lambda}+1)\\
&\quad+\frac{2}{\sin\left(\pi {\lambda}/2\right)}\left(\sum_{1\leq j\leq k}\zeta(2j)\zeta(2j-{\lambda})a^{2j-1}+a^{2k+1}\sum_{n=1}^{\infty}\sigma_{-{\lambda}}(n)\frac{n^{{\lambda}-2k}-a^{{\lambda}-2k}}{n^2-a^2}\right),\notag\end{aligned}$$where $$\begin{aligned}
A({\lambda}, a)&=\frac{a^{{\lambda}-1}}{\sin\left(\pi {\lambda}/2\right)}-(2\pi)^{1-{\lambda}}\Gamma({\lambda}),\nonumber\\
B({\lambda}, a)&=\frac{2}{a}(2\pi)^{-{\lambda}-1}\Gamma({\lambda}+1)-\frac{\pi a^{{\lambda}}}{\cos\left(\pi {\lambda}/2\right)}.\end{aligned}$$ Now as shown in [@bdrz p. 844], follows from the above result by first replacing $a$ by $i{\beta}$ for $-\pi<\arg{\beta}<\frac{1}{2}\pi$, then by $-i{\beta}$ for $-\frac{1}{2}\pi<\arg{\beta}<\pi$, and then by adding the resulting two identities. Now observe that compared to the representations for each of $\sum_{n=1}^{\infty} \sigma_{-{\lambda}}(n) n^{\frac{{\lambda}}{2}}
K_{{\lambda}}( 4 \pi e^{\pm\frac{\pi i}{4}} \sqrt{n{\beta}} )$, the one derived in for its left-hand side is much simpler and well-balanced.
Dixon and Ferrar’s principle was also known to Ramanujan, for, on page $336$ of his Lost Notebook [@lnb], we find the following identity: [$$\begin{aligned}
\label{ramin}
&\Gamma\left(s+\frac{1}{2}\right)
\bigg\{\frac{\zeta(1-s)}{(s-\frac{1}{2}){\beta}^{s-\frac{1}{2}}}+\frac{\zeta(-s)\tan\frac{1}{2}\pi s}{2{\beta}^{s+\frac{1}{2}}}\nonumber\\
&\quad+\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{2i}\left(({\beta}-in)^{-s-\frac{1}{2}}-({\beta}+in)^{-s-\frac{1}{2}}\right)\bigg\}
\nonumber\\
&=(2\pi)^s\bigg\{\frac{\zeta(1-s)}{2\sqrt{\pi{\beta}}}-2\pi\sqrt{\pi{\beta}}\zeta(-s)\tan\tfrac{1}{2}\pi s\nonumber\\
&\quad+\sqrt{\pi}\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{\sqrt{n}}e^{-2\pi\sqrt{2n{\beta}}}
\sin\left(\frac{\pi}{4}+2\pi\sqrt{2n{\beta}}\right)\bigg\}.\end{aligned}$$]{} Clearly, one can see powers of ${\beta}+in$ and ${\beta}-in$ both playing a role in the above identity than just one of them. Note that in the summands of the infinite series on the right-hand sides of each of and , one has $$\frac{1}{{\beta}^2+n^2}=\frac{1}{2{\beta}}\left(\frac{1}{{\beta}+in}+\frac{1}{{\beta}-in}\right).$$ The identity , however, incorrect since the series on its left-hand side diverges. We obtain a corrected version of in Corollary \[ramanujan-type\] of this paper. In fact, searching for a corrected version of was one of the chief motivations behind our work.
Also, among other things, we carry out an in-depth analysis of the series $$\begin{aligned}
\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{&I_{{\lambda}}\left(2\pi e^{-\frac{i\pi}{4}}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi e^{-\frac{i\pi}{4}}\sqrt{n} (\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&+I_{{\lambda}}\left(2\pi e^{\frac{i\pi}{4}}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi e^{\frac{i\pi}{4}}\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\bigg\},\end{aligned}$$ for $|\arg{\alpha}|<\pi/2$, $|\arg{\beta}|<\pi/2$ and Re$(\sqrt{{\alpha}})> $Re$(\sqrt{{\beta}})$, whose limiting case ${\alpha}\to{\beta}$ (after we divide it by $(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}$) is, as will be shown in the paper, the series on the left-hand side of . The convergence of the above series for the aforementioned values of ${\alpha}, {\beta}$ can be seen from the fact that for large $n$ $$\begin{aligned}
&I_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n} (\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&+I_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&\sim\frac{e^{-2\pi\sqrt{2n{\beta}}}}{2\pi\sqrt{n}\sqrt{{\alpha}-{\beta}}}\sin\left(\frac{\pi}{4}-2\pi\sqrt{2n{\beta}}\right),\end{aligned}$$ which, in turn, can be proved using and below.
Dixon and Ferrar [@dixfer3 p. 174] gave the following integral evaluation which corresponds to upon replacement of the kernel $J_{s}(\rho t)t^{s+1}$ by $tM_{0}(\rho t)$: $$\begin{aligned}
\label{st1}
&\int_{0}^{\infty}tM_{0}(\rho t)\left\{\frac{K_{\nu}\left(a\sqrt{\xi^2+t^2}\right)}{(\xi^2+t^2)^{\nu/2}}+\frac{K_{\nu}\left(a\sqrt{\xi^2-t^2}\right)}{(\xi^2-t^2)^{\nu/2}}\right\}\, dt\nonumber\\
&=\frac{i\xi^{1-\nu}}{a^{\nu}}\left\{\frac{K_{1-\nu}\left(\xi\sqrt{a^2+\rho^2}\right)}{(a^2+\rho^2)^{(1-\nu)/2}}+\frac{K_{1-\nu}\left(\xi\sqrt{a^2-\rho^2}\right)}{(a^2-\rho^2)^{(1-\nu)/2}}\right\}.\end{aligned}$$ where $\xi>0, a>0, \rho>0, a\neq \rho$, and Re $\nu>0$. (They also give different extensions of the above formula for values of $\nu, a$ and $\rho$ other than those considered above, which ensure its validity as it is, or with some modifications.) Here, of course, the path of integration is indented to avoid the singularity of the integrand. It ensures that $\arg(z^2-t^2)=-\pi$ when $t>z$. At the end of their paper, they also say that the indented path may be replaced by a straight-line path provided Re$(\nu)<1$ where they always retain the conventions $\arg(z^2-t^2)=0$ when $t<z$ and $\arg(z^2-t^2)=-\pi$ when $t>z$.
Though Dixon and Ferrar evaluated certain integrals with $tM_{0}(\rho t)$ as the kernel, they did not work with integrals having the general kernel $t^{s+1}M_{s}(\rho t)$ in their integrands. However, Koshliakov [@kosh1938 p. 425] obtained such a result[^4], namely, for $\rho>0, b>0, \xi>0$ and $s>-1$, $$\begin{aligned}
\label{koshcertain}
&\int_{0}^{\infty}t^{s+1}M_{s}(\rho t)\left\{\frac{K_{\nu}\left(b \sqrt{\xi^2+it^2}\right)}{(\xi^2+it^2)^{\nu/2}}+\frac{K_{\nu}\left(b \sqrt{\xi^2-it^2}\right)}{(\xi^2-it^2)^{\nu/2}}\right\}\, dt\nonumber\\
&=\frac{\xi^{1+s-\nu}\rho^{s}}{b^{\nu}}\left\{e^{-\frac{i\pi}{4}\left(\nu+s-1\right)}\frac{K_{1+s-\nu}(\xi e^{-\frac{i\pi}{4}} \sqrt{\rho^2+ib^2})}{(\rho^2+ib^2)^{\frac{1+s-\nu}{2}}}+e^{\frac{i\pi}{4}\left(\nu+s-1\right)}\frac{K_{1+s-\nu}(\xi e^{\frac{i\pi}{4}} \sqrt{\rho^2-ib^2})}{(\rho^2-ib^2)^{\frac{1+s-\nu}{2}}}\right\}.\end{aligned}$$ Henceforth, throughout the paper, we follow Koshliakov [@kosh1938] in using the notation[^5] $$\epsilon=e^{\frac{i\pi}{4}},\hspace{5mm} \overline{\epsilon}=e^{-\frac{i\pi}{4}},$$ for brevity and convenience. Note that in , Koshliakov has avoided the necessity of indenting the contour, which was the requirement in Dixon and Ferrar’s . Thus the path of integration in is the straight line from $0$ to $\infty$.
Main Results {#mr}
============
All of our results in this paper abide by the principle by Dixon and Ferrar of combining two integrals while working with the kernel $t^{s+1}F_{s}(t)$, which is, of course, a generalization of the kernel $tM_{0}(t)$ that Dixon and Ferrar work with in [@dixfer3].
An application of and leads to the following desired analogue of .
\[tfockathm\] Let $|\arg z|<\pi/4, |\arg w|<\pi/4$ and $(z)>$ $(w)$. Let $F_{s}(t)$ be defined in . Define $$\begin{aligned}
\label{funa}
\mathcal{A}(t):=\mathcal{A}(s, {\lambda}, z, w, t)&:=\frac{1}{\sqrt{z^2+it}\sqrt{w^2+it}}\left(\frac{1}{\sqrt{z^2+it}}+\frac{1}{\sqrt{w^2+it}}\right)^{2s}\left(\frac{\sqrt{z^2+it}-\sqrt{w^2+it}}{\sqrt{z^2+it}+\sqrt{w^2+it}}\right)^{{\lambda}}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{z^2+it}-\sqrt{w^2+it}}{\sqrt{z^2+it}+\sqrt{w^2+it}}\right)^2} \right)}}.\end{aligned}$$ Then for $\rho>0$, $(s)>-1$, $({\lambda})>-1/2$ and $(s+{\lambda})>-1$, we have $$\begin{aligned}
\label{tfockaeqn}
&\int_{0}^{\infty}t^{s+1}F_{s}(\rho t)\left(\mathcal{A}(t^2)+\mathcal{A}(-t^2)\right)\, dt\nonumber\\
&=\frac{{\Gamma}({\lambda}+1)}{{\Gamma}(s+{\lambda}+1)}(2\rho)^{s}\left\{I_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z+w)}{2}\right)+I_{{\lambda}}\left(\frac{\rho\epsilon(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\epsilon(z+w)}{2}\right)\right\}.\end{aligned}$$
Similar to , the above theorem also contains many important integral evaluations as its special cases. For example, letting $s={\lambda}$, dividing both sides by $(z-w)^{{\lambda}}$ and then letting $z\to w$ leads to the following analogue of Sonine’s integral for $\rho>0$, $|\arg w|<\pi/4$ and Re$({\lambda})>-1/2$: $$\begin{aligned}
&\int_{0}^{\infty}t^{{\lambda}+1}F_{{\lambda}}(\rho t)\left\{\frac{1}{(w^2+it)^{2{\lambda}+1}}+\frac{1}{(w^2-it)^{2{\lambda}+1}}\right\}\, dt\nonumber\\
&=\frac{1}{{\Gamma}(2{\lambda}+1)}\left(\frac{\rho}{2\sqrt{w}}\right)^{2{\lambda}}\left\{e^{\frac{i\pi{\lambda}}{4}}K_{{\lambda}}\left(e^{\frac{i\pi}{4}}\rho w\right)+e^{-\frac{i\pi{\lambda}}{4}}K_{{\lambda}}\left(e^{-\frac{i\pi}{4}}\rho w\right)\right\}.\end{aligned}$$ The transformation between two infinite series involving $\sigma_{s}(n)$ that we obtain upon using Theorem \[tfockathm\] in is now given. The corresponding result for $r_k(n)$, the number of representations of $n$ as sum of $k$ squares, $k\in\mathbb{N}$, was established in [@bdkz Theorem 1.6], although we emphasize that the level of difficulty in obtaining the result for $\sigma_s(n)$ is quite high as compared to the one for $r_k(n)$.
\[anathmbdkz\] Let $\mathcal{A}(s, {\lambda}, z, w, t)$ be defined in . Let $|\arg {\alpha}|<\pi/2, |\arg {\beta}|<\pi/2$ with $(\sqrt{{\alpha}})>$ $(\sqrt{{\beta}})$. Let $({\lambda})>0$. Define $h(s,{\lambda})$ by $$\label{hsl}
h(s, {\lambda}):=\begin{cases}
\hspace{6mm}0,\hspace{9mm}\text{if}\hspace{2mm}\textup{Re}(s+{\lambda})>0,\\
\displaystyle\frac{({\alpha}-{\beta})^{{\lambda}}}{2^{4{\lambda}+1}},\hspace{2mm}\text{if}\hspace{1mm}s=-{\lambda}.
\end{cases}$$ Then for all $s\in\mathbb{C}$ such that $(s+{\lambda})>0$ or $s=-{\lambda}$, we have $$\begin{aligned}
\label{anathmbdkzeqn}
&\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{{\Gamma}(s+{\lambda}+1)\zeta(1-s)}{2(8\pi)^s{\Gamma}({\lambda}+1)}h(s, {\lambda})-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\nonumber\\
&\quad\times\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&\quad+\frac{{\Gamma}({\lambda}+s+1)}{2^{3s+2}\pi^{s+1}{\Gamma}({\lambda}+1)}\sum_{n=1}^{\infty}\sigma_s(n)\left(\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)+\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)\right).\end{aligned}$$
We do not know if there exist transformations analogous to the above for the remaining values of $s$, that is, the ones not satisfying $(s+{\lambda})>0$ or $s=-{\lambda}$. However, we do show by analytic continuation that the above result can be extended to $({\lambda})>-1$. This is done in Section \[ac\]. Of course, one can analytically continue Theorem \[anathmbdkz\] to $(\nu)>-m$, where $m\in\mathbb{N}$, however, we refrain ourselves from doing this as the resulting transformation becomes complicated.
The above theorem as well as its analytic continuation given in Theorem \[anathmbdkzac\] give many interesting corollaries. We state below the ones resulting from Theorem \[anathmbdkz\] and reserve those resulting from Theorem \[anathmbdkzac\] for Section \[ac\].
\[dixmolcor\] Equation holds for $({\beta})>0$ and $({\lambda})>-1$.
Another important result is
\[dixmolcorgen\] For $({\beta})>0$ and $(s)>-\frac{1}{2}$, $$\begin{aligned}
\label{dixmolcorgeneqn}
&\sum_{n=1}^{\infty}\sigma_{s}(n)e^{-2\pi\sqrt{2n{\beta}}}\cos\left(2\pi\sqrt{2n{\beta}}\right)\nonumber\\
&=-\frac{1}{2}\zeta(-s)+\frac{{\Gamma}\left(s+\frac{3}{2}\right)\zeta(-s)}{2\sqrt{\pi}(2\pi{\beta})^{s+1}}+\frac{\sqrt{{\beta}}{\Gamma}\left(s+\frac{3}{2}\right)}{2^{s+2}\pi^{s+\frac{3}{2}}}\sum_{n=1}^{\infty}\sigma_s(n)\left\{({\beta}+in)^{-s-\frac{3}{2}}+({\beta}-in)^{-s-\frac{3}{2}}\right\}.\end{aligned}$$
The reason this identity is important because, specializing $s$ to be zero in the above identity gives an analogue of a result of Hardy [@hardyqjpam1915 Equation (2.12)] which he used in his study of the Gauss circle problem to prove that $$\label{harob}
\sum_{n\leq x}r_2(n)-\pi x=\Omega(x^{1/4}).$$ This analogue is $$\begin{aligned}
\label{dixmolcorgeneqns0}
\sum_{n=1}^{\infty}d(n)e^{-2\pi\sqrt{2n{\beta}}}\cos\left(2\pi\sqrt{2n{\beta}}\right)=\frac{1}{4}-\frac{1}{16\pi{\beta}}+\frac{\sqrt{{\beta}}}{8\pi}\sum_{n=1}^{\infty}d(n)\left\{({\beta}+in)^{-\frac{3}{2}}+({\beta}-in)^{-\frac{3}{2}}\right\}.\end{aligned}$$ We end this section with a rather exotic transformation that results from Theorem \[anathmbdkz\].
\[exotic\] For $|\arg {\alpha}|<\pi/2, |\arg {\beta}|<\pi/2$ and $(\sqrt{{\alpha}})>$ $(\sqrt{{\beta}})>0$, $$\begin{aligned}
\label{exoticeqn}
&\sum_{n=1}^{\infty}\sigma_{-1}(n)\bigg\{I_{1}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{1}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad\quad+I_{1}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{1}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{\pi^3({\alpha}-{\beta})}{48}+\frac{1}{2}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)\left\{{\gamma}+\log(2\pi\sqrt{{\alpha}{\beta}})-1-\left(\frac{\sqrt{{\alpha}}+\sqrt{{\beta}}}{\sqrt{{\alpha}}-\sqrt{{\beta}}}\right)^2\log\left(\frac{4\sqrt{{\alpha}{\beta}}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^2}\right)\right\}\nonumber\\
&\quad+\frac{1}{2}\sum_{n=1}^{\infty}\sigma_{-1}(n)\left(\frac{\sqrt{{\alpha}+in}-\sqrt{{\beta}+in}}{\sqrt{{\alpha}+in}+\sqrt{{\beta}+in}}+\frac{\sqrt{{\alpha}-in}-\sqrt{{\beta}-in}}{\sqrt{{\alpha}-in}+\sqrt{{\beta}-in}}\right).\end{aligned}$$
That the special case ${\lambda}=1$ of results from the above corollary after dividing both sides by $\sqrt{{\alpha}}-\sqrt{{\beta}}$ and then letting ${\alpha}\to{\beta}$ can then be seen without much effort.
This paper is organized as follows. In Section \[ex\], we record the well-known asymptotic expansions of Bessel functions as their argument $\xi\to0$ or as $|\xi|\to\infty$. Assuming that the parameters involved in and are complex, we find the exact conditions on them so that these equations are valid. This is crucial in deriving Theorem \[tfockathm\]. Sections \[proof1\] and \[proof2\] are devoted to the proofs of Theorems \[tfockathm\] and \[anathmbdkz\] respectively. Corollaries \[dixmolcor\], \[dixmolcorgen\] and \[exotic\] are proved in Section \[cor1\]. Theorem \[anathmbdkz\] admits analytic continuation to Re$(\nu)>-1$. This analytic continuation is obtained in Theorem \[anathmbdkzac\] of Section \[ac\]. The interesting corollaries that follow from this theorem, namely, Corollaries \[ramanujan-type\] - \[soninetrans\] are then derived. We end the paper with Section \[cr\] consisting of concluding remarks and directions for further work.
Extending the validity of certain integral evaluations {#ex}
======================================================
Here we extend the validity of certain integral evaluations needed in the sequel by relaxing the conditions on the parameters involved. This is extremely crucial, for, most of the times we do require the widest possible conditions under which these formulas are valid, to prove our results. To do that, however, we need the following asymptotic formulas of the Bessel functions $J_{s}(\xi), Y_{s}(\xi)$, and $K_{s}(\xi)$, as $|\xi|\to\infty$, given by [@watson-1966a p. 199 and p. 202] $$\begin{aligned}
J_{s}(\xi)&\sim \left(\frac{2}{\pi \xi}\right)^{{\tfrac}12}\bigg(\cos w\sum_{n=0}^{\infty}\frac{(-1)^n(s, 2n)}{(2\xi)^{2n}} -\sin w\sum_{n=0}^{\infty}\frac{(-1)^n(s, 2n+1)}{(2\xi)^{2n+1}}\bigg),\label{asymbess}\\
Y_{s}(\xi)&\sim \left(\frac{2}{\pi \xi}\right)^{{\tfrac}12}\bigg(\sin w\sum_{n=0}^{\infty}\frac{(-1)^n(s, 2n)}{(2\xi)^{2n}}+\cos w\sum_{n=0}^{\infty}\frac{(-1)^n(s, 2n+1)}{(2\xi)^{2n+1}}\bigg),\label{asymbess1}\\
K_{s}(\xi)&\sim \left(\frac{\pi}{2 \xi}\right)^{{\tfrac}12}e^{-\xi}\sum_{n=0}^{\infty}\frac{(s, n)}{(2\xi)^{n}},\label{asymbess2}\end{aligned}$$ for $|\arg \xi|<\pi$. Here $w=\xi-\tfrac{1}{2}\pi s-\tfrac{1}{4}\pi$ and $(s,n)=\frac{\Gamma(s+n+1/2)}{\Gamma(n+1)\Gamma(s-n+1/2)}$.
Also from [@koshkernel Equation (2.11)], $$\label{fasy}
F_s(\rho t)<\!\!<_{s}\begin{cases}
\textup{max}(1, |\log(\rho t)|), &\mbox{ if } \quad s=0, 0 < \rho t \le 1, \\
(\rho t)^{-|\textup{Re}(s)|}, &\mbox{ if } \quad s\neq 0, 0 < \rho t\le 1, \\
(\rho t)^{-1/2}, &\mbox{ if } \quad \rho t \ge 1.
\end{cases}$$ We begin with identity . In [@grn p. 693, formula 6.596.7] as well as in [@watson-1966a p. 416, Equation (2)], this result is stated to be valid for $a>0, \rho>0$, Re$(s)>-1$, and $|\arg\xi|<\pi/2$. However, in a footnote on page 416 in [@watson-1966a], Watson says that with certain limitations, one can take $a$ and $\rho$ to be complex. Since the precise conditions on $a$ and $\rho$ are not given there and since we require the one for $a$ in order to prove our results (see and below), we give them here.
Note that from , as $t\to\infty$, $$J_{s}(\rho t)=O\left(\sqrt{\frac{2}{\pi |\rho| t}}\left|\cos\left(\rho t-\frac{\pi s}{2}-\frac{\pi}{4}\right)\right|\right)=O_{s, \rho}\left(e^{|\textup{Im}(\rho)|t}\right),$$ where $|\arg(\rho)|=|\arg(\rho t)|<\pi$. Also, from , $$K_{\nu}\left(a\sqrt{t^2+\xi^2}\right)=O_{\nu, \xi}\left(e^{-\textup{Re}(a)t}\right)$$ as $t\to\infty$. Thus, to secure the convergence at the upper limit of integration, we require Re$(a)>|$$(\rho)|$.
Now for complex values of $\rho, b, s$ and $\xi$, the precise conditions for the validity of can be found to be Re$(\rho)>0$, $|\textup{Im}(\rho)|<$Re$(b\epsilon)$ and $|\textup{Im}(\rho)|<$Re$(b\overline{\epsilon})$, Re$(s)>-1$ and $|\arg(\xi)|<\frac{\pi}{4}$. However, since each of the variables $\rho, b$ and $\xi$ are positive in Theorem \[tfockathm\], we only show the validity of for Re$(s)>-1$, which is required, for example, in obtaining . The condition Re$(s)>-1$ is needed to secure the convergence of the integral in at the lower limit $0$. To see this, note that as $t\to 0$, we have [@stab p. 375, equations (9.6.9), (9.6.8)] $$\label{k0}
K_{s}(\rho t)\sim\begin{cases}
-\log(\rho t),\hspace{8mm}\text{if}\hspace{1mm}s=0,\\
\frac{1}{2}{\Gamma}(s)\left(\frac{\rho t}{2}\right)^{-s},\hspace{1mm}\text{if}\hspace{1.5mm}\textup{Re}(s)>0,
\end{cases}$$ and [@stab p. 360, equations (9.1.8), (9.1.9)] $$\label{y0}
Y_{s}(\rho t)\sim\begin{cases}
\frac{2}{\pi}\log(\rho t),\hspace{11mm}\text{if}\hspace{1mm}s=0,\\
-\frac{1}{\pi}{\Gamma}(s)\left(\frac{\rho t}{2}\right)^{-s},\hspace{1mm}\text{if}\hspace{1.5mm}\textup{Re}(s)>0.
\end{cases}$$ First let Re$(s)>0$. Note that $$\frac{K_{\nu}\left(b \sqrt{\xi^2+it^2}\right)}{(\xi^2+it^2)^{\mu/2}}+\frac{K_{\nu}\left(b \sqrt{\xi^2-it^2}\right)}{(\xi^2-it^2)^{\mu/2}}$$ approaches a constant depending on $b$, $\xi$ and $\mu$ as $t\to 0$, where $\xi\neq0$ if Re$(\nu)>0$. Also, from and , we deduce that $$M_{s}(\rho t)\sim\frac{2}{\pi}{\Gamma}(s)\left(\frac{\rho t}{2}\right)^{-s}.$$ Since the integrand on the left-hand side of contains $t^{s+1}$, convergence is clearly secured for Re$(s)>0$. In a similar way, it can be seen using [@dunster Equation (2.14)] and the corresponding asymptotic formulas for $J$ and $Y$-Bessel functions of purely imaginary order that the integral converges when Re$(s)=0$ too.
Now let $-1<$ Re$(s)<0$. Since the $K$-Bessel function is an even function of its order, we have from , $$\label{kbigo}
\frac{2}{\pi}K_{s}(\rho t)=\frac{2}{\pi}K_{-s}(\rho t)\sim\frac{1}{\pi}{\Gamma}(-s)\left(\frac{\rho t}{2}\right)^{s}.$$ Unfortunately, the relation between $Y_{-s}(\rho t)$ and $Y_{s}(\rho t)$ is not as straightforward as in the case of $K_{-s}(\rho t)$ and $K_{s}(\rho t)$, but using the standard formulas in the theory of Bessel functions [@grn Formulas 8.482, 8.483, 8.484] it can be seen that $$\begin{aligned}
\label{ynu-nu}
Y_{s}(\rho t)=Y_{-s}(\rho t)\sec(\pi s)-J_{s}(\rho t)\tan(\pi s).\end{aligned}$$ Now from , $$\begin{aligned}
\label{y0-}
Y_{-s}(\rho t)\sim-\frac{1}{\pi}{\Gamma}(-s)\left(\frac{\rho t}{2}\right)^{-s},\end{aligned}$$ whereas from [@stab p. 360, Formula 9.1.7], for $s\neq-1, -2, -3, \cdots$, $$\begin{aligned}
\label{j0}
J_{s}(\rho t)\sim\frac{\left(\rho t/2\right)^{s}}{{\Gamma}(s+1)}\end{aligned}$$ as $t\to 0$. Hence , and give $$\label{ybigo}
Y_{s}(\rho t)=O_{s, \rho}\left(t^{\textup{Re}(s)}\right).$$ Thus and give $$t^{s+1}M_{s}(\rho t)=O_{s, \rho}\left(t^{2\textup{Re}(s)+1}\right),$$ which implies that the convergence of the integral at $0$ is secured as long as Re$(s)>-1$. By [@temme p. 30, Theorem 2.3], the integral in represents a holomorphic function of $s$ for Re$(s)>-1$. Since the right-hand side is obviously analytic for Re$(s)>-1$, we conclude, by analytic continuation, that is valid as long as Re$(s)>-1$.
Proof of Theorem \[tfockathm\] {#proof1}
==============================
We first prove the result for $z>w>0$ and later extend it by analytic continuation to complex $z$ and $w$ such that $|\arg z|<\pi/4, |\arg w|<\pi/4$ and $(z)\geq$ $(w)>0$. Also from the hypotheses of the theorem, we have $\rho>0$.
Letting $\xi=we^{-i\pi/4}$, and $a=y e^{i\pi/4}$ with $y>0$ in , we see that for $\rho>0$ and Re$(s)>-1$, we have $$\begin{aligned}
\label{grn6931}
\int_{0}^{\infty}t^{s+1}J_{s}(\rho t)\frac{K_{\nu}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\nu/2}}\, dt=\frac{\overline{\epsilon}^{(1+\nu+s)}\rho^{s}w^{1+s-\nu}}{y^{\nu}}\frac{K_{1+s-\nu}\left(w\overline{\epsilon}\sqrt{\rho^2+iy^2}\right)}{(\rho^2+ iy^2)^{\frac{1+s-\nu}{2}}}.\end{aligned}$$ Similarly letting $\xi=we^{i\pi/4}$, and $a=y e^{-i\pi/4}$ with $y>0$, in , we see that for $\rho>0$ and Re$(s)>-1$, we have $$\begin{aligned}
\label{grn6932}
\int_{0}^{\infty}t^{s+1}J_{s}(\rho t)\frac{K_{\nu}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\nu/2}}\, dt=\frac{\epsilon^{1+\nu+s}\rho^{s}w^{1+s-\nu}}{y^{\nu}}\frac{K_{1+s-\nu}\left(w\epsilon\sqrt{\rho^2-iy^2}\right)}{(\rho^2- iy^2)^{\frac{1+s-\nu}{2}}}.\end{aligned}$$ Adding and , we deduce that for $y>0$, $\rho>0$ and Re$(s)>-1$, $$\begin{aligned}
\label{grn69312}
&\int_{0}^{\infty}t^{s+1}J_{s}(\rho t)\left\{\frac{K_{\nu}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\nu/2}}+\frac{K_{\nu}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\nu/2}}\right\}\, dt\nonumber\\
&=\frac{w^{1+s-\nu}\rho^{s}}{y^{\nu}}\left\{\overline{\epsilon}^{\left(\nu+s+1\right)}\frac{K_{1+s-\nu}(w\overline{\epsilon} \sqrt{\rho^2+iy^2})}{(\rho^2+iy^2)^{\frac{1+s-\nu}{2}}}+\epsilon^{\nu+s+1}\frac{K_{1+s-\nu}(w\epsilon \sqrt{\rho^2-iy^2})}{(\rho^2-iy^2)^{\frac{1+s-\nu}{2}}}\right\}.\end{aligned}$$ From , , and with $\xi$ and $b$ replaced by $w$ and $y$ respectively, we see that for $y>0$, $\rho>0$ and Re$(s)>-1$, $$\begin{aligned}
\label{grn69312f}
&\int_{0}^{\infty}t^{s+1}F_{s}(\rho t)\left\{\frac{K_{\nu}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\nu/2}}+\frac{K_{\nu}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\nu/2}}\right\}\, dt\nonumber\\
&=\frac{w^{1+s-\nu}\rho^{s}}{y^{\nu}}\left\{\frac{K_{1+s-\nu}(w\sqrt{y^2-i\rho^2})}{(y^2-i\rho^2)^{\frac{1+s-\nu}{2}}}+\frac{K_{1+s-\nu}(w\sqrt{y^2+i\rho^2})}{(y^2+i\rho^2)^{\frac{1+s-\nu}{2}}}\right\}.\end{aligned}$$ Now let $\nu=s+1/2$ in the above equation. Then, along with the fact [@grn p. 925, Formula **8.469.3**] that $$\label{khalf}
K_{\pm\frac{1}{2}}(x)=\sqrt{\frac{\pi}{2x}}e^{-x},$$ we deduce that for $y>0$, $\rho>0$ and Re$(s)>-1$, $$\begin{aligned}
&y^{s+\frac{1}{2}}\int_{0}^{\infty}t^{s+1}F_{s}(\rho t)\left\{\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}+\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\frac{s}{2}+\frac{1}{4}}}\right\}\, dt\nonumber\\
&=\sqrt{\frac{\pi}{2}}\rho^{s}\left\{\frac{\textup{exp}\left(-w\sqrt{y^2+(\rho\epsilon)^2}\right)}{\sqrt{y^2+(\rho\epsilon)^2}}+\frac{\textup{exp}\left(-w\sqrt{y^2+(\rho\overline{\epsilon})^2}\right)}{\sqrt{y^2+(\rho\overline{\epsilon})^2}}\right\}.\end{aligned}$$ Now multiply both sides of the above equation by $J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)$, where Re$({\lambda})>-1/2$, integrate both sides of the resulting equation with respect to $y$ from $0$ to $\infty$. The double integral on the left is absolutely convergent as can be seen from , , , and . Hence interchanging the order of integration we see that for $\rho>0$, Re$(s)>-1$ and Re$({\lambda})>-1/2$, $$\begin{aligned}
\label{tfocka}
&\int_{0}^{\infty}t^{s+1}F_{s}(\rho t)\int_{0}^{\infty}y^{s+\frac{1}{2}}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\left\{\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}+\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\frac{s}{2}+\frac{1}{4}}}\right\}\, dy\, dt\nonumber\\
&=\sqrt{\frac{\pi}{2}}\rho^{s}\int_{0}^{\infty}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\left\{\frac{\textup{exp}\left(-w\sqrt{y^2+(\rho\epsilon)^2}\right)}{\sqrt{y^2+(\rho\epsilon)^2}}+\frac{\textup{exp}\left(-w\sqrt{y^2+(\rho\overline{\epsilon})^2}\right)}{\sqrt{y^2+(\rho\overline{\epsilon})^2}}\right\}\, dy.\end{aligned}$$ Our next task is to evaluate the integral on the right-hand side of the above equation. To that end, note that from [@grn p. 708, formula (6.637.1)], we have for Re ${\alpha}>0$, Re ${\beta}>0, {\gamma}>0$, and Re $\nu>-1$, $$\label{grn708-1}
\int_{0}^{\infty}J_{\nu}({\gamma}x)\frac{\textup{exp}\left(-{\alpha}\sqrt{x^2+{\beta}^2}\right)}{\sqrt{x^2+{\beta}^2}}\, dx=I_{\frac{\nu}{2}}\left(\frac{{\beta}}{2}\left(\sqrt{{\alpha}^2+{\gamma}^2}-{\alpha}\right)\right)K_{\frac{\nu}{2}}\left(\frac{{\beta}}{2}\left(\sqrt{{\alpha}^2+{\gamma}^2}+{\alpha}\right)\right).$$ Now let ${\alpha}=w, {\gamma}=\sqrt{z^2-w^2}$ and $\nu=2{\lambda}$ in . Then replace ${\beta}$ first by $\rho\epsilon$ and then by $\rho\overline{\epsilon}$ and add the resulting equations to see that for $\rho>0$ and Re$({\lambda})>-1/2$, $$\begin{aligned}
\label{p11}
&\int_{0}^{\infty}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\left(\frac{\textup{exp}\left(-w\sqrt{(\rho\overline{\epsilon})^2+y^2}\right)}{\sqrt{(\rho\overline{\epsilon})^2+y^2}}+\frac{\textup{exp}\left(-w\sqrt{(\rho\epsilon)^2+y^2}\right)}{\sqrt{(\rho\epsilon)^2+y^2}}\right)\, dy\nonumber\\
&=I_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z+w)}{2}\right)+I_{{\lambda}}\left(\frac{\rho\epsilon(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\epsilon(z+w)}{2}\right).\end{aligned}$$ From and , we have for $\rho>0$, Re$(s)>-1$ and Re$({\lambda})>-1/2$, $$\begin{aligned}
\label{tfocka1}
&\int_{0}^{\infty}t^{s+1}F_{s}(\rho t)\int_{0}^{\infty}y^{s+\frac{1}{2}}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\left\{\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}+\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\frac{s}{2}+\frac{1}{4}}}\right\}\, dy\, dt\nonumber\\
&=\sqrt{\frac{\pi}{2}}\rho^{s}\left\{I_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\overline{\epsilon}(z+w)}{2}\right)+I_{{\lambda}}\left(\frac{\rho\epsilon(z-w)}{2}\right)K_{{\lambda}}\left(\frac{\rho\epsilon(z+w)}{2}\right)\right\}.\end{aligned}$$ To evaluate the inner integral on the left side of , we apply the following formula from [@pbm p. 365, formula **2.16.21.1**], valid for Re$(c)>|$Im$(b)|$ and Re$({\alpha}+u)>|$Re$(v)|$: $$\begin{aligned}
\int_{0}^{\infty}x^{{\alpha}-1}J_{u}(bx)K_{v}(cx)\, dx=2^{{\alpha}-2}b^{u}c^{-{\alpha}-u}\frac{{\Gamma}\left(\frac{{\alpha}+u+v}{2}\right){\Gamma}\left(\frac{{\alpha}+u-v}{2}\right)}{{\Gamma}(u+1)}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{\frac{{\alpha}+u+v}{2}, \frac{{\alpha}+u-v}{2}}
{u+1} \bigg| {-\frac{b^2}{c^2}} \right)}}.\end{aligned}$$ Now replace ${\alpha}$ by $s+3/2$, $u$ by $2{\lambda}$, $v$ by $s+1/2$, $b$ by $\sqrt{z^2-w^2}$ and $c$ by $\sqrt{w^2+it^2}$ in the above equation so that for Re$({\lambda})>-1/2$ and Re$(s+{\lambda})>-1$, we have [$$\begin{aligned}
\label{prudev1}
&\int_{0}^{\infty}y^{s+\frac{1}{2}}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}\, dy\nonumber\\
&=\frac{1}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}\bigg\{2^{s-\frac{1}{2}}(z^2-w^2)^{{\lambda}}(w^2+it^2)^{-\frac{1}{2}\left(s+\frac{3}{2}+2{\lambda}\right)}\frac{{\Gamma}(s+{\lambda}+1){\Gamma}\left({\lambda}+\frac{1}{2}\right)}{{\Gamma}(2{\lambda}+1)}\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{s+{\lambda}+1, {\lambda}+\frac{1}{2}}
{2{\lambda}+1} \bigg| {\frac{w^2-z^2}{w^2+it^2}} \right)}}\bigg\}.\end{aligned}$$]{} We first appropriately transform the ${}_2F_{1}$ in the above equation. By Pfaff’s transformation [@temme p. 110, Equation (5.5)], for $|\arg(1-\xi)|<\pi$, we have $$_2F_1\left(\begin{matrix} a,b\\c\end{matrix}\,\Bigr|\, \xi\right)=(1-\xi)^{-b}{}_2F_1\left(\begin{matrix} c-a,b\\c\end{matrix}\,\Bigr|\, \frac{\xi}{\xi-1}\right).$$ Hence $$\label{2f1pfaff}
{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{s+{\lambda}+1, {\lambda}+\frac{1}{2}}
{2{\lambda}+1} \bigg| {\frac{w^2-z^2}{w^2+it^2}} \right)}}=\left(\frac{z^2+it^2}{w^2+it^2}\right)^{-{\lambda}-\frac{1}{2}}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, {\lambda}+\frac{1}{2}}
{2{\lambda}+1} \bigg| {\frac{z^2-w^2}{z^2+it^2}} \right)}}.$$ Now for $|\xi|<1$, we have [@temme p. 130, Problem 5.7, Equation (3)], $$\label{temmehyp}
{}_2F_{1}\left(\begin{matrix} a,b\\2b\end{matrix}\,\Bigr|\, \xi\right)=\left(\frac{1+\sqrt{1-\xi}}{2}\right)^{-2a}{}_2F_{1}\left(\begin{matrix} a,a-b+\frac{1}{2}\\b+\frac{1}{2}\end{matrix}\,\Bigr|\, \left(\frac{1-\sqrt{1-\xi}}{1+\sqrt{1-\xi}}\right)^2\right).$$ Since $z>w>0$, we have $|z^2-w^2|<|z^2+it^2|$ for all $t>0$. Hence from , $$\begin{aligned}
\label{2f1qt}
{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, {\lambda}+\frac{1}{2}}
{2{\lambda}+1} \bigg| {\frac{z^2-w^2}{z^2+it^2}} \right)}}&=\frac{2^{2{\lambda}-2s}(z^2+it^2)^{{\lambda}-s}}{\left(\sqrt{z^2+it^2}+\sqrt{w^2+it^2}\right)^{2{\lambda}-2s}}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{z^2+it^2}-\sqrt{w^2+it^2}}{\sqrt{z^2+it^2}+\sqrt{w^2+it^2}}\right)^2} \right)}}.\end{aligned}$$ Substitute in to get $$\begin{aligned}
\label{2f1pfaffqt}
{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{s+{\lambda}+1, {\lambda}+\frac{1}{2}}
{2{\lambda}+1} \bigg| {\frac{w^2-z^2}{w^2+it^2}} \right)}}&=\frac{2^{2{\lambda}-2s}}{\left(\sqrt{z^2+it^2}+\sqrt{w^2+it^2}\right)^{2{\lambda}-2s}}\frac{(z^2+it^2)^{-s-\frac{1}{2}}}{(w^2+it^2)^{-{\lambda}-\frac{1}{2}}}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{z^2+it^2}-\sqrt{w^2+it^2}}{\sqrt{z^2+it^2}+\sqrt{w^2+it^2}}\right)^2} \right)}}.\end{aligned}$$ Next, substituting in gives for Re$({\lambda})>-1/2$ and Re$(s+{\lambda})>-1$, $$\begin{aligned}
\label{prudev1s}
\int_{0}^{\infty}y^{s+\frac{1}{2}}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2+it^2}\right)}{(w^2+it^2)^{\frac{s}{2}+\frac{1}{4}}}\, dy
=\frac{2^{-s-\frac{1}{2}}\sqrt{\pi}{\Gamma}(s+{\lambda}+1)}{{\Gamma}({\lambda}+1)}\mathcal{A}(t^2),\end{aligned}$$ upon simplification, where $\mathcal{A}(t)$ is defined in .
Similarly for Re$({\lambda})>-1/2$ and Re$(s+{\lambda})>-1$, we have $$\begin{aligned}
\label{prudev2s}
\int_{0}^{\infty}y^{s+\frac{1}{2}}J_{2{\lambda}}\left(y\sqrt{z^2-w^2}\right)\frac{K_{s+\frac{1}{2}}\left(y\sqrt{w^2-it^2}\right)}{(w^2-it^2)^{\frac{s}{2}+\frac{1}{4}}}\, dy=\frac{2^{-s-\frac{1}{2}}\sqrt{\pi}{\Gamma}(s+{\lambda}+1)}{{\Gamma}({\lambda}+1)}\mathcal{A}(-t^2).\end{aligned}$$ Add the corresponding sides of and and substitute the right-hand side of the resulting equation for the inner integral in so as to deduce for $z>w>0$ after simplification.
Now consider $z$ and $w$ to be complex numbers. The integrand of the left-hand side of is well-defined except for the singularities on the lines $\arg z=\pm\frac{\pi}{4}$ and $\arg w=\pm\frac{\pi}{4}$. The right-hand side is well-defined for all complex $z, w$ such that $z\neq\pm w$. Therefore, can be analytically continued in the region $|\arg z|<\pi/4, |\arg w|<\pi/4$ and $(z)>$ $(w)>0$.
Proof of Theorem \[anathmbdkz\] {#proof2}
===============================
We first prove the result for ${\alpha}>{\beta}>0$ and later extend it by analytic continuation to the region $(\sqrt{{\alpha}})\geq$ $(\sqrt{{\beta}})>0$. Note that $s, {\lambda}, {\alpha}$ and ${\beta}$ are fixed throughout the proof. In Theorem \[tfockathm\], let $\rho=4\pi\sqrt{x}$, $x>0$, replace $z, w$ and $t$ by $\sqrt{{\alpha}}, \sqrt{{\beta}}$ and $\sqrt{t}$ respectively. Define $$\label{f}
f(t):=\frac{1}{2\pi}t^{\frac{s}{2}}\left(\mathcal{A}(t)+\mathcal{A}(-t))\right),$$ where $\mathcal{A}(t)$ is the same as defined in except that we replace $z$ and $w$ by $\sqrt{{\alpha}}$ and $\sqrt{{\beta}}$ respectively.
Then from Theorem \[tfockathm\], for $x>0$, Re$(s)>-1$, Re$({\lambda})>-1/2$ and Re$(s+{\lambda})>-1$, $$\begin{aligned}
\label{g0}
2\pi \int_{0}^{\infty} f(t) F_s(4\pi\sqrt{xt}) \,dt&=\frac{2{\Gamma}({\lambda}+1)(8\pi\sqrt{x})^{s}}{{\Gamma}(s+{\lambda}+1)}\bigg(I_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{x}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{x}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&\quad+I_{{\lambda}}\left(2\pi\epsilon\sqrt{x}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon\sqrt{x}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\bigg).\end{aligned}$$ The function $f$ defined in satisfies the conditions given before so that Guinand’s result could be applied. To see this, first note ${\alpha}>{\beta}>0$ and Re$({\lambda})>0$. Since $f$ is infinitely differentiable, $f$ and $f'$ are integrals. Also, as $|\xi|\to \infty$, one has [@bdkz Equation (4.4)] $$\begin{aligned}
\label{asssy}
\mathcal{A}(-i\xi)\sim \frac{4^{s-{\lambda}}(\alpha-\beta)^{{\lambda}}}{\xi^{{\lambda}+s+1}},\end{aligned}$$ since the hypergeometric function tends to $1$. By the hypotheses of Theorem \[anathmbdkz\], we have Re$({\lambda})>0$, Re$(s+{\lambda})>0$ or $s=-{\lambda}$. Hence we conclude that $f$ tends to zero as $t\to\infty$. Further, doing the analysis similar to that done in [@bdkz p. 322] for $r_k(n)$, one can prove that $f(t)$, $tf'(t)$ and $t^2f''(t)$ belong to $L^{2}(0,\infty)$. Thus the hypotheses for applying Guinand’s result are satisfied.
Note that and imply $$\begin{aligned}
\label{g}
g(x)&=\frac{2{\Gamma}({\lambda}+1)(8\pi\sqrt{x})^{s}}{{\Gamma}(s+{\lambda}+1)}\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{x}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{x}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&\quad+I_{{\lambda}}\left(2\pi\epsilon\sqrt{x}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon\sqrt{x}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\bigg\}.\end{aligned}$$ Next, we need to evaluate the integrals $$\int_{0}^{\infty} y^{\pm\frac{s}{2}} f(y) \,dy\hspace{5mm}\text{and}\hspace{5mm}\int_{0}^{\infty} y^{\pm\frac{s}{2}} g(y) \,dy.$$ To that end, let $y\geq 0$. Then for Re$(\mu+{\lambda}+1)>0$, we have [@bdkz Equation (4.13)], $$\begin{aligned}
\label{itransform}
\int_{0}^{\infty}t^{\mu}I_{\lambda}(t)e^{-\frac{(2y+\alpha+\beta)t}{\alpha-\beta}}\,dt=&\frac{2^{-3\mu-1}}{\sqrt{y+\alpha}\sqrt{y+\beta}}\frac{(\alpha-\beta)^{\lambda+\mu+1}\Gamma(\lambda+\mu+1)}{(\sqrt{y+\alpha}+\sqrt{y+\beta})^{2\lambda}\Gamma(\lambda+1)}\left(\frac{1}{\sqrt{y+\alpha}}+\frac{1}{\sqrt{y+\beta}}\right)^{2\mu}\nonumber\\
& \times {\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-\mu, -\mu}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{y+\alpha}-\sqrt{y+\beta}}{\sqrt{y+\alpha}+\sqrt{y+\beta}}\right)^2} \right)}}.\end{aligned}$$ Let $\nu$ be a complex number such that Re$(\nu)>-1$. Then from , we see that for Re$(s+{\lambda}+1)>0$, [$$\begin{aligned}
\label{doublei}
\int_{0}^{\infty}y^{\nu}\mathcal{A}(-iy)\, dy&=\frac{2^{3s+1}(\alpha-\beta)^{-s-1}\Gamma(\lambda+1)}{\Gamma({\lambda}+s+1)}\int_0^{\infty}\int_0^{\infty}y^{\nu}t^se^{-\frac{(2y+\alpha+\beta)t}{\alpha-\beta}}I_{{\lambda}}(t)\,dt\,dy\nonumber\\
&=\frac{2^{3s+1}(\alpha-\beta)^{-s-1}\Gamma(\lambda+1)}{\Gamma({\lambda}+s+1)}\int_0^{\infty}t^sI_{{\lambda}}(t)e^{-\frac{(\alpha+\beta)t}{\alpha-\beta}}\left(\int_0^{\infty}y^{\nu}e^{-\frac{2yt}{\alpha-\beta}}\,dy\right)\,dt,\end{aligned}$$]{}where the interchange of the order of integration is justified since the double integral in the first line is absolutely convergent, for, as $t\to\infty$, we have [@temme p. 240, Equation (9.54)] $$\begin{aligned}
\label{iasy}
I_{{\lambda}}(t)\sim \frac{e^t}{\sqrt{2\pi t}},\end{aligned}$$ where Re$({\lambda})>0$ and $|{\lambda}|\to\infty$. Since Re$(\nu)>-1$, expressing the inner integral on the extreme right of in the form of a gamma function, we have for Re$({\lambda}+s-\nu)>0$, $$\begin{aligned}
\label{doubleis}
\int_{0}^{\infty}y^{\nu}\mathcal{A}(-iy)\, dy&=\frac{2^{3s-\nu}(\alpha-\beta)^{\nu-s}\Gamma(\lambda+1)\Gamma(\nu+1)}{\Gamma({\lambda}+s+1)}\int_0^{\infty}t^{s-\nu-1}I_{{\lambda}}(t)e^{-\frac{(\alpha+\beta)t}{\alpha-\beta}}\,dt\nonumber\\
&=\frac{4^{\nu+1}\Gamma(\lambda+s-\nu)\Gamma(\nu+1)}{\Gamma({\lambda}+s+1)}\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}\frac{1}{\sqrt{\alpha\beta}}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2(s-\nu-1)}\\\nonumber
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s+\nu+1, -s+\nu+1}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}},\end{aligned}$$ where in the last step, we again employed , this time with $y=0$ and $\mu=s-\nu-1$.
Let $C(\theta):=C(R, \theta):=\big\{\xi=Re^{i\theta}: 0\leq \theta\leq \frac{\pi}{2}, R>0\big\}$ be a circular arc in the first quadrant. Then $C(-\theta)$ is a circular arc in the fourth quadrant. From , clearly, $$\begin{aligned}
\int_{C(\theta)}\xi^{\nu}\mathcal{A}(-i\xi)\, d\xi,\quad \int_{C(-\theta)}\xi^{\nu}\mathcal{A}(-i\xi) d\xi\ll R^{-\text{Re }(s+{\lambda}-\nu+1)},\end{aligned}$$ and the latter tends to zero as $R\to \infty$ since Re$({\lambda}+s-\nu)>0$. Note that $\xi^{\nu}\mathcal{A}(\xi)$ is an analytic function for $\text{Re}(\xi)\geq 0$. Thus invoking Cauchy’s residue theorem twice, once along the closed contour consisting of $[0, R]$, $C(\theta)$ and $[iR, 0]$, and then, along its reflection along the real axis, we find that $$\begin{aligned}
\int_0^{i\infty}\xi^{\nu}\mathcal{A}(-i\xi)\, d\xi=\int_0^{\infty}\xi^{\nu}\mathcal{A}(-i\xi)\, d\xi=-\int_{-i\infty}^0\xi^{\nu}\mathcal{A}(-i\xi)\, d\xi.\end{aligned}$$ Therefore, $$\begin{aligned}
\label{hintfirst}
\int_0^{\infty}y^{\nu}\mathcal{A}(y)\, dy=i^{-\nu-1}\int_0^{\infty}y^{\nu}\mathcal{A}(-iy)\, dy\end{aligned}$$ and $$\begin{aligned}
\label{hintfourth}
\int_0^{\infty}y^{\nu}\mathcal{A}(-y)\, dy=i^{\nu+1}\int_0^{\infty}y^{\nu}\mathcal{A}(-iy)\, dy,\end{aligned}$$ so that from , , and , we have for $\text{Re}(\nu)>-1$ and $\text{Re}(s+{\lambda}-\nu)>0$, $$\begin{aligned}
\label{nuinte}
\int_{0}^{\infty}y^{\nu}(\mathcal{A}(y)+\mathcal{A}(-y))\, dy&=-\frac{2^{2\nu+3}\sin\left(\frac{\pi\nu}{2}\right)\Gamma(\lambda+s-\nu)\Gamma(\nu+1)}{\sqrt{\alpha\beta}\Gamma({\lambda}+s+1)}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2(s-\nu-1)}\nonumber\\
&\quad\times\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s+\nu+1, -s+\nu+1}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}.\end{aligned}$$ We are now ready to evaluate $\int_{0}^{\infty} y^{\pm\frac{s}{2}} f(y) \,dy$. To that end, let $\nu=s$ in so that for Re$(s)>-1$ and Re$({\lambda})>0$, we have using , $$\begin{aligned}
\label{f1}
\int_{0}^{\infty} y^{\frac{s}{2}} f(y) \,dy=-\frac{2^{2s}\Gamma({\lambda})\Gamma(s+1)\sin\left(\frac{\pi s}{2}\right)}{\pi\Gamma({\lambda}+s+1)}\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}.\end{aligned}$$ Now first let $\nu=0$ in so that for Re$(s+{\lambda})>0$, $$\begin{aligned}
\label{f2}
\int_{0}^{\infty} y^{-\frac{s}{2}} f(y) \,dy=0.\end{aligned}$$ We now evaluate the integral in when $s=-{\lambda}$. Note that in this case, implies that for $-1<$ Re$(\nu)<0$, $$\begin{aligned}
\label{nuinte1}
\int_{0}^{\infty}y^{\nu}(\mathcal{A}(y)+\mathcal{A}(-y))\, dy&=\frac{\pi 2^{2\nu+2}({\alpha}{\beta})^{{\lambda}+\nu+1/2}(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}{\cos\left(\frac{\pi\nu}{2}\right)(\sqrt{{\alpha}}+\sqrt{{\beta}})^{3{\lambda}+2\nu+2}}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{2{\lambda}+\nu+1, {\lambda}+\nu+1}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}},\end{aligned}$$ where we used a variant of the reflection formula for the gamma function [@temme p. 74], namely, $${\Gamma}\left(\frac{1}{2}+\theta\right){\Gamma}\left(\frac{1}{2}-\theta\right)=\frac{\pi}{\cos(\pi\theta)},\hspace{1.5mm}\text{where}\hspace{1.5mm}\theta-\frac{1}{2}\notin\mathbb{Z}.$$ Note that both sides of are analytic at $\nu=0$, hence letting $\nu=0$ in gives $$\begin{aligned}
\label{f21}
\int_{0}^{\infty} y^{-\frac{s}{2}} f(y) \,dy&=\frac{2({\alpha}{\beta})^{{\lambda}+\frac{1}{2}}(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^{3{\lambda}+2}}{\ensuremath{{}_{1}F_{0} \left( \genfrac{}{}{0pt}{}{2{\lambda}+1}
{-} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}\nonumber\\
&=\frac{2({\alpha}{\beta})^{{\lambda}+\frac{1}{2}}(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^{3{\lambda}+2}}\sum_{n=0}^{\infty}\frac{(2{\lambda}+1)_n}{n!}\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{2n}\nonumber\\
&=\frac{2({\alpha}{\beta})^{{\lambda}+\frac{1}{2}}(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^{3{\lambda}+2}}\frac{(4\sqrt{{\alpha}{\beta}})^{-2{\lambda}-1}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^{-4{\lambda}-2}}\nonumber\\
&=\frac{({\alpha}-{\beta})^{{\lambda}}}{2^{4{\lambda}+1}},\end{aligned}$$where in the third step, we employed the generalized binomial theorem [@temme p. 108, Equation (5.2)], namely, for $|\xi|<1$, $$(1-\xi)^{-a}=\sum_{n=0}^{\infty}\frac{(a)_n}{n!}\xi^{n}.$$ Combining and and using the definition of $h(s, {\lambda})$ in , we can write $$\label{fboth}
\int_{0}^{\infty} y^{-\frac{s}{2}} f(y) \,dy=h(s,{\lambda}).$$ To evaluate the remaining two integrals, namely, $\int_{0}^{\infty} y^{\pm\frac{s}{2}} g(y) \,dy$, we first note from [@dixfer4 Equation (24)][^6] (see also [@pbm p. 380-381, Equation **2.16.28.1**]) that for $|\textup{Re}(b)|<$Re$(c)$ and $|\textup{Re}(\nu)|<$ Re$(a+\nu)$, $$\begin{aligned}
\label{xik}
\int_0^{\infty}y^{a-1}I_{\nu}(by)K_{\nu}(cy)\, dy
&=2^{a-2}c^{-a-\nu}b^{\nu}\frac{{\Gamma}\left(a/2\right){\Gamma}\left(\nu+a/2\right)}{{\Gamma}(\nu+1)}{}_2F_{1}\left(\begin{matrix} \nu+\frac{a}{2},\frac{a}{2}\\\nu+1\end{matrix}\,\Bigr|\, \frac{b^2}{c^2}\right)\nonumber\\
&=\frac{(2c)^{a-2}(b/c)^{\nu}\Gamma(a/2)\Gamma(\nu+a/2)}{(c^2-b^2)^{a-1}\Gamma(\nu+1)}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{1+\nu-a/2, 1-a/2}
{1+\nu} \bigg| {\frac{b^2}{c^2}} \right)}},\end{aligned}$$ where the second equality follows from Euler’s transformation [@temme p. 110, Equation (5.5)] $$_2F_1\left(\begin{matrix} a,b\\c\end{matrix}\,\Bigr|\, \xi\right)=(1-\xi)^{c-a-b}{}_2F_1\left(\begin{matrix} c-a,c-b\\c\end{matrix}\,\Bigr|\, \xi\right).$$ In , replace $y$ by $\sqrt{y}$ and let $a=2(s+1)$, $b=2\pi\bar{\epsilon}(\sqrt{\alpha}-\sqrt{\beta})$, $c=2\pi\bar{\epsilon}(\sqrt{\alpha}+\sqrt{\beta})$ and $\nu=\lambda$ so that for $\text{Re}(s)>-1$ and $\text{Re}(s+{\lambda})>-1$, $$\begin{aligned}
\label{xik1}
&\int_0^{\infty}y^{s}I_{{\lambda}}(2\pi\bar{\epsilon}\sqrt{y(\alpha-\beta)})K_{{\lambda}}(2\pi\bar{\epsilon}\sqrt{y(\alpha+\beta)})\, dy\\\nonumber
&=\frac{i^{s+1}\Gamma(\lambda+s+1)\Gamma(s+1)}{\sqrt{\alpha\beta}(2\pi)^{2s+2}2^{2s+1}\Gamma({\lambda}+1)}\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}.\end{aligned}$$ Similarly for $\text{Re}(s)>-1$ and $\text{Re}(s+{\lambda})>-1$, $$\begin{aligned}
\label{xik2}
&\int_0^{\infty}y^{s}I_{{\lambda}}(2\pi{\epsilon}\sqrt{y(\alpha-\beta)})K_{{\lambda}}(2\pi{\epsilon}\sqrt{y(\alpha+\beta)})\, dy\\\nonumber
&=\frac{(-i)^{s+1}\Gamma(\lambda+s+1)\Gamma(s+1)}{\sqrt{\alpha\beta}(2\pi)^{2s+2}2^{2s+1}\Gamma({\lambda}+1)}\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}.\end{aligned}$$ From , and , we deduce that for $\text{Re}(s)>-1$ and $\text{Re}(s+{\lambda})>-1$, $$\begin{aligned}
\label{xik3}
\int_0^{\infty}y^{\frac{s}{2}}g(y)\, dy&=-\frac{(2\pi)^{-s-1}}{\pi\sqrt{{\alpha}{\beta}}}{\Gamma}(s+1)\sin\left(\frac{\pi s}{2}\right)\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2s}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}.\end{aligned}$$ Similarly from and , we find that for Re$({\lambda})>-1$, $$\begin{aligned}
\label{xik4}
&\int_0^{\infty}y^{-\frac{s}{2}}g(y)\, dy=0.\end{aligned}$$ Finally, from , , , , , , and , we deduce that for $-\frac{1}{2}<$ Re$(s)<\frac{1}{2}$, Re$({\lambda})>0$ and either Re$(s+{\lambda})>0$ or $s=-{\lambda}$, [$$\begin{aligned}
&\frac{1}{2\pi}\sum_{n=1}^{\infty}\sigma_s(n)(\mathcal{A}(n)+\mathcal{A}(-n))+\frac{2^{2s}\Gamma({\lambda})\Gamma(s+1)\sin\left(\frac{\pi s}{2}\right)}{\pi\Gamma({\lambda}+s+1)}\zeta(s+1)\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}-\zeta(1-s)h(s, {\lambda})\nonumber\\
&=\frac{2{\Gamma}({\lambda}+1)(8\pi)^{s}}{{\Gamma}(s+{\lambda}+1)}\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&\quad+I_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\bigg\}\nonumber\\
&\quad+\frac{(2\pi)^{-s-1}}{\pi\sqrt{{\alpha}{\beta}}}{\Gamma}(s+1)\sin\left(\frac{\pi s}{2}\right)\zeta(s+1)\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{\alpha}}+\frac{1}{\sqrt{\beta}}\right)^{2s}\nonumber\\
&\quad\times{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{\alpha}-\sqrt{\beta}}{\sqrt{\alpha}+\sqrt{\beta}}\right)^2} \right)}}.\end{aligned}$$]{} Now multiply both sides of the above equation by $\frac{{\Gamma}(s+{\lambda}+1)}{2{\Gamma}({\lambda}+1)(8\pi)^{s}}$, use the asymmetric form of the functional equation of $\zeta(-s)$ [@titch p. 25], namely, $$\label{asyfe}
\zeta(-s)=-2^{-s}\pi^{-s-1}{\Gamma}(1+s)\zeta(1+s)\sin\left(\frac{\pi s}{2}\right)$$ to simplify the resulting equation and use to arrive at for ${\alpha}>{\beta}>0$, $-\frac{1}{2}<$ Re$(s)<\frac{1}{2}$, Re$({\lambda})>0$ and either Re$(s+{\lambda})>0$ or $s=-{\lambda}$.
Now it can be seen that both sides of are analytic as long as Re$(\sqrt{{\alpha}})>$Re$(\sqrt{{\beta}})>0$, $|\arg{\alpha}|<\pi/2$, $|\arg{\beta}|<\pi/2$, Re$({\lambda})>0$, and $s\in\mathbb{C}$ such that either Re$(s+{\lambda})>0$ or $s=-{\lambda}$. To see why this is true even for $s=-1$, note that the power series expansions $$\begin{aligned}
\label{pp1}
-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}=\frac{1}{2{\lambda}(s+1)}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}-\frac{{\gamma}}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+O(|s+1|),\end{aligned}$$ and $$\begin{aligned}
\label{pp2}
&\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&=-\frac{1}{2{\lambda}(s+1)}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+c_{{\alpha}, {\beta}, {\lambda}}+O(|s+1|),\end{aligned}$$ where $c_{{\alpha}, {\beta}, {\lambda}}$ is some constant independent of $s$, clearly show that the principal parts of the left-hand sides of and completely cancel out.
Hence by analytic continuation, is valid for Re$(\sqrt{{\alpha}})>$Re$(\sqrt{{\beta}})>0$, $|\arg{\alpha}|<\pi/2$, $|\arg{\beta}|<\pi/2$, Re$({\lambda})>0$, and for all complex $s$ such that either Re$(s+{\lambda})>0$ or $s=-{\lambda}$.
Proofs of Corollaries of Theorem \[anathmbdkz\] {#cor1}
===============================================
First let Re$({\lambda})>0$. Let $s=-{\lambda}$ in Theorem \[anathmbdkz\]. Then divide both sides by $(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}$ and let ${\alpha}\to{\beta}$. This gives $$\begin{aligned}
\label{l0}
&\sum_{n=1}^{\infty}\sigma_{-{\lambda}}(n)\lim_{{\alpha}\to{\beta}}\frac{1}{({\alpha}-{\beta})
^{{\lambda}}}\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad\qquad\qquad\qquad\quad\quad+I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{\pi^{{\lambda}}{\beta}^{{\lambda}/2}\zeta({\lambda}+1)}{4{\Gamma}({\lambda}+1)}-\frac{\zeta({\lambda})}{{\lambda}2^{{\lambda}+1}{\beta}^{{\lambda}/2}}+\frac{\pi^{{\lambda}-1}{\beta}^{-1+{\lambda}/2}\zeta({\lambda})}{4{\Gamma}({\lambda}+1)}\nonumber\\
&\quad+\frac{1}{2^{2-3{\lambda}}\pi^{1-{\lambda}}{\Gamma}({\lambda}+1)}\sum_{n=1}^{\infty}\sigma_{-{\lambda}}(n)\left\{\frac{2^{{\lambda}}{\beta}^{{\lambda}/2}({\beta}+in)^{2{\lambda}-1}}{2^{4{\lambda}}({\beta}+in)^{2{\lambda}}}+\frac{2^{{\lambda}}{\beta}^{{\lambda}/2}({\beta}-in)^{2{\lambda}-1}}{2^{4{\lambda}}({\beta}-in)^{2{\lambda}}}\right\}.\end{aligned}$$ Let $L:=L(n, {\lambda}, {\beta})$ denote the limit in the above equation. We show that $$\label{l}
L=\frac{(\pi\sqrt{n})^{{\lambda}}}{{\Gamma}({\lambda}+1)}\left(e^{\frac{i\pi{\lambda}}{4}}K_{{\lambda}}\left(4\pi e^{\frac{i\pi}{4}}\sqrt{n{\beta}}\right)+e^{-\frac{i\pi{\lambda}}{4}}K_{{\lambda}}\left(4\pi e^{-\frac{i\pi}{4}}\sqrt{n{\beta}}\right)\right).$$ To that end, note that Re$(\sqrt{{\alpha}})>$Re$(\sqrt{{\beta}})$ implies $-\frac{\pi}{4}<\arg(2\pi{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}}))<\frac{3\pi}{4}$. First, consider $-\frac{\pi}{4}<\arg(2\pi{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}}))\leq\frac{\pi}{2}$. Then using and , we see that $$\begin{aligned}
\label{l1}
&\lim_{{\alpha}\to{\beta}}\frac{I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)}{(\sqrt{{\alpha}}-\sqrt{{\beta}})
^{{\lambda}}}\nonumber\\
&=e^{\frac{-\pi i{\lambda}}{2}}K_{{\lambda}}(4\pi{\epsilon}\sqrt{n{\beta}})\lim_{{\alpha}\to{\beta}}\frac{J_{{\lambda}}\left(2\pi e^{\frac{3\pi i}{4}}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)}{(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}\nonumber\\
&=e^{\frac{-\pi i{\lambda}}{2}}K_{{\lambda}}(4\pi{\epsilon}\sqrt{n{\beta}})\nonumber\\
&\quad\times\lim_{{\alpha}\to{\beta}}\frac{1}{(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}\left(\frac{e^{\frac{3\pi i{\lambda}}{4}}(\pi\sqrt{n})^{{\lambda}}({\alpha}-{\beta})^{{\lambda}}}{{\Gamma}({\lambda}+1)}+\sum_{m=1}^{\infty}\frac{(-1)^m\left(\pi e^{\frac{3\pi i}{4}}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)^{2m+{\lambda}}}{m!{\Gamma}(m+1+{\lambda})}\right)\nonumber\\
&=\frac{(\pi{\epsilon}\sqrt{n})^{{\lambda}}}{{\Gamma}({\lambda}+1)}K_{{\lambda}}(4\pi{\epsilon}\sqrt{n{\beta}}).\end{aligned}$$ Similarly, it can be shown that we get the same evaluation as above when $\frac{\pi}{2}<\arg(2\pi{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}}))<\frac{3\pi}{4}$. Also, one can show in the same way that $$\label{l2}
\lim_{{\alpha}\to{\beta}}\frac{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)}{(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}}=\frac{(\pi\overline{{\epsilon}}\sqrt{n})^{{\lambda}}}{{\Gamma}({\lambda}+1)}K_{{\lambda}}(4\pi\overline{{\epsilon}}\sqrt{n{\beta}}).$$ Thus and establish . Now substitute in , multiply both sides of the resulting equation by $\frac{2{\Gamma}({\lambda}+1)}{\pi^{{\lambda}}}$ and simplify so as to obtain for Re$({\lambda})>0$. The result extends to Re$({\lambda})>-1$ by analytic continuation.
This proof begins by first obtaining a result complementary to . To be more specific, let $s\neq-{\lambda}$ and Re$(s+{\lambda})>0$ in . Divide both sides of by by $(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}$, then let ${\alpha}\to{\beta}$ and then simplify so as to deduce $$\begin{aligned}
&\sum_{n=1}^{\infty} \sigma_{s}(n) n^{\frac{{\lambda}}{2}}
\left( e^{\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi {\epsilon}\sqrt{n{\beta}} ) +
e^{-\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi \overline{{\epsilon}} \sqrt{n{\beta}} ) \right)\nonumber\\
&=-\frac{\zeta(-s){\Gamma}({\lambda})}{2^{{\lambda}+1}\pi^{{\lambda}}{\beta}^{{\lambda}/2}}+\frac{\zeta(-s){\Gamma}({\lambda}+s+1)}{2(2\pi)^{s+{\lambda}+1}{\beta}^{s+\frac{{\lambda}}{2}+1}}\nonumber\\
&\quad+\frac{{\beta}^{\frac{{\lambda}}{2}}{\Gamma}({\lambda}+s+1)}{2(2\pi)^{s+{\lambda}+1}}\sum_{n=1}^{\infty}\sigma_s(n)\left\{\frac{1}{({\beta}+in)^{s+{\lambda}+1}}+\frac{1}{({\beta}-in)^{s+{\lambda}+1}}\right\}.\end{aligned}$$ Now follows from letting ${\lambda}=1/2$, using , and simplifying.
Let $s\to-1$ in Theorem \[anathmbdkz\]. This gives for Re$({\lambda})>1$ or ${\lambda}=1$, [$$\begin{aligned}
\label{exotics-1}
&\sum_{n=1}^{\infty}\sigma_{-1}(n)\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{2\pi^3}{3{\lambda}}h(-1, {\lambda})+\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}L_{1}({\alpha}, {\beta}, {\lambda})\nonumber\\
&\quad+\frac{2}{{\lambda}}\sum_{n=1}^{\infty}\sigma_{-1}(n)\left(\mathcal{A}\left(-1, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)+\mathcal{A}\left(-1, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)\right),\end{aligned}$$]{} where $$\begin{aligned}
\label{lalim}
L_{1}({\alpha}, {\beta}, {\lambda})&:=\lim_{s\to-1}\bigg\{-\frac{\zeta(-s)}{2{\lambda}}+\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\nonumber\\
&\quad\times\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\bigg\}.\end{aligned}$$ The above limit is now evaluated. Note that the Laurent series expansion of $\zeta(s)$ around $s=1$ gives $$\label{laurent}
\zeta(s)=\frac{1}{s-1}+{\gamma}+O(|s-1|).$$ Hence $$\begin{aligned}
\label{one}
-\frac{\zeta(-s)}{2{\lambda}}=\frac{1}{2{\lambda}(s+1)}-\frac{{\gamma}}{2{\lambda}}+O(|s+1|).\end{aligned}$$ as $s\to-1$. Also, $$\begin{aligned}
b^{s}&=\frac{1}{b}+\frac{\log b}{b}(s+1)+O(|s+1|^2),\label{bs}\\
{\Gamma}({\lambda}+s+1)&={\Gamma}({\lambda})+{\Gamma}'({\lambda})(s+1)+O(|s+1|^2),\label{gls}\end{aligned}$$ and $$\begin{aligned}
\label{hype}
&{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&=\frac{1}{1-\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2}+(s+1)\lim_{s\to-1}\left[\frac{d}{ds}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\right]+O(|s+1|^2).\end{aligned}$$ Next, $$\begin{aligned}
&\lim_{s\to-1}\left[\frac{d}{ds}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\right]\nonumber\\
&=\sum_{k=1}^{\infty}\frac{1}{k!(1+{\lambda})_k}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{2k}\lim_{s\to-1}\left[(-s)_k\frac{d}{ds}({\lambda}-s)_k+({\lambda}-s)_k\frac{d}{ds}(-s)_k\right]\end{aligned}$$ Using the fact $$\begin{aligned}
\frac{d}{d\xi}(\xi)_k=\frac{d}{d\xi}\frac{{\Gamma}(\xi+k)}{{\Gamma}(\xi)}=\frac{{\Gamma}(\xi+k)}{{\Gamma}(\xi)}\left(\psi(\xi+k)-\psi(\xi)\right),\end{aligned}$$ we find, after simplification, that $$\begin{aligned}
\label{derev}
&\lim_{s\to-1}\left[\frac{d}{ds}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\right]\nonumber\\
&=-\frac{(\sqrt{{\alpha}}+\sqrt{{\beta}})^2}{4\sqrt{{\alpha}{\beta}}}\left[\sum_{k=0}^{\infty}\frac{1}{1+{\lambda}+k}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{2k+2}-\log\left(\frac{4\sqrt{{\alpha}{\beta}}}{(\sqrt{{\alpha}}+\sqrt{{\beta}})^2}\right)\right].\end{aligned}$$ Now substitute in and use the resultant along with with $b=\frac{1}{8\pi}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^2$, and to obtain $$\begin{aligned}
\label{two}
&\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&=\frac{1}{4\pi\sqrt{{\alpha}{\beta}}{\Gamma}({\lambda}+1)}\left(\frac{1}{8\pi}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^2\right)^s{\Gamma}({\lambda}+s+1)\zeta(-s){\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&=-\frac{1}{2{\lambda}(s+1)}+\frac{{\gamma}}{2{\lambda}}+\frac{1}{2{\lambda}}\log(2\pi\sqrt{{\alpha}{\beta}})-\frac{1}{2{\lambda}}\psi({\lambda})+\frac{1}{2{\lambda}}\sum_{k=0}^{\infty}\frac{1}{1+{\lambda}+k}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{2k+2}\nonumber\\
&\quad+O(|s+1|)\end{aligned}$$ as $s\to-1$. Now add the corresponding sides of and so as to deduce from that $$\begin{aligned}
L_{1}({\alpha}, {\beta}, {\lambda})=\frac{1}{2{\lambda}}\left\{\sum_{k=0}^{\infty}\frac{1}{1+{\lambda}+k}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{2k+2}+\log(2\pi\sqrt{{\alpha}{\beta}})-\psi({\lambda})\right\}.\end{aligned}$$ Finally let ${\lambda}=1$ in , simplify the above limit, use the facts $h(-1,1)=({\alpha}-{\beta})/32$, where $h(s,{\lambda})$ is defined in , $\sum_{k=0}^{\infty}\xi^k/(k+2)=-1-\frac{\log(1-\xi)}{\xi}$ and $\mathcal{A}\left(-1, 1, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)=\frac{1}{4}\left(\frac{\sqrt{{\alpha}+in}-\sqrt{{\beta}+in}}{\sqrt{{\alpha}+in}+\sqrt{{\beta}+in}}\right)$ to arrive at after simplification.
Analytic continuation of Theorem \[anathmbdkz\] {#ac}
===============================================
Here we analytically continue Theorem \[anathmbdkz\] to complex values of ${\lambda}$ such that Re$({\lambda})>-1$. The special case ${\lambda}=-1/2$ of the resulting identity is then shown to have a form surprisingly similar to Ramanujan’s .
\[anathmbdkzac\] Let $\mathcal{A}(s, {\lambda}, z, w, t)$ be defined in . Let $|\arg {\alpha}|<\pi/2, |\arg {\beta}|<\pi/2$ with $(\sqrt{{\alpha}})>$ $(\sqrt{{\beta}})$. Let $({\lambda})>-1$. Then for all complex $s$ such that either $(s+{\lambda})>0$ or $s=-{\lambda}$, $$\begin{aligned}
\label{anathmbdkzeqnac}
&\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\nonumber\\
&\quad\times\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&\quad+\frac{\pi^{{\lambda}}({\alpha}-{\beta})^{{\lambda}}\zeta({\lambda}+1)\zeta(-s-{\lambda})}{2^{{\lambda}+1}{\Gamma}({\lambda}+1)}+\frac{{\Gamma}({\lambda}+s+1)}{2^{3s+2}\pi^{s+1}{\Gamma}({\lambda}+1)}\sum_{n=1}^{\infty}\sigma_s(n)\bigg[\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)\nonumber\\
&\quad+\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)+\frac{2^{2s-2{\lambda}+1}({\alpha}-{\beta})^{{\lambda}}}{n^{{\lambda}+s+1}}\sin\left(\tfrac{\pi}{2}({\lambda}+s)\right)\bigg].\end{aligned}$$
First let Re$({\lambda})>0$ and Re$(s+{\lambda})>0$. Rewrite in the form $$\begin{aligned}
\label{anathmbdkzeqno}
&\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{I_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+\frac{\zeta(-s)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\frac{{\Gamma}({\lambda}+s+1)}{{\Gamma}({\lambda}+1)}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}\nonumber\\
&\quad\times\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{{\lambda}-s, -s}
{{\lambda}+1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}+\frac{{\Gamma}({\lambda}+s+1)}{2^{3s+2}\pi^{s+1}{\Gamma}({\lambda}+1)}\nonumber\\
&\quad\times\sum_{n=1}^{\infty}\sigma_s(n)\bigg[\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)+\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)+\tfrac{2^{2s-2{\lambda}+1}({\alpha}-{\beta})^{{\lambda}}}{n^{{\lambda}+s+1}}\sin\left(\tfrac{\pi}{2}({\lambda}+s)\right)\bigg]\nonumber\\
&\quad-\frac{2^{-s-2{\lambda}-1}({\alpha}-{\beta})^{{\lambda}}{\Gamma}({\lambda}+s+1)}{\pi^{s+1}{\Gamma}({\lambda}+1)}\zeta({\lambda}+s+1)\zeta({\lambda}+1)\sin\left(\frac{\pi}{2}({\lambda}+s)\right),\end{aligned}$$ where in the last step, we used [@titch p. 8, Equation (1.3.1)] $$\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{n^{{\lambda}+s+1}}=\zeta({\lambda}+s+1)\zeta({\lambda}+1),$$ which is valid for Re$({\lambda})>0$ and Re$(s+{\lambda})>0$.
Now it can be easily seen that $$\begin{aligned}
\label{ae}
\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)+\mathcal{A}\left(s, {\lambda}, \sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)&=-\frac{2^{2s-2{\lambda}+1}({\alpha}-{\beta})^{{\lambda}}}{n^{{\lambda}+s+1}}\sin\left(\frac{\pi}{2}({\lambda}+s)\right)\nonumber\\
&\quad+O_{{\alpha}, {\beta}, {\lambda}, s}\left(n^{-\textup{Re}({\lambda})-\textup{Re}(s)-2}\right),\end{aligned}$$ which implies that the series on the right-hand side of is analytic for Re$({\lambda})>-1$. Also $$-\frac{2^{-s-2{\lambda}-1}({\alpha}-{\beta})^{{\lambda}}{\Gamma}({\lambda}+s+1)}{\pi^{s+1}{\Gamma}({\lambda}+1)}\zeta({\lambda}+s+1)\zeta({\lambda}+1)\sin\left(\frac{\pi}{2}({\lambda}+s)\right)$$ is analytic for Re$(s+{\lambda})>0$ and Re$({\lambda})>-1$ except for a simple pole at ${\lambda}=0$. However, the contribution of this pole is nullified by that of the simple pole of $-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}$ at ${\lambda}=0$. Thus by analytic continuation, we see that holds for Re$({\lambda})>-1$. Now use with $s$ replaced by $s+{\lambda}$ and simplify the right-hand side of . This proves for Re$({\lambda})>-1$ when Re$(s+{\lambda})>0$.
Now if $s=-{\lambda}$, note that the leading term in the asymptotic expansion in vanishes and so instead of , we work with only. As in the previous case, it can be shown by analytic continuation that for $s=-{\lambda}$, is actually valid for Re$({\lambda})>-1$. It is now pleasing to note that the term containing $h(s, {\lambda})$ in for $s=-{\lambda}$ is exactly the same as the term $$\frac{\pi^{{\lambda}}({\alpha}-{\beta})^{{\lambda}}\zeta({\lambda}+1)\zeta(-s-{\lambda})}{2^{{\lambda}+1}{\Gamma}({\lambda}+1)}$$ occurring in , as can be seen from the fact that $\zeta(0)=-1/2$. Hence we conclude that holds for Re$({\lambda})>-1$ even when $s=-{\lambda}$.
A few interesting corollaries of the above theorem are now given.
\[ramanujan-type\] Let $\textup{Re}({\beta})>0$. For $(s)>-1/2$, $$\begin{aligned}
\label{ramanujan-typecor}
&{\Gamma}\left(s+\frac{1}{2}\right)\bigg\{\frac{\zeta(-s)}{2{\beta}^{s+\frac{1}{2}}}+\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{2}\bigg[({\beta}-in)^{-s-\frac{1}{2}}+({\beta}+in)^{-s-\frac{1}{2}}\nonumber\\
&\qquad\qquad\quad+\frac{2}{n^{s+\frac{1}{2}}}\sin\bigg(\frac{\pi}{2}\bigg(s-\frac{1}{2}\bigg)\bigg)\bigg]\bigg\}\nonumber\\
&=(2\pi)^s\bigg\{-2\pi\sqrt{\pi {\beta}}\zeta(-s)-\sqrt{\frac{\pi}{2}}\zeta\left(\frac{1}{2}\right)\zeta\left(\frac{1}{2}-s\right)\nonumber\\
&\qquad\qquad+\sqrt{\pi}\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{\sqrt{n}}e^{-2\pi\sqrt{2n{\beta}}}
\sin\left(\frac{\pi}{4}-2\pi\sqrt{2n{\beta}}\right)\bigg\}.\end{aligned}$$
Let ${\lambda}=-1/2$ in Theorem \[anathmbdkzac\]. Using and [@grn p. 925, formulas **8.467**, **8.469.3**] $$I_{-\frac{1}{2}}(\xi)=\sqrt{\frac{2}{\pi\xi}}\cosh(\xi),$$ we see that the left-hand side of simplifies to $$\begin{aligned}
\label{lhs}
&\frac{1}{2\pi\sqrt{{\alpha}-{\beta}}}\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{\sqrt{n}}\bigg\{{\epsilon}\exp{\left(-2\pi\overline{{\epsilon}}\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)}\cosh\left(2\pi\overline{{\epsilon}}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)\nonumber\\
&\qquad\qquad\qquad\qquad\quad+\overline{{\epsilon}} \exp{\left(-2\pi{\epsilon}\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)}\cosh\left(2\pi{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)\bigg\},\end{aligned}$$ where as the right-hand side, upon using the identity [@pbm p. 389, Formula **7.3.1.106**] $${\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{-\frac{1}{2}-s, -s}
{\frac{1}{2}} \bigg| {\xi} \right)}}=\frac{1}{2}\left\{\left(1-\sqrt{\xi}\right)^{2s+1}+\left(1+\sqrt{\xi}\right)^{2s+1}\right\},$$ results in $$\begin{aligned}
\label{rhs}
&\frac{\zeta(-s)(\sqrt{{\alpha}}+\sqrt{{\beta}})}{\sqrt{{\alpha}-{\beta}}}+\frac{{\Gamma}\left(s+\frac{1}{2}\right)\zeta(-s)}{2^{s+2}\pi^{s+\frac{3}{2}}\sqrt{{\alpha}-{\beta}}}\left(\frac{1}{{\alpha}^{s+\frac{1}{2}}}+\frac{1}{{\beta}^{s+\frac{1}{2}}}\right)+\frac{\zeta\left(\frac{1}{2}\right)\zeta\left(\frac{1}{2}-s\right)}{\pi\sqrt{2}\sqrt{{\alpha}-{\beta}}}\nonumber\\
&+\frac{{\Gamma}\left(s+\frac{1}{2}\right)}{2^{s+2}\pi^{s+\frac{3}{2}}\sqrt{{\alpha}-{\beta}}}\sum_{n=1}^{\infty}\sigma_{s}(n)\bigg[\frac{1}{({\alpha}+in)^{s+\frac{1}{2}}}+\frac{1}{({\beta}+in)^{s+\frac{1}{2}}}\nonumber\\
&\quad+\frac{1}{({\alpha}-in)^{s+\frac{1}{2}}}+\frac{1}{({\beta}-in)^{s+\frac{1}{2}}}+\frac{4}{n^{s+\frac{1}{2}}}\sin\left(\frac{\pi}{2}\left(s-\frac{1}{2}\right)\right)\bigg].\end{aligned}$$ Now multiply the expressions in and by $\pi\sqrt{{\alpha}-{\beta}}$ and let ${\alpha}\to{\beta}$ in the resulting equality so as to get upon simplification $$\begin{aligned}
\label{rem3}
&\sum_{n=1}^{\infty}\frac{\sigma_s(n)}{\sqrt{n}}e^{-2\pi\sqrt{2n{\beta}}}\sin\left(\frac{\pi}{4}-2\pi\sqrt{2n{\beta}}\right)\nonumber\\
&=2\pi\sqrt{{\beta}}\zeta(-s)+\frac{{\Gamma}\left(s+\frac{1}{2}\right)\zeta(-s)}{\sqrt{2}(2\pi{\beta})^{s+\frac{1}{2}}}+\frac{1}{\sqrt{2}}\zeta\left(\frac{1}{2}\right)\zeta\left(\frac{1}{2}-s\right)\nonumber\\
&\quad+\frac{{\Gamma}\left(s+\frac{1}{2}\right)}{\sqrt{2}(2\pi)^{s+\frac{1}{2}}}\sum_{n=1}^{\infty}\sigma_s(n)\left[\frac{1}{({\beta}+in)^{s+\frac{1}{2}}}+\frac{1}{({\beta}-in)^{s+\frac{1}{2}}}+\frac{2}{n^{s+\frac{1}{2}}}\sin\left(\frac{\pi}{2}\left(s-\frac{1}{2}\right)\right)\right].\end{aligned}$$ Finally multiply both sides by $2^s\pi^{s+\frac{1}{2}}$ and rearrange so as to arrive at for Re$(s)>1/2$ or $s=1/2$. By analytic continuation, the result holds for Re$(s)>-1/2$.
Corollary \[ramanujan-type\] bears striking resemblance to Ramanujan’s incorrect identity . Even though the summands of the corresponding infinite series in and have wrong signs, the infinite series in remarkably have the same forms compared to those in . Same is the case with two of the residual terms in , which, except for the absence of the factor $\tan\left(\frac{1}{2}\pi s\right)$ in them, are exactly the same as the corresponding ones in . We thus claim that our Corollary \[ramanujan-type\] can be considered as the corrected version of Ramanujan’s identity .
It can be shown without much effort that Theorem from [@bdrz] is actually equivalent to Corollary \[ramanujan-type\]. However, because the former was left in an unsimplified form in general in [@bdrz], its closeness to Ramanujan’s incorrect identity that we have demonstrated above got unnoticed. The form was simplified only for $s=2m+1/2, m\in\mathbb{N}\cup\{0\}$, for example, see [@bdrz Theorem 5.2]. In light of this, we would like to emphasize here that Theorem from [@bdrz] is but a special case of our Theorem .
The case ${\lambda}=\frac{1}{2}\neq-s$ of Theorem \[anathmbdkz\] results in an identity which can be obtained by merely replacing ${\beta}$ by ${\alpha}$ in and then substracting as it is from the resulting former identity.
Corollary \[ramanujan-type\], in turn, gives the following new formula for $\zeta^{2}(1/2)$.
\[ramanujan-type0\] For $({\beta})>0$, $$\begin{aligned}
&\sum_{n=1}^{\infty}\frac{d(n)}{\sqrt{n}}e^{-2\pi\sqrt{2n{\beta}}}\sin\left(\frac{\pi}{4}-2\pi\sqrt{2n{\beta}}\right)\nonumber\\
&=\frac{1}{\sqrt{2}}\zeta^{2}\left(\frac{1}{2}\right)-\pi\sqrt{{\beta}}-\frac{1}{4\sqrt{{\beta}}}+\frac{1}{2}\sum_{n=1}^{\infty}d(n)\left(\frac{1}{\sqrt{{\beta}+in}}+\frac{1}{\sqrt{{\beta}-in}}-\frac{\sqrt{2}}{\sqrt{n}}\right).\end{aligned}$$
Set $s=0$ in Corollary \[ramanujan-type\], or equivalently, in , and simplify.
Another special case of Theorem \[anathmbdkzac\] is
\[7.4\] Let $|\arg {\alpha}|<\pi/2, |\arg {\beta}|<\pi/2$ with $(\sqrt{{\alpha}})>$$(\sqrt{{\beta}})$. Then $$\begin{aligned}
\label{anathmbdkzeqnac00}
&\sum_{n=1}^{\infty}d(n)\bigg\{I_{0}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{0}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{0}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{0}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=-\frac{{\gamma}}{2}+\frac{1}{2}\log\left(\tfrac{2}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)-\frac{1}{8\pi\sqrt{{\alpha}{\beta}}}+\frac{1}{4\pi}\sum_{n=1}^{\infty}d(n)\left\{\frac{1}{\sqrt{{\alpha}+in}\sqrt{{\beta}+in}}+\frac{1}{\sqrt{{\alpha}-in}\sqrt{{\beta}-in}}\right\}.\end{aligned}$$
Let ${\lambda}=0$ in Theorem \[anathmbdkzeqnac\]. This gives for Re$(s)>0$ or $s=0$, [$$\begin{aligned}
\label{anathmbdkzeqnac001}
&\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{I_{0}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{0}\left(2\pi\overline{\epsilon}(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\nonumber\\
&\qquad\qquad+I_{0}\left(2\pi\epsilon(\sqrt{n{\alpha}}-\sqrt{n{\beta}})\right)K_{0}\left(2\pi\epsilon(\sqrt{n{\alpha}}+\sqrt{n{\beta}})\right)\bigg\}\nonumber\\
&=L_2(s, {\alpha}, {\beta})+\frac{\zeta(-s){\Gamma}(s+1)}{2^{3s+2}\pi^{s+1}\sqrt{{\alpha}{\beta}}}\left(\frac{1}{\sqrt{{\alpha}}}+\frac{1}{\sqrt{{\beta}}}\right)^{2s}{\ensuremath{{}_{2}F_{1} \left( \genfrac{}{}{0pt}{}{-s, -s}
{1} \bigg| {\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^2} \right)}}\nonumber\\
&\quad+\frac{{\Gamma}(s+1)}{2^{3s+2}\pi^{s+1}}\sum_{n=1}^{\infty}\sigma_s(n)\left\{\mathcal{A}\left(s,0,\sqrt{{\alpha}}, \sqrt{{\beta}}, n\right)+\mathcal{A}\left(s,0,\sqrt{{\alpha}}, \sqrt{{\beta}}, -n\right)+\frac{2^{2s+1}}{n^{s+1}}\sin\left(\frac{\pi s}{2}\right)\right\},\end{aligned}$$]{} where $$\begin{aligned}
\label{ell1sab}
L_2(s, {\alpha}, {\beta}):=\lim_{{\lambda}\to 0}\left(-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}+\frac{\pi^{{\lambda}}({\alpha}-{\beta})^{{\lambda}}\zeta({\lambda}+1)\zeta(-s-{\lambda})}{2^{{\lambda}+1}{\Gamma}({\lambda}+1)}\right).\end{aligned}$$ We only need to evaluate the above limit. Now $$\begin{aligned}
\label{l11}
-\frac{\zeta(-s)}{2{\lambda}}\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)^{{\lambda}}=-\frac{\zeta(-s)}{2{\lambda}}-\frac{1}{2}\zeta(-s)\log\left(\frac{\sqrt{{\alpha}}-\sqrt{{\beta}}}{\sqrt{{\alpha}}+\sqrt{{\beta}}}\right)+O_{{\alpha}, {\beta}, s}(|{\lambda}|)\end{aligned}$$ as ${\lambda}\to0$. The power series expansion of $1/{\Gamma}(\xi)$ about $\xi=0$ [@wrench Equation (22)] implies $$\label{wrench}
\frac{1}{{\Gamma}(\xi)}=\xi+{\gamma}\xi^2+O(|\xi|^3).$$ Now , , along with the standard Taylor expansions of $\left(\frac{\pi}{2}({\alpha}-{\beta})\right)^{{\lambda}}$ and $\zeta(-s-{\lambda})$ about ${\lambda}=0$, together give $$\begin{aligned}
\label{l12}
\frac{\pi^{{\lambda}}({\alpha}-{\beta})^{{\lambda}}\zeta({\lambda}+1)\zeta(-s-{\lambda})}{2^{{\lambda}+1}{\Gamma}({\lambda}+1)}&=\frac{1}{2}\left\{1+{\lambda}\log\left(\tfrac{\pi}{2}({\alpha}-{\beta})\right)+O_{{\alpha},{\beta}}(|{\lambda}|^2)\right\}\left\{\tfrac{1}{{\lambda}}+{\gamma}+O(|{\lambda}|)\right\}\nonumber\\
&\quad\times\left\{\zeta(-s)-{\lambda}\zeta'(-s)+O_{s}(|{\lambda}|^2)\right\}\left\{1+{\gamma}{\lambda}+O(|{\lambda}|^2)\right\}\nonumber\\
&=\frac{\zeta(-s)}{2{\lambda}}+\frac{1}{2}\left(2{\gamma}+\log\left(\frac{\pi}{2}({\alpha}-{\beta})\right)\right)\zeta(-s)-\frac{1}{2}\zeta'(-s)\nonumber\\
&\quad+O_{s, {\alpha}, {\beta}}(|{\lambda}|).\end{aligned}$$ From and , $$\begin{aligned}
\label{l13}
L_2(s, {\alpha}, {\beta})=\left({\gamma}+\frac{1}{2}\log\left(\frac{\pi}{2}\left(\sqrt{{\alpha}}+\sqrt{{\beta}}\right)^2\right)\right)\zeta(-s)-\frac{1}{2}\zeta'(-s).\end{aligned}$$ Finally substitute in , then let $s=0$ and simplify to derive .
If we let ${\alpha}\to{\beta}$ in , use the fact that $I_{0}(0)=1$ and simplify, we obtain . Thus, is a new generalization of , different from .
The corollary of Theorem \[anathmbdkzac\] stated below gives a transformation for the series which is almost the same as the left-hand side of except that $\sigma_{-{\lambda}}(n)$ is replaced by $\sigma_{{\lambda}}(n)$.
\[soninetrans\] Let $(\beta)>0$. For $({\lambda})>0$ or ${\lambda}=0$, $$\begin{aligned}
&2 \sum_{n=1}^{\infty} \sigma_{{\lambda}}(n) n^{\frac{{\lambda}}{2}}
\left( e^{\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi e^{\frac{\pi i}{4}} \sqrt{n{\beta}} ) +
e^{-\frac{\pi i {\lambda}}{4}} K_{{\lambda}}( 4 \pi e^{-\frac{\pi i}{4}} \sqrt{n{\beta}} ) \right)\nonumber\\
&=-\frac{{\Gamma}({\lambda})\zeta(-{\lambda})}{(2\pi\sqrt{{\beta}})^{{\lambda}}}+\frac{{\Gamma}(2{\lambda}+1)\zeta(-{\lambda})}{(2\pi)^{2{\lambda}+1}{\beta}^{\frac{3{\lambda}}{2}+1}}+{\beta}^{\frac{{\lambda}}{2}}\zeta({\lambda}+1)\zeta(-2{\lambda})\nonumber\\
&\quad+\frac{{\beta}^{\frac{{\lambda}}{2}}{\Gamma}(2{\lambda}+1)}{(2\pi)^{2{\lambda}+1}}\sum_{n=1}^{\infty}\sigma_{{\lambda}}(n)\left\{\frac{1}{({\beta}+in)^{2{\lambda}+1}}+\frac{1}{({\beta}-in)^{2{\lambda}+1}}+\frac{2\sin(\pi{\lambda})}{n^{2{\lambda}+1}}\right\}.\end{aligned}$$
Let $s={\lambda}$ in Theorem \[anathmbdkzac\], divide both sides of the resulting identity by $(\sqrt{{\alpha}}-\sqrt{{\beta}})^{{\lambda}}$, then let ${\alpha}\to{\beta}$ and then simplify.
Letting ${\lambda}\to0$ in the above corollary and observing that $$\lim_{{\lambda}\to0}\left(-\frac{{\Gamma}({\lambda})\zeta(-{\lambda})}{(2\pi\sqrt{{\beta}})^{{\lambda}}}+{\beta}^{\frac{{\lambda}}{2}}\zeta({\lambda}+1)\zeta(-2{\lambda})\right)=-\frac{1}{2}{\gamma}-\frac{1}{4}\log{\beta},$$ which can be proved along similar lines as , we arrive at . Hence we obtain a yet another new generalization of , different from and .
If we let ${\lambda}=1/2$ in Corollary , we obtain the special case $s=1/2$ of Corollary .
Concluding Remarks {#cr}
==================
The integral has its origins in the work of Fock and Bursian on electromagnetism of alternating current in a circuit with two groundings. Number-theoretic applications of its generalization, due to Koshliakov, namely , were given in [@bdrz]. In particular, integrals with kernel $J_{0}(\rho t)$, or more generally $J_{s}(\rho t)$, can be used in the Vorono" summation formula for $r_k(n)$ to obtain results which are of importance in the study of Gauss circle problem. For example, one such result is [@bdkz Theorem 1.6], which gives as a special case, a well-known result of Hardy [@hardyqjpam1915 Equation (2.12)] which he used to obtain his famous omega bound for $\sum_{n\leq x}r_2(n)$, see .
Our work was motivated by the fact that since the Vorono" summation formula for $\sigma_s(n)$, which is instrumental in the study of the generalized Dirichlet divisor problem, involves an integral transform consisting of the kernel $F_{s}(\rho t)$ in , it may be of interest to seek an analogue of with $J_{s}(\rho t)$ replaced by this kernel. We highlight here that there aren’t many explicit integral valuations of integrals containing $F_{s}(\rho t)$ in its integrand, see for example, [@dixfer3 p. 161]. The aforementioned analogue is found in of this paper. It would be very interesting and important to see if has applications in physics and geology, just like its counterpart in , or more specifically, in .
Equation was also successfully used in this paper to obtain a general transformation . As shown in this paper, and its analytic continuation in are rich sources of important number-theoretic identities such as Corollary \[dixmolcorgen\] and Corollary \[ramanujan-type\]. Corollary \[exotic\] is remarkable and is the only instance among identities derived in this paper where the hypergeometric function in reduces to a non-trivial closed form. Of course, one can generalize our results by replacing $\sigma_s(n)$ by coefficients of those Dirichlet series which satisfy functional equations with two gamma factors.
An important special case of Theorem \[anathmbdkzac\] is Corollary \[ramanujan-type\]. This result comes remarkably close towards obtaining a corrected version of the first identity on page $336$ of Ramanujan’s Lost Notebook, that is, . Note that Ramanujan’s result has signs in the summands of the two infinite series exactly opposite to those in Corollary \[ramanujan-type\]. For example, the infinite series on the right-hand side of is the same as that on the right-hand side of but with the $+$ sign in the summand of the latter replaced by $-$. The former would then be a special case of the series $$\begin{aligned}
\sum_{n=1}^{\infty}\sigma_s(n)\bigg\{&I_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\overline{\epsilon}\sqrt{n} (\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\nonumber\\
&-I_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}-\sqrt{{\beta}})\right)K_{{\lambda}}\left(2\pi\epsilon\sqrt{n}(\sqrt{{\alpha}}+\sqrt{{\beta}})\right)\bigg\}\end{aligned}$$ when ${\lambda}=-1/2$, and so it looks like an analysis similar to that carried out in this paper would lead to a corrected version of Ramanujan’s formula that matches even in signs than the one we have given in . Indeed, the natural kernel to work with while tackling this series would be $\sin\left(\frac{\pi s}{2}\right)J_{s}(t)-\cos\left(\frac{\pi s}{2}\right)L_{s}(t)$, where $L_{s}(t):=-\frac{2}{\pi}K_{s}(t)-Y_s(t)$. However, the problem is, we do not know of any summation formula like which would involve this kernel. It would be worthwhile to look for applications of (resp. ) towards the Dirichlet divisor problem (resp. the generalized Dirichlet divisor problem). This is because, the $r_2(n)$-analogue of was successfully used by Hardy to derive his famous omega result. Note that Hardy [@ddp] and subsequently many others, including Soundararajan [@soundararajanimrn], base their study of the corresponding omega result for $d(n)$ on the conditionally convergent series $$\frac{{\beta}^{1/4}}{\pi\sqrt{2}}\sum_{n=1}^{\infty}\frac{d(n)}{n^{3/4}}\cos\left(4\pi\sqrt{n{\beta}}-\frac{\pi}{4}\right),$$ where as the series on the left-hand side of is different. To the best of our knowledge, no one has used an identity of the type in the study of Dirichlet divisor problem.
**Acknowledgements**
The authors sincerely thank Professor Anna Vishnyakova from V. N. Karazin Kharkiv National University for translating for them the last three pages of [@fockbursian]. They also sincerely thank Karrie Peterson, Head, MIT Libraries, for sending them a copy of [@schermann], and Nico M. Temme for informing them of the reference [@dunster]. The first author’s research is supported by the SERB-DST grant RES/SERB/MA/P0213/1617/0021. He sincerely thanks SERB-DST for the support.
[00]{}
M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, with Formulas, Graphs, and Mathematical Tables, 9th edition, Dover publications, New York, 1970.
B. C. Berndt, *Identities involving the coefficients of a class of Dirichlet series, III*, Trans. Amer. Math. Soc. **146** (1969), 323–348.
B. C. Berndt, O.-Y. Chan, S.-G. Lim and A. Zaharescu, *Questionable claims found in Ramanujan’s lost notebook*, in [*Tapas in Experimental Mathematics*]{}, T. Amdeberhan and V. H. Moll, eds., Contemp. Math., Vol. 457, American Mathematical Society, Providence, RI, 2008, pp. 69–98.
B. C. Berndt, A. Dixit, S. Kim and A. Zaharescu, *Sums of squares and products of Bessel functions*, Adv. Math. **338** (2018), 305–338.
B. C. Berndt, A. Dixit, A. Roy and A. Zaharescu, *New pathways and connections in number theory and analysis motivated by two incorrect claims of Ramanujan*, Adv. Math. **304** (2017), 809–929.
B. C. Berndt, S. Kim and A. Zaharescu, *The circle problem of Gauss and the divisor problem of Dirichlet - still unsolved*, Amer. Math. Monthly **125** No. 2 (2018), 99–114.
H. Cohen, *Some formulas of Ramanujan involving Bessel functions*, Publications Mathématiques de Besançon. Algèbre et Théorie des Nombres, 2010, 59–68.
A. Dixit and V. H. Moll, *Self-reciprocal functions, powers of the Riemann zeta function and modular-type transformations*, J. Number Theory **147** (2015), 211–249.
A. Dixit, N. Robles, A. Roy and A. Zaharescu, *Koshliakov kernel and identities involving the Riemann zeta function*, J. Math. Anal. Appl. **435** No. 2 (2016), 1107–1128.
A.L. Dixon and W.L. Ferrar, *Some summations over the lattice points of a circle* , Quart. J. Math. **1** (1934), 48–63.
A. L. Dixon and W. L. Ferrar, *Infinite integrals of Bessel functions*, Quart. J. Math. **1** (1935), 161–174.
A. L. Dixon and W. L. Ferrar, *Infinite integrals in the theory of Bessel functions*, Quart. J. Math. **1** (1930), 122–145.
T. M. Dunster, *Bessel functions of purely imaginary order, with an application to second-order linear differential equations having a large parameter*, SIAM. J. Math. Anal. **21** No. 4 (1990), 995–1018.
S. Endres and F. Steiner, *A simple infinite quantum graph*, Ulmer Seminare Funktionalanalysis und Differentialgleichungen **14** (2009), 187–200.
S. Egger and F. Steiner, *An exact trace formula and zeta functions for an infinite quantum graph with a non-standard Weyl asymptotics*, J. Phys. A: Math. Theor. **44**, No. 44 (2011), 185–202.
V. A. Fock, *Zur Berechnung des elektromagnetischen Wechselstromfeldes bei ebener Begrenzung*, Ann. Phys. **409** (4) (1933), 401–420.
V. A. Fock and V. R. Bursian, *Electromagnetic field of alternating current in a circuit with two groundings* (in Russian), J. Russian Phys.-Chem. Soc. (Zhurnal Russkago Fiziko-Khimicheskago Obshchestva) **58** No. 2 (1926), 355–363. (Available at <http://books.e-heritage.ru/book/10084021>)
A. P. Guinand, *Summation formulae and self-reciprocal functions* , Quart. J. Math. **10** (1939), 104–118.
A. P. Guinand, *Some rapidly convergent series for the Riemann $\xi$-function*, Quart.J. Math. (Oxford) **6** (1955), 156–160.
A. P. Guinand, *Concordance and the harmonic analysis of sequences*, Acta Math. **101**, No. 3 (1959), 235–271.
I. S. Gradshteyn and I. M. Ryzhik, eds., *Table of Integrals, Series, and Products*, 7th ed., Academic Press, San Diego, 2007.
G. H. Hardy, *On the expression of a number as the sum of two squares*, Quart. J. Pure Appl. Math. **46** (1915), 263–283.
G. H. Hardy, *On Dirichlet’s divisor problem*, Proc. London Math. Soc. (2) **15** (1916), 1–25.
N. S. Koshliakov, *On a certain definite integral connected with the cylindric function $J_{\mu}(x)$*, C. R. Acad. Sci. URSS **2** (1934), 145–147.
N. S. Koshliakov, *On Voronoi’s sum-formula*, Mess. Math. **58** (1929), 30–32.
N. S. Koshliakov, *Note on certain integrals involving Bessel functions*, Bull. Acad. Sci. URSS Ser. Math. **2** No. 4, 417–420; English text 421–425 (1938).
E. Landau, *Vorlesungen über Zahlentheorie*, Bd. 1, S. Hirzel, Leipzig, 1927.
A. P. Lursmana$\check{s}$vili, *On the number of lattice points in multidimensional spheres* (Russian), Akad. Nauk Gruzin. SSR. Trudy Tbiliss. Mat. Inst. Razmadze **19**, (1953), 79–120.
A. P. Lursmana$\check{s}$vili, *On the number of lattice points in multidimensional spheres of odd dimension* (Georgian), Soob$\check{s}$$\check{c}$eniya Akad. Nauk Gruzin. SSR. **14** (1953), 513–520.
A. I. Popov, *On some summation formulas* (in Russian), Bull. Acad. Sci. L’URSS, **7** (1934), 801–802.
A. I. Popov, *Bemerkung zur Arbeit von V. Fock “Zur Berechnung des elektromagnetischen Wechselstromfeldes bei ebener Begrenzung"* (in Russian), C. R. Acad. Sci. URSS **1** (2) (1934), 380–381.
A. I. Popov, *Über die zylindrische Funktionen enthaltenden Reihen* (in Russian), C. R. Acad. Sci. URSS **2** (1935), 96–99.
O. Praus, *Field of electric dipole above two-layer anisotropic medium*, Studia geoph. et geod. **9** No. 4 (1965), 359–380.
A. P. Prudnikov, Yu. A. Brychkov and O. I. Marichev, *Integrals and Series, Vol. 2, Special Functions*, Gordon and Breach, New York, 1986.
S. Ramanujan, *The Lost Notebook and Other Unpublished Papers*, Narosa, New Delhi, 1988.
A. Sommerfeld, *Über die Ausbreitung der Wellen in der drahtlosen Telegraphie*, Ann. Phys. **333** (4) (1909), 665–736..
K. Soni, *Some relations associated with an extension of Koshliakov’s formula*, Proc. Amer. Math. Soc. **17** No. 3 (1966), 543–551.
D. Schermann, *Das elektromagnetische Feld eines geneigten Rahmens* (Russian) (Electromagnetic field of an inclined frame), Journ. f. exp. u. theor. Phys. (Zhurnal eksperimental’noi i teoreticheskoi fiziki) **2** (1932), 129–139.
K. Soundararajan, *Omega results for the divisor and circle problems*, Int. Math. Res. Not. **2003** No. 36, 1987–1998.
N. M. Temme, *Special Functions: An Introduction to the Classical Functions of Mathematical Physics*, Wiley-Interscience, New York, 1996.
E. C. Titchmarsh, *The Theory of the Riemann Zeta Function*, Clarendon Press, Oxford, 1986.
G. F. Vorono", *Sur une fonction transcendante et ses applications à la sommation de quelques séries*, Ann. École Norm. Sup. (3) **21** (1904), 207–267, 459–533.
G. F. Vorono", *Sur la développement, $\grave{a}$ l’aide des fonctions cylindriques, des sommes doubles $\sum f(pm^2+2qmn+rn^2)$, oú $pm^2+2qmn+rn^2$ est une forme positive $\grave{a}$ coefficients entiers*, Verhandlungen des Dritten Internat. Math. Kong in Heidelberg, B. G. Teubner, Leipzig, 1905, 241–245.
G. N. Watson, *A Treatise on the Theory of Bessel Functions*, second ed., Cambridge University Press, London, 1966.
J. W. Wrench, *Concerning two series for the gamma function*, Math. Comp. **22** No. 103 (1968), 617–626.
[^1]: 2010 *Mathematics Subject Classification.* Primary 11M06, 33E20; Secondary 33C10.\
*Keywords and phrases.* Bessel functions, generalized sum-of-divisors function, Vorono" summation formula, analytic continuation
[^2]: These conditions, which are more general than the ones given in [@watson-1966a p. 416, Equation (2)] and [@grn p. 693, formula 6.596.7], are established in Section \[ex\].
[^3]: We have inserted some comments in square brackets and changed some notations to conform them with those used in our paper.
[^4]: There are two typos in Koshliakov’s version of , namely, in *his* version, the arguments of the $K$-Bessel functions inside the integral on the left contain $a$, and those on the right contain $x$. Both these typos are corrected, albeit with renaming of his parameters, in our version of his formula given in .
[^5]: Note that $\epsilon^{-\left(\nu-1\right)}=\overline{\epsilon}^{\left(\nu-1\right)}$, so there is no possibility of confusion unless, of course, one does not put the parentheses. However, we wish to keep the results in the sequel symmetric with respect to $\epsilon$ and $\overline{\epsilon}$ whenever we can. We will use, without mention, the fact that $\overline{\epsilon}=1/\epsilon$.
[^6]: As mentioned in [@dixfer4], the first equality in is an obvious modification of [@watson-1966a p. 410, Equation (13.45)].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'It is well known that the density and anisotropy profile in the inner regions of a stellar system with positive phase-space distribution function are not fully independent. Here we study the interplay between density profile and orbital anisotropy at large radii in physically admissible (consistent) stellar systems. The analysis is carried out by using two-component (,$\gau$) spherical self-consistent galaxy models, in which one density distribution follows a generalized $\gamma$ profile with external logarithmic slope $n$, and the other a standard $\gau$ profile (with external slope 4). The two density components have different “core” radii, the orbital anisotropy is controlled with the Osipkov-Merritt recipe, and for simplicity we assume that the mass of the $\gau$ component dominates the total potential everywhere. The necessary and sufficient conditions for phase-space consistency are determined analytically, also in presence of a dominant massive central black hole, and the analytical phase-space distribution function of (,1) models, and of models with a central black hole, is derived for $\gamma=0,1,2$. It is found that the density slope in the external regions of a stellar system can play an important role in determining the amount of admissible anisotropy: in particular, for fixed density slopes in the central regions, systems with a steeper external density profile can support more radial anisotropy than externally flatter models. This is quantified by an inequality formally identical to the “cusp slope-central anisotropy” theorem (An & Evans 2006), relating at all radii (and not just at the center) the density logarithmic slope and the anisotropy indicator in all Osipkov-Merritt systems.'
date: 'Submitted, October 1, 2008. Accepted version, November 3, 2008.'
title: 'Two–component galaxy models: the effect of density profile at large radii on the phase–space consistency'
---
stellar dynamics – galaxies: ellipticals – dark matter – black holes
Introduction
============
Observationally it is well established that elliptical galaxies have dark matter halos, and also host central supermassive black holes. These empirical facts motivate the study of multi-component dynamical models. When studying dynamical models of stellar systems (single or multi-component), the minimal requirement to be met by a physically acceptable model is the positivity of the phase-space distribution function (DF) of each distinct component. A model satisfying this essential requirement (which is much weaker than stability, but stronger than the fact that the Jeans equations have a physically acceptable solution) is called [*consistent*]{}; moreover, when the total gravitational potential is determined by the total density profile through the Poisson equation, the model is called [ *self–consistent*]{}. In other words, we call self–consistent a consistent self-gravitating system.
Two general strategies can be used to construct a (self) consistent model, or check whether a proposed model is (self) consistent: they are commonly referred to as the “$f$–to–$\rho$” and the “$\rho$–to–$f$” approaches, where $f$ is the model DF (e.g., see Bertin 2000, Binney & Tremaine 2008). An example of the first approach is the survey of self–consistent two–component galaxy models carried out by Bertin and co–workers, where the stellar and dark matter components are described by two DFs of the $f_{\infty}$ family (e.g., Bertin & Stiavelli 1984, Bertin et al. 1992); other well known examples are the King (1966) models and the $f_{\nu}$ models (Bertin & Trenti 2003). Unfortunately, the resulting spatial densities obtained by solving the associated Poisson equation are in general not expressible in terms of simple or even known functions, and so only numerical investigations are usually feasible. In the “$\rho$–to–$f$” approach the density distribution is given, and specific assumptions about the model internal dynamics are made; in special cases inversion formulae from the density to the DF can be obtained, usually in integral form or series expansion (see, e.g., Fricke 1952, Lynden–Bell 1962, Osipkov 1979, Merritt 1985, hereafter OM; Dejonghe 1986, 1987; Cuddeford 1991; Hunter & Qian 1993, Ciotti & Bertin 2005). In particular, in order to recover the DF of spherical models with orbital anisotropy, the OM technique has been developed from the Eddington (1916) inversion for isotropic systems, and widely used to study one and two–component models (see, e.g., Ciotti & Pellegrini 1992, hereafter CP92; Hiotelis 1994; Carollo et al. 1995; Ciotti & Lanzoni 1997; Ciotti 1996, 1999, hereafter C96, C99; Baes & Dejonghe 2004; Buyle et al. 2007). We remark that the OM parameterization is not necessarily the best description of real systems, however its simplicity and the fact that it captures the main features of models of galaxy formation, that are generally found nearly isotropic at the center and increasingly radially anisotropic in the outer envelope (e.g., van Albada 1982; Trenti, Bertin & van Albada 2005; Nipoti, Londrillo & Ciotti 2006; Binney & Tremaine 2008. But see Cuesta et al. 2008 and references therein), make it the natural choice for investigations as that presented in this paper.
In many cases, the difficulties inherent in the operation of recovering analytically the DF prevent a simple consistency analysis, and phase–space positivity must be investigated by numerical inspection of the inversion integral. In these cases the reasons underlying consistency or inconsistency of a proposed model tend to be obscured by the numerical nature of the solution. Fortunately, informations about consistency of multi–component OM systems can be obtained without recovering their DF, following the procedure described in CP92. This method uses the radial density profile of each component and the total potential of the system, and gives necessary and sufficient conditions for (self) consistency. Moreover, since only spatial differentiation and inequality checks are required, this method is best suited for analytical investigations. For example, C96 and C99 applied the CP92 technique to the general family of two–component, spherically symmetric and radially anisotropic $(\gau,\gad)$ models. This family is made of the superposition of two $\gamma$ models (Dehnen 1993, Tremaine et al. 1994) with different total masses, scale–lengths, inner density slopes, and OM radially anisotropic velocity dispersions. The possibility to investigate the combined effects of radial anisotropy and inner density slope on multi–component systems made the study of $(\gau,\gad)$ models interesting, as it is well known that the inner density profile sets important constraints on the amount of admissible radial anisotropy (e.g., Richstone & Tremaine 1984), and indeed in C96 and C99 analytical limitations on anisotropy as a function of the density slopes $\gau$ and $\gad$ were obtained. These models clarified the reasons behind the numerical findings of CP92, i.e. the difficulty of consistently superimposing a centrally peaked distribution such as the de Vaucouleurs (1948) profile to a centrally flat one, such as the King (1972) or quasi–isothermal density profile (even in the isotropic case). In fact, it was shown that the DF of the $\gau$ component in isotropic $(\gau,\gad)$ models is nowhere negative, independently of the mass and concentration of the $\gad$ component, whenever $0\leq\gad\leq\gau$ and $1\leq\gau <3$. On the contrary, a $\gau=0$ component becomes inconsistent when adding a $\gad=1$ halo with a small core radius. Thus, in two–component isotropic models, the component with the steeper central density distribution is usually the most robust against inconsistency. More recently, the importance of the central density slope in limiting the amount of possible radial anisotropy has also been quantified with the so called “cusp slope-central anisotropy” theorem (An & Evans 2006, hereafter AE06; see also eq. \[28\] in de Bruijne et al. 1996).
The previous investigations left however unexplored the importance of the *external* density slope in determining the model consistency. In fact, the phase–space density cannot be identified, in general, with any specific spatial position in the system, as (for example) stars of a given energy can span a large range of radial positions[^1] (systems made of circular orbits are an obvious exception). Therefore also the external regions of a density distribution can be important in limiting the maximum allowable anisotropy, but the $(\gau,\gad)$ models are of no help in the study of this issue, because the external density profiles of both components all decrease as $r^{-4}$. For these reasons here we focus on the phase–space properties of models, i.e. models similar to the standard $\gamma$ models in the inner regions, but with a density profile proportional to $r^{-n}$ (instead of $r^{-4}$) in the external regions; remarkably, several properties of models can be obtained from those of $\gamma$ models by differentiation with respect to their scale–length. In this notation, the 4-$\gamma$ models coincide with the standard $\gamma$ models. We also study the larger class of ($\ngam$,$\gau$) models, i.e. two–component systems in which a $\gau$ halo is added to a component. Thus, here we further explore the trends emerged in CP92, C96, and C99, determining the limits imposed by phase–space consistency on the parameters describing (,$\gau$) models, and models with a central BH \[hereafter (,BH) models\], with particular focus on the effects of the external slope parameter $n$. In specific cases (that we call [*halo-dominated*]{} models), the calculations are performed under the assumption that the mass of the halo component (or of the central BH) is dominant over the mass of the visible one. This assumption is mainly motivated by mathematical simplicity (see also Sect. 4), although this is not the only reason. In fact, for any given two–component model, it is expected that the DF properties are bracketed by those of the one component model and by those of the halo-dominated model (corresponding to the formal case of an infinite halo mass). Of course, while the case of dark matter dominated systems can be considered a viable representation of some real astrophysical systems, the case of a dominant BH is less natural, and it just gives the strongest possible limitations for consistency of systems with a central BH.
The paper is organized as follows. In Section 2 we briefly review the technique developed in CP92, and we prove that the necessary condition for consistency derived in CP92 for OM systems can be rewritten [ *exactly*]{} as the AE06 “cusp slope-central anisotropy” theorem, holding however at all radii and not just at the center. In Section 3 the one and two–component models are introduced, and the necessary and sufficient conditions imposed on the model parameters by phase–space consistency are derived for different values of the logarithmic density slope $n$. In Section 4 the DF of the component of halo-dominated ($\ngam$,1) models with $\gamma=0,1,2$, and of similar models with a dominant central BH, are derived explicitly for arbitrary (but integer) values of $n$, and the true boundaries of the consistency region in the parameter space are obtained. The main results of the investigation are summarized in Section 5. Finally, in the Appendix an easy method to solve analytically the Jeans equations in the general OM case for the wider class of ($n_1$-$\gau$,$n_2$-$\gad$) models is presented.
The consistency of multi–component OM systems
=============================================
A stellar system made of the sum of the density components $\roi$ is called consistent if each DF ($\fid$) is non–negative over the whole accessible phase–space; a consistent self–gravitating system is called self–consistent. The technique developed in CP92 permits us to check whether the DF of a multi–component spherical system, where the orbital anisotropy of each component is modeled according to the OM parameterization, is positive; in practice, only information about the radial trend of each density component and of the total integrated mass are used. In the OM formulation the DF of each component is obtained assuming $\fid=\fid(\qi)$, with $$\label{eq:Q}
\qi=\eps-\frac{J^2}{2\rai^2},$$ where $\eps$ and $J$ are respectively the relative energy and the angular momentum modulus per unit mass, $\rai$ is the so–called [ *anisotropy radius*]{}, and $\fid=0$ for $\qi\leq 0$ (e.g. see Binney & Tremaine 2008). The velocity dispersion tensor associated with eq. (\[eq:Q\]) is characterized by radial anisotropy increasing with the radius $r$, while in the limit $\rai\to\infty$ the system becomes globally isotropic. As well known, the DF of the density component $\roi$ is given by $$\begin{aligned}
\label{eq:fOM}
\fid(\qi)&=&\displaystyle\frac{1}{\sqrt{8}\pi^2}\frac{d}{d\qi}\int_0^{\qi}
\frac{d\roqi}{d\psit}\frac{d\psit}{\sqrt{\qi-\psit}}\nonumber \\
&=&\displaystyle\frac{1}{\sqrt{8}\pi^2}\int _0^{\qi} \frac{d^2\roqi}{
d\psit^2}\frac{d\psit}{\sqrt{\qi-\psit}}\mbox{,}\end{aligned}$$ where the [*augmented density*]{} is $$\label{eq:rhoOM}
\roqi(r)\equiv\left (1+\frac{r^2}{\rai^2}\right) \roi (r),$$ $\psit (r)=\sum_k\psii (r)$ is the total relative potential, $0\leq\qi\leq\psit (0)$, and in the integrals above it is assumed that the radius is eliminated from $\roqi$ in favour of $\psit$. It can be proved that the second equivalence in eq. (\[eq:fOM\]) holds for untruncated systems with finite total mass, as the models discussed here.
Necessary and sufficient conditions for consistency
---------------------------------------------------
Quite obviously, only a very limited number of density profiles among those expressible in analytic form admit an explicit DF, so that the study of phase–space consistency would appear restricted to such rare cases when conducted analytically, while all the remaining cases should be investigated numerically. Fortunately this is not true, as the CP92 technique is based on the verification (numerical or analytical) of the following : A [*necessary condition*]{} (NC) for the non–negativity of $\fid$ is $$\label{eq:NC}
\frac{d\roqi(r)}{dr}\leq 0,\quad 0\leq r \leq\infty .$$ If the NC is satisfied, a [*strong sufficient condition*]{} (SSC) for the non–negativity of $\fid$ is $$\label{eq:SSC}
\frac{d}{dr}\left[\frac{d\roqi(r)}{dr} \frac{r^2\sqrt {\psit(r)}}{\mt
(r)}\right]\geq 0, \quad 0\leq r\leq\infty .$$ Finally, a [*weak sufficient condition*]{} (WSC) for the non negativity of $\fid$ is $$\label{eq:WSC}
\frac{d}{dr}\left[ \frac{d\roqi(r)}{dr}\frac{r^2}{\mt (r)}\right]\geq
0, \quad 0\leq r\leq\infty .$$ : See CP92, C96, C99. The first consideration that follows from the conditions above is that the violation of the NC is connected only to the radial behavior of the augmented density $\roqi$, and so this condition applies independently of any other density component of the model. Obviously, a model failing the NC is [*certainly inconsistent*]{}, while a model satisfying the NC [*may be inconsistent*]{}, i.e. the $\fid$ may be negative even for values of model parameters allowed by the NC. The second consideration is that a model satisfying the WSC (or the more restrictive SSC) is [*certainly consistent*]{}, while a model failing the WSC (SSC) [*may be consistent*]{}. Therefore the consistency of an OM model satisfying the NC and failing the WSC (or the SSC) can be proved only by direct inspection of its DF.
### A density slope–OM anisotropy inequality
The NC can be recast into a simple inequality between the value of the density slope $$\gamma(r)\equiv-\frac{d\ln\rho}{d\ln r}$$ and the value of the orbital anisotropy indicator $$\label{betaOM}
\beta(r)\equiv 1-\frac{\sigt^2}{2\sigr^2}=\frac{r^2}{r^2+\ra^2}$$ that must hold at each radius for consistent OM systems. In the expression above $\sigr^2$ and $\sigt^2$ are the radial and tangential components of the velocity dispersion tensor of the system (e.g. Binney & Tremaine 2008), and in the following proof we restrict to natural systems, i.e. systems with monotonically decreasing density profile, so that $\gamma(r)>0$. The proof is trivial: in fact, it suffices to express the NC in terms of the logarithmic density slope as $$\frac{2}{\gamma(r)}-\frac{r^2}{r^2+\ra^2}\leq 0,$$ and from eq. (\[betaOM\]) one obtains the necessary condition $$\label{cuspslope}
\gamma(r)\geq 2\beta(r),$$ which must be verified by each OM component of a consistent multi–component system. Curiously, The condition above is formally [*identical*]{} to the inequality appearing in the “cusp slope-central anisotropy” theorem derived in AE06. This latter theorem was proved by using constant anisotropy systems (i.e., $\beta
(r)=\beta$, and in this case as well the inequality above holds at all radii in consistent models. See eq. \[10\] in AE06), and then it was convincingly argued that the inequality it holds asymptotically for the central regions of spherical systems with generic anisotropy distribution.
In the specific case of density distributions (see eq. \[\[eq:ng\]\]), the logarithmic density slope $$\label{gammarhong}
\gamma(r)=\frac{\gamma+n s}{1+s},\qquad s={r\over\rc},$$ and therefore eq. (\[cuspslope\]) shows that in general the NC is satisfied near the center of OM systems (where $\beta=0$ and $\gamma(r)=\gamma$), and at large radii (where $\beta =1$ and $\gamma(r) = n\geq3$), while critical behaviors may be expected at intermediate radii (see Sect. 4).
Classification of phase-space inconsistency as a function of the anisotropy radius
----------------------------------------------------------------------------------
Interestingly (but not unexpectedly) the particular functional form of the augmented density characteristic of OM parametrization, limits the possible manifestations of phase–space inconsistency to few general cases, that can be illustrated as follows. From eqs. (\[eq:fOM\])-(\[eq:rhoOM\]) it is apparent that the DF of each density component can be written as $$f(Q) =\fis(Q)+\frac{\fan(Q)}{\ra^2},
\label{eq:fSA}$$ where $\fis$ is the DF of the considered density component in the isotropic case (for simplicity, from now on we avoid the use of the subscript $k$). Let $A_+$ be the subset of phase–space defined by the property that $\fis$ is positive $\forall Q\in A_+$. Then, from eq. (\[eq:fSA\]) a first condition to be satisfied for consistency is $$\label{eq:sa-}
\ra\geq\racm \equiv\sqrt{\max\left\{0,{\rm sup}\left[-\frac{\fan(Q)}{
\fis(Q)}\right ]_{Q\in A_+}\right\}}.$$ Obviously, when $\fis>0$ over all the accessible phase–space (the common situation for realistic density distributions), inequality (\[eq:sa-\]) is also the [*only*]{} condition to be satisfied for the model consistency, and there is at most a lower bound for the anisotropy radius. For example this is the case for the families of one–component anisotropic $\gamma$ models (C99) and Sersic (1968) models (Ciotti 1991).
When the complement of $A_+$ is not empty, i.e. $\fis <0$ over some region $A_-$ of the accessible phase–space, a second inequality again derived from eq. (\[eq:fSA\]) must be verified for consistency: $$\label{eq:sa+}
\ra\leq\racp\equiv\sqrt{{\rm inf}\left[\frac{\fan (Q)}{
|\fis (Q)|}\right ]_{Q\in A_-}}.$$ Therefore, if there exists some $Q\in A_-$ for which $\fan <0$, then the proposed model is inconsistent[^2]. If $\fan$ is positive over $A_-$, from conditions (\[eq:sa-\])-(\[eq:sa+\]) it follows that $\racm<\ra<\racp$ for OM consistency, so that if $\racp<\racm$ the proposed model is inconsistent. Note that formally *identical* considerations hold when discussing the inequalities (\[eq:NC\]) and (\[eq:WSC\]), because from eq. (\[eq:rhoOM\]) it follows that the NC and WSC can be written in the same functional form of eq. (\[eq:fSA\]): of course, in these cases the sets $A_+$ and $A_-$ are to be intended as subsets of the radial range $0\leq r\leq\infty$.
In the following Section, after presenting the one and two–component models, the necessary and sufficient conditions for model consistency will be derived, also for the case of models with a massive central black hole.
The one and two–component models
================================
The one–component models are a natural generalization of the widely explored family of the so–called $\gamma$ models (Dehnen 1993, Tremaine et al. 1994), whose density, cumulative mass within $r$, and relative potential are given by $$\begin{aligned}
\label{eq:rhoG}
\rho(r)&=&\displaystyle\frac{3-\gamma}{4\pi}\frac{M\rc}
{r^{\gamma}(\rc+r)^{4-\gamma}},\\
M(r)&=&M\times\left(\frac{r}{\rc +r}\right)^{3-\gamma},\end{aligned}$$ $$\label{eq:PhiG}
\Psi(r)=\frac{GM}{\rc}
\begin{cases}\displaystyle{\frac{1}{(2-\gamma)}
\left [1-\left(\frac{r}{r+\rc}\right)^{2-\gamma}\right]}\cr
\displaystyle{\ln\frac{r+\rc}{r}},\quad(\gamma=2),\cr
\end{cases}$$ respectively.
In the models the logarithmic density slope for $r\gg\rc$ is not 4, but it is a free parameter $n>3$ (to ensure that their total mass is finite), so that these density profiles belong to the family considered by Zhao (1996). In the following, in order to exploit a useful analytical property of the density profiles, we will often assume $n$ restricted to integer values $\geq4$. In fact, the generic density profile of a model is given by $$\begin{aligned}
\label{eq:ng}
\rho_n(r)&\equiv&\frac{M\rc^{n-3}}
{4\pi\eulB(3-\gamma,n-3)r^{\gamma}(\rc+r)^{n-\gamma}}\nonumber\\&\\
&=&\frac{\rc^{n-3}(-1)^n}{\Gamma(n-3)}\frac{d^{n-4}}{d\rc^{n-4}}
\frac{\rho (r)}{\rc},\nonumber\end{aligned}$$ where $M$ is the total mass of the density distribution, $\eulB(x,y)=\Gamma(x)\Gamma(y)/\Gamma(x+y)$ and $\Gamma(x)$ are the complete Euler beta and gamma functions, respectively, and the first expression holds for any real number $n>3$. The second expression, based on repeated differentiation of eq. (\[eq:rhoG\]) with respect to $\rc$, holds instead for integer $n\geq4$. Of course, for $n=4$ the standard $\gamma$ model density profile is recovered. The radial behaviour of $\rho_n(r)$ is shown in Fig. \[f1\] (top panel) for $n$-1 models with increasing $n$. In the bottom panel we show the corresponding logarithmic density slopes, calculated accordingly to eq. (\[gammarhong\]). In particular, it is apparent how the inner slope is $\gamma$ (for $r\lsim\gamma\rc/n$), and the external is $n$ (for $r\gsim\rc$). As a consequence, while the density profile outside $\rc$ becomes more and more steep at increasing $n$, the innermost region where the density slope is $\gamma$ contracts near the center.
![The normalized density profile $\rho_n$ of the $n$-1 model (top panel), and its logarithmic slope (eq. \[\[gammarhong\]\], bottom panel), for $n=4,5$, and $6$.[]{data-label="f1"}](f1.eps){height="0.46\textheight" width="52.00000%"}
As anticipated, the possibility to write the density $\rho_n$ as a derivative with respect to the scale–length $\rc$ provides an easy way to determine analytical properties of the models for $n$ integer. Indeed, since the density enters linearly in the integrals of the total mass inside the radius $r$ and of the potential, from eq. (\[eq:ng\]) it follows that $$\begin{aligned}
\label{eq:Mng}
M_n(r)&=&\frac{\rc^{n-3}(-1)^n}{\Gamma(n-3)}\frac{d^{n-4}}
{d\rc^{n-4}}\frac{M(r)}{\rc},\\
\Psi_n(r)&=&\frac{\rc^{n-3}(-1)^n}{\Gamma(n-3)}
\frac{d^{n-4}}{d\rc^{n-4}}\frac{\Psi(r)}{\rc}.\label{eq:Phing}\end{aligned}$$ Instead, when $n$ is not at integer, $M_n$ and $\Psi_n$ are in general given by hypergeometric ${}_2F_1$ functions. Expressions similar to eqs. (\[eq:Mng\])-(\[eq:Phing\]) hold for any quantity that can be written as a linear functional of $\rho_n$, so that the surface density profile of models with $n$ integer can be obtained by repeated differentiation of the surface density profile of $\gamma$ models (when analytically available, e.g. see Binney & Merrifield 1998). This property will be exploited in Section 4 to obtain the explicit DF of halo dominated (,$\gau$) models; moreover, in Appendix B we show how quadratic functionals of $\rho_n$ (such as the gravitational energy and the velocity dispersions) can be also evaluated by using repeated differentiation with respect to $\rc$.
We are now in the position to introduce the two–component models used in this work. The most general family of two–component models, i.e. the ($n_1$-$\gau$,$n_2$-$\gad$) models, is made of the superposition of two models with different total masses, scale–lengths, internal and external density slopes, and finally two different anisotropy radii. For simplicity here we restrict to the case of (,$\gau$) models, where the “halo” density distribution is a standard $\gau$ model: we note however that some of the presented results can be generalized without much effort to the family of ($n_1$-$\gau$,$n_2$-$\gad$) models (see Appendix B). In the following, the total mass $\ms$ and the characteristic scale–length $\rs$ of the $\gau$ halo are adopted as normalization constants, so that the physical scales for density and potential are given by $\rhon=\ms/\rs^3$ and $\psin =G\ms/\rs$, while we define $s\equiv
r/\rs$, $\xi\equiv\rc/\rs$, $\mu\equiv M/\ms$. With this choice, the (,$\gau$) models are structurally determined by fixing the four independent parameters $(\ms,\rs,\mu,\xi)$, with the obvious conditions $\mu\geq 0$ and $\xi\geq 0$. We conclude this introductory discussion by noticing that, for reasons that will become apparent in Section 3.2 and 4, the present normalization differs from that adopted in C99, where the normalization mass and scale–length were those of the first component.
The necessary and sufficient conditions for one–component models
----------------------------------------------------------------
Before discussing the case of the two–component models, we consider the NC for anisotropic models, in order to determine analytically a critical value for the anisotropy radius such that a higher degree of radial OM anisotropy (i.e., a smaller $\ra$) would produce a negative DF for some permitted value of $Q$. We recall that the obtained anisotropy limit holds also when a second component is added (see Section 2), so that the present discussion is fully general. Moreover, we note that as the NC involves only the density distribution under scrutiny, we can use the first expression in eq. (\[eq:ng\]), and the following results hold for any $n>3$, not necessarily limited to integers. In the following the unit mass and unit length are the total mass $M$ and the scale–length $\rc$ of the model, with $\sa=\ra/\rc$.
In C99 the analytical expression for the critical $\sa(\gamma)$ was obtained for the whole family of $\gamma$-models, and here we derive its generalization $\sa(\gamma,n)$. In fact, eq. (A1) shows that for $2\leq\gamma <3$ and $n>3$ the NC is satisfied for $\sa\geq 0$, and the result of C99 is now obtained as the particular case for $n=4$. In other words, the NC leaves open the possibility of making models with $\gamma\geq 2$ using radial orbits only. In the range $0\leq \gamma <2$ the NC requires instead that $$\sa\geq \sM\sqrt{\frac{2-\gamma-(n-2)\sM}{\gamma+n\sM }},
\label{eq:ngNC}$$ where the explicit expression of $\sM(n,\gamma)$ is given by eq. (A2), and the inequality above reduces to eq. (13) of C99 when $n=4$. The NC then proves that models with $0\leq \gamma <2$ cannot be made of radial orbits only, independently of the value of the external slope $n$, and of the presence of any possible second component. This result extends the list of cases for which it has been proved that a density cusp shallower than $r^{-2}$ cannot be supported by radial orbits only (Richstone & Tremaine 1984, CP92, C99, AE06).
![Consistency limits on $\sa=\ra/\rc$ as a function of $\gamma$ in one–component models, for $n=4,5,6$. The NC limit is represented by the solid lines: all models with the pair ($\sa$,$\gamma$) in the triangular region below them are inconsistent. The dashed lines mark the WSC limit: all points above these lines correspond to consistent one–component models. For $\gamma=0,1,2$, the filled circles joined by the dotted line represent the more accurate limits obtained from the SSC. For the black-hole dominated ($n$-$\gamma$,[BH]{}) models the long-dashed lines (interrupted for $\gamma<1$, see Section 3.2), are the lower limit on $\sa$ given by the WSC, while the solid squares are the true limits derived from the DF.[]{data-label="f2"}](f2.eps){height="0.4\textheight" width="50.00000%"}
In Fig. \[f2\] the lower bound for the anisotropy radius given in eq. (\[eq:ngNC\]) is shown with the solid lines as a function of $\gamma$ for $n=4,5,6$: models (one or multi–component) with the pair ($\sa$,$\gamma$) in the nearly triangular region under the solid curves are inconsistent. In particular, for fixed $\gamma$ an increase of $n$ produces a decrease of the minimum $\sa$, i.e. we have here a first indication that a steepening of the density profile in the external regions of a system with fixed inner density slope can be effective in increasing the maximum amount of sustainable radial anisotropy. This behaviour is quantified by substitution in eq. (\[eq:ngNC\]) of the asymptotic expansion for $n\to\infty$ of $\sM(n,\gamma)$ in eq. (A2): $$\sM(n,\gamma)=\frac{1-2\gamma+\sqrt{1+4\gamma}}{2n}+{\rm O}(n^{-2}).
\label{eq:smAs}$$ As often happens in asymptotic analysis, even if the expansion above holds in principle only for very large values of $n$, the substitution of (\[eq:smAs\]) in (\[eq:ngNC\]) leads to percentual errors on $\sa$ less than 22%, 14%, and 10% for $n=4,5,6$ respectively (for the inner density slope $\gamma=1$).
We now move to discuss the WSC for one–component models: the obtained $\sa(n,\gamma)$ will mark a lower limit above which consistency (for the considered $n$ and $\gamma$) is guaranteed. Unfortunately, the WSC cannot be explored algebraically in the general case, because the resulting inequality (that for simplicity we do not report here) involves the solution of an equation of fifth degree for $n=5$, and the degree increases for increasing $n$. For $n=4$, instead, it is possible to treat the WSC analytically, since it reduces to the discussion of a cubic equation (see C99). Of course, the critical values for the anisotropy radius can be easily obtained solving numerically inequality (\[eq:WSC\]) in any specific case of interest, and the results are shown in Fig. \[f2\] for $n=4,5,6$ with dashed lines: all one–component models with the pair ($\sa$,$\gamma$) in the region above the dashed lines are consistent. Note how for increasing values of the external density slope $n$ more and more radially anisotropic orbital distributions can be supported, thus confirming the indications provided by the NC.
Values of $\sa$ nearer to the limits set by the DF are obtained by using the SSC. Inequality (\[eq:SSC\]) evaluated for a generic pair ($\gamma$,$n$) results in a transcendental equation that must be studied numerically (as already done for $n=4$ in C99), and the black dots joined by the dotted lines in Fig. \[f2\] represent the critical lower values of $\sa$ for one–component $\gamma$ models with $\gamma=0,1,2$, and $n=4,5,6$. As expected, the dotted lines are contained between the solid (NC) and the dashed (WSC) lines in each panel, and again they shift downward for increasing $n$.
Thus, from this preliminary investigation of one–component models we conclude that for fixed $\gamma$ an increase of $n$ corresponds to a decrease of the minimum admissible value of $\sa$, i.e. *steeper density distributions in the external regions can support more radial anisotropy*.
Sufficient conditions for halo–dominated (,$\gau$) models
---------------------------------------------------------
In order to proceed further with the preliminary discussion, we now apply the WSC to the component of a (,$\gau$) model, extending to this class of systems some of the results obtained in C96 and C99 for two–component $(\gau,\gad)$ models. In particular, the analytical study in C99 was restricted to some representative $(\gau,\gad)$ models, namely $a)$ isotropic two-component systems with inner slopes in the ranges $1\leq\gau<3$, and $0\leq\gad\leq\gau$ (i.e. the $\gad$ component is shallower in the central regions); $b)$ isotropic two–component (0,1) systems (i.e. the $\gad$ component is steeper in the central regions); and finally $c)$ anisotropic $\gau$ profiles with $1\leq\gau<3$ in the gravitational field of a dominant central black hole. Here, in order to obtain analytical results for the more general ($\ngam$,$\gau$) models, we assume that the $\gau$ component (the “halo”) dominates everywhere the gravitational field. Under this simplifying assumption, the following three results, corresponding to the points $a)$, $b)$, and $c)$ above, will be proved analytically:
1. In the case of halo–dominated [*isotropic*]{} (,$\gau$) models, with $1\leq\gamma <3$, $0\leq\gau\leq\gamma$, and $n>3$, the centrally more peaked component is consistent, independently of the value $\xi=\rc/\rs$ of its concentration relative to the centrally flatter $\gau$ halo. In the case of [*anisotropic*]{} (,$\gau$) models, the determination of a minimum anisotropy radius for consistency as a function of $n$,$\gamma$,$\gau$,$\xi$ reduces to the solution of an algebraic equation of sixth degree (for generic $n$), which is solved numerically. In the particular case of halo–dominated ($n$-2,1) models, the application of the WSC shows that for $\xi\leq (n-1)/2$ these models can be consistently assembled using radial orbits only.
2. In the case of halo–dominated [*isotropic*]{} ($n$-0,1) models, the WSC shows that for $\xi\leq (n+1)/2$ the $n$-0 density distribution is consistent. For broader $n$-0 density distributions, the models can be consistent only in presence of some amount of radial anisotropy.
3. In the case of [*anisotropic*]{} $\ngam$ models with a dominant black hole at their center it is possible to determine analytically a lower limit $\ra(n,\gamma)$ for consistency, and this limit decreases for increasing $n$.
![Minimum value of the anisotropy radius (normalized to the scale–length of the component, i.e. $\sa=\ra/\rc$) as a function of the relative concentration $\xi$, for consistency of halo dominated ($\ngam$,1) models. The external density slope $n$ increases as $n=4,5$, and $6$ from left to right panels, while the internal density slope as $\gamma=0,1$, and $2$ from top to bottom panels. Solid lines represent the limits derived from the WSC, while dashed lines are the $\sacm$ derived from the DF. Note that for ($n$-0,1) models also the $\sacp$ limits appear for $\xi>\xi_{\rm{c}}$ (see Sections 3.2 and 4.3 for details).[]{data-label="f3"}](f3.eps){height="0.4\textheight" width="50.00000%"}
The proof of the first result, which is an extension of the study mentioned in point $a)$, is conceptually straightforward but algebraically cumbersome, and only the main steps are reported in Appendix A (eq. \[A3\]). In particular, as this result holds also for $n>3$, while the external logarithmic slope of the $\gau$ halo is $4$, it means that a centrally steep density profile can be consistent in the gravitational field of a more massive (but centrally shallower) component, even when its external regions are less peaked. The situation is illustrated for some representative cases in Fig. \[f3\] (bottom and middle panels), where all points above the solid lines correspond to consistent halo–dominated ($n$-1,1) and ($n$-2,1) models, and apparently the isotropic limit ($\sa=\infty$) is allowed for any choice of $\xi$; the solid lines are obtained by solving numerically the corresponding WSC. A few additional trends are apparent. First, for a given central logarithmic density slope $\gamma$ and for a given relative concentration $\xi$, an increase of $n$ corresponds to a better and better ability to sustain radial anisotropy, and thus the trend found in one–component models is confirmed also in the two–component case. Second, at fixed $n$ and $\gamma$, the minimum anisotropy radius for consistency increases for increasing $\xi$. This trend was already found in C96 and C99: in practice, in a fixed potential broader density distributions are less and less able to sustain radial anisotropy. Third, at fixed external slope $n$ and relative concentration $\xi$, more centrally peaked systems are better able to support radial anisotropy. An additional comment concerns the specific case of ($n$-2,1) models (Fig. \[f3\], bottom panels). In fact, it is apparent how in presence of a dominant $\gau=1$ halo the WSC limits flatten to zero for relative concentration less than some critical value $\xi_{\rm{c}}$, and accordingly the model may be purely radially anisotropic. Remarkably, it is easy to show that the $n$-2 component can support purely radial orbits for relative concentrations $\xi\leq\xi_{\rm{c}}=(n-1)/2$ (see eq. \[A4\]), which are exactly the critical points in Fig. \[f3\]. Of course, this is just a sufficient condition, so we expect that the existence of a larger critical concentration for consistency in case of purely radial orbits will be revealed by direct inspection of the DF. A final comment is due. In fact, the adopted range of values for the internal density slopes $\gamma\geq\gamma_1$ imply that the halo density distribution is centrally less peaked than the $\ngamma$ component. Therefore, for large but [*finite*]{} total halo mass, the integrated mass of the halo is subdominant with respect to the stellar one for vanishing $r$, and the assumption of a dominant halo breaks down at the very center. This means that in an asymptotic sense, the analysis presented in the previous Section should be applied at the center of these finite-halo mass models.
Result (ii) extends the study mentioned in point $b)$ above, and it is proved in eq. (A5). The obtained limitation $\xi\leq\xi_{\rm{c}}=(n+1)/2$, shown by the solid line in Fig. \[f4\], suggests that the concentration of the $\gamma=0$ component must “adapt” to the dominant $\gau=1$ halo for phase–space consistency. Note that $\xi_{\rm{c}}$ increases for increasing $n$, i.e. a steeper external density profile can (partially) compensate for the effect of a central shallower density distribution. The appearence of the concentration limit $\xi_{\rm{c}}$ in the isotropic case manifests in the top panels of Fig. \[f3\]. In fact, as discussed in Section 2, in systems with positive isotropic functions (WSC or DF) only a lower limit for $\sa$ exists, as in the case of halo dominated ($n$-1,1) and ($n$-2,1) models. In the present case, instead, when for increasing $\xi$ the critical value $\xi_{\rm{c}}=2.5$ is crossed (in the $n=4$ panel), the additional condition $\sa<\sacp$ corresponding to the radial domain $A_-$ appears as the vertical asymptote of the solid line, and the parameter space for consistency progressively reduces, shrinking to zero for the value of $\xi$ where the two solid lines (obtained numerically) cross. A similar configuration repeats in the panels representing the $n=5$ and $n=6$ cases, where the consistency limits move, as expected, downward and rightward. Of course, these limits represent only a sufficient condition for the consistency, and when considering the DF in the limit $\mu\to 0$ (see Section 4) we expect to determine larger critical values of $\xi_{\rm{c}}$, and larger consistency regions in the ($\xi$,$\sa$) plane.
The result concerning the models with central black hole, related to point $c)$ above, is proved as follows. As well known, for $\rs\to 0$ the potential of the $\gau=1$ model becomes that of a point mass, and from the first result proven in this Section it follows that a black hole of any mass $M_{\rm BH}$ can be consistently added at the center of a globally isotropic model, when $1\leq\gamma<3$ and $n>3$. For this reason in the following we restrict to this range of slopes $\gamma$, and accordingly the long–dashed curves in Fig. \[f2\] interrupt for $\gamma<1$. Having reduced the study to models with a positive isotropic DF, we only have to estimate a lower limit of $\ra$. In Appendix A the WSC is applied to the anisotropic (,BH) models with $1\leq\gamma <3$; it can be discussed analytically in the special case of a [*dominant*]{} black hole, i.e. assuming in eq. (\[eq:WSC\]) $\mt =M_{\rm BH}$ (and so $\psit=GM_{\rm BH}/r$). As shown in eq. (A6), under these assumptions $$\label{eq:saBH}
\sa\geq\sM
\sqrt\frac{(3-\gamma)(\gamma-2)+\sM(n-2)[(6-2\gamma)-(n-3)\sM]}{
n(n-1)\sM^2+2n(\gamma-1)\sM+\gamma(\gamma-1)},$$ where $\sM=\sM(n,\gamma)$ is obtained by solving an algebraic equation of fourth degree: the obtained limits are shown in Fig. \[f2\] with the long–dashed lines, for $n=4,5,6$. For $n=4$ equation above coincide with that given in eq. (18) of C99, while asymptotic analysis proves that for $n\to\infty$ $$\label{eq:smBHas}
\sM(n,\gamma)=\frac{s_{\rm M 0}(\gamma)}{n}+{\rm O}(n^{-2}),$$ where $s_{\rm M0}(\gamma)$ is the larger real root of eq. (A9). We found that the substitution of eq. (\[eq:smBHas\]) in eq. (\[eq:saBH\]) leads to estimates of the minimum $\sa$ discrepant from the values obtained by the full formula by less than 50%, 33%, and 23% for $n=4,5,6$ respectively (for the inner density slope $\gamma=1$). As in the other cases, an increase of the external density slope $n$ makes it possible to sustain more radial orbits.
The DF of halo-dominated (,$\gau$) models
=========================================
It should be clear that the DF of two–component (,$\gau$) models cannot be obtained analytically, except for very special combinations of the values of $n,\gamma,$ and $\gau$, as for example the (1,1) models discussed in C96 (that would be referred as ($4$-$1$,$1$) models in the present notation). However, the search is not hopeless. In fact, from C99 it is known that the halo dominated DF of two–component (1,0) and (0,1) galaxy models can be obtained in terms of elementary functions, and a simple argument shows that this is also the case of (2,1) and (2,0) models. This fact, the linearity of the OM inversion formula with respect to $\rho$ in the non self–gravitating case (see eq. \[2\]), and finally the formula (\[eq:ng\]) for $n$ integer, prove that also the DF of the halo dominated (,1) and (,0) models can be expressed in terms of elementary functions, for any integer $n\geq4$ and $\gamma=0,1,2$. Note that, albeit their special nature, the study of halo–dominated models is useful because the formulae – expressible using elementary functions – can be studied very easily, making apparent the effects of the relative distribution of the investigated component and of the halo. In particular, in this Section we will determine the DF of the component of halo dominated (,1) models, and the exact phase–space constraints will be derived and compared with those obtained using the NC, WSC, and SSC in Section 3. We restrict to the $\gau=1$ case because the Hernquist (1990) potential is the simplest in the class of $\gamma$ models; furthermore, this choice allows us to investigate the consistency of models with central density slopes flatter ($\gamma=0$), equal ($\gamma=1$), and steeper ($\gamma=2$) than that of the halo. As a consequence, the results obtained for (,1) models should be representative of the whole situation.
With the normalization scales introduced at the beginning of Section 3, the relative potential of the Hernquist halo needed for the recovery of the DF is $$\label{eq:Her}
\Psi(r) =\displaystyle{\frac{\psin}{1+s}}=\psin \psitil(s),$$ while from eq. (\[eq:ng\]) $$\label{eq:rhoNor}
\rho_n (r)=\rhon \frac{\mu\xi^{n-3}(3-\gamma)(-1)^n}{4\pi\Gamma(n-3)s^{\gamma}}
\frac{d^{n-4}(\xi+s)^{\gamma-4}}{d\xi^{n-4}}.$$ As for the density and the potential, also for the DF it is useful to work with dimensionless functions, and we define $f=\fn\ftil(\Qtil)$ with $\fn=\rhon\psin^{-3/2}$, and $0\leq\Qtil=Q/\psin\leq1$: note that, at variance with C99, the normalization quantities used here are those of the halo, consistently with the halo dominated nature of the present models. In other words, the dominant density distribution sets the natural scales. The easiest way to compute the DF is to change the integration variable from the total potential to the radius in the first of the identities in eq. (\[eq:fOM\]). After normalization we obtain $$\label{eq:fOMnor}
f(Q)=\frac{\fn}{\sqrt{8}\pi^2}\left(\frac{d\Qtil}{d\nu}\right)^{-1}
\frac{d\Ftil(\nu)}{d\nu},\quad \nu=\nu (\Qtil),$$ where $$\begin{aligned}
\label{eq:fnu}
\Ftil(\nu)&=&-\int_{\nu}^{\infty}\displaystyle\frac{d\tilde\varrho_n}{ds}
\frac{ds}{\sqrt{\psitil(\nu)-\psitil(s)}}\nonumber \\
&=&\Ftilis(\nu)+\frac{\Ftilan (\nu)}{\sa^2},\end{aligned}$$ and from eq. (\[eq:Her\]) $$\label{eq:Qnu}
\Qtil=\frac{1}{1+\nu},\quad \left(\frac{d\Qtil}{
d\nu}\right)^{-1}=-(1+\nu)^2, \qquad 0\leq \nu\leq\infty.$$ In the equations above, a negative sign appears in front of the integral (\[eq:fnu\]) due to the monotonic decrease of the relative potential with increasing radius; $\tilde\varrho_n$ is the normalized augmented density associated with eq. (\[eq:rhoNor\]), while $\sa =\ra/\rs$ is the normalized anisotropy radius; finally, the subscripts of the functions $F$ refer to the isotropic and anisotropic parts of the DF, respectively. It is now evident how in halo dominated ($\ngam$,$\gau$) models the external slope parameter appears in the density $\tilde\varrho_n$ only, so that the DF of the component can be obtained by repeated differentiation with respect to $\xi$ of the simpler DF of the halo dominated ($\gamma$,$\gau$) model. In the following, we will also discuss black hole dominated ($\ngam$,BH) models: then, we will adopt as natural normalization scales the black hole mass $M_{\rm BH}$ and the scale–length $\rc$ of the component; moreover, $0\leq\Qtil\leq\infty$, since in this case $\psitil=1/s$, so that in eq. (\[eq:Qnu\]) $\Qtil=1/\nu$.
It is not difficult to show that, for generic values of $\gamma$, the integral (\[eq:fnu\]) can be expressed as a combination of hypergeometric ${}_2F_1$ functions when the halo is a $\gau=0$, or $1$ model, or a black hole. However, the resulting expressions are not more illuminating than the integral itself, so that we do not report them here. Instead, we prefer to show simple asymptotic expansions relative to a couple of interesting situations. The first concerns the behaviour of the DF of ($\ngam$,BH) models for $\Qtil\to\infty$, for which the leading term of the expansion (normalized to $\fn/\sqrt{8}\pi^2$) is $$\label{asymBH}
\ftil(\Qtil)\sim\gamma\,\left(\gamma-\frac{1}{2}\right)\Qtil^{\gamma-3/2}:$$ thus, the slope $\gamma=3/2$ marks the different behaviour of the models. We note also that $f<0$ for $\gamma<1/2$ and the model is inconsistent (Tremaine et al. 1994); furthermore, we note how in the leading term above the function $\Ftilan$ is producing no contribution. The other case of interest is the behaviour of the DF of isotropic models for $\Qtil\to 0$. We recall that this case is very general, as it applies to self-gravitating models but also to models in generic halos of finite total mass, because in all systems $\psit\sim G M_{\rm T}/s$ for $s\to\infty$. In this case we found $$\ftil(\Qtil)\sim n\,\left(n-\frac{1}{2}\right)\Qtil^{n-3/2},$$ an expression (obviously) formally identical to eq. (\[asymBH\]). Finally, we notice that the asymptotic expansion for $r\to\infty$ of the velocity dispersion of isotropic models (both single component and embedded in dark matter halos of finite total mass) is $\sigr^2\sim
1/(n+1)s$, normalized to $GM_{\rm{T}}/r_{\rm c}$, so that for $r\to\infty$ $\rho/\sigr^3\sim (n+1)^{-3/2} s^{3/2-n}$, in accordance with the comment in Footnote 1.
The halo dominated ($n$-2,1) Model
----------------------------------
Following the preliminary discussion, here we derive the explicit expression for the DF of the $n$-2 model with an arbitrary degree of OM orbital anisotropy, in the gravitational field of a dominant Hernquist halo. After partial fraction decomposition of eq. (\[eq:fnu\]), the isotropic and anisotropic components of the DF are $$\label{dfisg2}
\Ftilis(\nu)=\displaystyle\frac{(-1)^{n}\mu\sqrt{1+\nu}\,\xi^{n-3}}{2\pi(n-4)!}\,
\,\displaystyle\frac{d^{n-4}}{d\xi^{n-4}}\displaystyle
\left(\frac{G_3+G^0_3}{\xi^2}+\frac{G_2-G^0_2}{\xi^3}\right),$$ and $$\begin{aligned}
\label{dfang2}
\Ftilan(\nu)&=&\displaystyle\frac{(-1)^{n}\mu\sqrt{1+\nu}\,\xi^{n-3}}
{2\pi(n-4)!}\,
\displaystyle\frac{d^{n-4}G_3}{d\xi^{n-4}}\nonumber\\
&=&\displaystyle\frac{\mu\sqrt{1+\nu}(n-2)!\,\xi^{n-3}\,G_{n-1}}{4\pi(n-4)!},\end{aligned}$$ respectively, where for $k\geq2$ we define $$G_k(\xi,\nu)\equiv\int_{\nu}^{\infty}\sqrt{\frac{s+1}{s-\nu}}
\frac{ds}{(\xi+s)^k}, \qquad G^0_k(\nu)\equiv G_k(0,\nu).$$ For $\xi=1$ the function $G_k$ assume a very simple form: $$G_k(1,\nu)=\frac{\sqrt{\pi}\,\Gamma (k-1)}{\Gamma(k-1/2)
(1+\nu)^{k-1}},$$ while for $\xi\neq 1$ the formula $$\label{recG}
G_{k+1}(\xi,\nu)=-\frac{1}{k}\frac{dG_k(\xi,\nu)}{d\xi}=
\frac{(-1)^{k-1}}{k!}\frac{d^{k-1}G_2(\xi,\nu)}{d\xi^{k-1}}$$ holds, with $$\begin{aligned}
&G_2(\xi,\nu)=\displaystyle{\frac{1}{\xi+\nu}}&\nonumber\\
&+\displaystyle{\frac{1+\nu}{
(\xi+\nu)^{3/2}}}&\begin{cases}
\displaystyle{\frac{1}{\sqrt{1-\xi}}
\arctan \sqrt{\frac{1-\xi}{\xi +\nu}}},
&(0\leq\xi <1),\cr
\displaystyle{\frac{1}{\sqrt{\xi-1}}
\arctanh \sqrt{\frac{\xi -1}{\xi +\nu}}},
&(\xi >1).\cr
\end{cases}\end{aligned}$$ Note that the second expression in eq. (\[dfang2\]) has been obtained from the relation (\[recG\]).
![Upper limits for consistency on the concentration parameter $\xi=\rc/\rs$ of the fully isotropic halo-dominated ($n$-0,1) model, as a function of the external density slope $n$. The solid line is the limit derived from the WSC (Section 3.2), while the dotted line is the true limit obtained from the DF (Section 4.3).[]{data-label="f4"}](f4.eps){height="0.30\textheight" width="40.00000%"}
![From top to bottom: the dimensionless DF for the halo dominated ($n$-2,1), ($n$-1,1), and ($n$-0,1) models, as a function of $\Qtil$. The solid lines represent the isotropic DFs, while the dotted lines are the anisotropic DFs, with $\sa$ approaching the critical value for consistency. In all cases, $\xi=5$.[]{data-label="f5"}](f5.eps){height="0.45\textheight" width="55.00000%"}
From the derived formulae we determined the true anisotropy limit for the ($n$-2,1) models. We found that $\fis>0$ for all the explored values of $n$, i.e. the isotropic $n$-2 model in a dominant Hernquist halo is consistent independently of the halo scale–length, in accordance with the result in point $1)$ of Section 3.2, obtained with the WSC. Therefore the set $A_+$ in phase–space coincides with the whole accessible phase–space, and only the anisotropy limit $\sacm$ exists. In Fig. \[f5\] (top panel) we show the isotropic (solid) and strongly anisotropic (i.e., near to the consistency limit) DFs for $n=(4,5,6)$ and $\xi=5$. Note how the isotropic DFs are very similar, and monotonically decreasing. Instead, in the anisotropic cases the depression leading to inconsistency is apparent, clearly showing how phase–space inconsistency is always set outside the center. This feature is qualitatively similar to the others explored OM systems (e.g., see Fig. 2 in Ciotti & Lanzoni 1997, Figs. 2 and 3 in C99, Fig. 3 in Ciotti, Morganti & de Zeeuw 2008); in addition, the systematic shift of the DF depression towards high $Q$ values for increasing $n$ is also apparent: an argument supporting this phenomenon is given in Section 2.1.1. We notice that the cuspy dips shown by the DFs for values of $\sa$ near the consistency limit could be the source of kinetic instabilities, whose investigation is of course well beyond the framework of this paper. In any case, it is almost certain that any system would develop radial orbit instability for values of the anisotropy radius larger than the consistency limit (e.g., see the N-body experiments discussed in Nipoti, Londrillo & Ciotti 2002): therefore, the critical DFs in Fig. 5 describe equilibrium unstable systems.
The trend of the anisotropy limit $\sacm$ for different values of $n$ and increasing $\xi$ is shown in Fig. \[f3\] (bottom panel) with the dashed lines. First, for fixed $n$ and increasing $\xi$, $\sacm$ increases: in practice, for a given halo, a broader stellar density distribution is less and less able to sustain radial anisotropy. Second, at fixed $\xi$ a steeper external density slope (i.e. larger $n$) corresponds to higher amount of admissible radial anisotropy, thus showing that not only the inner density slope, but also the external density profile is important for phase–space consistency. As expected, the true limitation on $\sacm$ is less stringent than the corresponding limit derived from the WSC, and the dashed line is everywhere below the solid line. In particular, while from the WSC the critical concentrations $\xi_{\rm{c}}$ in order to have a purely radial model are $1.5,2,$ and $2.5$ (for $n=4,5,$ and $6$), from the DF we obtain $\xi_{\rm{c}}\simeq 2.8,3.7,$ and $4.5$.
It is also possible to determine the DF of ($n$-2,BH) models in presence of a dominant central black hole. The resulting formulae are identical to eqs. (\[dfisg2\])-(\[dfang2\]), where the quantity $\sqrt{1+\nu}$ at numerator is replaced by $\sqrt{\nu}$ and the $G$ functions are replaced by the corresponding $G_{\bullet}$ functions of same index, defined for $k\geq 2$ as $$G_{\bullet k}(\xi,\nu)\equiv\int_{\nu}^{\infty}\sqrt{\frac{s}{s-\nu}}
\frac{ds}{(\xi+s)^k},$$ and $$G^0_{\bullet k}(\nu)\equiv G_{\bullet k}(0,\nu)=\frac{\sqrt{\pi}\,\Gamma (k-1)}
{\Gamma(k-1/2)\nu^{k-1}}.$$ For $\xi>0$ the formula $$G_{\bullet k+1}(\xi,\nu)=-\frac{1}{k}\frac{d G_{\bullet k}(\xi,\nu)}{d\xi}=
\frac{(-1)^{k-1}}{k!}\frac{d^{k-1}G_{\bullet 2}(\xi,\nu)}{d\xi^{k-1}}$$ holds, where $$G_{\bullet 2}(\xi,\nu)=
\frac{1}{\xi+\nu}+\frac{\nu}{(\xi+\nu)^{3/2}\sqrt{\xi}}
\arctanh{\sqrt{\frac{\xi}{\xi+\nu}}}.$$
![Dimensionless DF for the black hole dominated ($n$-2,1) and ($n$-1,1) models, as a function of the relative energy $\Qtil$, for $n=4,5,$ and $6$. Normalization scales are the mass of the black hole and the scale–length of the component. Solid lines represent the isotropic DFs, while dotted lines the anisotropic DFs, with $\sa$ approaching the critical value for consistency.[]{data-label="f6"}](f6.eps){height="0.46\textheight" width="52.00000%"}
Note that after performing the required differentiations one must set $\xi=1$. In Fig. \[f6\] (upper panel) the obtained DFs are shown in the isotropic (solid lines) and strongly anisotropic (dotted lines) cases. Again, inconsistency due to OM anisotropy affects the DF at intermediate energies. The true anisotropy limits derived from the DF are given by solid squares in Fig. \[f2\]. Finally, note also the characteristic log-log linear trend of the DF for very high values of $Q$. This trend is easily explained by considering the asymptotic expansion given in eq. (\[asymBH\]).
The halo dominated ($n$-1,1) Model
----------------------------------
We now derive the explicit DF for the family of halo dominated ($n$-1,1) models, so that with this class of models we can study the effect of the external logarithmic slope when the two density profiles have the same inner slope. Of course, the (4-1,1) models are the halo dominated of (1,1) models discussed in C96.
Integration of eq. (\[eq:fnu\]) reveals that the functions involved in the expression of the DF are the same as for the ($n$-2,1) models. In particular, $$\begin{aligned}
\label{fisg1}
\Ftilis(\nu)&=&\displaystyle\frac{(-1)^{n}\mu\sqrt{1+\nu}\,
\xi^{n-3}}{2\pi(n-4)!}\,\nonumber\\
&&\displaystyle\frac{d^{n-4}}{d\xi^{n-4}}
\left(\frac{G^0_2-G_2}{\xi^{3}}-\frac{2G_3}{\xi^2}-\frac{3G_4}{\xi}\right),\end{aligned}$$ and $$\label{fang1}
\Ftilan(\nu)=\displaystyle\frac{(-1)^{n}\mu\sqrt{1+\nu}\,\xi^{n-3}}{2\pi(n-4)!}\,
\displaystyle\frac{d^{n-4}(2 G_3-3\xi G_4)}{d\xi^{n-4}}.$$ In Fig. \[f5\] (middle panel), the DF in the isotropic and strongly anisotropic cases is shown, for $n=4,5,6$. The behaviour is qualitatively similar to that of ($n$-2,1) models, i.e. while the isotropic systems display a monotonically decreasing DF, in the anisotropic case (when $\sacm$ approaches the consistency limit) the DF has a significant depression, that moves inwards for increasing $n$. In Fig. \[f3\] (middle row) we show the minimum value of the anisotropy radius normalized to $\rc$ as derived from the DF (dashed lines), and from the WSC. Again, in accordance with the preliminary study of Section 3.2, we found that $\fis>0$ independently of the relative concentration value $\xi$ and of the external slope $n$, so that only the limit $\sacm$ exists. All the trends exhibited by the ($n$-2,1) models are confirmed: in particular, $\sacm$ increases with $\xi$ for fixed $n$, while it decreases at fixed $\xi$ for increasing $n$. Note that the maximum radial anisotropy admissible for given $\xi$ and $n$ is smaller than for ($n$-2,1) models, as expected from the shallower central profile of $n$-1 models. Finally, at variance with the ($n$-2,1) models (and in agreement with the expectations of the preliminary analysis) no purely radial orbital configurations can be supported, independently of the value of the concentration $\xi$.
As for ($n$-2,BH) models, also in the case of the black hole dominated ($n$-1,BH) models the functions $\Ftilis$ and $\Ftilan$ are formally identical to those in eqs. (\[fisg1\])-(\[fang1\]), with each $G$ function replaced by the corresponding $G_{\bullet}$ function, and with the substitution $\sqrt{1+\nu}\to\sqrt{\nu}$; again, it is necessary to set $\xi=1$ after performing the required differentiations. While the true limits on the anisotropy radius are plotted in Fig. \[f2\] with the solid squares, the obtained DFs are shown in the isotropic (solid) and strongly anisotropic (dotted) cases in the bottom panel of Fig. \[f6\]; not unexpectedly, the inconsistency due to anisotropy manifests itself at intermediate energies. Finally, the behaviour of the DF for very high values of $Q$ is in accordance with the asymptotic expansion of the DF in eq. (\[asymBH\]).
The halo dominated ($n$-0,1) Model
----------------------------------
This case is expected to be the more complicated, because from C99 it is known that the (0,1) model presents a peculiar behaviour: in practice, there exists a range of $\xi$ where anisotropic models can be consistent when instead isotropic models are not, and this is due to the fact that $\fis$ becomes negative over the non empty set $A_-$, so that two limits on $\sa$ must be considered. Note that the discussion of the WSC in the case of ($n$-0,1) models in Section 3.2 already showed a similar behaviour.
As in the other cases, the integration of eq. (\[eq:fnu\]) can be done by using the $G_k$ functions only: $$\begin{aligned}
\label{fIg0}
\Ftilis(\nu)&=&\displaystyle\frac{(-1)^{n}3\mu\sqrt{1+\nu}\,\xi^{n-3}}
{\pi(n-4)!}\,
\frac{d^{n-4}G_5}{d\xi^{n-4}}\nonumber\\
&=&{n\choose4}\displaystyle\frac{3\mu\sqrt{1+\nu}\,\xi^{n-3}G_{n+1}}{\pi}\, ,\end{aligned}$$ $$\begin{aligned}
\Ftilan(\nu)&=&\displaystyle\frac{(-1)^{n}3\mu\sqrt{1+\nu}\,\xi^{n-3}}
{2\pi(n-4)!}\,\nonumber\\
&&\displaystyle\frac{d^{n-4}(2\xi^2 G_5-3\xi G_4+G_3)}{d\xi^{n-4}},\end{aligned}$$ where the second expression for $\Ftilis$ is obtained from the relation (\[recG\]) of the $G_{k}$ functions. Motivated by the remarks above, we start to study the sign of $\fis$, in order to determine the condition for the existence of the sets $A_+$ and $A_-$. Indeed, we found that for $\xi\lsim 5.233$ (in the $n=4$ case, see C99), $\xi\lsim 6.192$ (for $n=5$), and $\xi\lsim 7.166$ (for $n=6$), $\fis$ is positive everywhere over the accessible phase–space, a situation which is graphically represented with the dotted lines in Fig. \[f4\]. It follows that for sufficiently concentrated models only the anisotropy limit $\sacm$ exists, and in Fig. \[f5\] (bottom panel, solid lines) the isotropic DF is shown for ($n$-0,1) models with $\xi=5$ and $n=(4,5,6)$: note how, at variance with the other cases, the DF of ($n$-0,1) models decreases for increasing $Q$. The trend of $\sacm(\xi)$ is shown in Fig. \[f3\] (top panels, dashed lines): as for the models previously discussed, the $\sacm$ curve is similar (but displaced) with respect to that derived from the WSC. It is apparent how the $\sacm$ line lies above those of $n$-1 and $n$-2 models, as expected from the shallower density profile of $n$-0 models; moreover, we note how the minimum admissible values for $\sacm$ decrease for increasing $n$, as in the other cases. In Fig. \[f5\] the DFs for $\sa$ near the consistency limit $\sacm$ are shown with dotted lines for models having $\xi=5$, and as in the previous $\gamma=1$ and $\gamma=2$ models, inconsistency appears at intermediate values of $\Qtil$.
For values of $\xi$ larger than the critical values of $\xi_{\rm{c}}$ reported above, the set $A_-$ is not empty, because the isotropic component of the DF becomes negative at high relative energies. In fact, at variance with the ($n$-2,1) and ($n$-1,1) models, the DF of the $n$-0 model converges to a finite limit (that can be easily calculated analytically from eq. \[\[fIg0\]\]) for $\Qtil=1$; when increasing $\xi=\rh/\rs$, the value $\ftil(1)$ monotonically decreases, till it becomes negative for $\xi$ greater than $\xi_{\rm{c}}$. A similar behavior was found in the numerical investigation of consistency of King (1972) and quasi–isothermal halos added to a de Vaucouleurs (1948) density distribution (see CP92). Thus, while the effects of OM radial anisotropy appear to affect the DF at intermediate energies, the inconsistency of isotropic models embedded in a peaked halo seems to characterize the DF at very high (relative) energies. According to the discussion of Section 2, since the function $\fan$ is positive on $A_-$, the upper bound $\sacp$ can be determined: the appearence of this new value is manifested by the dotted lines in Fig. \[f3\], with the vertical asymptote corresponding to $\xi_{\rm{c}}$. Then, for $\xi$ greater than the critical limit, a range of concentrations exists where an anisotropy radius $\racm\leq\ra\leq\racp$ corresponds to consistent models; finally, for $\xi$ larger than the value where the dotted ($\sacp$) and dashed ($\sacm$) lines cross each other, $\sacm>\sacp$, and the models become inconsistent. Summarizing, consistent models correspond to points placed above the $\sacm$ line and on the left of the $\sacp$ line.
In the case of a central dominant black hole, the same comments as those for ($n$-2,BH) and ($n$-1,BH) models apply, and the functions $\Ftilis$ and $\Ftilan$ are obtained accordingly. Of course, the isotropic model is inconsistent, independently of the values of $n$, because the model with black hole would correspond to the formal limit $\xi\to\infty$ in ($n$-0,1) models, which is already excluded by the previous discussion. What happens when considering the effect of anisotropy? Is there any compensating effect? A numerical inspection shows that the anisotropic part of the DF is negative on $A_-$, so that not even orbital anisotropy can make the models consistent. Again, this result was expected from the study of ($n$-0,1) models, due to the crossing of the $\sacm$ and $\sacp$ critical lines at finite values of $\xi$.
Discussion and conclusions
==========================
In this paper, in a natural extension of previous investigations (CP92, C96, C99), we focused on the importance of the *external* density logarithmic slope on the determination of the phase–space consistency of one and two–component stellar systems with OM orbital anisotropy. In particular, the considered ($\ngam$,$\gau$) models generalize ($\gau$,$\gad$) models, as the density component is characterized by logarithmic slope $n$ instead of $4$ outside the core radius, while in the inner regions the logarithmic density slope is $\gamma$. The main results can be summarized as follows:
1. It is shown that, for $n$ integer, several structural and dynamical properties of the models can be obtained by differentiation, with respect to the scale–length, of the corresponding formulae valid for the standard $\gamma$ models. For example, it is shown how to construct explicit solutions of the Jeans equations for the class of two–component anisotropic ($n_1$-$\gau$,$n_2$-$\gad$) models once the solution is known for ($\gau$,$\gad$) models.
2. In one-component models, a lower limit for the anisotropy radius, so that smaller values certainly produce inconsistent models, is analytically derived following the technique introduced in CP92. It is found that for $0\leq\gamma<2$ this minimum anisotropy radius is strictly positive, but it decreases by increasing $n$, i.e. a larger amount of radial anisotropy can be supported by externally steeper density profiles. For $2\leq\gamma<3$ instead the necessary condition for consistency is satisfied $\forall\sa>0$, independently of the value of $\gamma$ and $n$. The minimum anisotropy radius so that larger values correspond to consistent models is then determined by using the strong and weak sufficient conditions for model consistency derived in CP92. As for the necessary condition, we find that the minimum value of the anisotropy radius decreases for increasing $\gamma$ and fixed $n$, and for increasing $n$ and fixed $\gamma$.
3. A similar analysis is then performed for two-component ($\ngam$,$\gau$) systems, in order to extend the study of C99. For simplicity we restrict to halo dominated models, i.e. we take into account only the gravitational field of the $\gau$ halo. In particular, we show that in the isotropic case, when $1\leq\gamma<3$ and $0\leq\gau\leq\gamma$ (i.e., the halo density is centrally less peaked than the stellar component), the DF of the component is nowhere negative, independently of the mass and concentration of the $\gau$ halo. This is true even when in the external parts the component is less peaked $(3<n<4)$ than the $\gau$ component. As a special application of this result, it follows that a black hole of any mass can be consistently added at the center of a globally isotropic model, when $1\leq\gamma
<3$. Moreover, in the case of ($n$-0,1) models, where instead the halo is centrally steeper than the stellar component, the sufficient condition applied to the $n$-0 density distribution reveals the existence of an upper limit $\xi\leq(n+1)/2$ of the relative concentration $\xi=\rh/\rs$ for model consistency in the isotropic case. Finally, in the case of anisotropic models with a dominant black hole at their center, we determined analytically a lower limit of the critical anisotropy radius for consistency as a function of $n$ and $\gamma$; this value decreases for increasing $\gamma$ and/or $n$.
4. The analytic expression for the DF of OM anisotropic halo dominated ($\ngam$,1) models, and (,BH) models with a dominant central BH is recovered in terms of elementary functions in the special cases when $\gamma=(0,1,2)$. It is found that, while for $\gamma=1,2$ the isotropic DF is positive independently of the model concentration, the $\gamma=0$ component becomes inconsistent when the halo is sufficiently concentrated, even in the isotropic case. In addition, the trend of the minimum value of the anisotropy radius as a function of the halo concentration confirms that steeper density profiles in the external region are consistent with a larger amount of radial anisotropy.
In conclusion, the explored family of models provide a direct indication that the density slope in the external regions of stellar systems can be important in determining the admissible radial anisotropy, and this in addition to the well known relevance of the central density slope. A quantification of this argument is embodied in a [*necessary*]{} inequality for model consistency, that we derived by expressing a previous finding (CP92) in terms of the logarithmic density slope. The obtained result holds for [*all*]{} consistent OM anisotropic systems, and relates the logarithmic density slope and the OM anisotropy indicator at [*each radius*]{}, in a way formally identical to the “cusp slope-central anisotropy” theorem by AE06, which is known to apply at the center of stellar systems with generic anisotropy distribution, and everywhere in constant anisotropy systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
We wish to thank the anonymous Referee for helpful suggestions that greatly improved the presentation of the paper.
Consistency requirements
========================
The NC for one–component models
-------------------------------
Simple algebra shows that the NC applied to anisotropic models can be written as $$\sa^2\geq \frac{2-\gamma-(n-2)s}{\gamma+ns}s^2, \quad 0\leq
s\leq\infty .$$ For $\gamma\geq2$ and $n\geq2$ the r.h.s. is everywhere negative, so that inequality (A1) is trivially satisfied for $\sa\geq0$. Instead, for $0\leq\gamma<2$ the parameter $\sa$ must be larger than the maximum of the expression of the r.h.s., which is reached at $$\sM (n,\gamma) =\frac{n-(2n-3)\gamma+
\sqrt{(n-\gamma)[n+(4n-9)\gamma]}}{2n(n-2)},$$ that after substitution in eq. (A1) gives eq. (\[eq:ngNC\]). Of course, for $n=4$ eq. (A2) in C99 is reobtained.
The WSC for halo–dominated (,$\gau$) models
-------------------------------------------
Here we apply the WSC to the component of globally isotropic, halo–dominated (,$\gau$) models, where $1\leq\gamma <3$, $0\leq\gau\leq\gamma$, and $n>3$. Under the assumption of a dominant halo, eq. (\[eq:WSC\]) reduces to investigate the positivity of an expression which factorizes in a strictly positive function and in an algebraic factor that after the natural substitutions $\gamma=1+\epsilon$ and $\gau=\gamma-\epu$ (with $0\leq\epsilon <2$ and $0\leq\epu\leq\gamma$) becomes $$\begin{aligned}
&&n(n-1)s^3+n[n+1-\epsilon+\epu+2\xi\epsilon]s^2\\
&&+\xi\{[(n+1)(2+\epsilon+\epu)
+\epsilon\epu-\epsilon^2]+\xi\epsilon\gamma\}s+
\xi^2\gamma(2+\epu),\nonumber\end{aligned}$$ whose positivity $\forall (s,\xi)\geq 0$ is easily proved. Therefore the WSC is verified for any value of the concentration parameter $\xi$. Note that the variables $\xi$ and $s$ adopted here correspond to $1/\xi$ and $s/\xi$ respectively in eq. (A13) of C99, due to the different normalization length.
Being the WSC of the isotropic models satisfied, the existence of a critical anisotropy radius $\sacm$ for consistency depends on the sign of the anisotropic part of the WSC, as discussed in Section 2.2. In particular, the positivity of the anisotropic WSC for ($n$-2,1) models reads $$s(n-3)+n-1-2\xi\geq 0,\quad 0\leq s\leq\infty,$$ and thus shows the existence of a critical value $\xi_{\rm{c}}=(n-1)/2$ of the concentration parameter, marking an upper limit on the values of $\xi$ for which $\sacm=0$, i.e. purely radial orbital distributions are allowed.
The discussion of the WSC for globally isotropic $n$-0 models in a dominant $\gamma=1$ halo is simple, reducing to the request that $$s(n-1)+n+1-2\xi\geq 0,\quad 0\leq s\leq\infty,$$ which is satisfied for $\xi\leq\xi_{\rm{c}}=(n+1)/2$. For $n=4$ we reobtain the result of C99.
In the case of central dominant BH, we assume in eq. (\[eq:WSC\]) $\mt=M_{\rm BH}$, and from the previous discussion we restrict to models with $1\leq\gamma <3$. The WSC then requires that $$\sa^2\geq
s^2\frac{(3-\gamma)(\gamma-2)+2(n-2)(3-\gamma)s-(n-2)(n-3)s^2}{
n(n-1)s^2+2n(\gamma-1)s+\gamma(\gamma-1)}.$$ After the differentiation of the r.h.s. and the successive study of a quartic equation, it can be proved that for $s\geq0$ eq. (A6) admits only one maximum, located at $\sM=\sM(n,\gamma)\geq 0$. The general expression for $\sM(n,\gamma)$ is not reported here, but it was used to produce the long–dashed lines in Fig. \[f2\]. In any case, $\sM(n,3)=0$, $$\sM(n,\gamma)=
\begin{cases}
\displaystyle{\frac{2}{n-3}},
&\gamma=1,\cr
\displaystyle{\frac{\sqrt[3]{s_0}}{3n(n-1)(n-3)}-
\frac{2(n-4)}{3(n-1)(n-3)}}\cr
+\displaystyle{\frac{4(n^3+n^2-20n+27)}{3 (n-1)(n-3)\sqrt[3]{s_0}}},
&\gamma=2,\cr
\end{cases}$$ where $$\begin{aligned}
s_0&=&n^2\left[73n^4-660n^3+2262n^2-3484n+2025\right.\nonumber\\
&+&\left.9\sqrt{\frac{(5n-9)(13n^2-59n+64)(n-1)^3(n-3)^3}{n}}\right],\end{aligned}$$ and for $n=4$ and $\gamma=1,2,3$ the values given in Table 1 of C99 are reobtained. We finally report the reduced asymptotic fourth–order equation needed to determine $\sM(n,\gamma)$ in the limit $n\to\infty$: remarkably, after the scaling $s=y/n$ and the successive limit $n\to\infty$, one obtains $$\begin{aligned}
&y^4&+2(2\gamma-3)y^3+6(\gamma-1)(\gamma-2)y^2\\
&+&2(\gamma-3)(\gamma-1)(2\gamma-1)y+(\gamma-3)(\gamma-2)(\gamma-1)\gamma=0.\nonumber\end{aligned}$$
The velocity dispersions and the virial quantities for models
=============================================================
Here we present an easy way to express analytically the main dynamical quantities of the two–component OM anisotropic ($n_1$-$\gau$,$n_2$-$\gad$) models for generic $\gau$ and $\gad$, and $n_1$ and $n_2$ integer $\geq 4$. The method is based on the evaluation of the similar quantities for the two–components $(\gau,\gad)$ models, and then on repeated differentiation with respect to the scale–lengths.
We start by considering the radial component $\sigr^2$ of the velocity dispersion $\sigma^2=\sigr^2+\sigt^2$ of a system with a density component $\rho$, which in the OM parameterization can be written as $$\label{bin&mer}
\rho(r)\sigr^2(r)=\frac{A(r)+\ra^2 I(r)}{r^2+\ra^2},$$ where $$A(r)=G\int_r^{\infty}\rho(r)\mt(r)dr,\quad
I(r)=G\int_r^{\infty}\frac{\rho(r)\mt(r)}{r^2}dr,$$ (Binney & Mamon 1982), and $\mt(r)$ is the total mass within $r$. Once $\sigr^2$ is known, the tangential velocity dispersion is obtained from eq. (\[betaOM\]).
As the method is general, we focus without loss of generality on the component “1” of a two–component model, for which the two functions $I$ and $A$ are given by the sum $\Is =\Iss+\Ish$ and $\As=\Ass+\Ash$, due to the linearity of mass in eqs. (B2), and where the meaning of the subscript indices is apparent. Other quantities of interest in applications are the global energies entering the scalar virial theorem, and for the component 1 the scalar virial theorem reads $2K_1=|W_{11}|+|\wsh|$, where $W_{11}=-4\pi G\int r\rhos\ms dr$ is the contribution due to the self–interaction, and $\wsh=-4\pi G\int
r\rhos M_{\rm2}dr$ is the interaction energy with the “halo” (e.g., Ciotti 2000).
The basic idea is to compute the integrals for the generic pair ($\rhos,M_{\rm2}$) relative to standard two–component $(\gau,\gad)$ models with different scale–lengths $\rcu$ and $\rcd$, to perform the required differentiations as prescribed by eqs. (\[eq:ng\]) and (\[eq:Mng\]), and finally (in the case of self–interactions) to set the two scale–lengths and the two slopes to the same value. From eqs. (\[eq:ng\]) and (\[eq:Mng\]) it follows $$\begin{aligned}
A_{12}(r)&=&G\frac{\rcu^{n_1-3}\rcd^{n_2-3}(-1)^{n_1+n_2}}
{\Gamma(n_1-3)\Gamma(n_2-3)}\nonumber\\
&&\frac{d^{n_1-4}}{d\rcu^{n_1-4}}\frac{d^{n_2-4}}{d\rcd^{n_2-4}}
\int_r^{\infty}\frac{\rho_{\gau}(r,\rcu)M_{\gamma_2}(r,\rcd)}
{\rcu\rcd}dr,\\
I_{12}(r)&=&G\frac{\rcu^{n_1-3}\rcd^{n_2-3}(-1)^{n_1+n_2}}
{\Gamma(n_1-3)\Gamma(n_2-3)}\nonumber\\
&&\frac{d^{n_1-4}}{d\rcu^{n_1-4}}\frac{d^{n_2-4}}{d\rcd^{n_2-4}}
\int_r^{\infty}\frac{\rho_{\gau}(r,\rcu)M_{\gamma_2}(r,\rcd)}
{\rcu\rcd r^2}dr,\\
W_{12}(r)&=&-4\pi G\frac{\rcu^{n_1-3}\rcd^{n_2-3}(-1)^{n_1+n_2}}
{\Gamma(n_1-3)\Gamma(n_2-3)}\nonumber\\
&&\frac{d^{n_1-4}}{d\rcu^{n_1-4}}\frac{d^{n_2-4}}{d\rcd^{n_2-4}}
\int_0^{\infty}\frac{\rho_{\gau}(r,\rcu)M_{\gamma_2}(r,\rcd)}
{\rcu\rcd}r dr;\end{aligned}$$ for generic values of $\gau$ and $\gad$ the integrals above involve hypergeometric ${}_2F_1$ functions, while for integer values of $\gau$ and $\gad$ the resulting formulae can be expressed in terms of elementary functions (e.g., see C96 and C99 for (1,0) and (1,1) models; Ciotti et al. 1996 for (2,1) and (2,2) models). Even simpler expressions can be obtained without difficulty for to the halo dominated ($n$-0,1), ($n$-1,1), and ($n$-2,1) models, by adopting as normalization constants the physical scales $\ms$ and $\rs$ of the $\gamma=1$ dominant component, and by fixing $n_2=4$, $\rcd=\rs$, $n_1=n$, $\xi\equiv\rcu/\rs$, $s\equiv
r/\rs$, $\mu\equiv M/\ms$ in eqs. (B3)-(B6), where $M$ and $\rcu$ are the mass and the scale–lenght of the $\ngam$ component. As the integration is trivial but the results not particularly illuminating, we do not show the explicit formulae.
[99]{} An, J.H., & Evans, W. 2006, ApJ, 642, 752 (AE06) Baes, M., & Dejonghe, H. 2004, MNRAS, 351, 18 Baes, M., Dejonghe, H., & Buyle, P. 2005, A&A, 432, 411 Bertin, G. 2000, Dynamics of Galaxies (Cambridge: Cambridge University Press) Bertin, G., & Stiavelli, M. 1984, A&A, 137, 26 Bertin, G., & Trenti, M. 2003, ApJ, 584, 729 Bertin, G., Saglia, R.P., & Stiavelli, M. 1992, ApJ, 384, 423 Binney, J., & Merrifield, M. 1998, Galactic Astronomy, (Princeton: Princeton University Press) Binney, J., & Mamon, G. 1982, MNRAS, 200, 361 Binney, J., & Tremaine, S. 2008, Galactic Dynamics (2nd Ed.; Princeton: Princeton University Press) de Bruijne, Jos H.J., van der Marel, R.P., & de Zeeuw, P.T. 1996, MNRAS, 282, 909 Buyle, P., Hunter, C., Dejonghe, H. 2007, MNRAS, 375, 773 Carollo, C.M., de Zeeuw, P.T., & van der Marel, R.P. 1995, MNRAS, 276, 1131 Ciotti, L. 1991, A&A, 249, 99 Ciotti, L. 1996, ApJ, 471, 68 (C96) Ciotti, L. 1999, ApJ, 520, 574 (C99) Ciotti, L. 2000, Lecture Notes on Stellar Dynamics (Pisa: Scuola Normale Superiore Ed.) Ciotti, L., & Pellegrini, S. 1992, MNRAS, 255, 561 (CP92) Ciotti, L., & Lanzoni, B. 1997, A&A, 321, 724 Ciotti, L., Lanzoni, B., & Renzini, A. 1996, MNRAS, 282, 1 Ciotti, L., & Bertin, G. 2005, A&A, 437, 419 Ciotti, L., Morganti, L., & de Zeeuw, P.T. 2008, MNRAS in press (arXiv:0809.4444) Cuddeford, P. 1991, MNRAS, 253, 414 Cuesta, A.J., Prada, F., Klypin, A., & Moles, M. 2008, MNRAS, 389, 385 de Vaucouleurs, G. 1948, Ann.d’Ap., 11, 247 Dehnen, W. 1993, MNRAS, 265, 250 Dejonghe, H. 1986, *Phys. Rep.*, 133, No. 3-4, 217 Dejonghe, H. 1987, MNRAS, 224, 13 Eddington, A.S. 1916, MNRAS, 76, 572 Fricke, W. 1952, Astron. Nachr., 280, 193 Hernquist, L. 1990, ApJ, 536, 359 Hiotelis, N. 1994, A&A, 291, 725 Hunter, C., & Qian, E. 1993, MNRAS, 262, 401 King, I. 1966, ApJ, 146, 372 King, I. 1972, ApJl, 174, L123 Lynden Bell, D. 1962, MNRAS, 123, 447 Merritt, D. 1985a, AJ, 90, 1027 Nipoti, C., Londrillo, P., & Ciotti, L. 2002, MNRAS, 332, 901 Nipoti, C., Londrillo, P., & Ciotti, L. 2006, MNRAS, 370, 681 Osipkov, L.P. 1979, Pis’ma Astron.Zh., 5, 77 Richstone, D.O., & Tremaine, S.D. 1984, ApJ, 286, 27 Sersic, J. L. 1968, Atlas de Galaxias Australes, Cordoba, Argentina: Observatorio Astronomico Tremaine, S.D., Richstone, D.O., Byun, Y.I., Dressler, A., Faber, S.M., Grillmair, C., Kormendy, J., & Lauer, T.R. 1994, AJ, 107, 634 Trenti, M, Bertin, G., & van Albada, T.S. 2005, A&A, 433, 57 van Albada, T.S. 1982, MNRAS, 201, 939 Zhao, H. S. 1996, MNRAS, 278, 488
[^1]: Incidentally, this implies that the use of $\rho/\sigma^3$ as a proxy for phase–space density, where $\sigma$ is the local value of the velocity dispersion, has no assignable meaning without an appropriate discussion. For example, in power–law isotropic systems with $\rho\propto r^{-\gamma}$, because the functions $\rho/\sigma^3\propto r^{\frac{\gamma-6}{2}}$ (for $\gamma>1$) and phase–space density $f(\eps)\propto\eps^{\frac{6-\gamma}{2(\gamma-2)}}$ (for $\gamma>2$, where $\eps=-E$ is the so–called relative energy, see Sect. 4 and Baes et al. 2005) are both power laws with respect to their arguments, the exponents are related in a simple way. However, the converse statement is not true: for example, in the Plummer (1911) sphere $f(\eps)\propto\eps^{7/2}$ is a power law at all energies, but $\rho/\sigma^3$ is not a power law of radius.
[^2]: In C99 and Ciotti (2000) it is erroneously stated that the model is inconsistent if $\fan <0$ $\forall Q\in A_-$. All the results presented therein are however correct.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Hancheng Guo, Jie Xiong and Jiayu Zheng[^1]'
title: '**Stochastic maximum principle for generalized mean-field delay control problem [^2]**'
---
**Abstract.** In this paper, we first give the existence and uniqueness theorems for generalized mean-filed delay stochastic differential equations (GMFDSDEs) and mean-field anticipated backward stochastic differential equations (MFABSDEs). Then we study the stochastic maximum principle for generalized mean-filed delay control problem. Since the state is distribution-depending, we define the adjoint equation as a MFABSDE, in which, all the derivatives of coefficients are in Fréchet sense. We deduce the stochastic maximum principle, and also obtain, under some additional assumptions, a sufficient condition for the optimality of the control.
**Keyword.** Existence and uniqueness, Stochastic maximum principle, mean-filed control problem, McKean-Vlasov equation, Fréchet derivative.\
**AMS subject classifications.** 93E20, 93E03, 60H10, 60H30
In this paper we discuss new types of differential equatios which we call mean-field anticipated backward stochastic differential equations (MFABSDEs): $$\begin{aligned}
\label{eqc2}
\begin{cases}
-dY_t = \EE^{'}[f(t, Y'_t, Z'_t, Y'_{t+\delta(t)}, Z'_{t+\zeta(t)}, Y_t, Z_t, Y_{t+\delta(t)}, Z_{t+\zeta(t)})]dt - Z_t dB_t , \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y_t = \xi_t, Z_t = \eta_t , \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [T, T+K],
\end{cases}\end{aligned}$$ where $B$ is a $d$-dimensional Brownian motion, $K$ is a constant. Precise assumptions on the coefficient $f$ and the definition of $\EE'$ are given in the following sections.
Actually, the above MFABSDE is inspired by the mean-field BSDEs $$\begin{aligned}
\label{eqq3}
Y_t = \xi_T + \int_t^T \EE^{'} \[f(s, Y_s, Z_s, Y_s^{'}, Z_s^{'})\]ds - \int_t^T Z_s dB_s, \ \ \ \ \ t \in [0, T]\end{aligned}$$ that is studied by [@BLP2009] and the anticipated BSDEs $$\begin{aligned}
\label{eqq4}
\begin{cases}
Y_t = \xi + \int_t^T f(s, Y_s, Z_s, Y_{s + \delta(s)}, Z_{s + \zeta(s)})ds - \int_t^T Z_s dB_s, \ \ \ \ t \in [0,T];
\cr
Y_t = \xi_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K];
\cr
Z_t = \eta_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K].
\end{cases}\end{aligned}$$ which is investigated by [@PY].
We consider the stochastic maximum principle for a generalized mean-field delay control problem, whose state equation is defined as $$\begin{aligned}
\label{2eq11}
\left\{
\begin{array}{l}
dX^v_t=b(t,X^v_t,X^v_{t-\delta}, P_ {{X^v_t}},P_ {{X^v_{t-\delta}}},v_t,v_{t-\delta}) dt+\sigma dB_t, \ \ \ t\in[0,T],\\
X^v_t=\xi_t, \ v_t=\eta_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t\in[-\delta,0],
\end{array}
\right.\end{aligned}$$ where $P_ {{X}}$ is the law of ${X}$, $\delta\in [0,T]$, $\sigma$ has the same structure as $b$, and the cost functional is defined as: $$\label{2eq12}
\begin{aligned}
J(v)=\mathbb{E}\left\{ \int^T_0 h(t,X^v_t,P_ {{X^v_t}},v_t,v_{t-\delta})dt+\Phi(X^v_T,P_ {{X^v_T}})\right\}.
\end{aligned}$$ The agent wishes to minimize his cost functional $J(v)$. Namely, an admissible control $u\in \mathcal{U}$ is said to be optimal if $$J(u)=\min_{v\in \mathcal{U}}J(v).$$ About stochastic maximum principle (SMP), some pioneering works have been done by Pontryagin et al. [@P1962], they obtained the Pontryagin’s maximum principle by using “spike variation”. Kushner [@K1965] [@K1972] studied the SMP in the framework when the diffusion coefficient does not depend on the control variable, and the cost functional consists of terminal cost only. Haussmann [@H1986] gave a version of SMP when the diffusion of the state does not depend on control item. Arkin and Saksonov [@AS1979], Bensoussan [@B1982] and Bismut [@B1978], proved different versions of SMP under various setups.\
Pardoux and Peng [@PP1990] introduced non-linear backward stochastic differential equations (BSDE) in 1990. They showed that under appropriate assumptions, BSDE admits an unique adapted solution, and the associated comparison theorem holds. An SMP was obtained by Peng [@P1990] in the same year. In that paper, first and second order variational inequalities are introduced, when the control domain need not to be convex, and the diffusion coefficient contains the control variable. The authors of [@BLP2009] obtained mean-field BSDE in a natural way as the limit of some high dimensional system of forward and backward stochastic differential equations. Li [@L2012] studied SMP for mean-filed controls, when the domain of the control is assumed to be convex. Under some additional assumptions, both necessary and sufficient conditions for the optimality of a control were proved.\
Buckdahn et al. [@BDL2011] considered an SMP for SDEs of mean-field type control problem when the coefficients depend on the state of the solution process as well as on its expected value. Moreover, the cost functional is also of mean-field type. Their system is defined as follows: $$\left\{
\begin{array}{l}
dX_t=b(t,X_t,\mathbb{E}[{{X_t}}],v_t) dt
+\sigma(t,X_t,\mathbb{E}[{{X_t}}],v_t) dB_t,\\
X_0=x.
\end{array}
\right.$$ and the cost/payoff functional is defined by: $$\begin{split}
\begin{aligned}
J(v)=\mathbb{E}\left\{ \int^T_0 h(t,X_t,\mathbb{E}[{{X_t}}],v_t)dt+\Phi(X_T,\mathbb{E}[{{X_T}}])\right\},
\end{aligned}
\end{split}$$ An SMP is derived, specifying the necessary conditions for the optimality. This maximum principle differs from the classical one in the sense that here the first order adjoint equation turns out to be a linear mean-field backward SDE, while the second order adjoint equation remains the same as in Peng’s SMP. About stochastic delay control problem, Chen and Wu [@CW2010] obtain the maximum principle for the optimal control of this problem by virtue of the duality method and the anticipated backward stochastic differential equations. The Authors of [@DHQ2013] develop this theory into classical mean-field type, which means the coefficients of the state depend on the expectation.\
Buckdahn et al. [@BLP2014] studied generalized mean-field stochastic differential equations and the associated partial differential equations (PDEs). “Generalized" means the coefficients depend on both the state process and its law. They proved that under appropriate regularity conditions on the coefficients, the SDE has the unique classical solution. In this paper, we study the optimal control when the state equation is in the controlled generalized mean-filed form.\
Preliminaries
=============
In this section, for the convenience of the reader, we state some results of Buckdahn et al. [@BLP2014] without proofs, which will be used in present work.
Let $(\Omega,{\cal{F}},P)$ be a probability space with filtration ${\cal{F}}_{t}$. Suppose that ${B}_t$ is a Brownian motion belongs to $(\Omega,{\cal{F}},P)$, where ${\mathcal{F}}$ is the filtration generated by ${B}_t$, and augmented by all $P$-null sets. Let $\mathcal{P}_2(\mathbb{R}^n)$ be the collection of all square integrable probability measures over $(\mathbb{R}^n,\mathcal{B}(\mathbb{R}^n))$, endowed with the 2-Wasserstein metric $W_2$, which is defined as $$W_2(P_{\mu },P_{\nu})=\inf\left\{ \left(\mathbb{E}[|\mu'-\nu'|^2]\right)^{\frac{1}{2}}\right \},$$ for all $\mu',\nu' \in L^2(\mathcal{F}_0;\mathbb{R}^d)$ with $P_{\mu'}=P_{\mu}, \ P_{\nu'}=P_{\nu}.$ Now let us introduce the following spaces: $$\begin{aligned}
L^p(\Omega, \mathcal{F}_T, P;\mathbb{R}^n)=\left\{\xi:\mathbb{R}^n\text{-valued} \ \mathcal{F}_T\text{-measurable\ r.v.} ; \mathbb{E}\left[|\xi|^p\right]<+\infty\right\},\end{aligned}$$ $$\begin{aligned}
L^0(\Omega, \mathcal{F}, P;\mathbb{R}^n)=\left\{\xi:\mathbb{R}^n\text{-valued} \ \mathcal{F}\text{-measurable\ random variables} \right\},\end{aligned}$$ $$\begin{aligned}
H^2_{\mathcal{F}}(s, r;\mathbb{R}^n)=&\bigg\{&({\varphi}_t)_{s\leq t\leq r}: \mathbb{R}^n\text{-valued} \ \mathcal{F}_t\text{-adapted stochastic process} ;\\
&&\mathbb{E}\left[\int^r_s |{\varphi}_t|^2dt\right]<+\infty\bigg\},\end{aligned}$$ $$\begin{aligned}
S^2_{\mathcal{F}}(s, r;\mathbb{R}^n)=&\bigg\{&({\varphi}_t)_{s\leq t\leq r}: \mathbb{R}^n\text{-valued} \ \mathcal{F}_t\text{-adapted stochastic process} ;\\
&&\mathbb{E}\left[\sup_{s\leq t\leq r} |{\varphi}_t|^2\right]<+\infty\bigg\}.\end{aligned}$$ $\mathcal{U}=H^2_{\mathcal{F}}(0,T;U)$ denotes the set of admissible controls of the following form: $$\begin{aligned}
v_t=\left\{
\begin{array}{l}
v_t\in H^2_{\mathcal{F}}(0, T;\mathbb{R}^n), \ t\in[0,T],\\
\gamma_t,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t\in[-\delta,0],
\end{array}
\right.\end{aligned}$$ where, $\gamma$ is square integrable on $[-\delta,0]$, $U$ is supposed to be a convex subset of $\mathbb{R}^k$. Given $b:[0,T]\times \mathbb{R}^n\times \mathbb{R}^n\times \ \mathcal{P}_2(\mathbb{R}^n)\times \ \mathcal{P}_2(\mathbb{R}^n) \times U\times U \longrightarrow \mathbb{R}^n,\ \sigma:[0,T]\times \mathbb{R}^n\times \mathbb{R}^n\times \ \mathcal{P}_2(\mathbb{R}^n)\times \ \mathcal{P}_2(\mathbb{R}^n) \times U\times U \longrightarrow \mathbb{R}^{n\times d}, \ h:[0,T]\times \mathbb{R}^n\times \ \mathcal{P}_2(\mathbb{R}^n) \times U \times U\longrightarrow \mathbb{R}, \ \Phi:\mathbb{R}^n\times \mathcal{P}_2(\mathbb{R}^n) \longrightarrow \mathbb{R}.$
About the deriavative with respect to measure, the following definition is taken from Cardaliaguet [@C2013].
A function $f: \mathcal{P}_2(\mathbb{R}^n)\longrightarrow \mathbb{R} $ is said to be differentiable in $\mu \in \mathcal{P}_2(\mathbb{R}^n)$ if, the function ${\widetilde}{f}: L^2(\mathcal{F};\mathbb{R}^n)\longrightarrow \mathbb{R}$ given by ${\widetilde}{f}(\mathfrak{v})=f(P_{\mathfrak{v}})$ is differentiable (in Fréchet sense) at $\mathfrak{v}_0$, defined by $P_{\mathfrak{v}_0}=\mu$, i.e. there exists a linear continuous mapping $D\title{f}(\mathfrak{v}_0):L^2( \mathcal{F} ;\mathbb{R}^n)\longrightarrow \mathbb{R},$ such that $${\widetilde}{f}(\mathfrak{v}_0+\eta)-{\widetilde}{f}(\mathfrak{v}_0)=D{\widetilde}{f}(\mathfrak{v}_0)(\eta)+o(|\eta|_{L^2}),$$ with $|\eta|_{L^2}\longrightarrow 0$ for $\eta \in L^2( \mathcal{F} ;\mathbb{R}^n). $
According to Riesz’ Representation Theorem, there exists a unique random variable $\theta_0\in L^2( \mathcal{F} ;\mathbb{R}^n)$ such that $D{\widetilde}{f}(\mathfrak{v}_0)(\eta)=(\theta_0,\eta)_{L^2}=\mathbb{E}[\theta_0\eta]$, for all $\eta\in L^2( \mathcal{F} ;\mathbb{R}^n).$ In [@C2013] it has been proved that there is a Borel function $h_0:\mathbb{R}^d\longrightarrow\mathbb{R}^d$ such that $\theta_0=h_0(\mathfrak{v}_0),\ \ \ a.s..$ Then, $$f(P_{\mathfrak{v}})-f(P_{\mathfrak{v}_0})=\mathbb{E}[h_0(\mathfrak{v}_0)(\mathfrak{v}-\mathfrak{v}_0)]+o(|\mathfrak{v}-\mathfrak{v}_0|_{L^2}),$$ $\mathfrak{v}\in L^2( \mathcal{F} ;\mathbb{R}^n).$\
We call $\partial_{\mu} f(\mu ,y):=h_0(y), \ y\in \mathbb{R}^n $, the derivative of $f: \mathcal{P}_2(\mathbb{R}^n)\longrightarrow \mathbb{R}^n$ at $\mu.$
For mean-field type SDE and BSDE, we introduce the following notations. Let $(\Omega',{\cal{F}}',P')$ be a copy of the probability space $(\Omega,{\cal{F}},P)$. For each random variable $\xi$ over $(\Omega,{\cal{F}},P)$ we denote by $\xi'$ a copy of $\xi$ defined over $(\Omega',{\cal{F}}',P')$. $\mathbb{E}'[\cdot]=\int_{\Omega'}(\cdot)dP'$ acts only over the variables from $(\Omega',{\cal{F}}',P')$.
Recall that for 2-Wasserstein metric $W_2(\cdot,\cdot),$ we have, $$W_2(P_{\mu },P_{\nu})=\inf\{ (\mathbb{E}[|\mu'-\nu'|^2])^{\frac{1}{2}} \},$$ for all $\mu',\nu' \in L^2(\mathcal{F}_0;\mathbb{R}^d)$ with $P_{\mu'}=P_{\mu}, \ P_{\nu'}=P_{\nu}.$
We say that $f\in C^{1,1}_b(\mathcal{P}_2(\mathbb{R}^d))$ (continuously differentiable over $\mathcal{P}_2(\mathbb{R}^d)$ with Lipschitz-continuous bounded derivative), if for all $\mathfrak{v}\in L^2(\mathcal{F},\mathbb{R}^d)$, there exists a $P_{\mathfrak{v}}$-modification of $\partial_{\mu }f(P_{\mathfrak{v}},\cdot)$, again denote by $\partial_{\mu }f(P_{\mathfrak{v}},\cdot)$, such that $\partial_{\mu }f:\mathcal{P}_2(\mathbb{R}^d)\times \mathbb{R}^d\longrightarrow \mathbb{R}^d$ is bounded and Lipschitz continuous, i.e., there is some real constant $C$ such that $$\begin{aligned}
i)\ &|\partial_{\mu }f(\mu ,x)|\leq C, \mu \in\mathcal{P}_2( \mathbb{R}^d), x\in \mathbb{R}^d,&\\
ii)\ &|\partial_{\mu }f(\mu ,x)-\partial_{\mu }f(\mu' ,x')|\leq C(W_2(\mu,\mu')+|x-x'|), \mu, \mu' \in\mathcal{P}_2( \mathbb{R}^d),&
\end{aligned}$$ $ x,x'\in \mathbb{R}^d.$ We consider this function $\partial_{\mu }f$ as the derivative of $f$.
Let us now consider a complete probability space $(\Omega,\mathcal{F},P)$ on which, we define a $d$-dimensional Brownian motion $B=(B^1,\cdots,B^d)=(B_t)_{t\in [0,T]}$, and $T\geq 0$ denotes an arbitrarily fixed time horizon. We make the following assumptions:\
There is a sub-$\sigma$-field $\mathcal{F}_0\subset\mathcal{F}$ such that\
i) the Brownian motion $B$ is independent of $\mathcal{F}_0$, and\
ii) $\mathcal{F}_0$ is “rich enough", i.e., $\mathcal{P}_2(\mathbb{R}^d)=\{ P_{\mathfrak{v}}, \mathfrak{v}\in L^2(\mathcal{F}_0; \mathbb{R}^d) \}.$\
By $\mathbb{F}=(\mathcal{F}_t)_{t\in [0,T]}$ we denote the filtration generated by $B$, completed and augmented by $\mathcal{F}_0$.\
Given deterministic Lipschitz functions $\sigma: \mathbb{R}^d\times\mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^{d\times d}$ and $b: \mathbb{R}^d\times\mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^{d},$ we consider for the initial state $(t,x)\in [0,T]\times \mathbb{R}^d$ and $\xi\in L^2(\mathcal{F}_t;\mathbb{R}^d)$ the stochastic differential equations (SDEs)
$$\begin{aligned}
X^{t,\xi}_s=\xi+\int^s_t \sigma(X^{t,\xi}_r,P_{X^{t,\xi}_r})dB_r+\int^s_t \sigma(X^{t,\xi}_r,P_{X^{t,\xi}_r})dr,\ s\in[t,T],
\end{aligned}$$
and $$\begin{aligned}
X^{t,x,\xi}_s=x+\int^s_t \sigma(X^{t,x,\xi}_r,P_{X^{t,\xi}_r})dB_r+\int^s_t \sigma(X^{t,x,\xi}_r,P_{X^{t,\xi}_r})dr,\ s\in[t,T].
\end{aligned}$$
We find out that under the assumptions above, both SDEs have a unique solution in $\mathcal{S}^2([t,T];\mathbb{R}^d),$ which is the space of $\mathbb{F}$-adapted continuous processes $Y=(Y_s)_{s\in[t,T]}$ with $\mathbb{E}[\sup_{s\in [t,T]}|Y_s|^2]\leq \infty.$
The couple of coefficient $(\sigma,b)$ belongs to $C^{1,1}_b(\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^{d\times d}\times \mathbb{R}^d),$ i.e., the components $\sigma_{i,j},b_j,\ 1\leq i,j \leq d,$ satisfy the following conditions:\
i) $\sigma_{i,j}(x,\cdot), b_{j}(x,\cdot)$ belong to $C^{1,1}_b( \mathcal{P}_2(\mathbb{R}^d))$, for all $x\in \mathbb{R}^d$\
ii) $\sigma_{i,j}(\cdot,\mu), b_{j}(\cdot,\mu)$ belong to $C^{1}_b(\mathbb{R}^d)$, for all $\mu\in \mathcal{P}_2(\mathbb{R}^d)$\
iii) The derivatives $\partial_x\sigma_{i,j}, \partial_xb_{j}:\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^d $, $\partial_{\mu}\sigma_{i,j}, \partial_{\mu}b_{j}:\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)\times \mathbb{R}^d\longrightarrow \mathbb{R}^d $, are bounded and Lipschitz continuous.
\[H22\] The couple of coefficient $(\sigma,b)$ belongs to $C^{2,1}_b(\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^{d\times d}\times \mathbb{R}^d),$ i.e., $(\sigma,b)\in C^{1,1}_b(\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)\longrightarrow \mathbb{R}^{d\times d}\times \mathbb{R}^d)$ and the components $\sigma_{i,j},b_j,\ 1\leq i,j \leq d,$ satisfies the following conditions:\
i) $\partial_{x_k}\sigma_{i,j}(\cdot,\cdot), \partial_{x_k} b_{j}(\cdot,\cdot)$ belong to $C^{1,1}_b( \mathbb{R}^d \times\mathcal{P}_2(\mathbb{R}^d))$, for all $1\leq k \leq d;$\
ii) $\partial_{\mu }\sigma_{i,j}(\cdot,\cdot,\cdot), \partial_{\mu }b_{j}(\cdot, \cdot,\cdot)$ belong to $C^{1,1}_b(\mathbb{R}^d \times \mathcal{P}_2(\mathbb{R}^d)\times \mathbb{R}^d)$, for all $\mu\in \mathcal{P}_2(\mathbb{R}^d)$\
iii) All the derivatives of $\sigma_{i,j}, b_{j}$, up to order 2 are bounded and Lipschitz continuous.
The following theorem is taken from [@BLP2014]. It gives the Itô’s formula related to a probability measure.
Let $\Phi \in C^{2,1}_b(\mathbb{R}^d\times \mathcal{P}_2(\mathbb{R}^d)).$ Then, under Hypothesis \[H22\], for all $0\leq t \leq s \leq T, x\in \mathbb{R}^d, \xi\in L^2(\mathcal{F}_t;\mathbb{R}^d)$ the Itô formula is satisfied as follow: $$\begin{aligned}
&&\Phi(X^{t,x,P_{\xi}}_s, P_{X^{t,\xi}_s})-\Phi(x,P_{\xi})\nonumber\\
&=&\int^s_t\big(\sum^d_{i=1}\partial_{x_i}\Phi(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})b_i(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})\nonumber\\
&&+\frac{1}{2} \sum^d_{i,j,k=1} \partial^2_{x_i,x_j}\Phi(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})(\sigma_{i,k}\sigma_{j,k})(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})\nonumber\\
&&+\mathbb{E}'\big[\sum^d_{i=1}(\partial_{\mu }\Phi)_i(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r},(X^{t,{\xi}}_r)')b_i((X^{t,{\xi}}_r)', P_{X^{t,\xi}_r})\nonumber\\
&&+\frac{1}{2} \sum^d_{i,j,k=1} \partial_{y_i}((\partial_{\mu }\Phi)_j(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r},(X^{t,{\xi}}_r)')(\sigma_{i,k}\sigma_{j,k})((X^{t,{\xi}}_r)', P_{X^{t,\xi}_r})\big]\big)dr\nonumber\\
&&+\int^s_t \sum^d_{i,j=1}\partial_{x_i}\Phi(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})\sigma_{i,j}(X^{t,x,P_{\xi}}_r, P_{X^{t,\xi}_r})dB^j_r, \ s\in [t,T].\end{aligned}$$
For mean-field type SDE and BSDE, we have still to introduce some notations. Let $({\Omega}',{\mathcal{F}}',{P}')$ be a copy of the probability space $(\Omega,{\mathcal{F}},P)$. For any random variable $\xi$ over $(\Omega,{\mathcal{F}},P)$, we denote by ${\xi}'$ its copy on ${\Omega}'$, respectively, which means that they have the same law as $\xi$, but defined over $({\Omega}',{\mathcal{F}}',{P}')$ . $\mathbb{{E}}'[\cdot]=\int_{{\Omega}'}(\cdot)d{P'}$ act only over the variables from ${\omega}'$. About stochastic delay and anticipated differential equations, we would like to introduce the following lemmas for the convenience of the readers. Our Lemma \[le3\] is Lemma 3.1 of Peng [@P2004]. Lemma \[le1\], which is Theorem 3.1 of Buckdahn [@BLP2009] , is a fundamental result of mean-filed BSDEs: an existence and uniqueness theorem. Lemma \[le2\] is the comparison theorem for solutions of mean-filed BSDEs that can be found in Buckdahn [@BLP2009].
\[le3\] For a fixed $\xi \in L^2(\mathscr{F}_T)$ and $g_0(\cdot)$ which is an $\mathscr{F}_t$-adapted process satisfying $\EE[(\int_0^T | g_0(t) |dt)^2] <\infty$, there exists a unique pair of processes $(y., z.) \in H^2_{\mathscr{F}}(0, T; \mathbb{R}^{1+d})$ satisfying the following BSDE: $$\begin{aligned}
y_t = \xi + \int_t^T g_0(s) ds - \int_t^T z_s dW_s, \quad t \in [0, T].\end{aligned}$$ If $g_0(\cdot) \in L_{\mathscr{F}^2(0, T)}$, then $(y., z.) \in S^2_{\mathscr{F}}(0, T) \times H^2_{\mathscr{F}}(0, T; \mathbb{R}^d).$ We have the following basic estimate: $$\begin{aligned}
\label{estimates1}
|y_t|^2 &+ \EE^{\mathscr{F}_t} \[ \int_t^T \( \frac{\beta}{2} |y_s|^2 + |z_s|^2 \) e^{\beta(s-t)} ds \] \nonumber \\
& \le \EE^{\mathscr{F}_t} \[ |\xi|^2 e^{\beta (T- t)} \] + \frac{2}{\beta} \EE^{\mathscr{F}_t} \[ \int_t^T |g_0(s)|^2 e^{\beta(s-t)} ds \]\end{aligned}$$ In particular, $$\begin{aligned}
\label{eatimates2}
|y_0|^2 &+ \EE \[ \int_0^T \( \frac{\beta}{2} |y_s|^2 + |z_s|^2 \) e^{\beta s} ds \] \nonumber \\
& \le \EE \[ |\xi|^2 e^{\beta T} \] + \frac{2}{\beta} \EE \[ \int_0^T |g_0(s)|^2 e^{\beta s} ds \]\end{aligned}$$ where $\beta >0$ is an arbitrary constant.
The following is a foundamental result for the existence of a unique solution to mean-filed BSDEs due to Buckdahn [@BLP2009].(Theorem 3.1)
\[le1\] Under the assumptions (H1) and (H2) of [@PY], and $\delta$, $\zeta$ satisfy (C1) and (C2). Then for any terminal conditions $\xi \in S^2_{\mathscr{F}}(T, T+K; \RR^m)$ and $\eta \in L^2_{\mathscr{F}}(T, T+K; \RR^{m \times d})$, the anticipated BSDE (\[eqq4\]) has a unique adapted solution $$\begin{aligned}
(Y_t, Z_t) \in \mathcal{S}^2_{\mathscr{F}}(0, T+K; \RR^m) \times H^2_{\mathscr{F}}(0, T+K; \RR^{m\times d}). \end{aligned}$$
We emphasize that, due to our notation, the driving coefficient of (\[eqq3\]) has to be interpreted as followings: $$\begin{aligned}
\mathbb{E}^{'}[f(s, Y_s^{'}, Z_s^{'}, Y_s, Z_s)](\omega) & = \mathbb{E}^{'}[f(s, Y_s^{'}, Z_s^{'}, Y_s(\omega), Z_s(\omega))] \nonumber \\
&= \int_{\Omega} f(\omega^{'}, \omega, s, Y_s^{'}(\omega^{'}), Z_s^{'}(\omega^{'}), Y_s(\omega), Z_s(\omega)) P(d \omega^{'}).\end{aligned}$$
The proof of the following comparison theorem for mean filed BSDE can be found in Buckdahn [@BLP2009].
\[le2\] Let ${\overline}{f}_{i}(t, y, z, y', z'), i=1, 2,$ be two drivers of mean-filed BSDEs satisfying the the assumptions (A3) and (A4) of [@BLP2009]. Moreover, suppose:
(i) One of the two coefficients is independent of $z'$.
(ii) One of the two coefficients is nondecreasing in $y'$.
Let $\xi_1, \xi_2 \in L^2(\Omega, \mathscr{F}_T, P)$ and denote by $(Y^1, Z^1)$ and $(Y^2, Z^2)$ the solution of the mean-field BSDE (\[eqq3\]) with data $(\xi_1, f_1)$ and $(\xi_2, f_2)$, respectively. Then of $\xi_1 \ge \xi_2,$ P-a.s., and $f_1 \le f_2,$ ${\overline}{P}$-a.s., it holds that also $Y_t^1 \le Y_t^2, t \in [0, T], $ P- a.s.
Basic properties of GMFDSDE and MFABSDE
=======================================
Existence and uniqueness theorems
---------------------------------
Consider equation (\[eqc2\]), where $\delta(\cdot)$ and $\zeta(\cdot)$ are two $\mathbb{R}^{+}$-valued continuous functions defined on $[0,T]$ such that:
(C1) There exists a constant $K \ge 0$ such that, for all $s \in [0, T]$, $$s+\delta(s) \le T+K; \quad s+ \zeta(s) \le T+K,$$ (C2) There exists a constant $L \ge 0$ such that, for all $t \in [0, T]$ and for all nonnegative and integrable $g(\cdot)$, $$\int_t^T g(s+\delta(s))ds \le L \int_t^{T+K} g(s) ds;$$ $$\int_t^T g(s+\zeta(s))ds \le L \int_t^{T+K} g(s) ds.$$
The setting of our problem is as follows: to find a pair of $\mathscr{F}$-adapted processes $(Y_., Z_.) \in S^2_{\mathscr{F}}(0, T+K; \mathbb{R}^m) \times H^2_{\mathscr{F}}(0, T+K; \mathbb{R}^{m \times d}) $ satisfying MFABSDE (\[eqc2\]).
Assume that for all $s\in [0,T]$, $f(s, \cdot): \mathbb{R}^m \times \mathbb{R}^{m \times d} \times H^2(\mathscr{F}_r; \mathbb{R}^m) \times H^2(\mathscr{F}_{{\overline}{r}}; \mathbb{R}^{m\times d} ) \times \mathbb{R}^m \times \mathbb{R}^{m \times d} \times H^2(\mathscr{F}_r; \mathbb{R}^m) \times H^2(\mathscr{F}_{{\overline}{r}}; \mathbb{R}^{m\times d}) \longrightarrow H^2(\mathscr{F}_s, \mathbb{R}^m)$, where $r, {\overline}{r} \in [s, T+K]$, and $f$ satisfies the following conditions:
(C3) There exists a constant $C >0$, such that for all $s \in [0,T], y_1, y_2, y'_1, y'_2 \in \mathbb{R}^m, z_1, z_2, z'_1, z'_2 \in \mathbb{R}^{m \times d}, \theta_{.,1}, \theta_{.,2}, \theta'_{.,1}, \theta'_{.,2} \in L^2_{\mathscr{F}}(s, T+K; \mathbb{R}^m), \gamma_{.,1}, \gamma_{.,2}, \gamma'_{.,1}, \gamma'_{.,2} \in L^2_{\mathscr{F}}(s, T+K; \mathbb{R}^{m \times d}), r, {\overline}{r} \in [s, T+K]$, we have $$\begin{aligned}
& |f(t, y_1, z_1,\theta_{r,1}, \gamma_{{\overline}{r},1}, y'_1, z'_1, \theta'_{r,1}, \gamma'_{{\overline}{r},1})
- f(t, y_2, z_2,\theta_{r,2}, \gamma_{{\overline}{r},2}, y'_2, z'_2, \theta'_{r,2}, \gamma'_{{\overline}{r},2}) | \nonumber \\
&\le C \[ |y_1 - y_2| + |z_1 - z_2| + \EE^{\mathscr{F}_s}\(|\theta_{r,1} - \theta_{r,2}| + |\gamma_{{\overline}{r},1} - \gamma_{{\overline}{r},2}|\) \right. \nonumber \\
&+ |y'_1 - y'_2| + |z'_1 - z'_2| + \EE^{\mathscr{F}_s}\( |\theta'_{r,1} - \theta'_{r,2}| + |\gamma'_{{\overline}{r},1} - \gamma'_{{\overline}{r},2}| \) \left. \];\end{aligned}$$
(C4) $\EE[\int_0^T |f(s, 0,0,0,0,0,0,0,0)|^2 ds] <\infty.$
Note that $f(s, \cdot, \cdot,\cdot,\cdot,\cdot,\cdot,\cdot,\cdot)$ is $\mathscr{F}_s$-measurable ensures that the solution to the mean-field anticipated BSDE is $\mathscr{F}_s$-adapted.
The following is the main result of this section: Two existence and uniqueness theorems for MFABSDEs and GMFDSDE, respectively.
\[unique\] Suppose that $f$ satiesfies (C3) and (C4), and $\delta$, $\zeta$ satisfy (C1) and (C2). Then for any given terminal conditions $\xi_. \in S^2_{\mathscr{F}}(T, T+K; \mathbb{R}^m)$ and $\eta_. \in H^2_{\mathscr{F}}(T, T+K; \mathbb{R}^{m \times d})$, the mean field anticipated BSDE (\[eqc2\]) has a unique solution, that is, there exists a unique pair of $\mathscr{F}_t$-adapted processes $(Y_., Z_.) \in S^2_{\mathscr{F}}(0, T+K; \mathbb{R}^m) \times H^2_{\mathscr{F}}(0, T+K; \mathbb{R}^{m \times d})$ satisfying (\[eqc2\]).
We first introduce a norm on the space $H^2_{\mathscr{F}}(0, T+K; \RR^{m}\times \RR^{m\times d})$ which is equivalent to the canonical norm: $$\begin{aligned}
\|v(\cdot)\|_{\beta} = \{\EE\int_0^{T+K} |v_s|^2 e^{\beta s } ds\}^{\frac{1}{2}}, \quad \beta >0 .\end{aligned}$$ The parameter $\beta$ will be specified later.
For any $(y,z) \in H^2_{\mathscr{F}}(0, T; \mathbb{R}^m \times \mathbb{R}^{m\times d})$, from Lemma \[le1\], there exists a unique solution $(Y, Z) \in S^2_{\mathscr{F}}(0, T+K; \mathbb{R}^m) \times H^2_{\mathscr{F}}(0, T+K; \mathbb{R}^{m \times d})$ to the following anticipated BSDE: $$\begin{aligned}
\label{eq3}
Y_t = \xi &+ \int_t^T \EE^{'}\[f(s, y'_s, z'_s, y'_{s+\delta(s)}, z'_{s+\zeta(s)}, Y_s, Z_s, Y_{s+\delta(s)}, Z_{s+\zeta(s)})\]ds \nonumber \\
& - \int_t^T Z_s dB_s, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0, T]. \end{aligned}$$
Define a mapping $h:H^2_{\mathscr{F}}(0, T+K; \RR^{m}\times\RR^{m \times d}) \longrightarrow H^2_{\mathscr{F}}(0, T+K; \RR^{m}\times\RR^{m \times d})$ such that $h[(y_., z_.)] = (Y_., Z_.)$. Now we prove that $h$ is a contraction mapping under the norm $\| \cdot \|_{\beta}$. For two arbitrary elements $(y_{.,1} , z_{.,1})$ and $(y_{.,2} , z_{.,2})$ in $H^2_{\mathscr{F}}(0, T+K; \RR^{m}\times\RR^{m \times d})$, set $(Y_{.,1} , Z_{.,1}) = h[(y_{.,1} , z_{.,1})]$ and $(Y_{.,2} , Z_{.,2}) = h[(y_{.,2} , z_{.,2})]$. Denote their diferences by $$(\hat{y}_{.}, \hat{z}_{.}) = ((y_1 - y_2)_{.}, (z_1 - z_2)_{.}), \quad (\hat{Y}_{.}, \hat{Z}_{.}) = ((Y_1 - Y_2)_{.}, (Z_1 - Z_2)_{.}).$$ Then, by appling Ito’s formula to $e^{\beta s} |\hat{Y}_s|^2$ and by Fubini Theorem, we get $$\begin{aligned}
& \EE \[\int_0^T \beta e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds \] + \EE\[\int_0^T e^{\beta s}|Z_{s,1} - Z_{s,2}|^2 ds\] \nonumber \\
=& \EE \int_0^T 2e^{\beta s}(Y_{s,1} - Y_{s,2}) \EE^{'}\[f\(s, y'_{1,s}, z'_{1,s}, y'_{1,s+\delta(s)}, z'_{1,s+\zeta(s)}, Y_{1,s}, Z_{1,s}, Y_{1,s+\delta(s)}, \right. \right. \nonumber\\
&Z_{1,s+\zeta(s)} \left. \) - f(s, y'_{2,s}, z'_{2,s}, y'_{2,s+\delta(s)}, z'_{2,s+\zeta(s)}, Y_{2,s}, Z_{2,s}, Y_{2,s+\delta(s)}, Z_{2,s+\zeta(s)}) \left. \] ds \nonumber \\
\le& C \EE \int_0^T 2e^{\beta s} |Y_{s,1} - Y_{s,2}| \( |y'_{s,1} - y'_{s,2}| + |z'_{s,1} - z'_{s,2}| + |Y_{s,1} - Y_{s,2}| \right. \nonumber \\
& + |Z_{s,1} - Z_{s,2}| + \EE^{\mathscr{F}_s}|y'_{s+\delta(s),1} - y'_{s+\delta(s),2}| + \EE^{\mathscr{F}_s}|z'_{s+\zeta(s),1} - z'_{s+\zeta(s),2}| \nonumber \\
& + \EE^{\mathscr{F}_s}|Y_{s+\delta(s),1} - Y_{s+\delta(s),2}| + \EE^{\mathscr{F}_s}|Z_{s+\zeta(s),1} - Z_{s+\zeta(s),2}| \left. \) ds \nonumber \\
& \le 2C \EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}||y_{s,1} - y_{s,2}| ds \nonumber \\
&+ 2C \EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}| |z_{s,1} - z_{s,2}| ds \nonumber \\
& + 2C \EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}| |y_{s+\delta(s),1} - y_{s+\delta(s),2}| ds \nonumber \\
&+ 2C \EE\int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}| |Z_{s,1} - Z_{s,2}|ds+ 2C\EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds \nonumber \\
&+ 2C \EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}||z_{s+\zeta(s),1} - z_{s+\zeta(s),2}| ds \nonumber \\
& + 2C\EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}| |Y_{s+\delta(s),1} - Y_{s+\delta(s),2}|ds \nonumber \\
&+ 2C\EE \int_0^{T} e^{\beta s} |Y_{s,1} - Y_{s,2}| |Z_{s+\zeta(s),1} - Z_{s+\zeta(s),2}| ds \nonumber \\
&= (1)+(2)+(3)+(4)+(5)+(6)+(7)+(8).\end{aligned}$$ Since $$\begin{aligned}
(1) &\le C \EE \int_0^{T+K} e^{\beta s} \( \frac{\beta}{8C} |Y_{s,1} -Y_{s,2}|^2 + \frac{8C}{\beta}|y_{s,1} - y_{s,2}|^2 \) ds \nonumber \\
&= \frac{\beta}{8} \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{8C^2}{\beta} \EE \int_0^{T+K} e^{\beta s} |y_{s,1} - y_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(2) &\le C \EE \int_0^{T+K} e^{\beta s} \( \frac{\beta}{8C} |Y_{s,1} -Y_{s,2}|^2 + \frac{8C}{\beta}|z_{s,1} - z_{s,2}|^2 \) ds \nonumber \\
&= \frac{\beta}{8} \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{8C^2}{\beta} \EE \int_0^{T+K} e^{\beta s} |z_{s,1} - z_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(3) &\le C \EE \int_0^{T+K} e^{\beta s} \( \frac{\beta}{8C} |Y_{s,1} -Y_{s,2}|^2 + \frac{8C}{\beta} L \cdot |y_{s,1} - y_{s,2}|^2 \) ds \nonumber \\
&= \frac{\beta}{8} \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{8C^2 L}{\beta} \EE \int_0^{T+K} e^{\beta s} |y_{s,1} - y_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(4) &\le C \EE \int_0^{T+K} e^{\beta s} \( 4C |Y_{s,1} -Y_{s,2}|^2 + \frac{1}{4C}|Z_{s,1} - Z_{s,2}|^2 \) ds \nonumber \\
&= 4C^2 \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{1}{4} \EE \int_0^{T+K} e^{\beta s} |Z_{s,1} - Z_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(6) &\le C \EE \int_0^{T+K} e^{\beta s} \( \frac{\beta}{8C} |Y_{s,1} -Y_{s,2}|^2 + \frac{8CL}{\beta}|z_{s,1} - z_{s,2}|^2 \) ds \nonumber \\
&= \frac{\beta}{8} \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{8C^2L}{\beta} \EE \int_0^{T+K} e^{\beta s} |z_{s,1} - z_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(7) & \le C \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + CL \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds \nonumber \\
& = C(1+L) \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds,\end{aligned}$$ $$\begin{aligned}
(8) &\le C \EE \int_0^{T} e^{\beta s} \( 4CL |Y_{s,1} -Y_{s,2}|^2 + \frac{1}{4CL}|Z_{s +\zeta(s),1} - Z_{s +\zeta(s),2}|^2 \) ds \nonumber \\
& \le 4C^2L \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{1}{4L} \EE \int_0^{T+K} e^{\beta s} L \cdot |Z_{s,1} - Z_{s,2}|^2 ds \nonumber \\
& = 4C^2L \EE \int_0^{T+K} e^{\beta s} |Y_{s,1} - Y_{s,2}|^2 ds + \frac{1}{4} \EE \int_0^{T+K} e^{\beta s} |Z_{s,1} - Z_{s,2}|^2 ds,\end{aligned}$$ consequently, $$\begin{aligned}
& \( \frac{\beta}{2}- 2C - 4C^2 - C(1+L) - 4C^2L \)\EE \[\int_0^{T+K} e^{\beta s} |\hat{Y}_s|^2 ds \] \nonumber \\
&\ \ \ \ \ + \frac{1}{2}\EE\[\int_0^{T+K} e^{\beta s}|\hat{Z}_s|^2 ds\] \le \frac{8C^2(L+1)}{\beta} \EE \[\int_0^{T+K} e^{\beta s} \( |\hat{y}_s|^2 + |\hat{z}_s|^2\) ds \].\end{aligned}$$ We choose $\beta = 32C^2L + 32C^2 + 6C + 2CL +1$, such that $$\begin{aligned}
& \EE \[\int_0^{T+K}\( |\hat{Y}_s |^2 + |\hat{Z}_s |^2 \) e^{\beta s} ds \]
\le \frac{1}{2} \EE \[\int_0^{T+K} \( |\hat{y}_s|^2 + |\hat{z}_s|^2\) e^{\beta s} ds\].\end{aligned}$$ Thus, $h$ is a contraction, and hence the conclusions of the theorem follows from Schauder’s fixed point theorem.
In addition to the above existence and uniqueness thoerem for (\[eqc2\]), we need to prove the same theorem for the following generalized mean-field delay stochastic differential equations(GMFDSDE): $$\begin{aligned}
\label{MFDSDE}
\begin{cases}
dX_t =b\(t, X_t, X_{t-\delta}, P_{X_t}, P_{X_{t-\delta}}\)dt +\sigma \(t, X_t, X_{t-\delta}, P_{X_t}, P_{X_{t-\delta}}\) dB_t , \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
X_t = \xi_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [-\delta, 0)
\end{cases}\end{aligned}$$ which will be applied to the control problem.
Assume that for all $t \in [-\delta, T]$, $b: [-\delta, T] \times \RR^m \times \RR^m \times \mathcal{P}_2(\RR^m)\times \mathcal{P}_2(\RR^m)\to H^2(-\delta, T; \RR^m)$, satisfies the following conditions:
(C5) There exists a constant $C \ge 0$, such that for all $t \in [-\delta, T]$, $x, x', x_{\delta}, x'_{\delta} \in \RR^{m}$, $\mu, \mu_{\delta}, \mu', \mu'_{\delta} \in \mathcal{P}_2(\RR^m)$, we have $$\begin{aligned}
& |b\(t, x, x_{\delta}, \mu, \mu_{\delta}\)- b\(t, x', x'_{\delta}, \mu', \mu'_{\delta}\)| \\
& \le C\(|x - x'| + |x_{\delta} - x'_{\delta}| + W_2(\mu, \mu') + W_2(\mu_{\delta}, \mu'_{\delta}) \)\end{aligned}$$ $\sigma$ satisfies the same condition as $b$.
(C6) $\sup \limits_{t \ge -\delta} \(|b(t, 0, 0, 0, 0)| + |\sigma(t, 0, 0, 0, 0)|\) < \infty. $
\[TMFDSDE\] Suppose that $b$ and $\sigma$ satiesfies (C5) and (C6), then for any given delay conditions $\xi_. \in H^2_{\mathscr{F}}(-\delta, 0; \mathbb{R}^m)$, the MFDSDE (\[MFDSDE\]) has a unique strong solution.
For any $\beta \ge 0$, we introduce a norm in the Banach space $H^2_{\mathscr{F}}(-\delta, T; \RR^{m})$: $$\| \nu(\cdot) \|_{\beta} = \( \EE \[\int_{-\delta}^T e^{- \beta t} | \nu _s |^2 dt \] \)^2$$ Clearly, it is equivalent to the original norm. We use this norm to construct a contraction mapping that allow us to apply the fixed point Theorem. Set $$\begin{aligned}
\begin{cases}
X_t = \xi_0 + \int_0^t b\(s, X_s, X_{s-\delta}, P_{X_s}, P_{X_{s-\delta}}\)ds + \int_0^s \sigma \(s, X_s, X_{s-\delta}, P_{X_s}, P_{X_{s-\delta}}\) dB_s , \nonumber \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
X_t = \xi_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [-\delta, 0). \nonumber
\end{cases}\end{aligned}$$ Given $x \in H^2_{\mathscr{F}}(-\delta, T; \RR^{m})$, we define $$\begin{aligned}
\begin{cases}
X_t = \xi_0 + \int_0^t b\(s, x_s, x_{s-\delta}, P_{x_s}, P_{x_{s-\delta}}\)ds + \int_0^s \sigma \(s, x_s, x_{s-\delta}, P_{x_s}, P_{x_{s-\delta}}\) dB_s , \nonumber \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
X_t = \xi_t, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [-\delta, 0). \nonumber
\end{cases}\end{aligned}$$ Then, $X \in H^2_{\mathscr{F}}(-\delta, T; \RR^{m})$. Denote $X = \Phi (x.)$, now we prove that $\Phi$ is a contraction mapping under the norm $\| \cdot \|_{\beta}$. For two arbitrary elements $x.$ and $x'.$ in $H^2_{\mathscr{F}}(-\delta, T; \RR^{m})$, set $x. = \Phi(x.)$, $x'. = \Phi(x'.)$. Denote their differences by ${\overline}{x} = x - x'$, ${\overline}{\Phi} = \Phi(x) -\Phi(x')$, ${\overline}{b} = b(t, x_t, x_{t-\delta}, P_{x_t}, P_{x_{t-\delta}}) - b(t, x'_t, x'_{t-\delta}, P_{x'_t}, P_{x'_{t-\delta}})$, ${\overline}{\sigma} = \sigma(t, x_t, x_{t-\delta}, P_{x_t}, P_{x_{t-\delta}}) - \sigma(t, x'_t, x'_{t-\delta}, P_{x'_t}, P_{x'_{t-\delta}})$ .
By applying Ito’s formula to $e^{\beta t} {\overline}{\Phi}_t^2$, we have $$\begin{aligned}
&e^{\beta t} {\overline}{\Phi}_t^2 + \beta \int_0^T e^{-\beta t} {\overline}{\Phi}_t^2 dt \\
=& 2\int_0^T e^{-\beta t} \langle {\overline}{\Phi}_t, {\overline}{b} \rangle dt + 2\int_0^T e^{-\beta t} \langle {\overline}{\Phi}_t, {\overline}{\sigma} \rangle dB_t + \int_0^T e^{-\beta t} tr \langle {\overline}{\sigma}_t, {\overline}{\sigma}^* \rangle dt\end{aligned}$$ Take expectation on both sides, then $$\begin{aligned}
\beta \EE \int_0^T e^{-\beta t} {\overline}{\Phi}_t^2 dt = 2 \EE \int_0^T e^{-\beta t} \langle {\overline}{\Phi}_t, {\overline}{b} \rangle dt + \EE \int_0^T e^{-\beta t} tr \langle {\overline}{\sigma}_t, {\overline}{\sigma}^* \rangle dt,\end{aligned}$$ by Cauchy-Schwartz enequality and condition (C5), $$\begin{aligned}
\beta \EE \int_0^T e^{-\beta t} {\overline}{\Phi}_t^2 dt \le \EE \int_0^T e^{-\beta t} {\overline}{\Phi}_t^2 dt + 2 C^2 \EE \int_0^T e^{-\beta t} |{\overline}{x}_t|^2 dt ,\end{aligned}$$ we choose $\beta = 1+ 4C^2$, such that $$\begin{aligned}
\EE \int_0^T e^{-\beta t} {\overline}{\Phi}_t^2 dt \le \frac{1}{2} \EE \int_0^T e^{-\beta t} |{\overline}{x}_t|^2 dt.\end{aligned}$$ Consequently, $\Phi$ is a strict contraction mapping, which complete the proof.
Comparison theorem for MFABSDEs
-------------------------------
Notice that the conditions on the driver $f$ which is needed for the comparison theorem for mean-field BSDEs and for the anticipated BSDEs are stronger than those needed for the existence and uniqueness theorem.
Let ${\overline}{f}_i = (t, y, z, y', z'), i = 1, 2,$ be two drivers of mean-field BSDEs, to derive the comparison principle for mean-field BSDEs, restrictions are forced on ${\overline}{f}_i, i = 1, 2,$ in [@BLP2009] as following:
(i) One of the two coefficients is independent of $z'$,
(ii) One of the two coefficients is nondecreasing in $y'$.
On the other hand, two example in [@PY] also given to demonstrate the comparison principle for the anticipated BSDEs (\[eqq4\]). Let $\hat{f}_i = (t, y, z, \theta, \gamma), i = 1, 2,$ be two drivers of (\[eqq4\]), if
(iii) $\hat{f}_2$ is increasing in the anticipated term of $Y.$
(iv) $\hat{f}_2$ indipendent of the anticipated term of $Z.$
then the comparison theorem holds for anticipated BSDEs.
Now we discuss the comparison principle for mean-filed anticipated BSDEs (\[eqc2\]), it is naturally to combine all the restrictions above both on ${\overline}{f}$ and $\hat{f}$. In addition, we force the other two restrictions on $f$:
(v) One of the two coefficients ($f_1$ or $f_2$) is independent of the anticipated term of $z'$,
(vi) $f_2$ is non-decreasing in the anticipated term of $Y'.$
Counterexample are given to show that if the driver $f$ of mean-field anticipated BSDEs depends on the anticipated term of $z'$ we can’t get the comparison theorem.
For $d = 1$ we consider the mean-field BSDE(\[eqc2\]) with time horizon $T = 1$, for all $t \in [0, T]$, with driver $f(t, y'_t, z'_t, y'_{t+\delta(t)}, z'_{t+\zeta(t)}, y_t, z_t, y_{t+\delta(t)},$\
$ z_{t+\zeta(t)}) = - z'_{t+\zeta(t)}$ and two different terminal values $\xi_1, \xi_2 \in L^2(\Omega, \mathcal{F}_T, P)$. Let us denotes the associated solutions by $(Y^1, Z^1)$ and $(Y^2, Z^2)$, respectively. Then, $$\begin{aligned}
\label{eqEoCT}
Y_t^i = \xi_i + \int_t^1 \EE[- Z_{s + \zeta(s)}^i]ds - \int_t^1 Z_s^i dW_s, \quad 0 \le t \le 1, i = 1, 2.\end{aligned}$$ Let $\xi_1 = - (B_1^{+})^3$, then by the Clark-Ocone formula we have: $$\begin{aligned}
- (B_1^{+})^3 = \xi_1 &=& \EE(\xi) + \int_0^1 \EE (D_s \xi | \mathscr{F}_s) dW_s \nonumber \\
&=& \EE(\xi) + \int_0^1 Z_s dW_s = -\frac{2}{\sqrt{2 \pi}} - \int_0^1 3(B_1^{+})^2 dW_s. \nonumber \end{aligned}$$ In addition, $$\begin{aligned}
Y_0^1 &=& - (B_1^+)^3 - \int_0^1 \( \EE(D_{s+\zeta(s)}\xi)1_{s + \zeta(s)\le 1} + \EE(D_1 \xi)1_{s + \zeta(s) >1} \) ds - \int_0^1 Z_sdW_s \nonumber \\
&=& -\frac{2}{\sqrt{2 \pi}} + \int_0^1 3 \EE (B_1^+)^2 ds = -\frac{2}{\sqrt{2 \pi}} + \frac{3}{2} >0. \nonumber\end{aligned}$$ Let now $\xi_2 = 0$. Then $(Y^2, Z^2) = (0, 0)$ is also a solution of (\[eqEoCT\]). Hence, we have $Y_0^1 > Y_0^2$ although $\xi_1 \le \xi_2$.
Without loss of generality, we assume that all the restrictions (i) - (vi) are satisfied by $f_2$. Let $(Y.^{(1)}, Z.^{(1)})$, $(Y.^{(2)}, Z.^{(2)})$ be respectively solutions of the following two mean-filed anticipated BSDEs: $$\begin{aligned}
\begin{cases}
Y_t^{(1)} =\xi_T^{(1)} + \int_t^T \EE^{'}[f_1(s, Y^{(1)}_s, Z^{(1)}_s, Y^{(1)}_{s+\delta(s)}, Z^{(1)}_{s+\zeta(s)}, Y'^{,(1)}_s, Z'^{,(1)}_s,Y'^{,(1)}_{s+\delta(s)}, Z'^{,(1)}_{s+\zeta(s)})]ds \nonumber \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ - \int_t^T Z^{(1)}_sdB_s , \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y_t^{(1)} = \xi_t^{(1)}, Z_t^{(1)} = \eta_t^{(1)} , \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K];
\end{cases}\end{aligned}$$ $$\begin{aligned}
\begin{cases}
Y_t^{(2)} =\xi_T^{(2)} + \int_t^T \EE^{'}[f_2(s, Y^{(2)}_s, Z^{(2)}_s, Y^{(2)}_{s+\delta(s)},Y'^{,(2)}_s, Y'^{,(2)}_{s+\delta(s)})]ds - \int_t^T Z^{(2)}_sdB_s, \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y^{(2)}_t = \xi_t^{(2)}, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K]; \nonumber
\end{cases}\end{aligned}$$
Assume that $f_2$ the above restrictions (i)-(vi), $\xi_{.}^{(1)}$, $\xi_{.}^{(2)} \in \mathcal{S}^2_{\mathscr{F}}(T, T+K)$, $\delta$, $\zeta$ satisfies (C1), (C2), and for all $t \in [0, T]$, $y \in \mathbb{R}^m$, $z \in \mathbb{R}^{m \times d}$, $f_2(t, y, z, \cdot, y', \cdot)$ is increasing, that is, $f_2(t, y, z, \theta_r, y', \theta'_r) \ge f_2(t, y, z, {\widetilde}{\theta}_r, y', {\widetilde}{\theta}'_r)$, if $\theta_r \ge {\widetilde}{\theta}_r$ and $\theta'_r \ge {\widetilde}{\theta}'_r$, $\theta_r, \theta'_r, {\widetilde}{\theta}_r, {\widetilde}{\theta}'_r \in H^2_{\mathscr{F}}(t, T+K), r \in [t, T+K]$. If $\xi_s^{(1)} \ge \xi_s^{(2)}, s \in [T, T+K]$ and $f_1(t, y, z, \theta_r, \gamma_{{\overline}{r}}, y', z', \theta'_r, \gamma'_{{\overline}{r}}) \ge f_2(t, y, z, \theta_r, y', \theta'_r), r, {\overline}{r} \in [t, T+K]$, then $$Y_t^{(1)} \ge Y_t^{(2)}, \quad \quad a.e., a.s.$$
Set $$\begin{aligned}
\begin{cases}
Y_t^{(3)} =\xi_T^{(2)} + \int_t^T \EE^{'}[f_2(s, Y^{(3)}_s, Z^{(3)}_s, Y^{(1)}_{s+\delta(s)},Y'^{,(3)}_s, Y'^{,(1)}_{s+\delta(s)})]ds - \int_t^T Z^{(3)}_sdB_s , \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y_t^{(3)} = \xi_t^{(2)}, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K].\nonumber
\end{cases}\end{aligned}$$ By Lemma \[le1\], we know that there exists a unique pair of $\mathscr{F}_t$-adapted processes $(Y.^{(3)}, Z.^{(3)}) \in S_{\mathscr{F}}^2(0, T+K, \RR^m) \times H_{\mathscr{F}}^2(0, T; \mathbb{R}^{m \times d})$ that satisfies the above BSDE. Since $f_1 \ge f_2$, $y \in \mathbb{R}^m, z \in \mathbb{R}^{m \times d}$, by Lemma \[le2\], we obtain $$\begin{aligned}
Y_t^{(1)} \ge Y_t^{(3)}, \quad a.s.\end{aligned}$$ Set $$\begin{aligned}
\begin{cases}
Y_t^{(4)} =\xi_T^{(2)} + \int_t^T \EE^{'}[f_2(s, Y^{(4)}_s, Z^{(4)}_s, Y^{(3)}_{s+\delta(s)},Y'^{,(4)}_s, Y'^{,(3)}_{s+\delta(s)})]ds - \int_t^T Z^{(4)}_sdB_s , \\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y_t^{(4)} = \xi_t^{(2)}, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K].\nonumber
\end{cases}\end{aligned}$$ Since for all $t \in [0, T], y \in \mathbb{R}^m, z \in \mathbb{R}^{m \times d}$, $f_2(t, Y_t, Z_t, \cdot, Y'_t, Z'_s, Y'_{s+\delta(s)}, Z'_{t+\zeta(t)})$ is increasing and $Y_t^{(1)} \ge Y_t^{(3)}$, by Lemma \[le2\], we know $$\begin{aligned}
Y_t^{(3)} \ge Y_t^{(4)}, \quad a.s.\end{aligned}$$ For $n = 5, 6, \cdot\cdot\cdot,$ we consider the following mean-field BSDE: $$\begin{aligned}
\begin{cases}
Y_t^{(n)} =\xi_T^{(2)} + \int_t^T \EE^{'}[f_2(s, Y^{(n)}_s, Z^{(n)}_s, Y^{(n-1)}_{s+\delta(s)},Y'^{,(n)}_s, Y'^{,(n-1)}_{s+\delta(s)})]ds\\
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ - \int_t^T Z^{(n)}_sdB_s , \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in [0,T];
\cr
Y_t^{(n)} = \xi_t^{(2)}, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K]. \nonumber
\end{cases}\end{aligned}$$ Similarly, we have $Y_t^{(4)} \ge Y_t^{(5)} \ge \cdot\cdot\cdot \ge Y_t^{(n)} \ge \cdot\cdot\cdot,$ a.s. We use $\|\nu(\cdot)\|_{\beta}$ in the proof of Theorem \[unique\] as the norm in the Banach space $H^2_{\mathscr{F}}(0, T+K; \RR^m) \times H^2_{\mathscr{F}}(0, T; \RR^{m\times d})$. Set $\hat{Y}_s^{(n)}= Y_s^{(n)} - Y_s^{(n-1)}$, $\hat{Z}_s^{(n)}= Z_s^{(n)} - Z_s^{(n-1)}$, $n \ge 4.$ Then by (\[eatimates2\]), we have $$\begin{aligned}
& \EE\[ \int_0^T \( \frac{\beta}{2} |\hat{Y}^{(n)}_{s}|^2 + |\hat{Z}^{(n)}_{s}|^2 \) e^{\beta s} ds \] \nonumber \\
\le &\frac{2}{\beta} \EE\[ \int_0^T \EE' |f_2(s, Y^{(n)}_s, Z^{(n)}_s, Y^{(n-1)}_{s+\delta(s)},Y'^{,(n)}_s, Y'^{,(n-1)}_{s+\delta(s)}) \right. \nonumber\\
&\qquad \qquad \quad - f_2(s, Y^{(n-1)}_s, Z^{(n-1)}_s, Y^{(n-2)}_{s+\delta(s)},Y'^{,(n-1)}_s, Y'^{,(n-2)}_{s+\delta(s)}) |^2 e^{\beta s}ds\left. \] \nonumber \\
\le &\frac{10C^2}{\beta} \EE \[ \int_0^T \(2 | \hat{Y}_s^{(n)}|^2 + |\hat{Z}_s^{(n)}|^2 + |\hat{Y}_{s+\delta{s}}^{(n-1)}|^2 + | \hat{Z}_{s+\delta{s}}^{(n-1)}|^2 \) e^{\beta s} ds \] \nonumber \\
\le& \frac{20C^2}{\beta} \EE \[ \int_0^T \( | \hat{Y}_s^{(n)}|^2 + |\hat{Z}_s^{(n)}|^2 \) e^{\beta s} ds \] + \frac{10C^2L}{\beta} \EE \[ \int_0^T \( |\hat{Y}_{s}^{(n-1)}|^2 + | \hat{Z}_{s}^{(n-1)}|^2\) e^{\beta s} ds \]. \nonumber \end{aligned}$$ Set $\beta = 60C^2L + 60C^2 +3$. Then $$\begin{aligned}
\frac{2}{3}\EE\[ \int_0^T \( |\hat{Y}^{(n)}_{s}|^2 + |\hat{Z}^{(n)}_{s}|^2 \) e^{\beta s} ds \] \le \frac{1}{3}\EE \[ \int_0^T \( |\hat{Y}_{s}^{(n-1)}|^2 + | \hat{Z}_{s}^{(n-1)}|^2\) e^{\beta s} ds \] \nonumber\end{aligned}$$ Hence, $$\begin{aligned}
\EE\[ \int_0^T \( |\hat{Y}^{(n)}_{s}|^2 + |\hat{Z}^{(n)}_{s}|^2 \) e^{\beta s} ds \] \le \(\frac{1}{2}\)^{n-4}\EE \[ \int_0^T \( |\hat{Y}_{s}^{(4)}|^2 + | \hat{Z}_{s}^{(4)}|^2\) e^{\beta s} ds \]. \nonumber\end{aligned}$$ It follows that $\(Y.^{(n)}\)_{n \ge 4}$ and $\(Z.^{(n)}\)_{n \ge 4}$ are Cauchy sequences in $H^2_{\mathscr{F}}(0, T+K; \RR^m) \times H^2_{\mathscr{F}}(0, T; \RR^{m\times d})$. Denote their limits by $Y.$ and $Z.$, respectively. Since $H^2_{\mathscr{F}}(0, T+K; \RR^m)$ and $H^2_{\mathscr{F}}(0, T; \RR^{m\times d})$ are both Banach spaces, we obtain $(Y. \times Z.) \in H^2_{\mathscr{F}}(0, T+K; \RR^m) \times H^2_{\mathscr{F}}(0, T; \RR^{m\times d})$. Note that for all $t \in[0, T]$, $$\begin{aligned}
&\EE\[ \int_t^T \EE' |f_2(s, Y^{(n)}_s, Z^{(n)}_s, Y^{(n-1)}_{s+\delta(s)},Y'^{,(n)}_s, Y'^{,(n-1)}_{s+\delta(s)}) - f_2(s, Y_s, Z_s, Y_{s+\delta(s)},Y'_s, Y'_{s+\delta(s)}) |^2 e^{\beta s}ds \] \nonumber \\
\le& 5C^2 \EE \[ \int_0^T \(2 | Y_s^{(n)} - Y_s|^2 + |Z_s^{(n)} - Z_s|^2 + L| Y_s^{(n-1)} - Y_s|^2 + L| Z_s^{(n-1)} - Z_s|^2 \) e^{\beta s} ds \] \to 0, \nonumber\end{aligned}$$ when $n \to \infty$. Therefore, $(Y., Z.)$ satisfies the following mean-field anticipated BSDE: $$\begin{aligned}
\begin{cases}
Y_t =\xi_T^{(2)} + \int_t^T \EE^{'}[f_2(s, Y_s, Z_s, Y_{s+\delta(s)},Y'_s, Y'_{s+\delta(s)})]ds- \int_t^T Z_sdB_s, \ \ t \in [0,T];
\cr
Y_t = \xi_t^{(2)},\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t \in (T, T+K]. \nonumber
\end{cases}\end{aligned}$$ By Theorem \[unique\], we know $$\begin{aligned}
Y_t = Y_t^{(2)}, \quad a.s.\end{aligned}$$ Since $Y_t^{(1)} \ge Y_t^{(3)} \ge Y_t^{(4)} \ge Y_t$, it holds immediately $$\begin{aligned}
Y_t^{(1)}\geq Y_t^{(2)}, \quad a.s.\end{aligned}$$
Formulation of the generalized mean-filed stochastic delay control problem
==========================================================================
In this section, we give the formulation of our generalized mean-field optimal control problem.\
We consider the generalized mean-field delay type optimal control system, with the state equation (\[2eq11\]) and the cost functional (\[2eq12\]). From Theorem \[TMFDSDE\], we know equation (\[2eq11\]) admits a unique solution. Recall that the agent wishes to minimize his cost functional, namely, an admissible control $u\in \mathcal{U}$ is said to be optimal if $$J(u)=\min_{v\in \mathcal{U}}J(v).$$Throughout this paper, we make the following assumptions on the coefficients:
\[2H1\] (1) The given functions $b, \sigma, h ,\Phi$ are differentiable with respect to $(x,x_{\delta},\mu,\mu_\delta, v,v_\delta).$\
(2) $b, \sigma$ are Lipschitz continuous w.r.t. $(x,x_\delta,\mu,\mu_\delta)$, The derivatives of $b, \sigma$ are Lipschitz continuous and bounded.\
(3) The derivatives of $h, \Phi$ are Lipschitz continuous and bounded by $C(1+|x|+|x_\delta|+|v|+|v_\delta| )$.\
We will make use of the following notations concerning matrices. We denote by $\mathbb{R}^{n\times d}$ the space of real matrices of $n\times d $-type, and by $\mathbb{R}^{n\times n}_d$ the linear space of the vectors of matrices $M=(M_1,\cdots, M_d)$, with $M_i\in \mathbb{R}^{n\times n}$, $1\leq i \leq d.$ Given any $\alpha, \beta\in \mathbb{R}^{n}$, $L, S \in \mathbb{R}^{n\times d}$, $\gamma \in \mathbb{R}^{d}$ and $M, N \in \mathbb{R}^{n\times n}_d$, we introduce the following notation: $\alpha \beta=\sum^{n}_{i=1}\alpha_i \beta_i \in \mathbb{R}$, $\alpha\times\beta=(\alpha_i \beta_j)_{1\leq i,j\leq n}$; $LS=\sum^{d}_{i=1} L_i S_i \in \mathbb{R}$, where $L=(L_1,\cdots, L_d), S=(S_1,\cdots, S _d)$; $ML=\sum^{d}_{i=1}M_i L_i \in \mathbb{R}^{n}$; $M\alpha \gamma=\sum^{d}_{i=1}(M_i \alpha) \gamma_i \in \mathbb{R}^n$; $MN=\sum^{d}_{i=1}M_i N_i \in \mathbb{R}^{n\times n}$; For simplicity, we use the following notations $$\Theta_t=(X^u_t,X^u_{t-\delta}, P_ {{X^u_t}},P_ {{X^u_{t-\delta}}},u_t,u_{t-\delta})$$; $$\Theta'_t=((X^u_t)',(X^u_{t-\delta})', P_ {{X^u_t}},P_ {{X^u_{t-\delta}}},(u_t)',(u_{t-\delta})').$$
Let us suppose that $u$ is an optimal control and $X^u$ the associated optimal trajectory. Then we introduce the convex perturbed control as follows: $$u^{\theta}_t=u_t+\theta(v_t-u_t),$$ where $\theta\geq 0$ is sufficiently small, and $v_t$ is an arbitrary element of $\mathcal{U}$, $X^{\theta}$ is the state under the control $u^{\theta}$. The convexity of $U$ guarantee that $u^{\theta}_t\in \mathcal{U}$, and obviously, $$0\leq J(u^{\theta})-J(u).$$
\[2lm31\] Under the Hypothesis \[2H1\], we have, $$\lim_{\theta\rightarrow0}\mathbb{E}[\sup_{0\leq t \leq T}|X^{\theta}_t-X^u_t|^2]=0.$$
Note that, for $\tau\in [0,T]$ $$\begin{aligned}
&&\mathbb{E}\left[\sup_{0\leq t \leq \tau}|X^{\theta}_t-X^u_t|^2\right]\nonumber \\
&\leq &C\mathbb{E}\int^\tau_0\left|b(t,X^\theta_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_t,u^\theta_{t-\delta})-b(s,\Theta_s)\right|^2ds\nonumber\\
&&+C\mathbb{E}\int^\tau_0\left|\sigma(t,X^\theta_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_t,u^\theta_{t-\delta})-\sigma(s,\Theta_s)\right|^2ds\nonumber\\
&\leq &C\mathbb{E}\int^\tau_0\sup_{0\leq s\leq r}\left|X^{\theta}_s-X^u_s\right|^2dr+\theta^2 C\mathbb{E}\int^T_0\left|v_s-u_s\right|^2ds.
\end{aligned}$$ From Gronwall’s inequality we have the desired result.
Next, we study the variational process of our state.
Let $K_t$ be the solution of the following linear equation:
$$\label{2eq31}
\left\{
\begin{aligned}
dK_t=&\bigg\{b_x(t,\Theta_t)K_t+\mathbb{E}'\left[b_{\mu }(t,\Theta_t, (X^u_t)')(K_t)'\right]+b_{x_\delta}(t,\Theta_t)K_{t-\delta}&\\
&+\mathbb{E}'\left[b_{\mu_\delta }(t,\Theta_t, (X^u_{t-\delta})')(K_{t-\delta})'\right]+b_v(t,\Theta_t)(v_t-u_t)&\\
&+b_{v_\delta}(t,\Theta_t)(v_{t-\delta}-u_{t-\delta})\bigg\}dt&\\
&+\bigg\{\sigma_x(t,\Theta_t)K_t+\mathbb{E}'\left[\sigma_{\mu }(t,\Theta_t, (X^u_t)')(K_t)'\right]+\sigma_{x_\delta}(t,\Theta_t)K_{t-\delta}&\\
&+\mathbb{E}'\left[\sigma_{\mu_\delta }(t,\Theta_t, (X^u_{t-\delta})')(K_{t-\delta})'\right]+\sigma_v(t,\Theta_t)(v_t-u_t)&\\
&+\sigma_{v_\delta}(t,\Theta_t)(v_{t-\delta}-u_{t-\delta})\bigg\}dB_t,\ \ \ \ t\in[0,T],&\\
K_0=&0,\ v_t=u_t,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ t\in[-\delta,0]. &
\end{aligned}
\right.$$
Then we have $$\lim_{\theta\rightarrow0}\mathbb{E}\left[\sup_{0\leq s \leq t}\left| \frac{X^{\theta}_s-X^u_s}{\theta}-K_s\right|^2\right]=0,$$ for all $t\in[0,T].$
From Theorem \[TMFDSDE\], we know equation (\[2eq31\]) admits a unique solution $K_t$. We set $$\eta_t=\frac{X^{\theta}_t-X^u_t}{\theta}-K_t, \ \ \ t\in [0,T].$$ Then we have $$\begin{aligned}
\eta_t&=&\frac{1}{\theta}\int^t_0\left[b(t,X^\theta_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_t,u^\theta_{t-\delta})-b(s,\Theta_s)\right]ds\nonumber\\
&&+\frac{1}{\theta}\int^t_0\left[\sigma(t,X^\theta_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_t,u^\theta_{t-\delta})-\sigma(s,\Theta_s)\right]dB_s\nonumber\\
&&-\int^t_0\bigg\{b_x(s,\Theta_t)K_s+\mathbb{E}'\left[b_{\mu }(s,\Theta_s, (X^u_s)')(K_s)'\right]+b_{x_\delta}(s,\Theta_s)K_{s-\delta}\nonumber\\
&&+\mathbb{E}'\left[b_{\mu_\delta }(s,\Theta_s, (X^u_{s-\delta})')(K_{s-\delta})'\right]+b_v(s,\Theta_s)(v_s-u_s)\nonumber\\
&&+b_{v_\delta}(s,\Theta_s)(v_{s-\delta}-u_{s-\delta})\bigg\}ds\nonumber\\
&&+\bigg\{\sigma_x(s,\Theta_s)K_s+\mathbb{E}'\left[\sigma_{\mu }(s,\Theta_s, (X^u_s)')(K_s)'\right]+\sigma_{x_\delta}(s,\Theta_s)K_{s-\delta}\nonumber\\
&&+\mathbb{E}'\left[\sigma_{\mu_\delta }(s,\Theta_s, (X^u_{s-\delta})')(K_{s-\delta})'\right]+\sigma_v(s,\Theta_s)(v_s-u_s)\nonumber\\
&&+\sigma_{v_\delta}(s,\Theta_s)(v_{s-\delta}-u_{s-\delta})\bigg\}dB_s.\nonumber\\\end{aligned}$$
Since for any $f\in C^{2,1}(\mathcal{P}_2(\mathbb{R}^d))$ $$\begin{aligned}
&&f(P_{\mu})-f(P_{\mu_0})=\int^1_0\frac{d}{d\lambda}f(P_{\mu_0+\lambda\eta})d\lambda\nonumber\\
& =&\int^1_0\mathbb{E}\big[f_{\mu}(P_{\mu_0+\lambda\eta},\mu_0+\lambda\eta)\cdot\eta\big]d\lambda,\end{aligned}$$
we notice that $$\begin{aligned}
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^\theta_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 b_x(s,X^u_s+\lambda\theta(\eta_s+K_s),X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&(\eta_s+K_s)d\lambda ds,\\
\\
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 b_{x_\delta}(s,X^u_s,X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}), P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&(\eta_s+K_s)d\lambda ds,\\
\\
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^u_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 \mathbb{E}'\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)(\eta_s+K_s)'\Big]d\lambda ds,\\
\\
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 \mathbb{E}'\Big[b_{\mu_{\delta }}\big(s,X^u_s,X^u_{s-\delta},P_ {{X^u_{s}}},P_ {X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta})},\\
&&u^\theta_s,u^\theta_{s-\delta},(X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}))'\big)(\eta_s+K_s)'\Big]d\lambda ds,\\
\\
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u^\theta_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 b_v(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s+\lambda\theta(v_s-u_s),u^\theta_{s-\delta})\\
&&(v_s-u_s)d\lambda ds,\\
\\
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u_{s-\delta})\Big]ds\\
&=&\int^t_0 \int^1_0 b_{v_\delta}(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u^\theta_{s-\delta}+\lambda\theta(v_{s-\delta}-u_{s-\delta}))\\
&&(v_{s-\delta}-u_{s-\delta})d\lambda ds.\\\end{aligned}$$
Similarly result can be obtained for $\sigma$. On the other hand, we have $$\begin{aligned}
&&\frac{1}{\theta}\int^t_0\Big[b(t,X^u_s,X^u_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\\
&&-b(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\Big]ds\\
&&-\int^t_0\mathbb{E}'\left[b_{\mu }(s,\Theta_s, (X^u_s)')(K_s)'\right]ds\\
&=&\int^t_0 \int^1_0 \mathbb{E}'\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)(\eta_s)'\Big]d\lambda ds\\
&&+\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))')-b_{\mu }(s,\Theta_s, (X^u_s)'\big)\Big](K_s)'\bigg\}d\lambda ds.\\\end{aligned}$$ Set $$\begin{aligned}
I^{\theta}_t&=&\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)-b_{\mu }\left(s,\Theta_s, (X^u_s)'\right)\Big](K_s)'\bigg\}d\lambda ds.\\\end{aligned}$$ Then, from Lemma \[2lm31\], Lipschitz continuity and the definition of 2-Wasserstein metric, we have $$\lim_{\theta\rightarrow0}\mathbb{E}\big[\sup_{0\leq s \leq T}|I^{\theta}_s|^2\big]=0.$$
Therefore, we have $$\begin{aligned}
&&\mathbb{E}\big[\sup_{0\leq s \leq t}|{\eta}_s|^2\big]\nonumber\\
&\leq &C \mathbb{E}\int^t_0 \int^1_0 \Big|b_x(s,X^u_s+\lambda\theta(\eta_s+K_s),X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&(\eta_s)\Big|^2 d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \Big| b_{x_\delta}(s,X^u_s,X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}), P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&(\eta_s)\Big|^2d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \mathbb{E}'\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)(\eta_s)'\Big]^2d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \mathbb{E}'\Big[b_{\mu_{\delta }}\big(s,X^u_s,X^u_{s-\delta},P_ {{X^u_{s}}},P_ {X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta})},\nonumber\\
&&u^\theta_s,u^\theta_{s-\delta},(X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}))'\big)(\eta_s)'\Big]^2d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \Big|\sigma_x(s,X^u_s+\lambda\theta(\eta_s+K_s),X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&(\eta_s)\Big|^2 d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \Big| \sigma_{x_\delta}(s,X^u_s,X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}), P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&(\eta_s)\Big|^2d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \mathbb{E}'\Big[\sigma_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\nonumber\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)(\eta_s)'\Big]^2d\lambda ds\nonumber\\
&&+C \mathbb{E}\int^t_0 \int^1_0 \mathbb{E}'\Big[\sigma_{\mu_{\delta }}\big(s,X^u_s,X^u_{s-\delta},P_ {{X^u_{s}}},P_ {X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta})},\nonumber\\
&&u^\theta_s,u^\theta_{s-\delta},(X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}))'\big)(\eta_s)'\Big]^2d\lambda ds\nonumber\\
&&+C\mathbb{E}\big[\sup_{0\leq s \leq t}|\beta^{\theta}_s|^2\big]\end{aligned}$$ where $$\begin{aligned}
\beta^{\theta}_t&=&\int^t_0 \int^1_0\Big[b_x(s,X^u_s+\lambda\theta(\eta_s+K_s),X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&-b_x(t,\Theta_t)\Big](K_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \Big[ b_{x_\delta}(s,X^u_s,X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}), P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&-b_{x_\delta}(t,\Theta_t)\Big](K_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[b_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\nonumber\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)-b_{\mu }(t,\Theta_t, (X^u_t)')\Big](K_s)'\bigg\}d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[b_{\mu_{\delta }}\big(s,X^u_s,X^u_{s-\delta},P_ {{X^u_{s}}},P_ {X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta})},\nonumber\\
&&u^\theta_s,u^\theta_{s-\delta},(X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}))'\big)-b_{\mu_\delta }(t,\Theta_t, (X^u_{t-\delta})')\Big](K_s)'\bigg\}d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 b_v(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s+\lambda\theta(v_s-u_s),u^\theta_{s-\delta})\nonumber\\
&&(v_s-u_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 b_{v_\delta}(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u^\theta_{s-\delta}+\lambda\theta(v_{s-\delta}-u_{s-\delta}))\nonumber\\
&&(v_{s-\delta}-u_{s-\delta})d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0\Big[\sigma_x(s,X^u_s+\lambda\theta(\eta_s+K_s),X^\theta_{s-\delta}, P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&-\sigma_x(t,\Theta_t)\Big](K_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \Big[ \sigma_{x_\delta}(s,X^u_s,X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}), P_ {{X^\theta_s}},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta})\nonumber\\
&&-\sigma_{x_\delta}(t,\Theta_t)\Big](K_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[\sigma_{\mu }\big(s,X^u_s,X^u_{s-\delta},P_ {X^u_s+\lambda\theta(\eta_s+K_s)},P_ {{X^\theta_{s-\delta}}},u^\theta_s,u^\theta_{s-\delta},\nonumber\\
&&(X^u_s+\lambda\theta(\eta_s+K_s))'\big)-\sigma_{\mu }(t,\Theta_t, (X^u_t)')\Big](K_s)'\bigg\}d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \mathbb{E}'\bigg\{\Big[\sigma_{\mu_{\delta }}\big(s,X^u_s,X^u_{s-\delta},P_ {{X^u_{s}}},P_ {X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta})},\nonumber\\
&&u^\theta_s,u^\theta_{s-\delta},(X^u_{s-\delta}+\lambda\theta(\eta_{s-\delta}+K_{s-\delta}))'\big)-\sigma_{\mu_\delta }(t,\Theta_t, (X^u_{t-\delta})')\Big](K_s)'\bigg\}d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \sigma_v(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s+\lambda\theta(v_s-u_s),u^\theta_{s-\delta})\nonumber\\
&&(v_s-u_s)d\lambda ds\nonumber\\
&&+\int^t_0 \int^1_0 \sigma_{v_\delta}(s,X^u_s,X^u_{s-\delta}, P_ {{X^u_s}},P_ {{X^u_{s-\delta}}},u_s,u^\theta_{s-\delta}+\lambda\theta(v_{s-\delta}-u_{s-\delta}))\nonumber\\
&&(v_{s-\delta}-u_{s-\delta})d\lambda ds\nonumber\\\end{aligned}$$ Proceeding as in the estimate of $I^{\theta}_t$, we can prove that $$\lim_{\theta\rightarrow0}\mathbb{E}\big[\sup_{0\leq s \leq T}|\beta^{\theta}_s|^2\big]=0.$$ Since the derivatives of $b,\sigma$ are bounded, we deduce that $$\mathbb{E}\big[\sup_{0\leq s \leq t}|\eta_s|^2\big]\leq C\mathbb{E}\int^t_0|\eta_s|^2ds+C\mathbb{E}\big[\sup_{0\leq s \leq t}|\beta^{\theta}_s|^2\big].$$ Finally, by Gronwall’s inequality, we complete the proof.
Let $u$ be an optimal control and $X^u_t$ be the corresponding optimal trajectory. Then, for any control $v\in \mathcal{U}$, we get $$\begin{aligned}
0&\leq &\mathbb{E}\big\{\Phi_x(X^u_T,P_{X^u_t})(K_T)+\mathbb{E}'[\Phi_{\mu}(X^u_T,P_{X^u_t},(X^u_T)')(K_T)']\big\}\nonumber\\
&&+\mathbb{E}\int^T_0\bigg\{ h_x(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta})(K_t)+\mathbb{E}'\big[h_{\mu}\big(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta},\nonumber\\
&&(X^u_T)'\big)(K_t)'\big]
+h_v(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta})(v_t-u_t)\nonumber\\
&&+h_{v_\delta}(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta})(v_{t-\delta}-u_{t-\delta})\bigg\} dt\end{aligned}$$
Since $u$ is an optimal control, we deduce $$\begin{aligned}
\label{2eq32}
0&\leq &J(u^{\theta}_t)-J(u_t)\nonumber\\
&=&\mathbb{E}\big[\Phi(X^{\theta}_T,P_{X^{\theta}_T})-\Phi(X^u_T,P_{X^u_T})\big]\nonumber\\
&&+\mathbb{E}\int^T_0\big[h(t,X^{\theta}_t,P_{X^{\theta}_t},u^{\theta}_t,u^{\theta}_{t-\delta})-h(t,X^u_t,P_{X^u_t},u^{\theta}_t,u^{\theta}_{t-\delta})\big]dt\nonumber\\
&&+\mathbb{E}\int^T_0\big[h(t,X^u_t,P_{X^u_t},u^{\theta}_t,u^{\theta}_{t-\delta})-h(t,X^u_t,P_{X^u_t},u_t,{u_{t-\delta}})\big]dt\nonumber\\
& =:&I_1+I_2+I_3.\end{aligned}$$
$$\begin{aligned}
I_1&=&\mathbb{E}\left[\int^1_0\Phi_x(X^u_T+\lambda\theta(\eta_T+K_T),P_{X^{\theta}_T})\theta(\eta_{T}+K_T)d\lambda\right]\nonumber\\
&&+\mathbb{E}\bigg\{\int^1_0 \mathbb{E}'\Big[ \Phi_{\mu}\left(X^u_T,P_{X^u_T+\lambda\theta(\eta_T+K_T)},(X^u_T+\lambda\theta(\eta_T+K_T))'\right)\nonumber\\
&&\theta(\eta_{T}+K_T)'\Big]d\lambda\bigg\},\nonumber\\
I_2&=&\mathbb{E}\left[\int^T_0\int^1_0 h_x(t,X^u_t+\lambda\theta(\eta_t+K_t),P_{X^{\theta}_t},u^{\theta}_t,u^{\theta}_{t-\delta})\theta(\eta_{t}+K_t)d\lambda dt\right]\nonumber\\
&&+\mathbb{E}\bigg\{ \int^T_0 \int^1_0 \mathbb{E}'\Big[ h_{\mu }\Big(t,X^u_t,P_{X^u_t+\lambda\theta(\eta_t+K_t)},u^{\theta}_t,u^{\theta}_{t-\delta},\nonumber\\
&&(X^u_t+\lambda\theta(\eta_t+K_t))'\Big)\theta(\eta_{t}+K_t)'\Big]d\lambda dt \bigg\},\nonumber\\
I_3&=&\mathbb{E}\bigg[\int^T_0\int^1_0 h_v\left(t,X^u_t,P_{X^u_t},u_t+\lambda\theta(v_t-u_t),u^{\theta}_{t-\delta}\right)\nonumber\\
&&\theta(v_{t}-u_t)d\lambda dt\bigg]\nonumber\\
&&+\mathbb{E}\bigg[\int^T_0\int^1_0 h_{v_\delta}\left(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta}+\lambda\theta(v_{t-\delta}-u_{t-\delta})\right)\nonumber\\
&&\theta(v_{t-\delta}-u_{t-\delta})d\lambda dt\bigg].\end{aligned}$$
From (\[2eq32\]), we get $$\begin{aligned}
0&\leq &\mathbb{E}\left[\int^1_0\Phi_x(X^u_T+\lambda\theta(\eta_T+K_T),P_{X^{\theta}_T})(K_T)d\lambda\right]\nonumber\\
&&+\mathbb{E}\bigg\{\int^1_0 \mathbb{E}'\Big[ \Phi_{\mu}\left(X^u_T,P_{X^u_T+\lambda\theta(\eta_T+K_T)},(X^u_T+\lambda\theta(\eta_T+K_T))'\right)\nonumber\\
&&(K_T)'\Big]d\lambda\bigg\},\nonumber\\
&&+\mathbb{E}\left[\int^T_0\int^1_0 h_x(t,X^u_t+\lambda\theta(\eta_t+K_t),P_{X^{\theta}_t},u^{\theta}_t,u^{\theta}_{t-\delta})(K_t)d\lambda dt\right]\nonumber\\
&&+\mathbb{E}\bigg\{ \int^T_0 \int^1_0 \mathbb{E}'\Big[ h_{\mu }\Big(t,X^u_t,P_{X^u_t+\lambda\theta(\eta_t+K_t)},u^{\theta}_t,u^{\theta}_{t-\delta},\nonumber\\
&&(X^u_t+\lambda\theta(\eta_t+K_t))'\Big)(K_t)'\Big]d\lambda dt \bigg\}\nonumber\\
&&+\mathbb{E}\bigg[\int^T_0\int^1_0 h_v\left(t,X^u_t,P_{X^u_t},u_t+\lambda\theta(v_t-u_t),u^{\theta}_{t-\delta}\right)\nonumber\\
&&(v_{t}-u_t)d\lambda dt\bigg]\nonumber\\
&&+\mathbb{E}\bigg[\int^T_0\int^1_0 h_{v_\delta}\left(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta}+\lambda\theta(v_{t-\delta}-u_{t-\delta})\right)\nonumber\\
&&(v_{t-\delta}-u_{t-\delta})d\lambda dt\bigg]\nonumber\\
&&+\rho^{\theta}_t.\end{aligned}$$
where
$$\begin{aligned}
\rho^{\theta}_t&= &\mathbb{E}\left[\int^1_0\Phi_x(X^u_T+\lambda\theta(\eta_T+K_T),P_{X^{\theta}_T})(\eta_T)d\lambda\right]\nonumber\\
&&+\mathbb{E}\bigg\{\int^1_0 \mathbb{E}'\Big[ \Phi_{\mu}\left(X^u_T,P_{X^u_T+\lambda\theta(\eta_T+K_T)},(X^u_T+\lambda\theta(\eta_T+K_T))'\right)\nonumber\\
&&(\eta_T)'\Big]d\lambda\bigg\},\nonumber\\
&&+\mathbb{E}\left[\int^T_0\int^1_0 h_x(t,X^u_t+\lambda\theta(\eta_t+K_t),P_{X^{\theta}_t},u^{\theta}_t,u^{\theta}_{t-\delta})(\eta_t)d\lambda dt\right]\nonumber\\
&&+\mathbb{E}\bigg\{ \int^T_0 \int^1_0 \mathbb{E}'\Big[ h_{\mu }\Big(t,X^u_t,P_{X^u_t+\lambda\theta(\eta_t+K_t)},u^{\theta}_t,u^{\theta}_{t-\delta},\nonumber\\
&&(X^u_t+\lambda\theta(\eta_t+K_t))'\Big)(\eta_t)'\Big]d\lambda dt \bigg\},\nonumber\\\end{aligned}$$
Since the derivatives of $\Phi,h$ are bounded and $$\lim_{\theta\rightarrow0}\mathbb{E}\big[\sup_{0\leq s \leq T}|\eta_s|^2\big]=0.$$ we get $$\lim_{\theta\rightarrow0}\rho^{\theta}_t=0.$$
From the fact $u^{\theta}_t\rightarrow u_t$ and Lipschitz continuity of $\Phi, h$ we obtain the result.
Necessary and sufficient conditions for the optimal control
===========================================================
It is well known that there is a duality relationship between the stochastic delay differential equations and the backward anticipated stochastic differential equations. In this section, we introduce the adjoint process with the help of mean-filed backward anticipated stochastic differential equation, then the variational inequality can be deduced.\
Let us consider the following adjoint equation: $$\label{2eq41}
\left\{
\begin{aligned}
-dp_t=& \bigg\{ b^*_x(t,\Theta_t)p_t+\sigma^*_x(t,\Theta_t)q_t+h_x(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta})&\\
&+\mathbb{E}'\Big[b^*_{\mu }(t,\Theta'_t,X^u_t)(p_t)'+\sigma^*_{\mu }(t,\Theta'_t,X^u_t)(q_t)'&\\
&+h_{\mu }(t,(X^u_t)',P_{X^u_t},(u_t)',(u_{t-\delta})',X^u_t)\Big]&\\
&+\mathbb{E}^{\mathcal{F}_t}\Big[b^*_{x_\delta}(t,\Theta_t)|_{t+\delta} p_{t+\delta}\Big]+\mathbb{E}^{\mathcal{F}_t}\Big[\sigma^*_{x_\delta}(t,\Theta_t)|_{t+\delta} q_{t+\delta}\Big]&\\
&+\mathbb{E}'\bigg[\mathbb{E}^{\mathcal{F}_t}\Big[b^*_{\mu_\delta}(t,\Theta'_t,X^u_t)|_{t+\delta} (p_{t+\delta})'\Big]&\\
&+\mathbb{E}^{\mathcal{F}_t}\Big[\sigma^*_{\mu_\delta}(t,\Theta'_t,X^u_t)|_{t+\delta} (q_{t+\delta})'\Big]\bigg]\bigg \} dt-q_tdB_t,&\\
p_T=&\Phi_x(X^u_t,P_{X^u_t})+\mathbb{E}'\big[\Phi_{\mu }((X^u_t)',(P_{X^u_t})',X^u_t)\big],&\\
p_t=&0, q_t=0, \ \ \ \ \ \ \ \ t\in(T,T+\delta],&
\end{aligned}
\right.$$ where $b^*$ denotes the transpose of $b$; $b_{x_\delta}(t,\Theta_t)|_{t+\delta}$ denotes the value of $b_{x_\delta}(t,\Theta_t)$ when $t$ replaced by $t+\delta$, other involved terms are defined similarly.\
From Theorem \[unique\], (\[2eq41\]) admits a unique adapted solution. Then, we get the main result of this paper which is stochastic maximum principle for generalized mean-field delay control problem.
(Necessary Conditions for the Optimal Control). Let $u$ be an optimal control, and $X^u_t$ denote the associated optimal trajectory. Let $(p,q)$ be the unique solution of equation (\[2eq41\]). Then the following integral stochastic maximum principle holds: for all $v\in \mathcal{U},$ $$\begin{aligned}
\label{2eq43}
\left<H_v(t,\Theta_t,p_t,q_t),v-u_t\right>+\left<\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],v-u_t\right>\geq 0,\end{aligned}$$ where, $$\begin{aligned}
H(t,\Theta,p,q)&=&b^*(t,x,x_\delta,\mu,\mu_\delta,v,v_\delta)p+\sigma^* (t,x,x_\delta,\mu,\mu_\delta,v,v_\delta)q\\
&&+h(t,x,\mu,v,v_\delta)\end{aligned}$$
By applying Itô’s formula to $\left<p_t,K_t\right>$, notice that $$\begin{aligned}
&&\mathbb{E}\int^T_0\bigg\{K^*_{t-\delta}b^*_{x_\delta}(t,\Theta_t)p_t-K^*_t\mathbb{E}^{\mathcal{F}_t}\left[b^*_{x_\delta}(t,\Theta_t)|_{t+\delta} p_{t+\delta}\right]\bigg\}dt \\
&=&\mathbb{E}\bigg\{\int^T_0K^*_{t-\delta}b^*_{x_\delta}(t,\Theta_t)p_tdt-\int_\delta^{T+\delta} K^*_{t-\delta}b^*_{x_\delta}(t,\Theta_t) p_{t}dt\bigg\} \\
&=&\mathbb{E}\bigg\{\int^\delta_0K^*_{t-\delta}b^*_{x_\delta}(t,\Theta_t)p_tdt-\int_T^{T+\delta} K^*_{t-\delta}b^*_{x_\delta}(t,\Theta_t) p_{t}dt\bigg\} \\
&=&0,\end{aligned}$$ and similarly results can be obtained for other terms, then we get $$\begin{aligned}
\label{2eq42}
&&\mathbb{E}\int^T_0\Big\{\left<H_v(t,\Theta_t,p_t,q_t),v_t-u_t\right>+\left<H_{v_\delta}(t,\Theta_t,p_t,q_t),v_{t-\delta}-u_{t-\delta}\right>\Big\}dt\nonumber\\
&=&\mathbb{E}\left<\Phi_x(X^u_t,P_{X^u_t})+\mathbb{E}'\big[\Phi_{\mu }((X^u_t)',(P_{X^u_t})',X^u_t)\big],K_T\right>\nonumber \\
&&+\mathbb{E}\int^T_0 \left<h_x(t,X^u_t,P_{X^u_t},u_t,u_{t-\delta}),K_t\right>dt\nonumber \\
&&+\mathbb{E}\int^T_0 \left<\mathbb{E}'\big[ h_{\mu }(t,(X^u_t)',P_{X^u_t}, (u_t)',(u_{t-\delta})',X^u_t)\big],K_t\right>dt\nonumber \\
&&+\mathbb{E}\int^T_0 \left< h_{v }(t,(X^u_t)',P_{X^u_t}, X^u_t, (u_t)',(u_{t-\delta})'),v_t-u_t\right>dt\nonumber \\
&&+\mathbb{E}\int^T_0 \left< h_{v_\delta }(t,(X^u_t)',P_{X^u_t}, X^u_t, (u_t)',(u_{t-\delta})'),v_{t-\delta}-u_{t-\delta}\right>dt\nonumber \\
&\geq&0.\end{aligned}$$ Set $$\begin{aligned}
v_s=\left\{
\begin{array}{l}
v_s, \ \ \ s\in[t,t+\varepsilon),\\
u_s, \ \ \ otherwise,
\end{array}
\right.\end{aligned}$$ where $t\in[0,T]$, $v\in \mathcal{U}$. Then (\[2eq42\]) leads to $$\begin{aligned}
&&\frac{1}{\varepsilon}\mathbb{E}\int^{t+\varepsilon}_t\left<H_v(s,\Theta_s,p_s,q_s),v_s-u_s\right>ds\nonumber\\
&&+\frac{1}{\varepsilon}\mathbb{E}\int^{t+\varepsilon+\delta}_{t+\delta}\left<H_{v_\delta}(s,\Theta_s,p_s,q_s),v_{s-\delta}-u_{s-\delta}\right>ds\geq0.\end{aligned}$$ That means $$\begin{aligned}
&&\frac{1}{\varepsilon}\mathbb{E}\int^{t+\varepsilon}_t\left<H_v(s,\Theta_s,p_s,q_s)+\mathbb{E}^{\mathcal{F}_s}\left[H_{v_\delta}(s,\Theta_s,p_s,q_s)|_{s+\delta}\right],v_s-u_s\right>ds\nonumber\\
&\geq& 0\end{aligned}$$ Letting $\varepsilon\rightarrow 0+$, by Lebesgue differential theorem, we have $$\begin{aligned}
&&\mathbb{E}\left<H_v(t,\Theta_t,p_t,q_t)+\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],v_t-u_t\right>\geq 0 \ \ a.e.\nonumber.\end{aligned}$$ Now, let $v\in \mathcal{U}$ be a selected element and $A$ an arbitrary element of $\sigma$-algebra $\mathcal{F}_t$, set $w_t=v\textbf{1}_A+u_t\textbf{1}_{A^c}$. Clearly, $w_t$ is an admissible control and for all $A\in \mathcal{F}_t,$ we obtain $$\begin{aligned}
&&\mathbb{E}\left<H_v(t,\Theta_t,p_t,q_t)+\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],w_t-u_t\right>\nonumber\\
&=&\mathbb{E}\left[\left<H_v(t,\Theta_t,p_t,q_t)+\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],v-u_t\right>\textbf{1}_A\right]\nonumber\\
&\geq& 0 \ \ a.e.\nonumber,\end{aligned}$$ which implies $$\begin{aligned}
&&\mathbb{E}\left[\left<H_v(t,\Theta_t,p_t,q_t)+\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],v-u_t\right>|\mathcal{F}_t\right]\nonumber\\
&=&\left<H_v(t,\Theta_t,p_t,q_t)+\mathbb{E}^{\mathcal{F}_t}\left[H_{v_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],v-u_t\right>\geq 0 \ \ a.e.\nonumber.\end{aligned}$$
Then we study the sufficient conditions.
(Sufficient Conditions for the Optimality of Control). Let Hypothesis \[2H1\] hold and let $u$ is the control satisfies (\[2eq43\]) and $(p,q)$ be the unique solution of (\[2eq41\]). We further assume $\Phi(x,\mu ), H(t,x,x_\delta,\mu,\mu_\delta, p_t,q_t,v,v_\delta)$ are convex respect to $(x,\mu)$ and $(x,x_\delta,\mu,\mu_\delta,v,v_\delta)$. Then $u$ is the optimal control of our control problem.
For any $v\in \mathcal{U}$, we have $$\begin{aligned}
&&J(v)-J(u)\nonumber \\
&=&\mathbb{E}\big[\Phi(X^v_T,P_{X^v_T})-\Phi(X^u_T,P_{X^u_T})\big]\nonumber\\
&&+\mathbb{E}\int^T_0 \big[h(t,X^v_t,P_{X^v_t},
v_t,v_{t-\delta} )-h(t,{X}^{u}_t,P_{X^u_t},
u_t ,u_{t-\delta})\big]dt.\end{aligned}$$ Since $\Phi$ is convex with respect to $x$. we get $$\begin{aligned}
&&\Phi(X^{v}_T,P_{X^v_T})-\Phi(X^{u}_T,P_{X^u_T})\nonumber \\
&\geq &\Phi_x(X^{u}_T,P_{X^u_T})(X^{v}_T-X^u_T)\nonumber\\
&&+\mathbb{E}'\big[\Phi_\mu(X^{u}_T,P_{X^u_T},(X^{u}_T)')(X^{v}_T-X^u_T)'\big]. \end{aligned}$$ Consequently $$\begin{aligned}
&&J(v)-J(u)\nonumber\\
&=&\mathbb{E}\big\{\Phi_x(X^{u}_T,P_{X^u_T})(X^{v}_T-X^u_T)\nonumber\\
&&+\mathbb{E}'\big[\Phi_x(X^{u}_T,P_{X^u_T},(X^{u}_T)')(X^{v}_T-X^u_T)'\big]\big\}\nonumber \\
&&+\mathbb{E}\int^T_0 \big[h(t,X^v_t,P_{X^v_t},
v_t,v_{t-\delta} )-h(t,{X}^{u}_t,P_{X^u_t},
u_t ,u_{t-\delta})\big]dt.\end{aligned}$$ By applying Itô’s formula to $\left<p_t,X^{v}_t-{X}^u_t\right>$ and taking the expectation, we obtain
$$\begin{aligned}
&&J(v)-J(u)\\
&\geq &\mathbb{E}\int^T_0\big [H(t,X^{v}_t,X^{v}_{t-\delta},P_{X^v_t},P_{X^v_{t-\delta}},
v_t,v_{t-\delta},p_t,q_t )-H(t,\Theta_t,p_t,q_t )\big]dt\nonumber \\
&&-\mathbb{E}\int^T_0 \bigg\{\left<H_x(t,\Theta_t,p_t,q_t),X^{v}_t-{X}^u_t\right>\nonumber \\
&&+\mathbb{E}'\left[\left< H_\mu(t,\Theta_t,({X}^{u}_t)',p_t,q_t,
),(X^{v}_t-{X}^u_t)'\right>\right]\bigg\}dt\nonumber\\
&&-\mathbb{E}\int^T_0 \bigg\{\left<\mathbb{E}^{\mathcal{F}_t}\left[H_{x_\delta}(t,\Theta_t,p_t,q_t)|_{t+\delta}\right],X^{v}_t-{X}^u_t\right>\nonumber \\
&&+\mathbb{E}'\left[\left<\mathbb{E}^{\mathcal{F}_t}\left[ H_{\mu_\delta}(t,\Theta_t,({X}^{u}_t)',p_t,q_t,
)|_{t+\delta}\right],(X^{v}_t-{X}^u_t)'\right>\right]\bigg\}dt\end{aligned}$$
Since $H$ is convex with respect to $(x,x_\delta,\mu,\mu_\delta,v,v_\delta)$ The use of Clark generalized gradient of $H$, evaluated at $({X}^u_t,{X}^u_{t-\delta},P_{{X}^u_t},P_{{X}^u_{t-\delta}},u_t,u_{t-\delta})$, yields
$$\begin{aligned}
&&H(t,X^{v}_t,X^{v}_{t-\delta},P_{X^v_t},P_{X^v_{t-\delta}},
v_t,v_{t-\delta},p_t,q_t )-H(t,\Theta_t,p_t,q_t )\nonumber \\
&\geq &\left<H_x(t,\Theta_t,p_t,q_t),X^{v}_t-{X}^u_t\right>\nonumber \\
&&+\mathbb{E}'\left[\left< H_\mu(t,\Theta_t,({X}^{u}_t)',p_t,q_t,
),(X^{v}_t-{X}^u_t)'\right>\right]\nonumber\\
&&+\left<H_{x_\delta}(t,\Theta_t,p_t,q_t),X^{v}_{t-\delta}-{X}^u_{t-\delta}\right>\nonumber \\
&&+\mathbb{E}'\left[\left< H_{\mu_\delta}(t,\Theta_t,({X}^{u}_t)',p_t,q_t
),(X^{v}_{t-\delta}-{X}^u_{t-\delta})'\right>\right]\nonumber\\
&&+ \left<H_v(t,\Theta_t,p_t,q_t),v_t-u_t\right>+\left<H_{v_\delta}(t,\Theta_t,p_t,q_t),v_{t-\delta}-u_{t-\delta}\right>.\nonumber\end{aligned}$$
Thus, by maximum condition (\[2eq43\]), we have $$\begin{aligned}
&&J(v)-J(u)\nonumber \\
&\geq &\mathbb{E}\int^T_0 \left<H_v(t,\Theta_t,p_t,q_t),v_t-u_t\right>dt\nonumber\\
&&+\mathbb{E}\int^T_0 \left<H_{v_\delta}(t,\Theta_t,p_t,q_t),v_{t-\delta}-u_{t-\delta}\right>dt\nonumber\\
&\geq& 0.\end{aligned}$$
The above equality complete the proof.
[99]{}
V. Arkin, I. Saksonov, Necessary optimality conditions for stochastic differential equations. Soviet Mathemetical Doklady, 20, 1–5, 1979.
A. Bensoussan, Lectures on stochastic control. In S. K. Mitter, A. Moro (Eds.), Springer lecture notes in mathematics: Vol. 972. Nonlinear filtering and stochastic control. Berlin: Springer, 1982.
J. M. Bismut, An introductory approach to duality in optimal stochastic control. SIAM Review, 20, 62–78, 1978.
R. Buckdahn, B. Djehiche, J. Li, [A General Stochastic Maximum Principle for SDEs of Mean-field Type]{}, Appl Math Optim 64: 197. doi:10.1007/s00245-011-9136-y, 2011.
R. Buckdahn, J. Li, S. Peng, [Mean-field backward stochastic differential equations and related partial differential equations]{}, Stoch. Proc. App. 119 3133-3154, 2009.
R. Buckdahn, J. Li, S. Peng, et al. Mean-field stochastic differential equations and associated PDEs\[J\]. arXiv preprint arXiv:1407.1215, 2014.
P. Cardaliaguet, [Weak solutions for first order mean field games with local coupling]{}, Analysis and Geometry in Control Theory and its Applications Volume 11 of the series Springer INdAM Series pp 111-158, 2015.
L. Chen, Z. Wu, Maximum principle for the stochastic optimal control problem with delay and application\[J\]. Automatica, 46(6): 1074-1080, 2010.
Du H, Huang J, Qin Y. A stochastic maximum principle for delayed mean-field stochastic differential equations and its applications\[J\]. IEEE Transactions on Automatic Control, 58(12): 3212-3217, 2013.
U. G. Haussmann, A stochastic maximum principle for optimal control of diffusions. Essex, UK: Longman Scientific and Technical, 1986.
H. J. Kushner, On the stochastic maximum principle: fixed time of control. Journal of Mathematical Analysis and Applications, 11, 78–92, 1965.
H. J. Kushner, Necessary conditions for continuous parameter stochastic optimization problems. SIAM Journal of Control, 10, 550–565, 1972.
J. Li, [Stochastic maximum principle in the mean-field controls]{}. Automatica, 48 366-373, 2012.
E. Pardoux, S. Peng, [Adapted solution of a backward stochastic differential equation]{}. System Control Lett. 14 (1-2) 61-74, 1990.
S. Peng, [A general stochastic maximum principle for optimal control problem]{}, SIAM J. Control and Optimization. Vol. 28, No. 4, 966-979, July, 1990.
S. Peng, Nonlinear expectations, nonlinear evaluations and risk measures. In Stochastic Methods in Finance. Lecture Notes in Math. 1856 165–253. Springer, Berlin. MR2113723, 2004.
S. Peng, Z. Yang, Anticipated backward stochastic differential equations. Ann. Probab. 37, No. 3, 877-902, 2009.
L. S. Pontrvagin, V. G. Boltyanskii, R. V. Gamkerlidze, E. F. Mischenko. The Mathematical Theory of Optimal Control Processes, John Wiley, New York, 1962.
[^1]: Department of Mathematics, Faculty of Science and Technology, University of Macau, Macau, China. E-mail: [email protected], [email protected] and jy\[email protected].
[^2]: Research supported partially by FDCT 025/2016/A1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'For a positive proportion of primes $p$ and $q$, we prove that ${\mathbb{Z}}$ is Diophantine in the ring of integers of ${\mathbb{Q}}(\sqrt[3]{p},\sqrt{-q})$. This provides a new and explicit infinite family of number fields $K$ such that Hilbert’s tenth problem for $O_K$ is unsolvable. Our methods use Iwasawa theory and congruences of Heegner points in order to obtain suitable rank stability properties for elliptic curves.'
address:
- ' Departamento de MatemáticasPontificia Universidad Católica de ChileFacultad de Matemáticas4860 Av. Vicuña MackennaMacul, RM, Chile'
- ' Departamento de MatemáticasPontificia Universidad Católica de ChileFacultad de Matemáticas4860 Av. Vicuña MackennaMacul, RM, Chile'
author:
- 'Natalia Garcia-Fritz'
- Hector Pasten
title: 'Towards Hilbert’s Tenth Problem for rings of integers through Iwasawa theory and Heegner points'
---
[^1]
Introduction
============
A celebrated result due to Matiyasevich [@Matiyasevich] after the work of Davis, Putnam, and Robinson [@DPR], shows that computably enumerable sets over ${\mathbb{Z}}$ are exactly the same as Diophantine sets over ${\mathbb{Z}}$. As a consequence, Hilbert’s tenth problem is unsolvable: namely, there is no algorithm (Turing machine) that takes as input polynomial equations over ${\mathbb{Z}}$ and decides whether they have integer solutions.
A natural extension is the analogue of Hilbert’s tenth problem for rings of integers of number fields, which remains as one of the main open problems in the area. In the direction of a negative solution, the following conjecture formulated by Denef and Lipshitz [@DenLip] is widely believed (the notion of Diophantine set is recalled in Section \[SecDioph\]):
\[MainConj\] Let $K$ be a number field $K$. Then ${\mathbb{Z}}$ is a Diophantine subset of $O_K$.
By a standard argument, if ${\mathbb{Z}}$ is Diophantine in $O_K$ for a number field $K$, then the analogue of Hilbert’s tenth problem for $O_K$ has a negative solution, and moreover, the Diophantine sets over $O_K$ are the same as computably enumerable subsets. Thus, the efforts have concentrated in proving new cases of Conjecture \[MainConj\].
However, progress on Conjecture \[MainConj\] has been slow and extremely difficult. First, one has the following known cases dating back to the eighties:
- $K$ is totally real or a quadratic extension of a totally real field [@DenLip; @Denef]
- $[K:{\mathbb{Q}}]=4$, $K$ is not totally real, and $K/{\mathbb{Q}}$ has a proper intermediate field [@DenLip]
- $K$ has exactly one complex place [@Pheidas; @ShlH10; @Videla]
- $K$ is contained in a number field $L$ such that ${\mathbb{Z}}$ is Diophantine in $O_L$; in particular, this holds when $K/{\mathbb{Q}}$ is abelian [@ShaShl]
Later work by Poonen [@Poonen], Cornelissen-Pheidas-Zahidi [@CPZ], and Shlapentokh [@ShlE] reduced Conjecture \[MainConj\] to the existence of elliptic curves preserving its (positive) rank in suitable extensions of number fields (cf. Section \[SecDioph\]). This allowed to verify Conjecture \[MainConj\] on a case-by-case basis for concrete examples of number fields by exhibiting a suitable elliptic curve (see the Paragraph \[SubSecIntDioph\] for the worked-out example $K={\mathbb{Q}}(\sqrt[5]{2})$). Besides specific examples, the elliptic curve criteria also led to the following more recent result by Mazur and Rubin:
- For any number field $F$ for which ${\mathbb{Z}}$ is Diophantine in $O_F$, there is a positive proportion of primes $\ell$ such that for all $n\ge 1$, there are infinitely many choices of $K$ as a cyclic $\ell^n$-extension of $F$ for which ${\mathbb{Z}}$ is Diophantine in $O_K$. (cf. [@MazRubDS]; see also [@MazRubH10]).
This concludes our summary of known cases of Conjecture \[MainConj\].
In this work we develop a method to attack Conjecture \[MainConj\] which allows us to prove it in several new cases. Our main result is the following:
\[ThmMainIntro\] There are explicit Chebotarev sets of primes ${\mathscr{P}}$ and ${\mathscr{Q}}$ having positive density in the primes, such that for all $p\in {\mathscr{P}}$ and all $q\in {\mathscr{Q}}$ we have that ${\mathbb{Z}}$ is Diophantine in the ring of integers of ${\mathbb{Q}}(\sqrt[3]{p}, \sqrt{-q})$. In particular, for these fields $K={\mathbb{Q}}(\sqrt[3]{p}, \sqrt{-q})$, the analogue of Hilbert’s tenth problem for $O_K$ is unsolvable.
The sets ${\mathscr{P}}$ and ${\mathscr{Q}}$ can be chosen to have densities $5/16>31\%$ and $1/12>8\%$ respectively.
The sets of prime numbers ${\mathscr{P}}$ and ${\mathscr{Q}}$ from our construction will be made explicit in Section \[SecProof\].
Theorem \[ThmMainIntro\] precisely covers one of the simplest kinds of number fields which is out of the scope of all the available results on Hilbert’s tenth problem for rings of integers —see Paragraph \[SubsecComparison\] for details.
Besides the fact that we prove new cases of Conjecture \[MainConj\], the main novelty in our work is the method of proof for Theorem \[ThmMainIntro\]. Using Shlapentokh’s elliptic curve criterion and the result of Pheidas, Shlapentokh and Videla quoted above in item (iii), we reduce the problem to the existence of a suitable elliptic curve $E$ over ${\mathbb{Q}}$ with ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[3]{p}))=0$ and ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt{-q}))>0$. The first condition is achieved by a study of the variation of cyclotomic Iwasawa invariants under base-change, while the second one is achieved by means of a recent result by Kriz and Li [@KrizLi] on congruences of Heegner points in the context of the celebrated Goldfeld’s conjecture. Both methods —Iwasawa theory and Heegner points— impose a series of technical conditions on the admissible elliptic curves $E$, and a critical aspect of the proof of Theorem \[ThmMainIntro\] is to make sure that there exist elliptic curves that are admissible for both methods simultaneously.
Let us remark that the approach by Mazur and Rubin (cf. Item (v) above) is substantially different to ours: they achieve the necessary rank-stability conditions by delicate cohomological computations to modify Selmer ranks under quadratic twists.
We conclude this introduction by recalling that there is strong evidence for Conjecture \[MainConj\], at least if one assumes some standard conjectures on elliptic curves. On the one hand, Mazur and Rubin [@MazRubH10] proved that it follows from the squareness conjecture for the $2$-torsion of Shafarevich-Tate groups of elliptic curves over number fields. On the other hand, Murty and Pasten [@MurtyPasten] proved that it follows from rank aspects of the Birch and Swinnerton-Dyer conjecture.
Preliminaries on Iwasawa theory
===============================
Our main results are relevant on logic aspects of number theory but part of our arguments require some Iwasawa theory. So, it might be useful to include here a brief reminder of the latter subject. Other than for the purpose of checking the notation, experts can safely skip this section.
In this section $\ell$ denotes a prime number. We specialize to the case of cyclotomic Iwasawa theory for elliptic curves, which suffices for our purposes. All results discussed in this section can be found in [@MazurIwa] and [@Greenberg].
Algebra {#ParAlg}
-------
It is a standard fact that if $X$ is a finitely generated torsion ${\mathbb{Z}}_{\ell}[[T]]$-module, then $X$ is pseudo-isomorphic[^2] to $$\bigoplus_{i}{\mathbb{Z}}_{\ell}[[T]]/(\ell^{\mu_i})\oplus \bigoplus_j {\mathbb{Z}}_{\ell}[[T]]/(f_j^{m_j})$$ where $\mu_i$ are positive integers and $f_j\in{\mathbb{Z}}_{\ell}[T]$ are monic, irreducible polynomials such that all non-leading coefficients are in $\ell{\mathbb{Z}}_{\ell}$.
The above decomposition is unique up to order, and one defines the following invariants of $X$:
- $\mu_X=\sum_i \mu_i$ the $\mu$-invariant
- $f_X= \ell^{\mu_X}\prod_j f_j^{m_j} $ the characteristic polynomial
- $\lambda_X=\deg f_X$ the $\lambda$-invariant.
Iwasawa algebras and modules {#ParGal}
----------------------------
For a number field $k$, we let $k_\infty$ be the maximal cyclotomic ${\mathbb{Z}}_\ell$-extension of $k$. For each $m\ge 0$ we let $k_m/k$ for the unique cyclotomic ${\mathbb{Z}}/\ell^m{\mathbb{Z}}$-extension of $k$ contained in $k_\infty$, and we observe that $k_0=k$ and $k_\infty=\cup_m k_m$. The Galois group $\Gamma_k$ of $k_\infty/k$ is (non-canonically) isomorphic to ${\mathbb{Z}}_{\ell}$ as a profinite group. The *Iwasawa algebra* $\Lambda_k={\mathbb{Z}}_{\ell}[[\Gamma_k]]$ is the profinite completion of the group algebra ${\mathbb{Z}}_{\ell}[\Gamma_k]$. The choice of a topological generator $\gamma\in\Gamma_k$ determines a continuous ${\mathbb{Z}}_{\ell}$-algebra isomorphism $\Psi_{k,\gamma}:\Lambda_k\to {\mathbb{Z}}_{\ell}[[T]]$ by the rule $\gamma\mapsto 1+T$.
Thus, one can study finitely generated torsion $\Lambda_k$-modules by means of the classification theorem of the previous paragraph. Using Selmer groups of elliptic curves, one can construct certain finitely generated modules $X(E/k_\infty)$ that are believed to be $\Lambda_k$-torsion.
Selmer groups
-------------
Let $k$ be a number field and let $E$ be an elliptic curve over $k$. For any algebraic extension $L/k$ (not necessarily finite) the $\ell$-primary part of the Selmer group of $E$ over $L$ is denoted by ${\mathrm{Sel}}_{\ell^\infty}(E/L)$. It fits in the exact sequence $$\label{EqnSha}
0\to E(L)\otimes_{\mathbb{Z}}{\mathbb{Q}}_{\ell}/{\mathbb{Z}}_{\ell}\to {\mathrm{Sel}}_{\ell^\infty}(E/L)\to \Sha(E/L)[\ell^\infty]\to 0$$ where $\Sha(E/L)$ is the Shafarevich-Tate group.
It is known that ${\mathrm{Sel}}_{\ell^\infty}(E/k)\simeq ({\mathbb{Q}}_\ell/{\mathbb{Z}}_\ell)^{\rho_\ell(E/k)}\oplus G$ for certain integer $\rho_\ell(E/k)\ge 0$, where $G$ is a finite group. The number $\rho_\ell(E/k)$ is the co-rank of ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ and the previous exact sequence gives ${\mathrm{rank}\,}E(k)\le \rho_\ell(E/k)$. Of course, if $\Sha(E/k)[\ell^\infty]$ is finite (as conjectured) then ${\mathrm{rank}\,}E(k)= \rho_\ell(E/k)$.
A foundational result of Mazur allows one to recover valuable arithmetic information about $E$ over all the number fields $k_m$ from the $\Gamma_k$-action on ${\mathrm{Sel}}_{\ell^\infty}(E/k_\infty)$.
Let $E$ be an elliptic curve over $k$ having good ordinary reduction at each prime of $k$ above $\ell$. The natural maps $${\mathrm{Sel}}_{\ell^\infty}(E/k_m)\to {\mathrm{Sel}}_{\ell^\infty}(E/k_\infty)^{{\rm Gal}(k_\infty/k_m)}$$ have finite kernels and cokernels, whose orders are bounded as $m$ varies.
Iwasawa modules attached to elliptic curves
-------------------------------------------
The group ${\mathrm{Sel}}_{\ell^\infty}(E/k_\infty)$ has a natural $\Lambda_k$-module structure and it is topologically discrete. Giving ${\mathbb{Q}}_{\ell}/{\mathbb{Z}}_{\ell}$ the trivial $\Gamma_k$-action, we have an induced $\Lambda_k$-module structure on $$X(E/k_\infty):={\mathrm{Hom}}_{\mathrm{cont}} ({\mathrm{Sel}}_{\ell^\infty}(E/k_\infty), {\mathbb{Q}}_{\ell}/{\mathbb{Z}}_{\ell}).$$ By Pontryagin duality, $X(E/k_\infty)$ is compact. It is a standard result that $X(E/k_\infty)$ is finitely generated as a $\Lambda_k$-module.
If $E$ has good ordinary reduction at each prime of $k$ above $\ell$, then $X(E/k_\infty)$ is a torsion $\Lambda_k$-module.
The general case of this conjecture remains open. For us it suffices to know the following result, which is a consequence of Mazur’s control theorem.
\[ThmMazurFin\] If $E$ has good ordinary reduction at each prime of $k$ above $\ell$, and if ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is finite, then $X(E/k_\infty)$ is a torsion $\Lambda_k$-module.
Iwasawa invariants
------------------
Suppose that $X(E/k_\infty)$ is a torsion $\Lambda_k$-module and fix a topological generator $\gamma\in \Gamma_k$. Then we can regard $X(E/k_\infty)$ as a finitely generated torsion ${\mathbb{Z}}_{\ell}[[T]]$-module via $\Psi_{k,\gamma}$ (cf. Paragraph \[ParGal\]), and we have the associated invariants $\mu_{E/k}:=\mu_{X(E/k_\infty)}$, $\lambda_{E/k}:=\lambda_{X(E/k_\infty)}$, and $f_{E/k}:=f_{X(E/k_\infty)}$. The $\mu$ and $\lambda$ invariants are well-defined integers that only depend on $\ell$ and $E/k$. However, the characteristic polynomial $f_{E/k}\in{\mathbb{Z}}_\ell[T]$ also depends on the choice of topological generator $\gamma\in\Gamma_k$.
Directly from the definitions one has
\[Lemmalm\] Suppose that $X(E/k_\infty)$ is a torsion $\Lambda_k$-module. Then:
- $\mu_{E/k}=0$ if and only if $X(E/k_\infty)$ is finitely generated as a ${\mathbb{Z}}_{\ell}$-module,
- $\mu_{E/k}=\lambda_{E/k}=0$ if and only if $X(E/k_\infty)$ is a finite group.
It is quite common that the $\mu$-invariant vanishes, but this is not always the case. On the other hand, we will be mostly interested on the $\lambda$-invariant due to the following important consequence of Mazur’s control theorem (cf. Theorem 1.9 in [@Greenberg]).
\[ThmLambdaRk\] Suppose that $E$ has good ordinary reduction at each prime of $k$ above $\ell$ and that $X(E/k_\infty)$ is a torsion $\Lambda_k$-module. Then for each $m\ge 0$ we have $${\mathrm{rank}\,}E(k_m)\le \rho_\ell(E/k_m)\le \lambda_{E/k}.$$
The value of $f_{E/k}(0)$
-------------------------
We now discuss the characteristic polynomial $f_{E/k}$ assuming, of course, that $X(E/k_\infty)$ is $\Lambda_k$-torsion and that a topological generator $\gamma\in\Gamma_k$ is chosen.
\[Propf0\] Suppose that $E$ has good ordinary reduction at each prime of $k$ above $\ell$ and that $X(E/k_\infty)$ is a torsion $\Lambda_k$-module. We have that ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is finite if and only if $f_{E/k}(0)\ne 0$.
By Mazur’s control theorem, ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is finite if and only if ${\mathrm{Sel}}_{\ell^\infty}(E/k_\infty)^{\Gamma_k}$ is finite. Choosing a topological generator $\gamma\in\Gamma_k$ and recalling that $\Psi_{k,\gamma}(\gamma)=1+T$, we see that the latter is equivalent to the finiteness of $X(E/k_\infty)/T\cdot X(E/k_\infty)$ by duality between ${\mathrm{Sel}}_{\ell^\infty}(E/k_\infty)$ and $X(E/k_\infty)$. Finally, we note that $X(E/k_\infty)/T\cdot X(E/k_\infty)$ is finite if and only if $T\nmid f_{E/k}$ (cf. Paragraph \[ParAlg\]).
It is of great interest to evaluate $f_{E/k}(0)\in {\mathbb{Z}}_{\ell}$ when it is non-zero. The precise value depends on the choice of topological generator $\gamma\in\Gamma_k$, but nevertheless the $\ell$-adic valuation of $f_{E/k}(0)$ turns out to depend only on $\ell$, the number field $k$, and the elliptic curve $E/k$. Before giving the formula, we need some notation.
Given a non-archimedean place $v$ of $k$, we write ${\mathbb{F}}_v$ for the residue field and $k_v$ for the completion of $k$ at $v$. If $E$ is an elliptic curve over $k$, the reduction of $E$ at $v$ is denoted by $\tilde{E}_v$ and its non-singular locus is $\tilde{E}_v^{ns}$. We write $E(k_v)_0$ for the subgroup of $E(k_v)$ consisting of local points whose reduction at $v$ is in $\tilde{E}_v^{ns}$. The Tamagawa factor at $v$ is defined by $c_v(E/k)=[E(k_v):E(k_v)_0]$, and the the product $\prod_v c_v(E/k)$ over all non-archimedean places $v$ of $k$ is denoted by ${\mathrm{Tam}}(E/k)$. Let us write $a\sim_{\ell} b$ if $a,b\in {\mathbb{Q}}_{\ell}$ have the same $\ell$-adic valuation.
Note that the finiteness assumption on ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ not only implies $f_{E/k}(0)\ne 0$ (cf. Theorem \[ThmMazurFin\] and Proposition \[Propf0\]), but also, it implies that both $\Sha(E/k)[\ell^{\infty}]$ and $E(k)$ are finite (cf. the exact sequence ). So the formula in the next result only involves non-zero finite numbers. See also [@PR] and [@Schneider].
\[Thmf0\] Suppose that $E$ has good ordinary reduction at each prime of $k$ above $\ell$ and that ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is finite. Then $f_{E/k}(0)\in{\mathbb{Z}}_\ell$ is non-zero and it satisfies $$f_{E/k}(0)\sim_{\ell} \frac{{\mathrm{Tam}}(E/k)\cdot \#\Sha(E/k)[\ell^{\infty}] }{\#E(k)^2} \cdot \prod_{v|\ell} \#\tilde{E}_v({\mathbb{F}}_v)^2.$$
The proof of Theorem \[Thmf0\] relies on the theory of the cyclotomic Euler characteristic, which we don’t review here.
A general construction of integrally Diophantine extensions {#SecDioph}
===========================================================
Integrally Diophantine extensions {#SubSecIntDioph}
---------------------------------
Given a commutative unitary ring $A$ and a positive integer $n$, let us recall that a subset $S\subseteq A^n$ is *Diophantine* over $A$ if for some $m\ge 0$ there are polynomials $F_1,\ldots, F_k \in A[x_1,...,x_n,y_1,...,y_m]$ such that $$S=\{{\bf a}\in A^n : \exists {\bf b}\in A^m \mbox{ such that for each }1\le j\le k \mbox{ we have } F_j({\bf a},{\bf b})=0\}.$$
An extension of number fields $K/F$ is *integrally Diophantine* if $O_F$ is Diophantine in $O_K$. It is a standard fact that if $K/F/L$ is a tower of number fields and both $F/L$ and $K/F$ are integrally Diophantine, then so is $K/L$. In particular,
\[LemmaTransfer\] Let $K/F$ be an extension of number fields. If $K/F$ is integrally Diophantine and ${\mathbb{Z}}$ is Diophantine in $O_F$, then ${\mathbb{Z}}$ is Diophantine in $O_K$.
As discussed in the introduction, there are elliptic curve criteria developed by Poonen, Cornelissen-Pheidas-Zahidi, and Shlapentokh to prove that a given extension of number fields is integrally Diophantine. For our purposes, let us recall here Shlapentokh’s criterion.
\[ThmShl\] Let $K/F$ be an extension of number fields. Suppose that there is an elliptic curve $E$ defined over $F$ such that ${\mathrm{rank}\,}E(K)={\mathrm{rank}\,}E(F)>0$. Then $K/F$ is integrally Diophantine.
This result can be applied on a case-by-case basis to concrete examples of number fields, but it is not known at present how to systematically apply it in general —at least, not without assuming some conjecture on elliptic curves (cf. [@MazRubH10] and [@MurtyPasten]). For instance, the number field ${\mathbb{Q}}(\sqrt[5]{2})$ is not covered by the results (i)-(v) quoted in the introduction. However, the elliptic curve $E$ of affine equation $y^2 + xy = x^3 - x^2 - x + 1$ (Cremona label $58a1$) satisfies $${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[5]{2}))={\mathrm{rank}\,}E({\mathbb{Q}})=1,$$ as it can be readily checked on Sage. By Shlapentokh’s theorem we get that ${\mathbb{Q}}(\sqrt[5]{2})/{\mathbb{Q}}$ is integrally Diophantine (in fact, already the results of Poonen [@Poonen] or Cornelissen-Pheidas-Zahidi [@CPZ] suffice here). That is, Conjecture \[MainConj\] holds for $K={\mathbb{Q}}(\sqrt[5]{2})$.
A general construction
----------------------
The general case of Conjecture \[MainConj\] —and in fact, of Hilbert’s tenth problem for rings of integers of number fields— remains open, as discussed in the introduction. It is therefore of great importance to develop tools that permit to construct integrally Diophantine extensions. The next result, despite its simple proof, gives a general way to construct such extensions and to prove new cases of Conjecture \[MainConj\].
\[PropMain\] Let $F/M$ and $L/M$ be extensions of number fields with $L/M$ quadratic. Let $K=F.L$ be the compositum of $F$ and $L$ over $M$. Suppose that there is an elliptic curve $E$ over $M$ satisfying the conditions:
- ${\mathrm{rank}\,}E(F)=0$
- ${\mathrm{rank}\,}E(L)>0$.
Then $K/F$ is integrally Diophantine.
Let $E^L$ be the elliptic curve over $M$ defined as the quadratic twist of $E$ by $L$. By (i) and (ii) we see that $L$ is not contained in $F$, so the extension $K/F$ is quadratic and $${\mathrm{rank}\,}E^L(K)={\mathrm{rank}\,}E(F)+ {\mathrm{rank}\,}E^L(F)={\mathrm{rank}\,}E^L(F).$$ Since ${\mathrm{rank}\,}E^L(K) ={\mathrm{rank}\,}E(K)\ge {\mathrm{rank}\,}E(L)>0$ we can apply Theorem \[ThmShl\] to the extension $K/F$ with the elliptic curve $E^L$ over $F$.
As an example of how this construction readily leads to new cases of Conjecture \[MainConj\], let us point out the following simple consequence.
\[PropEasy\] Let $p$ be a prime of the form $p\equiv 3\bmod 4$. Then ${\mathbb{Z}}$ is Diophantine in the ring of integers of ${\mathbb{Q}}(\sqrt[5]{2}, \sqrt{p})$.
We apply Proposition \[PropMain\] with $M={\mathbb{Q}}$, $F={\mathbb{Q}}(\sqrt[5]{2})$, and $L={\mathbb{Q}}(\sqrt{p})$ for $p\equiv 3\bmod 4$ a prime number. Consider the elliptic curve $E$ over ${\mathbb{Q}}$ with affine equation $y^2=x^3-4x$; this is the twist by $2$ of the celebrated Congruent Number elliptic curve $y^2=x^3-x$. By classical results due to Heegner [@Heegner] and Birch [@Birch] (see also [@Monsky]) we have that $2p$ is a congruent number when $p\equiv 3\bmod 4$. Hence ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt{p}))>0$ for these primes $p$. On the other hand, a direct computation on Sage shows that ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[5]{2}))=0$. Therefore $K={\mathbb{Q}}(\sqrt[5]{2}, \sqrt{p})$ is integrally Diophantine over $F={\mathbb{Q}}(\sqrt[5]{2})$. In Paragraph \[SubSecIntDioph\] we already checked that ${\mathbb{Q}}(\sqrt[5]{2})/{\mathbb{Q}}$ is integrally Diophantine, hence $K/{\mathbb{Q}}$ also is (cf. Lemma \[LemmaTransfer\]).
In Proposition \[PropEasy\], all the number fields that we obtain are quadratic extensions of a single number field —namely, of ${\mathbb{Q}}(\sqrt[5]{2})$— and of course one can obtain many other results of this sort thanks to Proposition \[PropMain\]. Theorem \[ThmMainIntro\] instead is much more delicate and the proof lies deeper, as it concerns number fields of the form ${\mathbb{Q}}(\sqrt[3]{p},\sqrt{-q})$ where the two parameters $p$ and $q$ can be chosen independently of each other.
Comparison with other available results {#SubsecComparison}
---------------------------------------
The following elementary lemma will help us to check that the families of number fields $K$ for which we prove that ${\mathbb{Z}}$ is Diophantine in $O_K$ are indeed new.
\[LemmaNew\] Let $\ell>2$ be a prime and let $F/{\mathbb{Q}}$ be a degree $\ell$ extension which is not totally real. Then $F$ is not contained in a quadratic extension of a totally real number field.
Suppose that $H/N$ is a quadratic extension with $N$ a totally real number field, such that $F\subseteq H$. Observe that $N$ and $F$ are linearly disjoint over ${\mathbb{Q}}$, for otherwise an element $\gamma\in F-{\mathbb{Q}}$ would be conjugate to an element of $N$, but $F={\mathbb{Q}}(\gamma)$ is not totally real. It follows that the compositum $F.N$ has degree $\ell$ over $N$, but this is not possible since $F.N\subseteq H$ and $[H:N]=2$.
For instance, let us verify that the number fields in Proposition \[PropEasy\] are not included in number fields covered by the results in items (i), (ii), (iii) from the introduction: For a prime $p$, the number field ${\mathbb{Q}}(\sqrt[5]{2}, \sqrt{p})$ has degree $10$ over ${\mathbb{Q}}$ and it has $4$ complex places, hence $K$ is not contained in fields for which (ii) or (iii) apply. By Lemma \[LemmaNew\], $K$ is not contained in a quadratic extension of a totally real field, so (i) does not apply.
Regarding item (v), we note that the quadratic extensions ${\mathbb{Q}}(\sqrt[5]{2}, \sqrt{p})/{\mathbb{Q}}(\sqrt[5]{2})$ are obtained by adjoining the square root of a rational prime, as opposed to using unrestricted primes from ${\mathbb{Q}}(\sqrt[5]{2})$. So the method of proof in [@MazRubH10; @MazRubDS] (cf. item (v) above) does not apply either, as the auxiliary primes required by the methods of *loc. cit.* cannot be guaranteed to be chosen in ${\mathbb{Q}}$.
Finally, we remark that Lemma \[LemmaNew\] allows us to check in a similar way that the number fields considered in our main result Theorem \[ThmMainIntro\] are out of the scope of the available results in the literature. In fact, if $p$ and $q$ are prime numbers, the number field $\mathbb{Q}(\sqrt[3]{p},\sqrt{-q})$ has degree $6$ over $\mathbb{Q}$, it has three complex places, and it contains the non-totally real number field $\mathbb{Q}(\sqrt[3]{p})$.
Preserving rank zero: Iwasawa theory
====================================
For a positive integer $n$, the set of complex $n$-th roots of $1$ is denoted by $\mu_n$. The purpose of this section is to prove the following result
\[ThmRkZero\] Let $\ell>2$ be a prime. Let $E$ be an elliptic curve over ${\mathbb{Q}}$ of conductor $N$ and assume the following:
1. \[A\] $E$ has good ordinary reduction at $\ell$
2. \[B\] the residual Galois representation $\rho_{E[\ell]}:G_{\mathbb{Q}}\to {\mathrm{Aut}}(E[\ell])$ is surjective
3. \[C\] ${\mathrm{rank}\,}E({\mathbb{Q}}(\mu_\ell))=0$
4. \[D\] $\Sha(E/{\mathbb{Q}}(\mu_\ell))[\ell]=(0)$
5. \[E\] $\ell\nmid {\mathrm{Tam}}(E/{\mathbb{Q}}(\mu_\ell))\cdot \#\tilde{E}_\ell({\mathbb{F}}_\ell)$.
Consider the set of prime numbers $${\mathscr{P}}(E,\ell)=\{p : p\nmid N\mbox{, }p \equiv 1\bmod \ell\mbox{, and }a_p(E)\not\equiv 2\bmod \ell\}.$$ Then ${\mathscr{P}}(E,\ell)$ is a Chebotarev set of primes of density $$\delta({\mathscr{P}}(E,\ell))=\frac{\ell^2-\ell-1}{(\ell-1)(\ell^2-1)}>0$$ and for every $\ell$-power free integer $a>1$ supported on ${\mathscr{P}}(E,\ell)$ we have that ${\mathrm{Sel}}_{\ell^\infty}(E/{\mathbb{Q}}(\mu_\ell,\sqrt[\ell]{a}))$ is finite. In particular, for these integers $a$ we have ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[\ell]{a}))={\mathrm{rank}\,}E({\mathbb{Q}}(\mu_\ell,\sqrt[\ell]{a}))=0$.
For our applications on Hilbert’s tenth problem we (crucially) need the case $\ell=3$. Nevertheless, this more general theorem can be of independent interest. We also remark that the various hypotheses in the statement are amenable for computations; a concrete example is presented in Section \[SecProof\]. Condition is convenient for showing that ${\mathscr{P}}(E,\ell)$ has positive density, but it can certainly be relaxed —however, it is enough for our purposes.
We remark that a related result is sketched as Theorem 18 in [@VladD] for primes $\ell\ge 5$ without verifying the existence of infinitely many integers $a$. For $\ell=3$, the special case of $E=X_1(11)$ is worked out in [@TimD] where the existence of infinitely many integers $a$ is shown by explicit computations with an equation for the modular curve $X_1(11)$.
Chao Li has pointed out to us another possible approach to achieve rank $0$ over cubic extensions, at least for a particular kind of elliptic curves. Namely, given a Mordell elliptic curve $E: y^2=x^3+k$ (i.e. an elliptic curve of $j$-invariant $0$) defined over ${\mathbb{Q}}$ with ${\mathrm{rank}\,}E({\mathbb{Q}})=0$, we have ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[3]{p}))=0$ if ${\mathrm{rank}\,}E^{[p]}({\mathbb{Q}})=0$ and ${\mathrm{rank}\,}E^{[p^2]}({\mathbb{Q}})=0$, where $E^{[d]}:y^2=x^3+d^2k$ is the cubic twist by $d$. The theory of [@KrizLi] allows one to achieve ${\mathrm{rank}\,}E^{[p]}({\mathbb{Q}})=0$ for many primes $p$ under suitable conditions. For our applications it would be of great interest to extend this theory in order to control the two relevant cubic twists simultaneously. We point out that both cubic twists $E^{[p]}$ and $E^{[p^2]}$ are needed; for instance, for $E: y^2=x^3+1$ and $p=5$ we have ${\mathrm{rank}\,}E({\mathbb{Q}})=0$, ${\mathrm{rank}\,}E^{[5]}({\mathbb{Q}})=0$ but ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[3]{5}))=1$.
Variation of the $\lambda$-invariant
------------------------------------
In view of Theorem \[ThmLambdaRk\], we will be interested in the variation of the $\lambda$-invariant of an elliptic curve under base change. An important result by Hachimori and Matsuno [@HaMa] gives a precise formula under suitable assumptions. Here we state a special case. As usual, for a number field $k$ the set of places of $k$ is denoted by $M_k$.
\[ThmHM\] Let $\ell>2$ be a prime and let $k$ be a number field. Let $E$ be an elliptic curve over $k$ with good ordinary reduction at all primes of $k$ above $\ell$. Suppose that $X(E/k_\infty)$ is a torsion $\Lambda_k$-module and that $\mu_{E/k}=0$.
Let $k'/k$ be an $\ell$-power Galois extension of number fields such that $E$ does not have additive reduction at the primes that ramify in $k'/k$. Then $\mu_{E/k'}=0$ and $X(E/k'_\infty)$ is a torsion $\Lambda_{k'}$-module. Furthermore, $$\lambda_{E/k'}=[k'_\infty:k_\infty]\lambda_{E/k} + \sum_{w\in P_1} (e_{k'_\infty/k_\infty}(w)-1) + 2 \sum_{w\in P_2} (e_{k'_\infty/k_\infty}(w)-1)$$ where $$\begin{aligned}
P_1&= \{ w\in M_{k'} : w\nmid \ell \mbox{ and } E \mbox{ has split multiplicative reduction at }w\} \\
P_2&= \{ w\in M_{k'} : w\nmid \ell \mbox{, } E \mbox{ has good reduction at } w \mbox{, and }E(k'_{\infty,w})[\ell^\infty]\ne (0)\}
\end{aligned}$$ and $e_{k'_\infty/k_\infty}(w)$ denotes the corresponding ramification index.
The following lemma allows one to give a simple alternative description of the set of places $P_2$ of the previous theorem.
\[LemmaDD\] Let $\ell>2$ be a prime and let $k'/k$ be an $\ell$-power Galois extension of number fields. Let $E$ be an elliptic curve over $k$ with good reduction at a prime $v\nmid \ell$ of $k$ and let $w|v$ be a prime of $k'$. Then, $E(k'_{\infty,w})[\ell^\infty]= (0)$ if and only if $\tilde{E}_v({\mathbb{F}}_v)[\ell]= (0)$.
Theorem \[ThmHM\] leads to a simple criterion to ensure that, under favorable conditions, the finiteness of the Iwasawa module $X(E/k_\infty)$ is preserved in prime-power degree field extensions (see also Corollary 3.20 in [@DD]).
\[PropPreserveXfin\] Let $\ell>2$ be a prime and let $k$ be a number field. Let $E$ be an elliptic curve over $k$ with good ordinary reduction at all primes of $k$ above $\ell$. Suppose that $X(E/k_\infty)$ is finite.
Let $k'/k$ be a $\ell$-power Galois extension of number fields satisfying the following conditions:
- $E$ has good reduction at each prime $v$ of $k$ that ramifies in $k'/k$, and
- for each prime $v\nmid \ell$ of $k$ that ramifies in $k'/k$, we have $\tilde{E}_v({\mathbb{F}}_v)[\ell]=(0)$.
Then $X(E/k'_\infty)$ is finite and ${\mathrm{rank}\,}E(k')=0$.
Since $X(E/k_\infty)$ is finite, it is a torsion $\Lambda_k$-module and $\mu_{E/k}=\lambda_{E/k}=0$ (cf. Lemma \[Lemmalm\]). By (i), we can apply Theorem \[ThmHM\] and we obtain that $X(E/k'_\infty)$ is $\Lambda_{k'}$-torsion and $\mu_{E/k'}=0$. Furthermore, $e_{k'_\infty/k_\infty}(w)=1$ for each $w\in P_1$ by our assumption (i), and $e_{k'_\infty/k_\infty}(w)=1$ for each $w\in P_2$ by (ii) and Lemma \[LemmaDD\]. Therefore $\lambda_{E/k'}=[k'_\infty : k_\infty]\lambda_{E/k} =0$.
Since $X(E/k'_\infty)$ is $\Lambda_{k'}$-torsion and $\mu_{E/k'}=\lambda_{E/k'}=0$, we obtain that $X(E/k'_\infty)$ is finite (cf. Lemma \[Lemmalm\]). In particular, since $\lambda_{E/k'}=0$ we get ${\mathrm{rank}\,}E(k')=0$ (cf. Theorem \[ThmLambdaRk\]).
In view of Proposition \[PropPreserveXfin\], we need a test for finiteness of $X(E/k_\infty)$.
Testing finiteness of $X(E/k_\infty)$
-------------------------------------
The following test for finiteness of $X(E/k_\infty)$ follows from the theory of the cyclotomic Euler characteristic and it is well-known to experts.
\[PropTestFinX\]Let $\ell$ be a prime and let $k$ be a number field. Let $E$ be an elliptic curve over $k$ with good ordinary reduction at all primes of $k$ above $\ell$. Assume that $E(k)$ is finite and that that $\Sha(E/k)[\ell^\infty]$ is also finite —the latter happens, for instance, if $\Sha(E/k)[\ell]=(0)$. Then $X(E/k_\infty)$ is $\Lambda_k$-torsion. Furthermore, consider the quantity $$\varpi(E/k):= \frac{{\mathrm{Tam}}(E/k)\cdot \# \Sha(E/k)[\ell^\infty] }{\#E(k)^2} \cdot \prod_{v|\ell} \#\tilde{E}_v({\mathbb{F}}_v)^2.$$ Then $\varpi(E/k)\in {\mathbb{Z}}_\ell$, and if $\varpi(E/k)$ is an $\ell$-adic unit, then $X(E/k_\infty)$ is finite.
Since $E(k)$ and $\Sha(E/k)[\ell^\infty]$ are finite, the exact sequence $$0\to E(k)\otimes_{\mathbb{Z}}{\mathbb{Q}}_\ell/{\mathbb{Z}}_\ell\to {\mathrm{Sel}}_{\ell^\infty}(E/k)\to \Sha(E/k)[\ell^\infty]\to 0$$ gives that ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is also finite —in fact, we get ${\mathrm{Sel}}_{\ell^\infty}(E/k)\simeq \Sha(E/k)[\ell^\infty]$. In particular, Theorem \[ThmMazurFin\] gives that $X(E/k_\infty)$ is $\Lambda_k$-torsion.
Thus, we can consider the characteristic polynomial $f_{E/k}\in{\mathbb{Z}}_\ell[T]$ (for a fixed choice of topological generator $\gamma\in\Gamma_k$). By Proposition \[Propf0\] and the fact that ${\mathrm{Sel}}_{\ell^\infty}(E/k)$ is finite, we get that $f_{E/k}(0)$ is a non-zero $\ell$-adic integer. Theorem \[Thmf0\] gives $\varpi(E/k)\sim_\ell f_{E/k}(0)$, so $\varpi(E/k)\in{\mathbb{Z}}_\ell$.
Suppose that $\varpi(E/k)$ is an $\ell$-adic unit. Let us recall that $f_{E/k}=\ell^{\mu_{E/k}}\prod_j f_j^{m_j}$ for certain monic irreducible polynomials $f_j$ having all their non-leading coefficients in $\ell{\mathbb{Z}}_\ell$. Since $f_{E/k}(0)\sim_\ell \varpi(E/k)$ is an $\ell$-adic unit, we deduce that in fact $\mu_{E/k}=0$ and that the product $\prod_j f_j^{m_j}$ equals $1$ because it is empty —for otherwise, the constant term of each $f_j$ would be divisible by $\ell$. Hence $\lambda_{E/k}=0$. Since $\mu_{E/k}=\lambda_{E/k}=0$, we get that $X(E/k_\infty)$ is finite (cf. Lemma \[Lemmalm\]).
Chebotarev sets of primes
-------------------------
A set of prime numbers ${\mathscr{S}}$ is called a *Chebotarev set* if there is a Galois extension $K/{\mathbb{Q}}$ and a conjugacy-stable set $C\subseteq \mathrm{Gal}(K/{\mathbb{Q}})$ such that $S$ agrees with the set $\{p : {\mathrm{Frob}}_p\in C\}$ up to a finite set. Finite unions, finite intersections, and complements of Chebotarev sets are again Chebotarev. Let us recall that the Chebotarev density theorem states that if ${\mathscr{S}}$ arises from $K$ and $C$ as above, then the limit $$\delta({\mathscr{S}})=\lim_{x\to \infty}\frac{\#{\mathscr{S}}\cap[1,x]}{\pi(x)}$$ exists and equals $\#C/[K:{\mathbb{Q}}]$, where $\pi(x)$ is the number of prime numbers $p\le x$. The quantity $\delta({\mathscr{S}})$ is called the *density* of ${\mathscr{S}}$.
Auxiliary primes
----------------
The next result ensures that we have enough primes for the constructions required in the proof of Theorem \[ThmRkZero\]. As usual, if $E$ is an elliptic curve over ${\mathbb{Q}}$ and $p$ is a prime of good reduction, we write $a_p(E)=p+1-\# \tilde{E}_p({\mathbb{F}}_p)$.
\[PropCheb\] Let $\ell$ be a prime number. Let $E$ be an elliptic curve defined over ${\mathbb{Q}}$ of conductor $N$. Suppose that the residual Galois representation $\rho_{E[\ell]}$ is surjective. Consider the set of prime numbers $${\mathscr{P}}(E,\ell)=\{p : p\nmid N\mbox{, }p \equiv 1\bmod \ell\mbox{, and }a_p(E)\not\equiv 2\bmod \ell\}.$$ Then ${\mathscr{P}}(E,\ell)$ is a Chebotarev set of primes with density $$\delta({\mathscr{P}}(E,\ell))=\frac{\ell^2-\ell-1}{(\ell-1)(\ell^2-1)}>0.$$ Furthermore, for each $p\in{\mathscr{P}}(E,\ell)$ and each place $v| p$ of ${\mathbb{Q}}(\mu_\ell)$ we have that $\tilde{E}_v({\mathbb{F}}_v)[\ell]=(0)$.
The Weil pairing shows that the $\ell$-division field $K={\mathbb{Q}}(E[\ell])$ contains ${\mathbb{Q}}(\mu_\ell)$. The cyclotomic character $\chi:{\mathrm{Gal}}({\mathbb{Q}}(\mu_\ell)/{\mathbb{Q}})\to {\mathbb{F}}_\ell^\times$ extends to $\chi:{\mathrm{Gal}}(K/{\mathbb{Q}})\to {\mathbb{F}}_\ell^\times$ by means of the quotient ${\mathrm{Gal}}(K/{\mathbb{Q}})\to {\mathrm{Gal}}({\mathbb{Q}}(\mu_\ell)/{\mathbb{Q}})$. We recall that the residual Galois representation $\rho_{E[\ell]}:{\mathrm{Gal}}(K/{\mathbb{Q}})\to GL_2({\mathbb{F}}_\ell)$ is injective and for all $p\nmid \ell N$ it satisfies
- $\det(\rho_{E[\ell]}({\mathrm{Frob}}_p))=\chi({\mathrm{Frob}}_p)=p\bmod \ell$, and
- ${\mathrm{tr}}(\rho_{E[\ell]}({\mathrm{Frob}}_p))=a_p(E)\bmod \ell$.
The map $\rho_{E[\ell]}:{\mathrm{Gal}}(K/{\mathbb{Q}})\to GL_2({\mathbb{F}}_\ell)$ is in fact an isomorphism by our surjectivity assumption. Let $C=\{\gamma\in GL_2({\mathbb{F}}_\ell) : \det(\gamma)=1\mbox{ and }{\mathrm{tr}}(\gamma)\ne 2\}\subseteq GL_2({\mathbb{F}}_\ell)$. The set $C$ is conjugacy-stable and it is non-empty (e.g. $-I\in C$ when $\ell>2$). In fact, it is an elementary exercise to check that $\#\{\gamma\in SL_2({\mathbb{F}}_\ell) : {\mathrm{tr}}(\gamma)=2\}=\ell^2$, which gives $$\# C =\#SL_2({\mathbb{F}}_\ell)-\#\{\gamma\in SL_2({\mathbb{F}}_\ell) : {\mathrm{tr}}(\gamma)=2\}= \ell(\ell^2-1)- \ell^2=\ell(\ell^2-\ell-1).$$ Since we can write $${\mathscr{P}}(E,\ell)=\{p : p\nmid N \mbox{ and }\rho_{E[\ell]}({\mathrm{Frob}}_p)\in C\}$$ we see that ${\mathscr{P}}(E,\ell)$ is a Chebotarev set of density $$\delta({\mathscr{P}}(E,\ell))=\frac{\#C}{\# GL_2({\mathbb{F}}_\ell)} = \frac{\ell(\ell^2-\ell-1)}{(\ell^2-1)(\ell^2-\ell)}=\frac{\ell^2-\ell-1}{(\ell-1)(\ell^2-1)}>0.$$ Finally, let $p\in {\mathscr{P}}(E,\ell)$. Since $p\equiv 1\bmod \ell$ we have that $p$ splits completely in ${\mathbb{Q}}(\mu_\ell)$, hence, for each place $v|p$ of ${\mathbb{Q}}(\mu_\ell)$ we have that the residue field of ${\mathbb{Q}}(\mu_\ell)$ at $v$ is ${\mathbb{F}}_v={\mathbb{F}}_p$. Therefore $\tilde{E}_v({\mathbb{F}}_v)\simeq \tilde{E}_p({\mathbb{F}}_p)$ because $E$ has good reduction at $p\nmid N$, and we get $$\# \tilde{E}_v({\mathbb{F}}_v) = p+1-a_p(E)\equiv 1+1-a_p(E)\not\equiv 0\bmod \ell$$ because $p\in{\mathscr{P}}(E,\ell)$. Thus, the $\ell$-torsion of $\tilde{E}_v({\mathbb{F}}_v)$ is trivial.
Rank zero over extensions
-------------------------
Let us keep the notation of assumptions of Theorem \[ThmRkZero\].
The fact that ${\mathscr{P}}(E,\ell)$ is a Chebotarev set of primes of the asserted density follows from Proposition \[PropCheb\] and our Condition .
Let $k={\mathbb{Q}}(\mu_\ell)$. By Condition , $E$ has good ordinary reduction at each place $v|\ell$ of $k$. Furthermore, $\ell$ ramifies completely in $k/{\mathbb{Q}}$, thus, there is only one place $v|\ell$ and the residue field satisfies ${\mathbb{F}}_v={\mathbb{F}}_\ell$, which in particular gives $\tilde{E}_v({\mathbb{F}}_v)\simeq \tilde{E}_\ell({\mathbb{F}}_\ell)$ because $E$ has good reduction at $\ell$.
By Condition we get $\Sha(E/k)[\ell^\infty]=(0)$, which together with Condition gives that the integer $${\mathrm{Tam}}(E/k)\cdot \# \Sha(E/k)[\ell^\infty] \cdot \prod_{v|\ell} \#\tilde{E}_v({\mathbb{F}}_v)^2$$ is not divisible by $\ell$. By Conditions , , and we can apply Proposition \[PropTestFinX\] and we deduce that $X(E/k_{\infty})$ is a finite group because the $\ell$-adic integer $\varpi(E/k)$ is not divisible by $\ell$.
Let $a>1$ be an $\ell$-power free integer supported on ${\mathscr{P}}(E,\ell)$ and let $k'={\mathbb{Q}}(\mu_\ell, \sqrt[\ell]{a})=k(\sqrt[\ell]{a})$. The extension $k'/k$ is Galois of degree $\ell$. Note that $\ell>2$, $E$ has good ordinary reduction at each prime of $k$ above $\ell$ (by Condition ), and $X(E/k_\infty)$ is finite, so that we can apply Proposition \[PropPreserveXfin\]. For this, note that if a prime $v\in M_k$ ramifies in $k'/k$ then it divides $a$, hence, $v$ divides a prime $p\in{\mathscr{P}}(E, \ell)$ and we deduce that (cf. Proposition \[PropCheb\]):
- $E$ has good reduction at $v$ (as $p\nmid N$ by definition of ${\mathscr{P}}(E,\ell)$), and
- $\tilde{E}_v({\mathbb{F}}_v)[\ell]=(0)$.
Therefore, Proposition \[PropPreserveXfin\] gives that $X(E/k'_\infty)$ is finite. In particular $\lambda_{E/k'}=0$ (cf. Lemma \[Lemmalm\]), ${\mathrm{Sel}}_{\ell^\infty}(E/k')$ is finite, and ${\mathrm{rank}\,}E(k')=0$ (cf. Theorem \[ThmLambdaRk\]).
Achieving positive rank: Heegner points {#SecKL}
=======================================
In this section we present a theorem by Kriz and Li [@KrizLi] regarding congruences of Heegner points on quadratic twists of elliptic curves, along with some additional facts tailored for our intended applications. The results of Kriz and Li are particularly convenient for producing many explicit quadratic twists with positive rank for a given elliptic curve over ${\mathbb{Q}}$. We also work-out a concrete example that will be needed later.
Heegner points {#SecHeegner}
--------------
Let $N$ be a positive integer. Attached to $N$ there is the modular curve $X_0(N)$ over ${\mathbb{Q}}$ whose non-cuspidal points classify cyclic degree $N$ isogenies of elliptic curves. The cusp $i\infty\in X_0(N)({\mathbb{Q}})$ determines an embedding of the modular curve into its Jacobian $j_N: X_0(N)\to J_0(N)$ defined over ${\mathbb{Q}}$.
A quadratic imaginary field $K$ is said to satisfy the *Heegner hypothesis* for $N$ if each prime $p|N$ splits in $K$. If $K$ satisfies the Heegner condition for $N$ we can choose a factorization $(N)={\mathfrak{n}}{\mathfrak{n}}'$ in $O_K$ with ${\mathfrak{n}}'$ the complex conjugate of the ideal ${\mathfrak{n}}$, and one has $O_K/{\mathfrak{n}}\simeq {\mathbb{Z}}/N{\mathbb{Z}}$. The map $${\mathbb{C}}/O_K\to {\mathbb{C}}/{\mathfrak{n}}^{-1}$$ is a cyclic degree $N$ isogeny of complex elliptic curves, and it defines a non-cuspidal point $Q_{N,K, {\mathfrak{n}}}\in X_0(N)(H)$ where $H$ is the Hilbert class field of $K$. Mapping to $J_0(N)$ and taking trace for the extension $H/K$, we obtain the $K$-rational point $$P_{N,K,{\mathfrak{n}}}=\sum_{\sigma\in {\mathrm{Gal}}(H/K)} j_N(Q_{N,K, {\mathfrak{n}}}^\sigma)\in J_0(N)(K).$$ The point $P_{N,K,{\mathfrak{n}}}$ is the *Heegner point in $J_0(N)$ attached to $K$*. It is independent of the choice of factorization $N={\mathfrak{n}}{\mathfrak{n}}'$ up to sign and adding torsion, and we will denote it by $P_{N,K}$ because this ambiguity is irrelevant for our discussion.
Modular parametrizations {#SecModular}
------------------------
Let $E$ be an elliptic curve over ${\mathbb{Q}}$ of conductor $N$. The modularity theorem [@BCDT; @TaylorWiles; @Wiles] gives a non-constant map $\varphi_E:X_0(N)\to E$ defined over ${\mathbb{Q}}$, with $\varphi_E(i\infty)=0_E$ where $0_E$ is the neutral point of $E$. Such a map $\varphi_E$ is called modular parametrization. It induces a morphism of abelian varieties $\pi_E : J_0(N)\to E$ defined over ${\mathbb{Q}}$ satisfying $\varphi_E=\pi_E\circ j_N$. Of course $\varphi_E$ is not unique; however, if the degree of $\varphi_E$ is assumed to be minimal (which we don’t) then $\varphi_E$ is determined up to multiplication by $-1$ in $E$.
Let $\omega_E\in H^0(E,\Omega^1_{E/{\mathbb{Q}}})$ be a global Néron differential on $E$; it is unique up to sign. Let $f(q)=q+a_2q^2+...\in S_2(\Gamma_0(N))$ be the unique Fourier-normalized Hecke newform attached to $E$ by the modularity theorem. Then there is a rational number $c(\varphi_E, \omega_E)$ such that $$\varphi_E^*\omega_E=c(\varphi_E,\omega_E)\cdot f(q)\frac{dq}{q}.$$ The absolute value of $c(\varphi_E,\omega_E)$ only depends on $\varphi_E$ and it will be denoted by $c(\varphi_E)$. By considerations on the formal completion of the standard integral model of $X_0(N)$ at $i\infty$, it is easily seen that $c(\varphi_E)$ is an integer. In concrete examples, the Manin constant of $\varphi_E$ can be computed.
A modular parametrization $\varphi_E:X_0(N)\to E$ is *optimal* if the kernel of the induced map $\pi_E:J_0(N)\to E$ is connected. If an optimal modular parametrization exists for $E$ we say that $E$ is optimal (sometimes referred to as strong Weil curve). Every $E$ is isogenous over ${\mathbb{Q}}$ to an optimal elliptic curve. It is a conjecture of Manin that if $\varphi_E:X_0(N)\to E$ is optimal, then $c(\varphi_E)=1$. Manin’s conjecture is proved in a number of cases, e.g. when the conductor $N$ is odd and squarefree (this is Corollary 4.2 in [@MazurIsog] together with Théorème A in [@AbbesUllmo]). See [@ARS] and the references therein for more details on Manin’s conjecture.
Logarithms {#SecLog}
----------
Let $E$ be an elliptic curve over ${\mathbb{Q}}$ and let $p$ be a prime. By integration, every differential $\omega\in H^0(E,\Omega^1_{E/{\mathbb{Q}}})$ uniquely determines a logarithm map $$\log_{p,\omega} : E({\mathbb{Q}}_p)\to {\mathbb{Q}}_p$$ which satisfies the following:
- For all $a,b\in E({\mathbb{Q}}_p)$, $\log_{p,\omega}(a+b)=\log_{p,\omega}(a)+\log_{p,\omega}(b)$.
- If $\omega\ne 0$ then the kernel of $\log_{p,\omega}$ is $E({\mathbb{Q}}_p)_{tor}$.
Concretely, the construction is as follows: Choose a global minimal Weierstrass equation over ${\mathbb{Z}}$ $$\label{EqGenW}
y^2+a_1xy+a_3y=x^3+a_2x^2+a_4x+a_6$$ and let $t=-x/y$. Then $t$ is a local parameter at the neutral element $0_E\in E({\mathbb{Q}})$ and, moreover, a local point $a\in E({\mathbb{Q}}_p)$ is congruent to $0_E$ modulo $p$ if and only if $t(a)\in p{\mathbb{Z}}_p$. Let us write $\omega= (b_0+b_1t+b_2t^2+...)dt$. Given any $a\in E({\mathbb{Q}}_p)$, choose $m$ such that $ma$ is congruent to $0_E$ modulo $p$. Then $$\log_{p,\omega}(a)=\frac{1}{m}\sum_{r= 1}^\infty \frac{b_{r-1}\cdot t(ma)^r}{r}$$ where the series converges because $t(ma)\in p{\mathbb{Z}}_p$. The value of $\log_{p,\omega}(a)$ is independent of the choice of $m$. We note that the series is the formal integral of $\omega$.
We will consider in the following situation: $K/{\mathbb{Q}}$ is a quadratic field such that $p$ splits in $K$ and $a\in E(K)$ is a point (in fact, we will take $p=2$). The choice of a prime ${\mathfrak{p}}|p$ in $O_K$ is equivalent to the choice of an embedding $\sigma:K\to {\mathbb{Q}}_p$. Such a $\sigma$ induces an inclusion $\sigma: E(K)\to E({\mathbb{Q}}_p)$ and the quantity $\log_{p,\omega} \sigma(a)$ is thus defined.
For explicit computations with logarithms, it will be convenient to recall the following formulas from Section IV.1 in [@SilvermanAdv]. The global Néron differential over ${\mathbb{Z}}$ for $E$ is (in affine coordinates of the model ) given by $$\omega_E=\frac{dx}{2y+a_1x +a_3}$$ where $a_j\in{\mathbb{Z}}$ are the coefficients in the global minimal Weierstrass equation . This differential expressed in terms of the local parameter $t=-x/y$ at $0_E$ is given by $$\label{EqDiff}
\omega_E=(1+a_1t+(a_1^2+a_2)t^2 + (a_1^3+2a_1a_2+2a_3)t^3+...)dt\in {\mathbb{Z}}[[t]]dt$$ where the coefficients of the power series are integers that can be computed in terms of the coefficients $a_j$ from . For our purposes these first few coefficients suffice.
Auxiliary primes
----------------
For an elliptic curve $E/{\mathbb{Q}}$ of conductor $N$ and a quadratic field $K$, let us define the sets of prime numbers $$\begin{aligned}
{\mathscr{Q}}(E,K)&=\{q : q\nmid 2N,\, q\mbox{ splits in }K, \mbox{ and } a_q(E)\equiv 1\bmod 2 \}\\
{\mathscr{Q}}_+(E,K)&=\{q : q\in {\mathscr{Q}}(E,K)\mbox{ and }q\equiv 1\bmod 4\}\\
{\mathscr{Q}}_-(E,K)&=\{q : q\in {\mathscr{Q}}(E,K)\mbox{ and }q\equiv -1\bmod 4\}
\end{aligned}$$ and observe that ${\mathscr{Q}}(E,K)={\mathscr{Q}}_+(E,K)\cup {\mathscr{Q}}_-(E,K)$.
\[LemmaPrimes\] The sets ${\mathscr{Q}}(E,K)$, ${\mathscr{Q}}_+(E,K)$, and ${\mathscr{Q}}_-(E,K)$ are Chebotarev sets of primes. If $E[2]({\mathbb{Q}})\ne (0)$ then the three sets are empty.
Suppose that $E[2]({\mathbb{Q}})=(0)$ and $K\ne {\mathbb{Q}}(\sqrt{-1})$.
- If ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\simeq {\mathbb{Z}}/3{\mathbb{Z}}$ then $$\delta({\mathscr{Q}}(E,K))=1/3,\quad \delta({\mathscr{Q}}_+(E,K))=1/6, \mbox{ and }\, \delta({\mathscr{Q}}_-(E,K))=1/6.$$
- If ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\simeq S_3$ and ${\mathbb{Q}}(E[2])$ does not contain the fields ${\mathbb{Q}}(\sqrt{-1})$ and $K$, then $$\delta({\mathscr{Q}}(E,K))=1/6,\quad \delta({\mathscr{Q}}_+(E,K))=1/12, \mbox{ and }\, \delta({\mathscr{Q}}_-(E,K))=1/12.$$
- If ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\simeq S_3$ and ${\mathbb{Q}}(E[2])$ contains the field ${\mathbb{Q}}(\sqrt{-1})$, then ${\mathscr{Q}}_-(E,K)$ is empty, ${\mathscr{Q}}_+(E,K)={\mathscr{Q}}(E,K)$, and $\delta({\mathscr{Q}}(E,K))=1/6$.
The condition $a_q(E)\equiv 1\bmod 2$ at a prime $q\nmid N$ is equivalent to the condition $${\mathrm{tr}}(\rho_{E[2]}({\mathrm{Frob}}_q))=1\in {\mathbb{F}}_2$$ for the residual Galois representation $\rho_{E[2]}:{\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\to GL_2({\mathbb{F}}_2)\simeq S_3$. The only elements of $GL_2({\mathbb{F}}_2)$ with trace equal to $1$ are $$\left[\begin{array}{cc} 1&1\\1&0\end{array}\right]\, \mbox{ and }\, \left[\begin{array}{cc} 0&1\\1&1\end{array}\right]$$ i.e. the two $3$-cycles in $S_3$.
The three sets of primes are empty if there is a rational $2$-torsion point, since in this case every prime $q\nmid 2N$ satisfies $a_q(E)\equiv {\mathrm{tr}}(\rho_{E[2]}({\mathrm{Frob}}_q))\equiv 0\bmod 2$.
In case (i), since ${\mathbb{Q}}(E[2])$ does not contain quadratic fields, the result follows from the Chebotarev density theorem and the fact that the congruence conditions modulo $4$ are splitting conditions for the extension ${\mathbb{Q}}(\sqrt{-1})/{\mathbb{Q}}$.
In cases (ii) and (iii) we have $\delta({\mathscr{Q}}(E,K))=1/6$ by the Chebotarev density theorem. Case (ii) follows by linear disjointness of the relevant fields.
In case (iii) it only remains to show that ${\mathscr{Q}}_-(E,K)$ is empty. Note that ${\mathbb{Q}}(E[2])$ does not contain $K$ because a hexic extension of ${\mathbb{Q}}$ contains a unique quadratic subfield. The only non-trivial morphism $s: GL_2({\mathbb{F}}_2)\simeq S_3\to \mu_2$ is the sign map, and it sends the trace $1$ matrices (i.e. $3$-cycles of $S_3$) to $1\in \mu_2$. By uniqueness of $s$, the composition $s\circ \rho_{E[2]}:{\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\to \mu_2$ is the same as the map induced by the non-trivial quadratic character $\chi:{\mathrm{Gal}}({\mathbb{Q}}(\sqrt{-1})/{\mathbb{Q}})\to\mu_2$ via the restriction map ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\to {\mathrm{Gal}}({\mathbb{Q}}(\sqrt{-1})/{\mathbb{Q}})$. Since each prime $q\equiv -1\bmod 4$ has $\chi({\mathrm{Frob}}_q)=-1$, we get that ${\mathscr{Q}}_-(E,K)$ is empty.
Case (iii) in Lemma \[LemmaPrimes\] can in fact happen and we must avoid it for our intended applications (cf. the remarks after Corollary \[CoroKL\]). For instance, the elliptic curve $E$ of Cremona label $121a1$ has the property that ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\simeq S_3$ and ${\mathbb{Q}}(\sqrt{-1})\subseteq {\mathbb{Q}}(E[2])$. Here is the data for this $E$ and the first $20$ primes $q\nmid 2N$ $${\tiny
\begin{array}{c|rrrrrrrrrrrrrrrrrrrr}
q \nmid 2\cdot 121 & 3 & 5 & 7 & 13 & 17 & 19 & 23 & 29 & 31 & 37 & 41 & 43 & 47 & 53 & 59 & 61 & 67 & 71 & 73 & 79 \\
q\bmod 4 & -1 & 1 & -1 & 1 & 1 & -1 & -1 & 1 & -1 & 1 & 1 & -1 & -1 & 1 & -1 & 1 & -1 & -1 & 1 & -1 \\
a_q(E) & 2 & 1 & 2 & -1 & 5 & -6 & 2 & -9 & -2 & -3 & 5 & 0 & 2 & 9 & 8 & -6 & 2 & 12 & 2 & 10
\end{array}}$$ Already for these first few values we see that when $q\equiv -1\bmod 4$ we get $a_q(E)$ even, while for $q\equiv 1\bmod 4$ we can have $a_q(E)$ even or odd depending on whether $\rho_{E[2]}({\mathrm{Frob}}_q)\in GL_2({\mathbb{F}}_2)\simeq S_3$ is the identity or a $3$-cycle respectively.
Let us point out that in Lemma \[LemmaPrimes\] the condition that $K$ is not contained in ${\mathbb{Q}}(E[2])$ will not be a problem for us, due to the following observation.
Let $E$ be an elliptic curve over ${\mathbb{Q}}$ of conductor $N$ satisfying ${\mathrm{Gal}}({\mathbb{Q}}(E[2])/{\mathbb{Q}})\simeq S_3$. There is only one quadratic field contained in ${\mathbb{Q}}(E[2])$ and it does not satisfy the Heegner hypothesis for $2N$.
Since $[{\mathbb{Q}}(E[2]):{\mathbb{Q}}]=6$, there is only one quadratic field $F\subseteq {\mathbb{Q}}(E[2])$. The primes that ramify in $F$ divide $2N$, hence the result.
Congruences of Heegner points, after Kriz and Li
------------------------------------------------
The following is a re-statement of Theorem 4.3 in [@KrizLi], keeping track of some choices that are made along the proof in *loc. cit.*
\[ThmKL\] Suppose $E/\mathbb{Q}$ is an elliptic curve with $E(\mathbb{Q})[2]=(0)$. Consider a modular parametrization $\varphi_E:X_0(N)\to E$ and the corresponding quotient $\pi_E:J_0(N)\to E$. Let $K$ be an imaginary quadratic field satisfying the Heegner hypothesis for $2N$ and let $\sigma:K\to \mathbb{Q}_2$ be an embedding. Assume that $$\label{CondKL}
\frac{\#\tilde{E}_2^{ns}(\mathbb{F}_2)}{2\cdot c(\varphi_E)}\cdot \log_{2,\omega_E}(\sigma(\pi_E(P_{N,K})))\in {\mathbb{Z}}_2^\times.$$ Then for each squarefree integer $d\equiv 1\bmod 4$ supported on ${\mathscr{Q}}(E,K)$ the following holds:
- Both $E$ and $E^{d}$ have rank at most $1$ over ${\mathbb{Q}}$, and the rank part of the Birch and Swinnerton-Dyer conjecture holds for them.
- ${\mathrm{rank}\,}E({\mathbb{Q}})\ne {\mathrm{rank}\,}E^d({\mathbb{Q}})$ if and only if $\psi_d(-N)=-1$, where $\psi_d$ is the quadratic Dirichlet character attached to the quadratic field ${\mathbb{Q}}(\sqrt{d})$.
Let us make some observations regarding condition .
- $\log_{2,\omega_E}$ depends on the choice of the Néron differential $\omega_E$ only up to sign, which is irrelevant for condition .
- The choice of $\varphi_E$ affects both $c(\varphi_E)$ and $\pi_E(P_{N,K})$, but the quantity in remains the same up to sign.
- $P_{N,K}\in J_0(N)(K)$ also depends on the choice of a factorization $(N)={\mathfrak{n}}{\mathfrak{n}}'$ in $O_K$, but only up to multiplication by $-1$ and adding torsion. Hence $\log_{2,\omega_E}(\sigma(\pi_E(P_{N,K})))$ remains the same up to sign.
- The proof in [@KrizLi] also shows that the quantity in is a $2$-adic integer. So, the condition actually is about indivisibility by $2$.
It is worth pointing out that Theorem \[ThmKL\] (cf. Theorem 4.3 in [@KrizLi]) is a consequence of a very general congruence formula for Heegner points. In fact, the theory in [@KrizLi] (for $p=2$) actually shows that the quantity in is congruent modulo $2$ to the analogous quantity for the twisted elliptic curve $E^d$. Thus, condition implies that the analogous indivisibility holds for $E^d$ and, in particular, the corresponding Heegner points in $E(K)$ and $E^d(K)$ are non-torsion (as their logarithms are non-zero). This, together with known results on the Birch and Swinnerton-Dyer conjecture (due to Kolyvagin [@Kol], Gross-Zagier [@GZ], Murty-Murty [@MM], and Bump-Friedberg-Hoffstein [@BFH]) implies item (i) in Theorem \[ThmKL\].
Regarding item (ii) of Theorem \[ThmKL\], the condition $\psi_d(-N)=-1$ ensures that the global root numbers of $E$ and $E^d$ are different, so that (ii) follows from (i). The next result is Corollary 5.11 in [@KrizLi], which shows that this parity condition takes a particularly simple form when holds and the Tamagawa factor $c_2(E/{\mathbb{Q}})$ is odd.
\[CoroKL\] Let us keep the same notation and assumptions of Theorem \[ThmKL\] and let $\Delta_E\in{\mathbb{Z}}$ be the minimal discriminant of $E$. Assume condition for $E$ and that $c_2(E/{\mathbb{Q}})$ is odd. We have ${\mathrm{rank}\,}E({\mathbb{Q}})\ne {\mathrm{rank}\,}E^d({\mathbb{Q}})$ if and only if $\Delta_E>0$ and $d<0$.
In particular, if ${\mathrm{rank}\,}E({\mathbb{Q}})=0$, then Corollary \[CoroKL\] can only ensure ${\mathrm{rank}\,}E^d({\mathbb{Q}})>0$ for negative values of $d$, under favorable conditions. Later we will be interested in taking $d=-q$ for $q$ a prime. So, it will be crucial that the set ${\mathscr{Q}}_-(E,K)$ be non-empty; we must avoid case (iii) in Lemma \[LemmaPrimes\].
Ranks and Hilbert’s tenth problem {#SecProof}
=================================
In this section we prove Theorem \[ThmMainIntro\]. For this, we will need an auxiliary elliptic curve to which we will apply Theorem \[ThmRkZero\] (with $\ell=3$) and Theorem \[ThmKL\]. The chosen elliptic curve is 557b1 in Cremona’s notation, which has minimal Weierstrass equation over ${\mathbb{Z}}$ $$y^2+y=x^3-x^2-268x+1781.$$
When using this elliptic curve, we will simply indicate “$E=557b1$”.
Compatibility in the Birch and Swinnerton-Dyer conjecture {#SecBSD}
---------------------------------------------------------
The Birch and Swinnerton-Dyer conjecture for abelian varieties over number fields (as formulated in [@Tate]) enjoys a number of compatibility properties, such as compatibility under product, isogenies [@Cassels; @Tate; @MilneADT], and base change [@MilneInv]. The same compatibilities hold regarding the finiteness of the $p$-primary part of the Shafarevich-Tate group and the $p$-adic valuation of the conjectural special value formula (cf. Remark 7.4 in [@MilneADT]). We will use these compatibilities to reduce the verification of hypothesis in Theorem \[ThmRkZero\] for $\ell=3$ to a computation involving elliptic curves over ${\mathbb{Q}}$.
This has a practical purpose, since exact numerical computations with elliptic curves over number fields are more difficult than over ${\mathbb{Q}}$. More concretely, we will need to check that $\Sha(E/{\mathbb{Q}}(\sqrt{-3})[3]$ is trivial in a specific case, which involves a $3$-descent to compute the relevant $3$-Selmer group. However, the algorithm to compute $3$-Selmer groups seems to be implemented in Magma only for elliptic curves over ${\mathbb{Q}}$.
We recall that if $E$ is an elliptic curve over ${\mathbb{Q}}$ with $L$-function $L(E,s)$ and real period $\Omega_E$, then the modularity of $E$ implies that $L(E,s)$ is entire and that $L(E,1)/\Omega_E$ is a rational number that can be exactly, explicitly, and efficiently computed by the theory of modular symbols: it has small denominator (a divisor of $2\cdot c(\varphi_E)\cdot \# E({\mathbb{Q}})_{tor}$ for any modular parametrization $\varphi_E$), so a good numerical approximation suffices.
\[PropBSD\] Let $E$ be a semi-stable elliptic curve over ${\mathbb{Q}}$ with good reduction at the prime $3$. Let $E'=E^{-3}$ be the quadratic twist of $E$ by $-3$. Suppose that
- ${\mathrm{rank}\,}E({\mathbb{Q}})=0$ and ${\mathrm{rank}\,}E'({\mathbb{Q}})=0$.
- $\Sha(E/{\mathbb{Q}})[3^\infty]$ and $\Sha(E'/{\mathbb{Q}})[3^\infty]$ are finite.
- The $3$-adic valuation of the rational numbers $L(E,1)/\Omega_E$ and $L(E',1)/\Omega_{E'}$ is as predicted by the $3$-part of the Birch and Swinnerton-Dyer conjecture.
Then ${\mathrm{rank}\,}E({\mathbb{Q}}(\mu_3))=0$, the group $\Sha(E/{\mathbb{Q}}(\mu_3))[3^\infty]$ is finite, and we have $$\frac{L(E,1)}{\Omega_E}\cdot \frac{L(E',1)}{\Omega_{E'}}\sim_3 \frac{{\mathrm{Tam}}(E/{\mathbb{Q}}(\mu_3))\cdot \# \Sha(E/{\mathbb{Q}}(\mu_3))[3^\infty]}{\#E({\mathbb{Q}}(\mu_3))_{tor}^2}.$$ Furthermore, if $3\nmid \#\tilde{E}_3({\mathbb{F}}_3)$, then $3\nmid \#E({\mathbb{Q}}(\mu_3))_{tor}$.
Note that ${\mathbb{Q}}(\mu_3)={\mathbb{Q}}(\sqrt{-3})$. By (i) we have ${\mathrm{rank}\,}E({\mathbb{Q}}(\mu_3))={\mathrm{rank}\,}E({\mathbb{Q}})+{\mathrm{rank}\,}E'({\mathbb{Q}})=0$.
Let $A$ be the abelian surface over ${\mathbb{Q}}$ obtained as Weil restriction of scalars of $E/{\mathbb{Q}}(\mu_3)$ to ${\mathbb{Q}}$. Then $A$ is isogenous over ${\mathbb{Q}}$ to $E\times E'$. By (ii) and compatibilities, we get that $\Sha(E/{\mathbb{Q}}(\mu_3))[3^\infty]$ is finite.
The period of $E/{\mathbb{Q}}(\mu_3)$ can be computed using a global Néron differential $\omega_E$ of a minimal model of $E$ over ${\mathbb{Q}}$ (as $E$ is semi-stable), which up to powers of $2$ gives $|\int_{E({\mathbb{C}})}\omega_E\wedge\overline{\omega_E}|$. Up to powers of $2$, this is equal to $\Omega_E\Omega_{E'}\sqrt{3}$ because $E$ has good reduction at $3$ (cf. [@Pal]).
By (iii) and compatibilities, the $3$-part of the Birch and Swinnerton-Dyer conjecture holds for $E/{\mathbb{Q}}(\mu_3)$. Thus we have $$\frac{{\mathrm{Tam}}(E/{\mathbb{Q}}(\mu_3))\cdot \# \Sha(E/{\mathbb{Q}}(\mu_3))[3^\infty]}{\#E({\mathbb{Q}}(\mu_3))_{tor}^2}\sim_3 \frac{L(E/{\mathbb{Q}}(\mu_3),1)\sqrt{3}}{|\int_{E({\mathbb{C}})}\omega_E\wedge\overline{\omega_E}|}$$ where the $\sqrt{3}$ comes from the discriminant of ${\mathbb{Q}}(\mu_3)$ (see [@DDbsd] for an explicit statement of the conjectural Birch and Swinnerton-Dyer formula for abelian varieties over number fields). Artin formalism gives $L(E/{\mathbb{Q}}(\mu_3),s) = L(E,s)L(E',s)$, and the claimed formula follows.
Finally, let $v$ be the only place of ${\mathbb{Q}}(\mu_3)$ over $3$, and note that $v|3$ is ramified. Since $E$ has good reduction at $3$, we have that $E({\mathbb{Q}}(\mu_3))_{tor}$ injects in $\tilde{E}_v({\mathbb{F}}_v)\simeq \tilde{E}_3({\mathbb{F}}_3)$, proving the final claim.
We remark that the semi-stability condition is just for convenience and it can be relaxed with a more careful analysis; for our purposes this is enough.
Keeping rank zero in many cubic extensions
------------------------------------------
\[LemmaRkZero557b1\] Let $E=557b1$. Define the set of primes $${\mathscr{P}}= \{p : p\equiv 1\bmod 3 \mbox{ and } a_p(E)\not\equiv 2\bmod 3\}.$$ Then ${\mathscr{P}}$ is a Chebotarev set of primes of density $5/16$ and for each $p\in{\mathscr{P}}$ we have $${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[3]{p}))=0.$$
It suffices to check the conditions - in Theorem \[ThmRkZero\] for $\ell=3$.
In lmfdb.org [@LMFDB] one checks that $E$ has good ordinary reduction at $3$ and that $\rho_{E[3]}:G_{\mathbb{Q}}\to GL_2({\mathbb{F}}_3)$ is surjective, thus Conditions and are satisfied.
Consider the quadratic twist $E'=E^{-3}$, which has Cremona label 5013a1. Let us apply Proposition \[PropBSD\]. The following code in Magma checks that both ${\mathrm{Sel}}_3(E/{\mathbb{Q}})$ and ${\mathrm{Sel}}_3(E'/{\mathbb{Q}})$ are trivial:
E1:=EllipticCurve("557b1");
E2:=EllipticCurve("5013a1");
Order(ThreeSelmerGroup(E1));
Order(ThreeSelmerGroup(E2));
1
1
In particular, we obtain:
- ${\mathrm{rank}\,}E({\mathbb{Q}})=0$ and ${\mathrm{rank}\,}E'({\mathbb{Q}})=0$
- $\Sha(E/{\mathbb{Q}})[3]$ and $\Sha(E/{\mathbb{Q}})[3]$ are trivial. Hence, $\Sha(E/{\mathbb{Q}})[3^\infty]$ and $\Sha(E/{\mathbb{Q}})[3^\infty]$ are also trivial.
The exact analytic order of the Shafarevich-Tate groups of $E$ and $E'$ over ${\mathbb{Q}}$ is $1$ (see for instance lmfdb.org). This is the order predicted by the BSD conjecture, and the computation is exact since the rank is $0$.
Therefore, conditions (i), (ii) and (iii) in Proposition \[PropBSD\] hold. In addition, $\# \tilde{E}_3({\mathbb{F}}_3)= 2$ so $3\nmid \#E({\mathbb{Q}}(\mu_3))_{tor}$ and we get $$\frac{L(E,1)}{\Omega_E}\cdot \frac{L(E',1)}{\Omega_{E'}}\sim_3 {\mathrm{Tam}}(E/{\mathbb{Q}}(\mu_3))\cdot \#\Sha(E/{\mathbb{Q}}(\mu_3))[3^\infty].$$ The number on the left is known to be a rational number of small denominator (cf. Paragraph \[SecBSD\]), so the approximation $$\frac{L(E,1)}{\Omega_E}\cdot \frac{L(E',1)}{\Omega_{E'}}\approx \frac{4.14294}{4.14294}\cdot \frac{1.36207 }{0.68103} \approx 2.00001$$ implies that the number is in fact $2$. Thus, Proposition \[PropBSD\] and the fact that $\# \tilde{E}_3({\mathbb{F}}_3)= 2$ show that conditions , , and of Theorem \[ThmRkZero\] hold for $E$ and $\ell=3$.
Increasing the rank in many quadratic extensions
------------------------------------------------
Our next goal is to produce many twists of $E=557b1$ having positive rank over ${\mathbb{Q}}$. For this we will use the results in Section \[SecKL\], which in particular involve the computation of certain logarithm on $p$-adic points of elliptic curves. The next observation will allow us to truncate the relevant power series with controlled $p$-adic precision in our computations.
\[LemmaConv\] Let $F(t)=b_0+b_1 t+b_2 t^2+...\in {\mathbb{Z}}_p[[t]]$ and let $a\in p{\mathbb{Z}}_p$. Suppose that for certain $n\ge 1$ the integer $$m=v_p\left(ab_0 + \frac{a^2b_1}{2} +...+ \frac{a^{n}b_{n-1}}{n}\right)$$ satisfies $m<n-(\log n)/(\log p)$. Then $$v_p\left(\int_0^a F(t)dt\right) =m.$$
Observe that for all $r\ge 1$ we have $$v_p(a^rb_{r-1}/r)\ge r + 0 - v_p(r)\ge r-\frac{\log r}{\log p}$$ because $a\in p{\mathbb{Z}}_p$ and $b_j\in {\mathbb{Z}}_p$. Using this for $r>n$ we get $$v_p\left(ab_0 + \frac{a^2b_1}{2} +...+ \frac{a^{n}b_{n-1}}{n}-\int_0^a F(t)dt\right) >m,$$ hence the result.
Next we produce the required twists of positive rank for $E=557b1$.
\[LemmaRkOne557b1\] Let $E=557b1$. Define the set of primes $${\mathscr{Q}}= \{ q : q\equiv -1\bmod 4,\ (q/7)=1,\ a_q(E)\equiv 1\bmod 2\}.$$ Then ${\mathscr{Q}}$ is a Chebotarev set of primes of density $1/12$ and for each $q\in{\mathscr{Q}}$ we have $${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt{-q}))=1.$$
We are going to apply Theorem \[ThmKL\] and Corollary \[CoroKL\] to $E$ and $K={\mathbb{Q}}(\sqrt{-7})$. Then we will use Lemma \[LemmaPrimes\] to show that the set of primes ${\mathscr{Q}}$ has the required properties.
First, we note that $K$ satisfies the Heegner hypothesis for $2N=2\cdot 557$. This is because a prime $\ell$ splits in $K$ if and only if $\ell$ is a quadratic residue modulo $7$.
The elliptic curve $E$ has good reduction at $2$ and $\#\tilde{E}_2({\mathbb{F}}_2)=1$. It admits an optimal modular parametrization $\varphi: X_0(557)\to E$ (see lmfdb.org) and for this $\varphi$ the Manin constant is $c(\varphi)=1$ as $557$ is odd and squarefree, cf. Paragraph \[SecModular\] (alternatively, there are efficient algorithms to directly compute $c(\varphi)$). Let $\pi: J_0(557)\to E$ be the corresponding optimal quotient. Choosing an embedding $\sigma: K\to {\mathbb{Q}}_2$, we need to show that Condition holds, i.e. $$\label{CondKL557b1}
v_2\left(\log_{2,\omega_E} \sigma(\pi(P_{N,K}))\right)=1$$ where $v_2$ is the $2$-adic valuation in ${\mathbb{Q}}_2$. According to Table 2 in [@KrizLi] this condition indeed holds. However, the details are not included in *loc. cit.* so we do the computation here. For this we use the explicit description of $\log_{2,\omega_E}$ recalled in Paragraph \[SecLog\].
The equation $$\label{EqW}
y^2+y=x^3-x^2-268x+1781$$ is the reduced global minimal Weierstrass model for $E$ over ${\mathbb{Z}}$. The global Néron differential for this model as a power series on $t=-x/y$ (local parameter at $0_E$) is $$\omega_E= (1-t^2+2t^3+...)dt$$ where we used for the model to compute the first few coefficients. All the coefficients in this power series are integers.
Consider the $2$-adic neighborhood of $0_E$ given by $U=\{a\in E({\mathbb{Q}}_2) : t(a)\in 2{\mathbb{Z}}_2\}$. The logarithm map $\log_{2,\omega_E}:U\to{\mathbb{Q}}_2$ is given by $$\label{EqLog557b1}
\log_{2,\omega_E}(a)=t(a) -\frac{1}{3}t(a)^3 + \frac{1}{2}t(a)^4 +...$$ The Hilbert class-field of $K={\mathbb{Q}}(\sqrt{-7})$ is $K$ itself since the class number is $1$ so, in fact, we have (cf. Paragraph \[SecHeegner\]) $$\pi(P_{N,K})=\varphi(Q_{N,K,{\mathfrak{n}}})\in E(K).$$ Using Sage, we do the following: For a suitable choice of ${\mathfrak{n}}$, we compute the point $P=\varphi(Q_{N,K,{\mathfrak{n}}})$ in coordinates over $K$ for the model . Then $t=-x(P)/y(P)$ is computed and we choose a valuation $v$ on $K$ extending the $2$-adic valuation of ${\mathbb{Q}}$ —this is the same as choosing the embedding $\sigma: K\to {\mathbb{Q}}_2$. With all of this, we compute $v(t(P))$ and $v(t(P)-t(P)^3/3)$. Here is the code:
E=EllipticCurve('557b1');
P=E.heegner_point(-7).point_exact(100);
t=-P[0]/P[1];
K=t.parent();
u=QQ.valuation(2);
vK=u.extensions(K); v=vK[0];
v(t); v(t-t^3/3)
1
1
Since $v(t(P))=1$ we see that $\pi(P_{N,K})\in U$ and we can use to compute the logarithm of $P$. Since $v(t(P)-t(P)^3/3)=1$ we can apply Lemma \[LemmaConv\] with $p=2$, $F(t)=\omega_E$, and $n=3$, obtaining $v(\log_{2,\omega_E}(P))=1$. This proves that holds.
Since $E({\mathbb{Q}})[2]=(0)$ and holds, we can apply Theorem \[ThmKL\]. Furthermore, we have $c_2(E)=1$ because $E$ has good reduction at $2$ ($N$ is odd), so we can apply Corollary \[CoroKL\]. We have that ${\mathrm{rank}\,}E({\mathbb{Q}})=0$ and $\Delta_E=557>0$. Therefore, for all squarefree integers $d\equiv 1 \bmod 4$ supported on ${\mathscr{Q}}(E,K)$ with $d<0$ we have ${\mathrm{rank}\,}E^d({\mathbb{Q}})=1$.
We observe that ${\mathscr{Q}}= {\mathscr{Q}}_-(E,K)$ and for each $q\in {\mathscr{Q}}$, the integer $d=-q$ satisfies the previously required conditions, hence ${\mathrm{rank}\,}E^{-q}({\mathbb{Q}})=1$. Since ${\mathrm{rank}\,}E({\mathbb{Q}})=0$ we deduce ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt{-q})=1$ for each $q\in {\mathscr{Q}}$.
It only remains to check that ${\mathscr{Q}}$ (i.e. ${\mathscr{Q}}_-(E,K)$) is a Chebotarev set of density $1/12$ (in particular, that it is non-empty!). We apply Lemma \[LemmaPrimes\]. The degree of the field $L={\mathbb{Q}}(E[2])$ and the discriminant of its only quadratic subfield can be computed on Sage as follows:
E=EllipticCurve('557b1');
L.<a> = E.division_field(2);
L.degree();
L.subfields(2)[0][0].discriminant()
6
557
Thus, we are in case (ii) of Lemma \[LemmaPrimes\], which proves what we wanted.
Proof of the main result
------------------------
Consider the elliptic curve $E=557b1$ and let ${\mathscr{P}}$ and ${\mathscr{Q}}$ be the sets of primes given by Lemmas \[LemmaRkZero557b1\] and \[LemmaRkOne557b1\]. For each $p\in{\mathscr{P}}$ and each $q\in{\mathscr{Q}}$ we have that ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt[3]{p}))=0$ and ${\mathrm{rank}\,}E({\mathbb{Q}}(\sqrt{-q}))=1$. By Proposition \[PropMain\] with $M={\mathbb{Q}}$, $F={\mathbb{Q}}(\sqrt[3]{p})$ and $L={\mathbb{Q}}(\sqrt{-q})$ we get that the extension ${\mathbb{Q}}(\sqrt[3]{p}, \sqrt{-q})/{\mathbb{Q}}(\sqrt[3]{p})$ is integrally Diophantine.
Since ${\mathbb{Q}}(\sqrt[3]{p})$ has exactly one complex place, the extension ${\mathbb{Q}}(\sqrt[3]{p})/{\mathbb{Q}}$ is integrally Diophantine by the Pheidas-Shlapentokh-Videla theorem [@Pheidas; @ShlH10; @Videla]. Therefore, Lemma \[LemmaTransfer\] gives that ${\mathbb{Z}}$ is Diophantine in ${\mathbb{Q}}(\sqrt[3]{p},\sqrt{q})$.
Acknowledgments
===============
The first author was supported by the FONDECYT Iniciación en Investigación grant 11170192, and the CONICYT PAI grant 79170039. The second author was supported by FONDECYT Regular grant 1190442.
We gratefully acknowledge the hospitality of the IMJ-PRG in Paris during our visit in February 2019, where the final ideas of the project were developed. We heartily thank Loïc Merel for this invitation and for several enlightening discussions on elliptic curves.
The final technical details of this work were worked out and presented in May 2019 at the AIM meeting “Definability and Decidability problems in Number Theory”. We thank the AIM for their hospitality and the participants for their valuable feedback. Specially, we thank Karl Rubin and Alexandra Shlapentokh for their interest and technical remarks.
Comments by Matias Alvarado, Chao Li, and Barry Mazur on an earlier version of this manuscript are gratefully acknowledged.
[9]{}
A. Abbes, E. Ullmo, *À propos de la conjecture de Manin pour les courbes elliptiques modulaires*. Compositio Math. 103 (1996), no. 3, 269-286.
A. Agashe, K. Ribet, W. Stein, *The Manin constant*. Pure Appl. Math. Q. 2 (2006), no. 2, Special Issue: In honor of John H. Coates. Part 2, 617-636.
B. J. Birch, *Diophantine analysis and modular functions*. 1969 Algebraic Geometry (Internat. Colloq., Tata Inst. Fund. Res., Bombay, 1968) pp. 35-42 Oxford Univ. Press, London.
C. Breuil, B. Conrad, F. Diamond, R. Taylor, *On the modularity of elliptic curves over ${\mathbb{Q}}$: wild $3$-adic exercises*. J. Amer. Math. Soc. 14 (2001), no. 4, 843-939.
D. Bump, S. Friedberg, J. Hoffstein, *Eisenstein series on the metaplectic group and non-vanishing theorems for automorphic $L$-functions and their derivatives*. Annals of Math. 131, 1990, 53-127.
J. Cassels, *Arithmetic on curves of genus 1*. VIII. On conjectures of Birch and Swinnerton-Dyer. J. Reine Angew. Math. 217 (1965) 180-199.
G. Cornelissen, T. Pheidas, K. Zahidi, *Division-ample sets and the Diophantine problem for rings of integers*. J. Théor. Nombres Bordeaux 17 (2005), no. 3, 727-735.
M. Davis, H. Putnam, J. Robinson, *The decision problem for exponential diophantine equations*. Ann. of Math. (2) 74 (1961) 425-436.
J. Denef, *Diophantine sets over algebraic integer rings*. II. Trans. Amer. Math. Soc. 257 (1980), no. 1, 227-236.
J. Denef, L. Lipshitz, *Diophantine sets over some rings of algebraic integers*. J. London Math. Soc. (2) 18 (1978), no. 3, 385-391.
T. Dokchitser, *Ranks of elliptic curves in cubic extensions*. Acta Arith. 126 (2007), no. 4, 357-360.
V. Dokchitser, *Root numbers of non-abelian twists of elliptic curves*. With an appendix by Tom Fisher. Proc. London Math. Soc. (3) 91 (2005), no. 2, 300-324.
T. Dokchitser, V. Dokchitser, *Computations in non-commutative Iwasawa theory*. With an appendix by J. Coates and R. Sujatha. Proc. Lond. Math. Soc. (3) 94 (2007), no. 1, 211-272.
T. Dokchitser, V. Dokchitser, *On the Birch-Swinnerton-Dyer quotients modulo squares*. Ann. of Math. (2) 172 (2010), no. 1, 567-596.
B. Gross, D. Zagier, *Heegner points and derivatives of $L$-series*. Inv. Math. 84, 1986, 225-320.
R. Greenberg, *Introduction to Iwasawa theory for elliptic curves*. Arithmetic algebraic geometry (Park City, UT, 1999), 407-464, IAS/Park City Math. Ser., 9, Amer. Math. Soc., Providence, RI, 2001.
Y. Hachimori, K. Matsuno, *An analogue of Kida’s formula for the Selmer groups of elliptic curves*. J. Algebraic Geom. 8 (1999), no. 3, 581-601.
K. Heegner, *Diophantische Analysis und Modulfunktionen*. Math. Z. 56, (1952). 227-253.
V. Kolyvagin, *Euler Systems*. The Grothendieck Festschrift, Eds. P. Cartier, et al., vol. II, Progr. in Math. 87, Birkhäuser, 1990, 435-483.
D. Kriz, C. Li, *Goldfeld’s conjecture and congruences between Heegner points*. Forum of Mathematics, Sigma, 7, E15 (2019).
The LMFDB Collaboration, *The L-functions and Modular Forms Database*, <http://www.lmfdb.org>
Y. Matiyasevich, *The Diophantineness of enumerable sets*. (Russian) Dokl. Akad. Nauk SSSR 191 (1970) 279-282.
B. Mazur, *Rational points of abelian varieties with values in towers of number fields*. Invent. Math. 18 (1972), 183-266.
B. Mazur, *Rational isogenies of prime degree* (with an appendix by D. Goldfeld). Invent. Math. 44 (1978), no. 2, 129-162.
B. Mazur, K. Rubin, *Ranks of twists of elliptic curves and Hilbert’s tenth problem*. Invent. Math. 181 (2010), no. 3, 541-575.
B. Mazur, K. Rubin, *Diophantine stability*. With an appendix by Michael Larsen. Amer. J. Math. 140 (2018), no. 3, 571-616.
J. Milne, *On the arithmetic of abelian varieties*. Invent. Math. 17 (1972), 177-190.
J. Milne, *Arithmetic duality theorems*. Second edition. BookSurge, LLC, Charleston, SC, 2006. P. Monsky, *Mock Heegner points and congruent numbers*. Math. Z. 204 (1990), no. 1, 45-67.
M. R. Murty, V. K. Murty, *Mean values of derivatives of modular $L$-series*, Annals of Math. 133, 1991, 447-475.
M. R. Murty, H. Pasten, *Elliptic curves, L-functions, and Hilbert’s tenth problem*. J. Number Theory 182 (2018), 1-18.
V. Pal, *Periods of quadratic twists of elliptic curves*. With an appendix by Amod Agashe. Proc. Amer. Math. Soc. 140 (2012), no. 5, 1513-1525.
B. Perrin-Riou, *Arithmétique des courbes elliptiques et théorie d’Iwasawa*. (French) \[Arithmetic of elliptic curves and Iwasawa theory\] Mém. Soc. Math. France (N.S.) No. 17 (1984), 130 pp.
T. Pheidas, *Hilbert’s tenth problem for a class of rings of algebraic integers*. Proc. Amer. Math. Soc. 104 (1988), no. 2, 611-620.
B. Poonen, *Using elliptic curves of rank one towards the undecidability of Hilbert’s tenth problem over rings of algebraic integers*. Algorithmic number theory (Sydney, 2002), 33-42, Lecture Notes in Comput. Sci., 2369, Springer, Berlin, 2002.
P. Schneider, *Iwasawa L-functions of varieties over algebraic number fields. A first approach*. Invent. Math. 71 (1983), no. 2, 251-293.
J. Silverman, *Advanced topics in the arithmetic of elliptic curves*. Graduate Texts in Mathematics, 151. Springer-Verlag, New York, 1994. H. Shapiro, A. Shlapentokh, *Diophantine relationships between algebraic number fields*. Comm. Pure Appl. Math. 42 (1989), no. 8, 1113-1122.
A. Shlapentokh, *Extension of Hilbert’s tenth problem to some algebraic number fields*. Comm. Pure Appl. Math. 42 (1989), no. 7, 939-962.
A. Shlapentokh, *Elliptic curves retaining their rank in finite extensions and Hilbert’s tenth problem for rings of algebraic numbers*. Trans. Amer. Math. Soc. 360 (2008), no. 7, 3541-3555.
J. Tate, *On the conjectures of Birch and Swinnerton-Dyer and a geometric analog*. Séminaire Bourbaki, Vol. 9, Exp. No. 306, 415-440, Soc. Math. France, Paris, 1965.
R. Taylor, A. Wiles, *Ring-theoretic properties of certain Hecke algebras*. Ann. of Math. (2) 141 (1995), no. 3, 553-572.
C. Videla, *Sobre el décimo problema de Hilbert*. Atas da Xa Escola de Algebra, Vitoria, ES, Brasil. Colecao Atas 16 Sociedade Brasileira de Matematica (1989), 95-108.
A. Wiles, *Modular elliptic curves and Fermat’s last theorem*. Ann. of Math. (2) 141 (1995), no. 3, 443-551.
[^1]: N. G.-F. was supported by the FONDECYT Iniciación en Investigación grant 11170192, and the CONICYT PAI grant 79170039. H.P. was supported by FONDECYT Regular grant 1190442.
[^2]: A pseudo isomorphism is a morphism with finite kernel and cokernel.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In [@5] we proved that generically functions defined in any open set can be approximated by a sequense of their Padé approximants, in the sense of uniform convergence on compacta. In this paper we examine a more particular space, $A^{\infty}(\Omega)$ , and prove that we can obtain similar approximation results with functions smooth on the boundary.'
author:
- 'G. Fournodavlos'
title: |
**Generic Approximation of functions\
by their Padé approximants, II**
---
[*Subject Classification MSC2010*]{}: primary 41A21, 30K05 secondary 30B10, 30E10, 30K99, 41A10, 41A20.\
[*Key words*]{}: Padé approximant, Taylor series, Baire’s theorem, Runge’s theorem, generic property.
Introduction {#sec1}
============
On a disc every holomorphic function $f$ can be approximated by the partial sums of its Taylor expansion. It is also true that generically in a simply connected domain every holomorphic function is the uniform on compacta limit of a subsequense of the partial sums of its Taylor expansion. The partial sums are polynomials and thus by the maximum principle we are led to uniform approximation on compact sets with connected complement. If we replace the partial sums by the Padé approximants $[p/q]_f$, which are rational functions with poles, then we obtain approximation on compact sets with arbitrary connectivity ([@4], [@5]).
In [@2] it was proved that generically every entire function can be approximated uniformly on compacta by a subsequense $[p_n/q_n]_f$ of its Padé approximants, provided $p_n{\;\rightarrow\;}+\infty$ and $p_n-q_n{\;\rightarrow\;}+\infty$. In [@5] we weakened the previous assumption to $p_n{\;\rightarrow\;}+\infty$ only and we extended the result to any simply connected domain. We also obtained the same approximation on any open subset of ${\mathbb{C}}$ (of arbitrary connectivity) under the assumption $p_n{\;\rightarrow\;}+\infty$ and $q_n{\;\rightarrow\;}+\infty$.
In the present paper we obtain similar results using smooth functions; that is, holomorphic functions on ${{\varOmega}}$ such that every derivative $f^{(l)}$ extends continuously on ${\overline{\varOmega}}$ $(f\in A^{\infty}({{\varOmega}}))$. In the case of a domain ${{\varOmega}}\subseteq{\mathbb{C}}$ such that $({\mathbb{C}}\cup\{\infty\}{\smallsetminus}{\overline{\varOmega}}$ is connected we obtain the result provided $p_n{\;\rightarrow\;}+\infty$; this result is generic in a subset of $A^{\infty}({{\varOmega}})$, which is the closure of the set of polynomials, under the natural topology. We do not know in general if this subspace is the whole $A^{\infty}({{\varOmega}})$. If there exists a finite constant $M<\infty$ so that all points $A,B\in {\overline{\varOmega}}$ can be joined in ${\overline{\varOmega}}$ by a curve ${{\varGamma}}$ with length $|{{\varGamma}}|\le M$ , then the polynomials are dense in $A^{\infty}({{\varOmega}})$ ([@11]). In the case $q=0$ the generic result in the closure of polynomials in $A^{\infty}({{\varOmega}})$ is known ([@9]).
Finally in the general case of an open subset ${{\varOmega}}\subseteq{\mathbb{C}}$ we obtain a similar generic result in the closure in $A^{\infty}({{\varOmega}})$ of holomorphic functions in some varying neighborhood of ${{\varOmega}}$, provided $p_n{\;\rightarrow\;}+\infty$ and $q_n{\;\rightarrow\;}+\infty$. Our method of proof is based on Baire’s Category theorem ([@7], [@8]) and extends the methods of [@4] and [@13].
Preliminaries {#sec2}
=============
Let ${{\varOmega}}\subseteq{\mathbb{C}}$ be an open set and let us consider the set $A^{\infty}({{\varOmega}})=\{ f\in H({{\varOmega}})$: $f^{(l)}$ extends continuously on ${\overline{\varOmega}}$, $l=0,1,{\ldots}$}, where $f^{(l)}, l=0,1,{\ldots}$ denote the derivatives of the holomorphic function $f$.
We define the following metric $\rho$ on $A^{\infty}({{\varOmega}})$: $$\rho(f,g)=\sum^\infty_{l=0}\sum^\infty_{n=1}\frac{1}{2^{l+n}}\min\{\|f-g\|_{l,n},1\},$$ where $\|f-g\|_{l,n}={\displaystyle}\sup_{z\in {\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,n)}}|f^{(l)}-g^{(l)}|$, $l=0,1,{\ldots}$. It is easy to see that a sequense in $A^{\infty}({{\varOmega}})$, $(f_m)_{m\in{\mathbb{N}}}$ converges $f_m\overset{\rho}{\longrightarrow}f\in A^{\infty}({{\varOmega}})$, if and only if ${f_m}^{(l)}{\;\rightarrow\;}f^{(l)}$ uniformly on each compact subset of ${\overline{\varOmega}}$, for every $l$. The space $(A^{\infty}({{\varOmega}}),\rho)$ is compete.
Let $f$ be a function holomorphic in a neighborhood of 0 and let $f(z)={\sum\limits}^\infty_{v=0}a_vz^v$ its Taylor series. A Padé approximant $[p/q]_f$ of $f$, $p,q\in\{0,1,2,{\ldots}\}$, is a rational function of the form $$\frac{{\sum\limits}^p_{v=0}n_vz^v}{{\sum\limits}^q_{v=0}d_vz^v}, \ \ d_0=1.$$ such that its Taylor series ${\sum\limits}^\infty_{v=0}b_vz^v$ coincides with ${\sum\limits}^\infty_{v=0}a_vz^v$ up to the first $p+q+1$ terms; that is $b_v=a_v$ for $v=0,{\ldots},p+q$ ([@1]).
We notice that in case of $q=0$ there exists always a unique Padé approximant of $f$ and $[p/q]_f(z)= S_p(z)$, where $S_p(z)= {\displaystyle}\sum^p_{v=0}a_vz^v$. For $q\ge 1$ it is true that there exists a unique Padé approximant of $f$, if and only if the following determinant is not zero: $$\det\left|\begin{array}{cccc}
a_{p-q+1} & a_{p-q+2} & \cdots & a_p \\
a_{p-q+2} & a_{p-q+3} & \cdots & a_{p+1} \\
\vdots & \vdots & & \vdots \\
a_p & a_{p+1} & \cdots & a_{p+q-1} \\
\end{array}\right|\neq0, \ \ a_i=0, \ \ \text{when} \ \ i<0.
\eqno(\ast)$$ Then we write $f\in D_{p,q}$.
If $f\in D_{p,q}$, then $[p/q]_f$ $(q\ge1)$ is given by the Jacobi explicit formula: $$[p/q]_f=\frac{\det\left|\begin{array}{cccc}
z^qS_{p-q}(z) & z^{q-1}S_{p-q+1} & \cdots & S_p(z) \\
a_{p-q+1} & a_{p-q+2} & \cdots & a_{p+1} \\
\vdots & \vdots & & \vdots \\
a_p & a_{p+1} & \cdots & a_{p+q} \\
\end{array}\right|}{\det\left|\begin{array}{cccc}
z^q & z^{q-1} & \cdots & 1 \\
a_{p-q+1} & a_{p-q+2} & \cdots & a_{p+1} \\
\vdots & \vdots & & \vdots \\
a_p & a_{p+1} & \cdots & a_{p+q} \\
\end{array}\right|},$$ with $S_k(z)=\left\{\begin{array}{cc}
{\sum\limits}^k_{v=0}a_vz^v, & k\ge0 \\
0, & k<0. \\
\end{array}\right.$
\[rem2.1\] If all of the coefficients $\frac{f^{(v)}(0)}{v!}=a_v$, $v=0,1,{\ldots},p+q$, involved in the determinant $(\ast)$ depend linearly on $d\in{\mathbb{C}}$, $a_v=c_v\cdot d+{\tau }_v$, such that $c_v=0$, when $v<p$ and $c_p\neq0$, then the determinant is a polynomial in $d$ of degree $q$ and hence only for finite values of $d$ the determinant is zero.
If $L$ is any set we write $h\in H(L)$ if $h$ is holomorphic in some open set containing $L$. We also denote $\|h\|_L={\displaystyle}\sup_{z\in L}|h(z)|$, for every function $h:L\rightarrow{\mathbb{C}}$ on the set $L$.
\[lem2.2\] Let $r>0$, $p,q,s\in{\mathbb{N}}$ and $K\subseteq{\mathbb{C}}$ a compact set. If $f\in\ H(\overline{{{\varDelta}}(0,r)})$, $f\in D_{p,q}$ such that its Padé approximant $[p/q]_f$ has no poles in $K$, then for every ${\varepsilon }>0$ there exists ${\delta }>0$ such that for every $g\in\ H(\overline{{{\varDelta}}(0,r)})$ with $\|g-f\|_{\overline{{{\varDelta}}(0,r)}}<{\delta }$ it holds $g\in\ D_{p,q}$ and $\|[p/q]^{(l)}_g-[p/q]^{(l)}_f\|_K<{\varepsilon }$, $\forall l\in\{0,1,{\ldots},s\}$.
Let ${\varepsilon }>0$. Observe that the determinant $(\ast)$ and the coefficients of the numerator and the denominator of $[p/q]^{(l)}_f$, $l=0,1,{\ldots},s$ depend polynomially on $\frac{f^{(v)}(0)}{v!}$, $v=0,1,{\ldots},p+q$. This implies that there exists $\widetilde{{\delta }}$ such that for every $g\in H(\overline{{{\varDelta}}(0,r)})$ with $\big|\frac{g^{(v)}(0)}{v!}-\frac{f^{(v)}(0)}{v!}\big|<\widetilde{{\delta }}$, $v=0,1,{\ldots},p+q$ it holds $g\in D_{p,q}$ and $\|[p/q]^{(l)}_g-[p/q]^{(l)}_f\|_K<{\varepsilon }$, $l=0,1,{\ldots},s$.
If $0<{\delta }<\min\{r^v\cdot\widetilde{{\delta }}\;|\; v=0,1,{\ldots},p+q\}$ and $\|g-f\|_{\overline{{{\varDelta}}(0,r)}}<{\delta }$, then by Cauchy’s estimates we obtain: $$\bigg|\frac{g^{(v)}(0)}{v!}-\frac{f^{(v)}(0)}{v!}\bigg|=
\bigg|\frac{(g-f)^{(v)}(0)}{v!}\bigg|\le\frac{\|g-f\|_{\overline{{{\varDelta}}(0,r)}}}{r^v}<\frac{{\delta }}{r^v}<\widetilde{{\delta }}.
\qquad \text{{$\quad\blacksquare$}}$$
\[rem2.3\] It follows from Lemma \[lem2.2\] that $D_{p,q}\cap\ A^{\infty}({{\varOmega}})$ is open ($0\in {{\varOmega}}$).
A special case {#sec3}
==============
Let ${{\varOmega}}\subseteq{\mathbb{C}}$ be an open set containing 0, such that $({\mathbb{C}}\cup\infty){\smallsetminus}{\overline{\varOmega}}$ is connected. Also, let $F\subseteq{\mathbb{N}}\times{\mathbb{N}}$ which contains a sequence $(\widetilde{p}_m,\widetilde{q}_m)_{m\in{\mathbb{N}}}$, such that $\widetilde{p}_m{\;\rightarrow\;}+\infty$. We define
- $B_F=\{f\in A^{\infty}({{\varOmega}})$: there exists $(p_m,q_m)_{m\in{\mathbb{N}}}$ in $F$ such that $f\in D_{p_m,q_m}$, for all $m\in{\mathbb{N}}$ and for every $K\subseteq{\overline{\varOmega}}$ compact $[p_m/q_m]^{(l)}_f{\;\rightarrow\;}f^{(l)}$ uniformly on $K$, for each $l=0,1,{\ldots}\}$.
- $E(n,s,(p,q))=\{f\in A^{\infty}({{\varOmega}}):f\in D_{p,q}$ and $\|[p/q]_f-f\|_{l,n}<1/s, l=0,1,{\ldots},s\}$, $n,s\in{\mathbb{N}}, (p,q)\in F$.
\[lem3.1\] $B_F=\bigcap\limits^\infty_{n,s=1}\bigcup\limits_{(p,q)\in
F}E(n,s,(p,q)$.
It is standard and is omitted. \[A similar proof can be found in [@12]\]. [$\quad\blacksquare$]{}
\[lem3.2\] $E(n,s,(p,q))$ is open.
$D_{p,q}\cap{\mathbb{A}}^{\infty}({{\varOmega}})$ is open (Remark \[rem2.3\]) and similarly to the proof of the Lemma \[lem2.2\], we can prove that the map $f\mapsto\|[p/q]_f-f\|_{l,n}$ is continuous, for any $l$. [$\quad\blacksquare$]{}
We will now focus our attention on a more accessible space, $H({\overline{\varOmega}})$, which is a subspace of $A^{\infty}({{\varOmega}})$ and is considered with its relative topology.
\[lem3.3\] The polynomials are dense in $H({\overline{\varOmega}})$.
Let $f\in H({\overline{\varOmega}})$ and ${\varepsilon }>0$. It suffices to show that for $N=N({\varepsilon })\in{\mathbb{N}}$ and $L=L({\varepsilon })\in{\mathbb{N}}$ there exists a polynomial $P$ such that $\|P-f\|_{l,N}<{\varepsilon }$, $\forall l\le L$. Observe that $({\mathbb{C}}\cup\infty){\smallsetminus}({\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,N)})$ is connected.
- $f\in H({\overline{\varOmega}})$, thus, there exists $U\subseteq{\mathbb{C}}$ open such that $f\in H(U)$ and ${\overline{\varOmega}}\subseteq U$.
- It is true that we can find $V\subseteq{\mathbb{C}}$ open, such that ${\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,N)}\subseteq V\subseteq U$ and $({\mathbb{C}}\cup\infty){\smallsetminus}V$ connected (in other words $V$ is simply connected) ([@3], [@6]).
By Runge’s theorem there exists a sequense of polynomials $(P_i)_{i\in{\mathbb{N}}}$, such that $P^{(l)}_i{\;\rightarrow\;}f^{(l)}$ uniformly on each compact subset of $V$ for every $l$, which completes the proof. [$\quad\blacksquare$]{}
\[thm3.4\] $B_F\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is $G_{\delta }$ and dense in $cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$. (Hence $B_F\neq\emptyset$).
Lemma \[lem3.2\] implies that $\bigcup\limits_{(p,q)\in
F}E(n,s(p,q))\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is open in $cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$. By Lemma \[lem3.1\] $B_F\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is $G_{\delta }$ in $cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$. We claim that $\bigcup\limits_{(p,q)\in F}E(n,s,(p,q))\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is dense in $cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$. If this is true, then Baire’s Category theorem completes the proof. By Lemma \[lem3.3\] it suffices to prove that for every polynomial $P$ and ${\varepsilon }>0$ there exists $f\in\bigcup\limits_{(p,q)\in F}E(n,s,(p,q))\cap H({\overline{\varOmega}})$ such that $\|P-f\|_{l,N}<{\varepsilon }$, for every $l\le L=L({\varepsilon })\in{\mathbb{N}}$, where $N=N({\varepsilon })\in{\mathbb{N}}$.
- Let $P$ be a polynomial and ${\varepsilon }>0$. There exists $(p,q)\in F$ such that $p>\text{deg}P$.
If $q=0$, define $f(z)=P(z)+dz^p$, $d\in{\mathbb{C}}{\smallsetminus}\{0\}$. It is immediate that $f\in D_{p,q}$ and $[p/q]_f=f$. It follows $f\in
E(n,s,(p,q))\cap H({\overline{\varOmega}})$. In addition, $\|f-P\|_{l,N}=|d|\cdot\|z^p\|_{l,N}<{\varepsilon }$, $\forall l\le L$, when $0<|d|<{\varepsilon }/{\displaystyle}\max_{0\le l\le L}\|z^p\|_{l,N}$.
If $q\ge1$, we define $\widetilde{f}_j(z)=\frac{P(z)+d_jz^p}{1-(c_jz)^q}$, $c_j,d_j\in{\mathbb{C}}{\smallsetminus}\{0\}$, where $d_j$ and $c_j$ will be determined later on, $j\in{\mathbb{N}}$.
- Let ${\lambda }>\max\{n,N\}$. We have ${\displaystyle}\inf_{z\in\overline{{{\varDelta}}(0,{\lambda })}}|1-(c_jz)^q|\ge1-|c_j|^q\cdot\|z\|^q_{\overline{{{\varDelta}}(0,{\lambda })}}>\frac{1}{2}$, when $0<|c_j|<\frac{1}{2^{\frac{1}{q}}\|z\|_{\overline{{{\varDelta}}(0,{\lambda })}}}$, $j\in{\mathbb{N}}$.
- We have $\|\widetilde{f}_j(z)-P(z)\|_{\overline{{{\varDelta}}(0,{\lambda })}}=\big\|\frac{P(z)+d_jz^p}{1-(c_jz)^q}\big\|_{\overline{{{\varDelta}}(0,{\lambda })}}\le
2(\|P(z)\|_{\overline{{{\varDelta}}(0,{\lambda })}}\cdot|c_j|^q\cdot\|z\|^q_{\overline{{{\varDelta}}(0,{\lambda })}}+|d_j|\cdot\|z\|^q_{\overline{{{\varDelta}}(0,{\lambda })}})$. Thus, there exists ${\delta }_j>0$, $j\in{\mathbb{N}}$, such that: $\|\widetilde{f}_j-P\|_{\overline{{{\varDelta}}(0,{\lambda })}}<1/j$, when $0<|c_j|<{\delta }_j<
\frac{1}{2^{\frac{1}{q}}\|z\|_{\overline{{{\varDelta}}(0,{\lambda })}}}$ and $0<|d_j|<{\delta }_j$. Hence, $\widetilde{f}_j{\;\rightarrow\;}P$ uniformly on ${\overline{{{\varDelta}}(0,{\lambda })}}$ and so $\widetilde{f}^{(l)}_j{\;\rightarrow\;}P^{(l)}$ uniformly on ${\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,N)}$ (which is contained in ${{\varDelta}}(0,{\lambda })$), for every $l$. Therefore, there exists $j_0\in{\mathbb{N}}$ such that $\|\widetilde{f}_{j_0}-P\|_{l,N}<{\varepsilon }/2$, $l=0,1,{\ldots},L$.
- We fix $c_{j_0}$ satisfying the above. Around 0, $\widetilde{f}_{j_0}(z)=P(z)+d_{j_0}z^p+P(z)\cdot(c_{j_0}z)^q+d_{j_0}z^p\cdot(c_{j_0}z)^q+\cdots$. According to Remark \[rem2.1\] we can choose $0<|d_{j_0}|<{\delta }_{j_0}$, such that $\widetilde{f}_{j_0}\in D_{p,q}$. By the uniqueness of the Padé approximant of $\widetilde{f}_{j_0}$ we obtain $[p/q]_{\widetilde{f}_{j_0}}=\widetilde{f}_{j_0}$.
- Let $r>0$: $\overline{{{\varDelta}}(0,r)}\subseteq{{\varOmega}}\cap{{\varDelta}}(0,{\lambda })$. By Lemma \[lem2.2\] there exists ${\delta }>0$ such that for every $f\in H(\overline{{{\varDelta}}(0,r)})$ with $\|f-\widetilde{f}_{j_0}\|_{\overline{{{\varDelta}}(0,r)}}<{\delta }$ it holds $\|[p/q]_f-[p/q]_{\widetilde{f}_{j_0}}\|_{l,n}<1/2s$, $\forall l\le s$. Also, we demand $0<{\delta }<\min\{1/2s,{\varepsilon }/2\}$.
- The Taylor series of $\widetilde{f}_{j_0}$ around 0 has radius of convergence greater than $\frac{1}{|c_{j_0}|}>{\lambda }$. It follows that its partial sums $\big(\sum^k_{v=0}\frac{\widetilde{f}^{(v)}_{j_0}(0)}{v!}z^v\big)^{(l)}{\;\rightarrow\;}\widetilde{f}^{(l)}_{j_0}$ uniformly on ${\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,{\lambda })}$, for every $l$. Hence, there exists a partial sum $f(z)=\sum^{k_0}_{v=0}\frac{\widetilde{f}^{(v)}_{j_0}(0)}{v!}z^v$ such that $\|f-\widetilde{f}_{j_0}\|_{l,{\lambda }}<{\delta }$, $\forall l\le\max\{s,L\}$.
- $f$ satisfies: $\|[p/q]_f-f\|_{l,n}\le\|[p/q]_f-[p/q]_{\widetilde{f}_{j_0}}\|_{l,n}
+\|\widetilde{f}_{j_0}-f\|_{l,n}<1/2s+{\delta }<1/s$, $\forall l\le s$. It follows that $f\in E(n,s,(p,q))\cap H({\overline{\varOmega}})$. Also, it holds $\|f-P\|_{l,N}\le\|f-\widetilde{f}_{j_0}\|_{l,N}+\|\widetilde{f}_{j_0}-P\|_{l,N}<
{\varepsilon }/2+{\delta }<{\varepsilon }$, $\forall l\le L$.
This completes the proof. [$\quad\blacksquare$]{}
The general case {#sec4}
================
Let ${{\varOmega}}\subseteq{\mathbb{C}}$ be an open set containing 0. Also, let $F\subseteq{\mathbb{N}}\times{\mathbb{N}}$ which contains a sequence $(\widetilde{p}_m,\widetilde{q}_m)_{m\in{\mathbb{N}}}$ such that $\widetilde{p}_m{\;\rightarrow\;}+\infty$ and $\widetilde{q}_m{\;\rightarrow\;}+\infty$. We define $B_F$ and $E(n,s,(p,q))$ similarly as in Section \[sec3\].
The analogue of Lemmas \[lem3.1\], \[lem3.2\] hold in this case also. Like before we concentrate on $H({\overline{\varOmega}})$ and its closure in $A^{\infty}({{\varOmega}})$.
\[lem4.1\] The rational functions with poles off ${\overline{\varOmega}}$ are dense in $H({\overline{\varOmega}})$.
Let $f\in H({\overline{\varOmega}})$. There exists $U\subseteq{\mathbb{C}}$ open (depending on $f$) such that ${\overline{\varOmega}}\subseteq U$ and $f\in H(U)$. By Runge’s theorem there exists a sequense $(R_i)_{i\in{\mathbb{N}}}$ of rational functions with poles in $({\mathbb{C}}\cup\infty){\smallsetminus}U$, hence $R_i\in H(U)\subseteq H({\overline{\varOmega}})$, $\forall i\in{\mathbb{N}}$, such that $R_i{\;\rightarrow\;}f$ uniformly on each compact set of $U$. Similarly to the Lemma \[lem3.3\], for a given ${\varepsilon }$ there exists $i_0$ such that $\|R_i-f\|_{l,N}<{\varepsilon }/2$, $\forall l\le L$, where $L=L({\varepsilon })\in{\mathbb{N}}$, $N=N({\varepsilon })\in{\mathbb{N}}$ are chosen so that $\rho(f,R_{i_0})<{\varepsilon }$. [$\quad\blacksquare$]{}
\[thm4.2\] $B_F\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is $G_{\delta }$ and dense in $cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$. (Hence $B_F\neq\emptyset$).
Since $\bigcup\limits_{(p,q)\in F}E(n,s,(p,q))$ is open, it follows that $B_F\cap cl_{A^{\infty}({{\varOmega}})}H({\overline{\varOmega}})$ is $G_{\delta }$ in the subspace. By Baire’s Category theorem the proof would be complete if the set $\bigcup\limits_{(p,q)\in F}E(n,s,(p,q))\cap H({\overline{\varOmega}})$ was dense in $H({\overline{\varOmega}})$, $n,s\in{\mathbb{N}}$. By Lemma \[lem4.1\] it suffices to show that for every rational function $R$ with poles off ${\overline{\varOmega}}$ (or $R\in H({\overline{\varOmega}})$) and every ${\varepsilon }>0$ there exists $f\in\bigcup\limits_{(p,q)\in F}E(n,s,(p,q))\cap H({\overline{\varOmega}})$ such that $\|f-R\|_{l,N}<{\varepsilon }$, $\forall l\le L=L({\varepsilon })\in{\mathbb{N}}$, $N=N({\varepsilon })\in{\mathbb{N}}$.
- Let $R(z)=\frac{A(z)}{B(z)}$ be a rational function with poles only in $({\mathbb{C}}\cup\{\infty\}){\smallsetminus}{\overline{\varOmega}}$, where $A$, $B$ are polynomials and let ${\varepsilon }>0$. There exists $(p,q)\in F$ such that $p>\text{deg}A$ and $q>\text{deg}B$. We define $\widetilde{f}_j(z)=\frac{A(z)+d_jz^p}{B(z)-(c_jz)^q}$, $c_j,d_j\in{\mathbb{C}}{\smallsetminus}\{0\}$, where $c_j$ and $d_j$ will be determined later on, $j\in{\mathbb{N}}$.
- Since $R$ has no poles in ${\overline{\varOmega}}$, there exists $U\subseteq{\mathbb{C}}$ open such that $R\in H(U)$ and ${\overline{\varOmega}}\subseteq U$. Also, there exists $K\subseteq U$ compact such that $K^0\supseteq{\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,{\lambda })}$, where ${\lambda }=\max\{n,N\}$, and every component of $({\mathbb{C}}\cup\infty){\smallsetminus}K$ contains at least one component of $({\mathbb{C}}\cup\infty){\smallsetminus}U$ ([@14]).
- We have $B(0)\neq 0$ and ${\displaystyle}\inf_{z\in K}|B(z)|>0$. Furthermore, ${\displaystyle}\inf_{z\in K}|B(z)-(c_jz)^q|\ge{\displaystyle}\inf_{z\in K}|B(z)|-|c_j|^q\cdot\|z\|^q_K>0$, when $0<|c_j|<\big(\frac{\inf_K|B(z)|}{\|z\|^q_K}\big)^{1/q}$, $\forall j\in{\mathbb{N}}$. Thus, $$\|\widetilde{f}_j(z)-R(z)\|_K= \bigg\|\frac{A(z)(c_jz)^q+B(z)d_jz^p}{B(z)(B(z)-(c_jz)^q)}\bigg\|_K\le
\frac{\|A(z)\|_K\cdot|c_j|^q\cdot\|z\|^q_K+\|B(z)\|_K\cdot|d_j|\cdot\|z\|^p_K}{{\displaystyle}\inf_{z\in K}|B(z)|\cdot{\displaystyle}\inf_{z\in K}|B(z)-(c_jz)^q|}.$$
- There exists ${\delta }_j>0$, $j\in{\mathbb{N}}$, such that $\|\widetilde{f}_j-R\|_K<1/j$, whenever $|c_j|<{\delta }_j<\big(\frac{\inf_K|B(z)|}{\|z\|^q_K}\big)^{1/q}$ and $|d_j|<{\delta }_j$, $\forall j\in{\mathbb{N}}$. Hence, $\widetilde{f}_j{\;\rightarrow\;}R$ uniformly on $K$ and $\widetilde{f}^{(l)}_j{\;\rightarrow\;}R^{(l)}$ on each compact subset of $K^0$, for every $l$. This implies that there exists $j_0\in{\mathbb{N}}$, such that $\|\widetilde{f}_{j_0}-R\|_{l,N}<{\varepsilon }/2$, $\forall l\le L$.
- We fix $c_{j_0}$ satisfying the above. Around 0 we have: $\widetilde{f}_{j_0}(z)=B^{-1}(0)A(z)+B^{-1}(0)d_{j_0}z^p-B^{-1}(0)A(z)\cdot(\widetilde{B}(z)-1)-
B^{-1}(0)d_{j_0}z^p\cdot(\widetilde{B}(z)-1)+ \cdots$, where $\widetilde{B}(z)=B^{-1}(0)B(z)-B^{-1}(0)(c_{j_0}z)^q$. By Remark \[rem2.1\] we can choose $0<|d_{j_0}|<{\delta }_{j_0}$ such that $\widetilde{f}_{j_0}\in D_{p,q}$. Thus, there exists a unique Padé approximant of $\widetilde{f}_{j_0}$ and $\widetilde{f}_{j_0}=\frac{B^{-1}(0)A(z)+B^{-1}(0)d_{j_0}z^p}{B^{-1}(0)B(z)-B^{-1}(0)(c_{j_0}z)^q}$ satisfies $[p/q]_{\widetilde{f}_{j_0}}=\widetilde{f}_{j_0}$.
- There exists $r>0$: $\overline{{{\varDelta}}(0,r)}\subseteq{{\varOmega}}\cap{{\varDelta}}(0,{\lambda })\subseteq\ K^0$. Lemma \[lem2.2\] provides $0<{\delta }<\min\{1/2s,{\varepsilon }/2\}$ such that for every $f\in H(\overline{{{\varDelta}}(0,r)})$ with $\|f-\widetilde{f}_{j_0}\|_{\overline{{{\varDelta}}(0,r)}}<{\delta }$, it follows $f\in D_{p,q}$ and $\|[p/q]_f-[p/q]_{\widetilde{f}_{j_0}}\|_{l,n}<1/2s$, $\forall l\le s$.
- By Runge’s theorem there exists a sequense of rational functions, $(R_i)_{i\in{\mathbb{N}}}$, with poles off $K$ and more particularly (see previous property of $K$) off $U$, such that $R_i{\;\rightarrow\;}\widetilde{f}_{j_0}$ uniformly on $K$. This implies that $R^{(l)}_i{\;\rightarrow\;}\widetilde{f}^{(l)}_{j_0}$ uniformly on each compact subset of $K^0$. Hence, there exists $f=R_{i_0}\in H(U)\subseteq H({\overline{\varOmega}})$ such that $\|f-\widetilde{f}_{j_0}\|_{l,{\lambda }}<{\delta }$, $\forall l\le\max\{s,L\}$, because ${\overline{\varOmega}}\cap\overline{{{\varDelta}}(0,{\lambda })}\subseteq K^0$.
- It follows that $\|[p/q]_f-f\|_{l,n}\le\|[p/q]_f-[p/q]_{\widetilde{f}_{j_0}}\|_{l,n}+\|\widetilde{f}_{j_0}-f\|_{l,n}<1/2s+{\delta }<1/s$, $\forall l\le s$. Thus, $f\in E(n,s,(p,q))\cap H({\overline{\varOmega}})$. Moreover, it holds $\|f-R\|_{l,N}\le\|f-\widetilde{f}_{j_0}\|_{l,N}+\|\widetilde{f}_{j_0}-R\|_{l,N}<{\delta }+{\varepsilon }/2<{\varepsilon }$, $\forall l\le L$.
This completes the proof. [$\quad\blacksquare$]{}
[14]{} G. A. Baker, Jr. and P. R. Graves-Morris: Padé Approximants, Vol. 1 and 2, (Encyclopedia of Mathematics and its Applications), Cambridge Un. Press, 2010. P. B. Borwein: The usual behaviour of rational approximants, Canad. Math. Bull. Vol. 26, 1983, p. 317-323. G. Costakis: Some remarks on universal functions and Taylor series, Math. Proc. Cambr. Philos. Soc., Vol. 128, 2000, p. 157-175. N. J. Daras and V. Nestoridis: Universal Padé approximation, arXiv: 1102.4782v1, 2011. G. Fournodavlos: Generic Approximation of functions by their Padé approximants I, arXiv: 1103.3841, 2011. K.-G. Grosse-Erdmann: Holomorphe Monster und universelle Funktionen, Mitt. Math. Sem. Giessen, Vol. 176, 1987, p. 1-84. K.-G. Grosse-Erdmann: Universal families and hypercyclic operators, Bull. Amer. math. Soc., Vol. 36, 1999, p. 345-381. J.-P. Kahane: Baire’s Category theorem and trigonometric series, J. Anal. Math., Vol. 80, 2000, p. 143-182. Ch. Kariofillis, Ch. Konstandaki and V. Nestoridis: Smooth universal Taylor series, Monatsh. Math., Vol. 147, Issue 3, 2006, p. 249-257. A. Melas and V. Nestoridis: Universality of Taylor series as a generic property of holomorphic functions, Adv. Math. Vol. 157, 2001, p. 138-176. A. Melas and V. Nestoridis: On various types of universal Taylor series, Complex Variables Theory, Vol. 44, Issue 3, 2001, p. 245-258. V. Nestoridis: Universal Taylor series, Annales de l’ Institute Fourier, Vol. 46, 1996, p. 1293-1306. V. Nestoridis: A strong notion of universal Taylor series, J. London Math. Soc., Vol. 68, Issue 2, 2003, p. 712-724. W. Rudin, Real and Complex Analysis, McGraw-Hill, 1986.
Department of Mathematics,\
University of Athens\
Panepistemiopolis, 157 84 Athens, Greece\
e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.